Springer Texts in Statistics Advisors: George Casella
Stephen Fienberg
Ingram Olkin
Springer Texts in Statistics Al...

Author:
Jun Shao

This content was uploaded by our users and we assume good faith they have the permission to share this book. If you own the copyright to this book and it is wrongfully on our website, we offer a simple DMCA procedure to remove your content from our site. Start by pressing the button below!

Springer Texts in Statistics Advisors: George Casella

Stephen Fienberg

Ingram Olkin

Springer Texts in Statistics Alfred: Elements of Statistics for the Life and Social Sciences Berger: An Introduction to Probability and Stochastic Processes Bilodeau and Brenner: Theory of Multivariate Statistics Blom: Probability and Statistics: Theory and Applications Brockwell and Davis: Introduction to Times Series and Forecasting, Second Edition Chow and Teicher: Probability Theory: Independence, Interchangeability, Martingales, Third Edition Christensen: Advanced Linear Modeling: Multivariate, Time Series, and Spatial Data: Nonparametric Regression and Response Surface Maximization, Second Edition Christensen: Log-Linear Models and Logistic Regression, Second Edition Christensen: Plane Answers to Complex Questions: The Theory of Linear Models, Third Edition Creighton: A First Course in Probability Models and Statistical Inference Davis: Statistical Methods for the Analysis of Repeated Measurements Dean and Voss: Design and Analysis of Experiments du Toit, Steyn, and Stumpf: Graphical Exploratory Data Analysis Durrett: Essentials of Stochastic Processes Edwards: Introduction to Graphical Modelling, Second Edition Finkelstein and Levin: Statistics for Lawyers Flury: A First Course in Multivariate Statistics Jobson: Applied Multivariate Data Analysis, Volume I: Regression and Experimental Design Jobson: Applied Multivariate Data Analysis, Volume II: Categorical and Multivariate Methods Kalbfleisch: Probability and Statistical Inference, Volume I: Probability, Second Edition Kalbfleisch: Probability and Statistical Inference, Volume II: Statistical Inference, Second Edition Karr: Probability Keyfitz: Applied Mathematical Demography, Second Edition Kiefer: Introduction to Statistical Inference Kokoska and Nevison: Statistical Tables and Formulae Kulkarni: Modeling, Analysis, Design, and Control of Stochastic Systems Lange: Applied Probability Lehmann: Elements of Large-Sample Theory Lehmann: Testing Statistical Hypotheses, Second Edition Lehmann and Casella: Theory of Point Estimation, Second Edition Lindman: Analysis of Variance in Experimental Design Lindsey: Applying Generalized Linear Models (continued after index)

Jun Shao

Mathematical Statistics Second Edition

Jun Shao Department of Statistics University of Wisconsin, Madison Madison, WI 53706-1685 USA [email protected]

Editorial Board George Casella Department of Statistics University of Florida Gainesville, FL 32611-8545 USA

Stephen Fienberg Department of Statistics Carnegie Mellon University Pittsburgh, PA 15213-3890 USA

Ingram Olkin Department of Statistics Stanford University Stanford, CA 94305 USA

With 7 figures.

Library of Congress Cataloging-in-Publication Data Shao, Jun. Mathematical statistics / Jun Shao.—2nd ed. p. cm.— (Springer texts in statistics) Includes bibliographical references and index. ISBN 0-387-95382-5 (alk. paper) 1. Mathematical statistics. I. Title. II. Series. QA276.S458 2003 519.5—dc21 2003045446 ISBN 0-387-95382-5

Printed on acid-free paper.

ISBN-13 978-0-387-95382-3 © 2003 Springer Science+Business Media, LLC. All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC., 233 Spring St., New York, N.Y., 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed in the United States of America. th

9 8 7 6 5 4 (corrected printing as of 4 printing, 2007) springer.com

To Guang, Jason, and Annie

Preface to the First Edition This book is intended for a course entitled Mathematical Statistics offered at the Department of Statistics, University of Wisconsin-Madison. This course, taught in a mathematically rigorous fashion, covers essential materials in statistical theory that a first or second year graduate student typically needs to learn as preparation for work on a Ph.D. degree in statistics. The course is designed for two 15-week semesters, with three lecture hours and two discussion hours in each week. Students in this course are assumed to have a good knowledge of advanced calculus. A course in real analysis or measure theory prior to this course is often recommended. Chapter 1 provides a quick overview of important concepts and results in measure-theoretic probability theory that are used as tools in mathematical statistics. Chapter 2 introduces some fundamental concepts in statistics, including statistical models, the principle of sufficiency in data reduction, and two statistical approaches adopted throughout the book: statistical decision theory and statistical inference. Each of Chapters 3 through 7 provides a detailed study of an important topic in statistical decision theory and inference: Chapter 3 introduces the theory of unbiased estimation; Chapter 4 studies theory and methods in point estimation under parametric models; Chapter 5 covers point estimation in nonparametric settings; Chapter 6 focuses on hypothesis testing; and Chapter 7 discusses interval estimation and confidence sets. The classical frequentist approach is adopted in this book, although the Bayesian approach is also introduced (§2.3.2, §4.1, §6.4.4, and §7.1.3). Asymptotic (large sample) theory, a crucial part of statistical inference, is studied throughout the book, rather than in a separate chapter. About 85% of the book covers classical results in statistical theory that are typically found in textbooks of a similar level. These materials are in the Statistics Department’s Ph.D. qualifying examination syllabus. This part of the book is influenced by several standard textbooks, such as Casella and vii

viii

Preface to the First Edition

Berger (1990), Ferguson (1967), Lehmann (1983, 1986), and Rohatgi (1976). The other 15% of the book covers some topics in modern statistical theory that have been developed in recent years, including robustness of the least squares estimators, Markov chain Monte Carlo, generalized linear models, quasi-likelihoods, empirical likelihoods, statistical functionals, generalized estimation equations, the jackknife, and the bootstrap. In addition to the presentation of fruitful ideas and results, this book emphasizes the use of important tools in establishing theoretical results. Thus, most proofs of theorems, propositions, and lemmas are provided or left as exercises. Some proofs of theorems are omitted (especially in Chapter 1), because the proofs are lengthy or beyond the scope of the book (references are always provided). Each chapter contains a number of examples. Some of them are designed as materials covered in the discussion section of this course, which is typically taught by a teaching assistant (a senior graduate student). The exercises in each chapter form an important part of the book. They provide not only practice problems for students, but also many additional results as complementary materials to the main text. The book is essentially based on (1) my class notes taken in 1983-84 when I was a student in this course, (2) the notes I used when I was a teaching assistant for this course in 1984-85, and (3) the lecture notes I prepared during 1997-98 as the instructor of this course. I would like to express my thanks to Dennis Cox, who taught this course when I was a student and a teaching assistant, and undoubtedly has influenced my teaching style and textbook for this course. I am also very grateful to students in my class who provided helpful comments; to Mr. Yonghee Lee, who helped me to prepare all the figures in this book; to the Springer-Verlag production and copy editors, who helped to improve the presentation; and to my family members, who provided support during the writing of this book. Madison, Wisconsin January 1999

Jun Shao

Preface to the Second Edition In addition to correcting typos and errors and making a better presentation, the main effort in preparing this new edition is adding some new material to Chapter 1 (Probability Theory) and a number of new exercises to each chapter. Furthermore, two new sections are created to introduce semiparametric models and methods (§5.1.4) and to study the asymptotic accuracy of confidence sets (§7.3.4). The structure of the book remains the same. In Chapter 1 of the new edition, moment generating and characteristic functions are treated in more detail and a proof of the uniqueness theorem is provided; some useful moment inequalities are introduced; discussions on conditional independence, Markov chains, and martingales are added, as a continuation of the discussion of conditional expectations; the concepts of weak convergence and tightness are introduced; proofs to some key results in asymptotic theory, such as the dominated convergence theorem and monotone convergence theorem, the L´evy-Cram´er continuity theorem, the strong and weak laws of large numbers, and Lindeberg’s central limit theorem, are included; and a new section (§1.5.6) is created to introduce Edgeworth and Cornish-Fisher expansions. As a result, Chapter 1 of the new edition is self-contained for important concepts, results, and proofs in probability theory with emphasis in statistical applications. Since the original book was published in 1999, I have been using it as a textbook for a two-semester course in mathematical statistics. Exercise problems accumulated during my teaching are added to this new edition. Some exercises that are too trivial have been removed. In the original book, indices on definitions, examples, theorems, propositions, corollaries, and lemmas are included in the subject index. In the new edition, they are in a separate index given in the end of the book (prior to the author index). A list of notation and a list of abbreviations, which are appendices of the original book, are given after the references.

ix

x

Preface to the Second Edition

The most significant change in notation is the notation for a vector. In the text of the new edition, a k-dimensional vector is denoted by c = (c1 , ..., ck ), whether it is treated as a column or a row vector (which is not important if matrix algebra is not considered). When matrix algebra is involved, any vector c is treated as a k × 1 matrix (a column vector) and its transpose cτ is treated as a 1 × k matrix (a row vector). Thus, for c = (c1 , ..., ck ), cτ c = c21 + · · · + c2k and ccτ is the k × k matrix whose (i, j)th element is ci cj . I would like to thank reviewers of this book for their constructive comments, the Springer-Verlag production and copy editors, students in my classes, and two teaching assistants, Mr. Bin Cheng and Dr. Hansheng Wang, who provided help in preparing the new edition. Any remaining errors are of course my own responsibility, and a correction of them may be found on my web page http://www.stat.wisc.edu/˜ shao. Madison, Wisconsin April, 2003

Jun Shao

Contents Preface to the First Edition

vii

Preface to the Second Edition

ix

Chapter 1. Probability Theory

1

1.1 Probability Spaces and Random Elements . . . . . . . . . . .

1

1.1.1 σ-fields and measures . . . . . . . . . . . . . . . . . .

1

1.1.2 Measurable functions and distributions . . . . . . . .

6

1.2 Integration and Differentiation . . . . . . . . . . . . . . . . .

10

1.2.1 Integration . . . . . . . . . . . . . . . . . . . . . . . .

10

1.2.2 Radon-Nikodym derivative . . . . . . . . . . . . . . .

15

1.3 Distributions and Their Characteristics . . . . . . . . . . . .

17

1.3.1 Distributions and probability densities

. . . . . . . .

17

1.3.2 Moments and moment inequalities . . . . . . . . . . .

28

1.3.3 Moment generating and characteristic functions . . .

32

1.4 Conditional Expectations . . . . . . . . . . . . . . . . . . . .

37

1.4.1 Conditional expectations . . . . . . . . . . . . . . . .

37

1.4.2 Independence

. . . . . . . . . . . . . . . . . . . . . .

41

1.4.3 Conditional distributions . . . . . . . . . . . . . . . .

43

1.4.4 Markov chains and martingales . . . . . . . . . . . . .

45

1.5 Asymptotic Theory . . . . . . . . . . . . . . . . . . . . . . .

49

1.5.1 Convergence modes and stochastic orders . . . . . . .

50

1.5.2 Weak convergence . . . . . . . . . . . . . . . . . . . .

56

1.5.3 Convergence of transformations . . . . . . . . . . . .

59

1.5.4 The law of large numbers . . . . . . . . . . . . . . . .

62

1.5.5 The central limit theorem . . . . . . . . . . . . . . . .

67

xi

xii

Contents

1.5.6 Edgeworth and Cornish-Fisher expansions . . . . . .

70

1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

74

Chapter 2. Fundamentals of Statistics

91

2.1 Populations, Samples, and Models . . . . . . . . . . . . . . .

91

2.1.1 Populations and samples . . . . . . . . . . . . . . . .

91

2.1.2 Parametric and nonparametric models . . . . . . . . .

94

2.1.3 Exponential and location-scale families . . . . . . . .

96

2.2 Statistics, Sufficiency, and Completeness . . . . . . . . . . . . 100 2.2.1 Statistics and their distributions . . . . . . . . . . . . 100 2.2.2 Sufficiency and minimal sufficiency

. . . . . . . . . . 103

2.2.3 Complete statistics . . . . . . . . . . . . . . . . . . . 109 2.3 Statistical Decision Theory . . . . . . . . . . . . . . . . . . . 113 2.3.1 Decision rules, loss functions, and risks . . . . . . . . 113 2.3.2 Admissibility and optimality . . . . . . . . . . . . . . 116 2.4 Statistical Inference . . . . . . . . . . . . . . . . . . . . . . . 122 2.4.1 Point estimators . . . . . . . . . . . . . . . . . . . . . 122 2.4.2 Hypothesis tests . . . . . . . . . . . . . . . . . . . . . 125 2.4.3 Confidence sets

. . . . . . . . . . . . . . . . . . . . . 129

2.5 Asymptotic Criteria and Inference . . . . . . . . . . . . . . . 131 2.5.1 Consistency . . . . . . . . . . . . . . . . . . . . . . . 132 2.5.2 Asymptotic bias, variance, and mse . . . . . . . . . . 135 2.5.3 Asymptotic inference . . . . . . . . . . . . . . . . . . 139 2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 Chapter 3. Unbiased Estimation

161

3.1 The UMVUE . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 3.1.1 Sufficient and complete statistics . . . . . . . . . . . . 162 3.1.2 A necessary and sufficient condition . . . . . . . . . . 166 3.1.3 Information inequality . . . . . . . . . . . . . . . . . . 169 3.1.4 Asymptotic properties of UMVUE’s . . . . . . . . . . 172 3.2 U-Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 3.2.1 Some examples . . . . . . . . . . . . . . . . . . . . . . 174 3.2.2 Variances of U-statistics . . . . . . . . . . . . . . . . . 176 3.2.3 The projection method . . . . . . . . . . . . . . . . . 178

Contents

xiii

3.3 The LSE in Linear Models . . . . . . . . . . . . . . . . . . . 182 3.3.1 The LSE and estimability . . . . . . . . . . . . . . . . 182 3.3.2 The UMVUE and BLUE . . . . . . . . . . . . . . . . 186 3.3.3 Robustness of LSE’s . . . . . . . . . . . . . . . . . . . 189 3.3.4 Asymptotic properties of LSE’s . . . . . . . . . . . . 193 3.4 Unbiased Estimators in Survey Problems . . . . . . . . . . . 195 3.4.1 UMVUE’s of population totals . . . . . . . . . . . . . 195 3.4.2 Horvitz-Thompson estimators . . . . . . . . . . . . . 199 3.5 Asymptotically Unbiased Estimators . . . . . . . . . . . . . . 204 3.5.1 Functions of unbiased estimators . . . . . . . . . . . . 204 3.5.2 The method of moments . . . . . . . . . . . . . . . . 207 3.5.3 V-statistics . . . . . . . . . . . . . . . . . . . . . . . . 210 3.5.4 The weighted LSE . . . . . . . . . . . . . . . . . . . . 213 3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 Chapter 4. Estimation in Parametric Models

231

4.1 Bayes Decisions and Estimators . . . . . . . . . . . . . . . . 231 4.1.1 Bayes actions . . . . . . . . . . . . . . . . . . . . . . . 231 4.1.2 Empirical and hierarchical Bayes methods . . . . . . 236 4.1.3 Bayes rules and estimators . . . . . . . . . . . . . . . 239 4.1.4 Markov chain Monte Carlo . . . . . . . . . . . . . . . 245 4.2 Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 4.2.1 One-parameter location families . . . . . . . . . . . . 251 4.2.2 One-parameter scale families . . . . . . . . . . . . . . 255 4.2.3 General location-scale families . . . . . . . . . . . . . 257 4.3 Minimaxity and Admissibility . . . . . . . . . . . . . . . . . 261 4.3.1 Estimators with constant risks . . . . . . . . . . . . . 261 4.3.2 Results in one-parameter exponential families . . . . 265 4.3.3 Simultaneous estimation and shrinkage estimators . . 267 4.4 The Method of Maximum Likelihood . . . . . . . . . . . . . 273 4.4.1 The likelihood function and MLE’s . . . . . . . . . . 273 4.4.2 MLE’s in generalized linear models . . . . . . . . . . 279 4.4.3 Quasi-likelihoods and conditional likelihoods . . . . . 283 4.5 Asymptotically Efficient Estimation . . . . . . . . . . . . . . 286 4.5.1 Asymptotic optimality . . . . . . . . . . . . . . . . . 286

xiv

Contents

4.5.2 Asymptotic efficiency of MLE’s and RLE’s . . . . . . 290 4.5.3 Other asymptotically efficient estimators . . . . . . . 295 4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299 Chapter 5. Estimation in Nonparametric Models 5.1 Distribution Estimators . . . . . . . . . . . . . . . 5.1.1 Empirical c.d.f.’s in i.i.d. cases . . . . . . . 5.1.2 Empirical likelihoods . . . . . . . . . . . . 5.1.3 Density estimation . . . . . . . . . . . . . . 5.1.4 Semi-parametric methods . . . . . . . . . . 5.2 Statistical Functionals . . . . . . . . . . . . . . . . 5.2.1 Differentiability and asymptotic normality 5.3

5.4

5.5

5.6

. . . . . . .

. . . . . . .

. . . . . . .

319 319 320 323 330 333 338 338

5.2.2 L-, M-, and R-estimators and rank statistics . . . Linear Functions of Order Statistics . . . . . . . . . . . . 5.3.1 Sample quantiles . . . . . . . . . . . . . . . . . . . 5.3.2 Robustness and efficiency . . . . . . . . . . . . . . 5.3.3 L-estimators in linear models . . . . . . . . . . . . Generalized Estimating Equations . . . . . . . . . . . . . 5.4.1 The GEE method and its relationship with others 5.4.2 Consistency of GEE estimators . . . . . . . . . . . 5.4.3 Asymptotic normality of GEE estimators . . . . . Variance Estimation . . . . . . . . . . . . . . . . . . . . . 5.5.1 The substitution method . . . . . . . . . . . . . . 5.5.2 The jackknife . . . . . . . . . . . . . . . . . . . . . 5.5.3 The bootstrap . . . . . . . . . . . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

343 351 351 355 358 359 360 363 367 371 372 376 380 383

. . . . . . . .

393 393 394 397 401 404 404 406 410

Chapter 6. Hypothesis Tests 6.1 UMP Tests . . . . . . . . . . . . . . . . . . . 6.1.1 The Neyman-Pearson lemma . . . . . 6.1.2 Monotone likelihood ratio . . . . . . . 6.1.3 UMP tests for two-sided hypotheses . 6.2 UMP Unbiased Tests . . . . . . . . . . . . . 6.2.1 Unbiasedness, similarity, and Neyman 6.2.2 UMPU tests in exponential families . 6.2.3 UMPU tests in normal families . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . structure . . . . . . . . . . . .

. . . . . . . .

. . . . . . . .

xv

Contents

6.3 UMP Invariant Tests . . . . . . . . . . . . . . . . . . . . . . 417 6.3.1 Invariance and UMPI tests . . . . . . . . . . . . . . . 417 6.3.2 UMPI tests in normal linear models . . . . . . . . . . 422 6.4 Tests in Parametric Models . . . . . . . . . . . . . . . . . . . 428 6.4.1 Likelihood ratio tests . . . . . . . . . . . . . . . . . . 428 6.4.2 Asymptotic tests based on likelihoods . . . . . . . . . 431 6.4.3 χ2 -tests . . . . . . . . . . . . . . . . . . . . . . . . . . 436 6.4.4 Bayes tests . . . . . . . . . . . . . . . . . . . . . . . . 440 6.5 Tests in Nonparametric Models . . . . . . . . . . . . . . . . . 442 6.5.1 Sign, permutation, and rank tests . . . . . . . . . . . 442 6.5.2 Kolmogorov-Smirnov and Cram´er-von Mises tests . . 446 6.5.3 Empirical likelihood ratio tests . . . . . . . . . . . . . 449 6.5.4 Asymptotic tests . . . . . . . . . . . . . . . . . . . . . 452 6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454 Chapter 7. Confidence Sets

471

7.1 Construction of Confidence Sets . . . . . . . . . . . . . . . . 471 7.1.1 Pivotal quantities . . . . . . . . . . . . . . . . . . . . 471 7.1.2 Inverting acceptance regions of tests . . . . . . . . . . 477 7.1.3 The Bayesian approach . . . . . . . . . . . . . . . . . 480 7.1.4 Prediction sets . . . . . . . . . . . . . . . . . . . . . . 482 7.2 Properties of Confidence Sets . . . . . . . . . . . . . . . . . . 484 7.2.1 Lengths of confidence intervals . . . . . . . . . . . . . 484 7.2.2 UMA and UMAU confidence sets . . . . . . . . . . . 488 7.2.3 Randomized confidence sets

. . . . . . . . . . . . . . 491

7.2.4 Invariant confidence sets . . . . . . . . . . . . . . . . 493 7.3 Asymptotic Confidence Sets

. . . . . . . . . . . . . . . . . . 495

7.3.1 Asymptotically pivotal quantities . . . . . . . . . . . 495 7.3.2 Confidence sets based on likelihoods . . . . . . . . . . 497 7.3.3 Confidence intervals for quantiles . . . . . . . . . . . 501 7.3.4 Accuracy of asymptotic confidence sets . . . . . . . . 503 7.4 Bootstrap Confidence Sets . . . . . . . . . . . . . . . . . . . 505 7.4.1 Construction of bootstrap confidence intervals . . . . 506 7.4.2 Asymptotic correctness and accuracy . . . . . . . . . 509 7.4.3 High-order accurate bootstrap confidence sets . . . . 515

xvi

Contents

7.5 Simultaneous Confidence Intervals . . . . . 7.5.1 Bonferroni’s method . . . . . . . . . 7.5.2 Scheff´e’s method in linear models . 7.5.3 Tukey’s method in one-way ANOVA 7.5.4 Confidence bands for c.d.f.’s . . . . 7.6 Exercises . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . models . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

519 519 520 523 525 527

References

543

List of Notation

555

List of Abbreviations

557

Index of Definitions, Main Results, and Examples

559

Author Index

571

Subject Index

575

Chapter 1

Probability Theory Mathematical statistics relies on probability theory, which in turn is based on measure theory. The present chapter provides some principal concepts and notational conventions of probability theory, and some important results that are useful tools in statistics. A more complete account of probability theory can be found in a standard textbook, for example, Billingsley (1986), Chung (1974), or Lo`eve (1977). The reader is assumed to be familiar with set operations and set functions (mappings) in advanced calculus.

1.1 Probability Spaces and Random Elements In an elementary probability course, one defines a random experiment to be an experiment whose outcome cannot be predicted with certainty, and the probability of A (a collection of possible outcomes) to be the fraction of times that the outcome of the random experiment results in A in a large number of trials of the random experiment. A rigorous and logically consistent definition of probability was given by A. N. Kolmogorov in his measure-theoretic fundamental development of probability theory in 1933 (Kolmogorov, 1933).

1.1.1 σ-fields and measures Let Ω be a set of elements of interest. For example, Ω can be a set of numbers, a subinterval of the real line, or all possible outcomes of a random experiment. In probability theory, Ω is often called the outcome space, whereas in statistical theory, Ω is called the sample space. This is because in probability and statistics, Ω is usually the set of all possible outcomes of a random experiment under study. 1

2

1. Probability Theory

A measure is a natural mathematical extension of the length, area, or volume of subsets in the one-, two-, or three-dimensional Euclidean space. In a given sample space Ω, a measure is a set function defined for certain subsets of Ω. It is necessary for this collection of subsets to satisfy certain properties, which are given in the following definition. Definition 1.1. Let F be a collection of subsets of a sample space Ω. F is called a σ-field (or σ-algebra) if and only if it has the following properties. (i) The empty set ∅ ∈ F. (ii) If A ∈ F , then the complement Ac ∈ F . (iii) If Ai ∈ F, i = 1, 2, ..., then their union ∪Ai ∈ F. A pair (Ω, F ) consisting of a set Ω and a σ-field F of subsets of Ω is called a measurable space. The elements of F are called measurable sets in measure theory or events in probability and statistics. Since ∅c = Ω, it follows from (i) and (ii) in Definition 1.1 that Ω ∈ F if F is a σ-field on Ω. Also, it follows from (ii) and (iii) that if Ai ∈ F , i = 1, 2, ..., and F is a σ-field, then the intersection ∩Ai ∈ F . This can be c shown using DeMorgan’s law: (∩Ai ) = ∪Aci . For any given Ω, there are two trivial σ-fields. The first one is the collection containing exactly two elements, ∅ and Ω. This is the smallest possible σ-field on Ω. The second one is the collection of all subsets of Ω, which is called the power set and is the largest σ-field on Ω. Let us now consider some nontrivial σ-fields. Let A be a nonempty proper subset of Ω (A ⊂ Ω, A 6= Ω). Then (verify) {∅, A, Ac , Ω}

(1.1)

is a σ-field. In fact, this is the smallest σ-field containing A in the sense that if F is any σ-field containing A, then the σ-field in (1.1) is a subcollection of F. In general, the smallest σ-field containing C, a collection of subsets of Ω, is denoted by σ(C) and is called the σ-field generated by C. Hence, the σ-field in (1.1) is σ({A}). Note that σ({A, Ac }), σ({A, Ω}), and σ({A, ∅}) are all the same as σ({A}). Of course, if C itself is a σ-field, then σ(C) = C. On the real line R, there is a special σ-field that will be used almost exclusively. Let C be the collection of all finite open intervals on R. Then B = σ(C) is called the Borel σ-field. The elements of B are called Borel sets. The Borel σ-field B k on the k-dimensional Euclidean space Rk can be similarly defined. It can be shown that all intervals (finite or infinite), open sets, and closed sets are Borel sets. To illustrate, we now show that, on the real line, B = σ(O), where O is the collection of all open sets. Typically, one needs to show that σ(C) ⊂ σ(O) and σ(O) ⊂ σ(C). Since an open interval is an open set, C ⊂ O and, hence, σ(C) ⊂ σ(O) (why?). Let U be an open set. Then U can be expressed as a union of a sequence of finite open

1.1. Probability Spaces and Random Elements

3

intervals (see Royden (1968, p.39)). Hence, U ∈ σ(C) (Definition 1.1(iii)) and O ⊂ σ(C). By the definition of σ(O), σ(O) ⊂ σ(C). This completes the proof. Let C ⊂ Rk be a Borel set and let BC = {C ∩ B : B ∈ B k }. Then (C, BC ) is a measurable space and BC is called the Borel σ-field on C. Now we can introduce the notion of a measure. Definition 1.2. Let (Ω, F ) be a measurable space. A set function ν defined on F is called a measure if and only if it has the following properties. (i) 0 ≤ ν(A) ≤ ∞ for any A ∈ F. (ii) ν(∅) = 0. (iii) If Ai ∈ F , i = 1, 2, ..., and Ai ’s are disjoint, i.e., Ai ∩ Aj = ∅ for any i 6= j, then ! ∞ ∞ [ X ν Ai = ν(Ai ). i=1

i=1

The triple (Ω, F, ν) is called a measure space. If ν(Ω) = 1, then ν is called a probability measure and we usually denote it by P instead of ν, in which case (Ω, F, P ) is called a probability space. Although measure is an extension of length, area, or volume, sometimes it can be quite abstract. For example, the following set function is a measure: ∞ A ∈ F, A 6= ∅ ν(A) = (1.2) 0 A = ∅.

Since a measure can take ∞ as its value, we must know how to do arithmetic with ∞. In this book, it suffices to know that (1) for any x ∈ R, ∞+x = ∞, x ∞ = ∞ if x > 0, x ∞ = −∞ if x < 0, and 0 ∞ = 0; (2) ∞ + ∞ = ∞; and (3) ∞a = ∞ for any a > 0. However, ∞ − ∞ or ∞/∞ is not defined. The following examples provide two very important measures in probability and statistics. Example 1.1 (Counting measure). Let Ω be a sample space, F the collection of all subsets, and ν(A) the number of elements in A ∈ F (ν(A) = ∞ if A contains infinitely many elements). Then ν is a measure on F and is called the counting measure. Example 1.2 (Lebesgue measure). There is a unique measure m on (R, B) that satisfies m([a, b]) = b − a (1.3)

for every finite interval [a, b], −∞ < a ≤ b < ∞. This is called the Lebesgue measure. If we restrict m to the measurable space ([0, 1], B[0,1]), then m is a probability measure.

4

1. Probability Theory

If Ω is countable in the sense that there is a one-to-one correspondence between Ω and the set of all integers, then one can usually consider the trivial σ-field that contains all subsets of Ω and a measure that assigns a value to every subset of Ω. When Ω is uncountable (e.g., Ω = R or [0, 1]), it is not possible to define a reasonable measure for every subset of Ω; for example, it is not possible to find a measure on all subsets of R and still satisfy property (1.3). This is why it is necessary to introduce σ-fields that are smaller than the power set. The following result provides some basic properties of measures. Whenever we consider ν(A), it is implicitly assumed that A ∈ F. Proposition 1.1. Let (Ω, F, ν) be a measure space. (i) (Monotonicity). If A ⊂ B, then ν(A) ≤ ν(B). (ii) (Subadditivity). For any sequence A1 , A2 , ..., ! ∞ ∞ [ X ν Ai ≤ ν(Ai ). i=1

i=1

(iii) (Continuity). If A1 ⊂ A2 ⊂ A3 ⊂ · · · (or A1 ⊃ A2 ⊃ A3 ⊃ · · · and ν(A1 ) < ∞), then ν lim An = lim ν (An ) , n→∞

where

lim An =

n→∞

∞ [

n→∞

Ai

or =

i=1

∞ \

Ai

i=1

!

.

Proof. We prove (i) only. The proofs of (ii) and (iii) are left as exercises. Since A ⊂ B, B = A ∪ (Ac ∩ B) and A and Ac ∩ B are disjoint. By Definition 1.2(iii), ν(B) = ν(A) + ν(Ac ∩ B), which is no smaller than ν(A) since ν(Ac ∩ B) ≥ 0 by Definition 1.2(i). There is a one-to-one correspondence between the set of all probability measures on (R, B) and a set of functions on R. Let P be a probability measure. The cumulative distribution function (c.d.f.) of P is defined to be F (x) = P ((−∞, x]) ,

x ∈ R.

(1.4)

Proposition 1.2. (i) Let F be a c.d.f. on R. Then (a) F (−∞) = limx→−∞ F (x) = 0; (b) F (∞) = limx→∞ F (x) = 1; (c) F is nondecreasing, i.e., F (x) ≤ F (y) if x ≤ y; (d) F is right continuous, i.e., limy→x,y>x F (y) = F (x). (ii) Suppose that a real-valued function F on R satisfies (a)-(d) in part (i). Then F is the c.d.f. of a unique probability measure on (R, B).

1.1. Probability Spaces and Random Elements

5

The Cartesian product of sets (or collections of sets) Γi , i ∈ I = {1, ..., k} (or {1, 2, ...}) is defined asQ the set of all (a1 , ..., ak ) (or (a1 , a2 , ...)), ai ∈ Γi , i ∈ I, and is denoted by i∈I Γi = Γ1 × · · Q · × Γk (or Γ1 × Γ2 × · · ·). Let (Ωi , Fi ), i ∈ I, be measurable spaces. Since a σi∈I Fi is not necessarily Q Q F is called the product σ-field on the product space field, σ i i∈I Ωi Q Q i∈I Q is denoted by and Ω , σ F (Ω , F ). As an example, i i i i i∈I i∈I i∈I consider (Ωi , Fi ) = (R, B), i = 1, ..., k. Then the product space is Rk and it can be shown that the product σ-field is the same as the Borel σ-field on Rk , which is the σ-field generated by the collection of all open sets in Rk . In Example 1.2, the usual length of an interval [a, b] ⊂ R is the same as the Lebesgue measure of [a, b]. Consider a rectangle [a1 , b1 ] × [a2 , b2 ] ⊂ R2 . The usual area of [a1 , b1 ] × [a2 , b2 ] is (b1 − a1 )(b2 − a2 ) = m([a1 , b1 ])m([a2 , b2 ]),

(1.5)

i.e., the product of the Lebesgue measures of two intervals [a1 , b1 ] and [a2 , b2 ]. Note that [a1 , b1 ] × [a2 , b2 ] is a measurable set by the definition of the product σ-field. Is m([a1 , b1 ])m([a2 , b2 ]) the same as the value of a measure defined on the product σ-field? The following result answers this question for any product space generated by a finite number of measurable spaces. (Its proof can be found in Billingsley (1986, pp. 235-236).) Before introducing this result, we need the following technical definition. A measure ν on (Ω, F ) is said to be σ-finite if and only if there exists a sequence {A1 , A2 , ...} such that ∪Ai = Ω and ν(Ai ) < ∞ for all i. Any finite measure (such as a probability measure) is clearly σ-finite. The Lebesgue measure in Example 1.2 is σ-finite, since R = ∪An with An = (−n, n), n = 1, 2, .... The counting measure in Example 1.1 is σ-finite if and only if Ω is countable. The measure defined by (1.2), however, is not σ-finite. Proposition 1.3 (Product measure theorem). Let (Ωi , Fi , νi ), i = 1, ..., k, be measure spaces with σ-finite measures, where k ≥ 2 is an integer. Then there exists a unique σ-finite measure on the product σ-field σ(F1 ×· · ·×Fk ), called the product measure and denoted by ν1 × · · · × νk , such that ν1 × · · · × νk (A1 × · · · × Ak ) = ν1 (A1 ) · · · νk (Ak ) for all Ai ∈ Fi , i = 1, ..., k. In R2 , there is a unique measure, the product measure m× m, for which m × m([a1 , b1 ] × [a2 , b2 ]) is equal to the value given by (1.5). This measure is called the Lebesgue measure on (R2 , B 2 ). The Lebesgue measure on (R3 , B 3 ) is m × m × m, which equals the usual volume for a subset of the form [a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]. The Lebesgue measure on (Rk , B k ) for any positive integer k is similarly defined. The concept of c.d.f. can be extended to Rk . Let P be a probability

6

1. Probability Theory

measure on (Rk , B k ). The c.d.f. (or joint c.d.f.) of P is defined by F (x1 , ..., xk ) = P ((−∞, x1 ] × · · · × (−∞, xk ]) ,

xi ∈ R.

(1.6)

Again, there is a one-to-one correspondence between probability measures and joint c.d.f.’s on Rk . Some properties of a joint c.d.f. are given in Exercise 10 in §1.6. If F (x1 , ..., xk ) is a joint c.d.f., then Fi (x) =

lim

xj →∞,j=1,...,i−1,i+1,...,k

F (x1 , ..., xi−1 , x, xi+1 , ..., xk )

is a c.d.f. and is called the ith marginal c.d.f. Apparently, marginal c.d.f.’s are determined by their joint c.d.f. But a joint c.d.f. cannot be determined by k marginal c.d.f.’s. There is one special but important case in which a joint c.d.f. F is determined by its k marginal c.d.f. Fi ’s through F (x1 , ..., xk ) = F1 (x1 ) · · · Fk (xk ),

(x1 , ..., xk ) ∈ Rk ,

(1.7)

in which case the probability measure corresponding to F is the product measure P1 × · · · × Pk with Pi being the probability measure corresponding to Fi . Proposition 1.3 can be extended to cases involving infinitely many measure spaces (Billingsley, 1986). In particular, if (Rk , B k , Pi ), i = 1, 2, ..., are probability spaces, then there is a product probability measure P on Q ∞ k k k i=1 (R , B ) such that for any positive integer l and Bi ∈ B , i = 1, ..., l, P (B1 × · · · × Bl × Rk × Rk × · · ·) = P1 (B1 ) · · · Pl (Bl ).

1.1.2 Measurable functions and distributions Since Ω can be quite arbitrary, it is often convenient to consider a function (mapping) f from Ω to a simpler space Λ (often Λ = Rk ). Let B ⊂ Λ. Then the inverse image of B under f is f −1 (B) = {f ∈ B} = {ω ∈ Ω : f (ω) ∈ B}. The inverse function f −1 need not exist for f −1 (B) to be defined. The reader is asked to verify the following properties: (a) f −1 (B c ) = (f −1 (B))c for any B ⊂ Λ; (b) f −1 (∪Bi ) = ∪f −1 (Bi ) for any Bi ⊂ Λ, i = 1, 2, .... Let C be a collection of subsets of Λ. We define f −1 (C) = {f −1 (C) : C ∈ C}. Definition 1.3. Let (Ω, F ) and (Λ, G) be measurable spaces and f a function from Ω to Λ. The function f is called a measurable function from (Ω, F) to (Λ, G) if and only if f −1 (G) ⊂ F.

1.1. Probability Spaces and Random Elements

7

If Λ = R and G = B (Borel σ-field), then f is said to be Borel measurable or is called a Borel function on (Ω, F) (or with respect to F). In probability theory, a measurable function is called a random element and denoted by one of X, Y , Z,.... If X is measurable from (Ω, F ) to (R, B), then it is called a random variable; if X is measurable from (Ω, F ) to (Rk , B k ), then it is called a random k-vector. If X1 , ..., Xk are random variables defined on a common probability space, then the vector (X1 , ..., Xk ) is a random k-vector. (As a notational convention, any vector c ∈ Rk is denoted by (c1 , ..., ck ), where ci is the ith component of c.) If f is measurable from (Ω, F) to (Λ, G), then f −1 (G) is a sub-σ-field of F (verify). It is called the σ-field generated by f and is denoted by σ(f ). Now we consider some examples of measurable functions. If F is the collection of all subsets of Ω, then any function f is measurable. Let A ⊂ Ω. The indicator function for A is defined as 1 ω∈A IA (ω) = 0 ω 6∈ A. For any B ⊂ R,

∅ A −1 IA (B) = Ac Ω

0 6∈ B, 1 6∈ B 0 6∈ B, 1 ∈ B 0 ∈ B, 1 6∈ B 0 ∈ B, 1 ∈ B.

Then σ(IA ) is the σ-field given in (1.1). If A is a measurable set, then IA is a Borel function. Note that σ(IA ) is a much smaller σ-field than the original σ-field F. This is another reason why we introduce the concept of measurable functions and random variables, in addition to the reason that it is easy to deal with numbers. Often the σ-field F (such as the power set) contains too many subsets and we are only interested in some of them. One can then define a random variable X with σ(X) containing subsets that are of interest. In general, σ(X) is between the trivial σ-field {∅, Ω} and F, and contains more subsets if X is more complicated. For the simplest function IA , we have shown that σ(IA ) contains only four elements. The class of simple functions is obtained by taking linear combinations of indicators of measurable sets, i.e., ϕ(ω) =

k X

ai IAi (ω),

(1.8)

i=1

where A1 , ..., Ak are measurable sets on Ω and a1 , ..., ak are real numbers. One can show directly that such a function is a Borel function, but it

8

1. Probability Theory

follows immediately from Proposition 1.4. Let A1 , ..., Ak be a partition of Ω, i.e., Ai ’s are disjoint and A1 ∪ · · · ∪ Ak = Ω. Then the simple function ϕ given by (1.8) with distinct ai ’s exactly characterizes this partition and σ(ϕ) = σ({A1 , ..., Ak }). Proposition 1.4. Let (Ω, F) be a measurable space. (i) f is Borel if and only if f −1 (a, ∞) ∈ F for all a ∈ R. (ii) If f and g are Borel, then so are f g and af + bg, where a and b are real numbers; also, f /g is Borel provided g(ω) 6= 0 for any ω ∈ Ω. (iii) If f1 , f2 , ... are Borel, then so are supn fn , inf n fn , lim supn fn , and lim inf n fn . Furthermore, the set o n A = ω ∈ Ω : lim fn (ω) exists n→∞

is an event and the function limn→∞ fn (ω) h(ω) = f1 (ω)

ω∈A ω 6∈ A

is Borel. (iv) Suppose that f is measurable from (Ω, F ) to (Λ, G) and g is measurable from (Λ, G) to (∆, H). Then the composite function g◦f is measurable from (Ω, F) to (∆, H). (v) Let Ω be a Borel set in Rp . If f is a continuous function from Ω to Rq , then f is measurable. Proposition 1.4 indicates that there are many Borel functions. In fact, it is hard to find a non-Borel function. The following result is very useful in technical proofs. Let f be a nonnegative Borel function on (Ω, F ). Then there exists a sequence of simple functions {ϕn } satisfying 0 ≤ ϕ1 ≤ ϕ2 ≤ · · · ≤ f and limn→∞ ϕn = f (Exercise 17 in §1.6). Let (Ω, F , ν) be a measure space and f be a measurable function from (Ω, F) to (Λ, G). The induced measure by f , denoted by ν◦f −1 , is a measure on G defined as ν ◦ f −1 (B) = ν(f ∈ B) = ν f −1 (B) , B ∈ G. (1.9)

It is usually easier to deal with ν ◦ f −1 than to deal with ν since (Λ, G) is usually simpler than (Ω, F ). Furthermore, subsets not in σ(f ) are not involved in the definition of ν ◦ f −1 . As we discussed earlier, in some cases we are only interested in subsets in σ(f ). If ν = P is a probability measure and X is a random variable or a random vector, then P ◦ X −1 is called the law or the distribution of X and

1.1. Probability Spaces and Random Elements

9

is denoted by PX . The c.d.f. of PX defined by (1.4) or (1.6) is also called the c.d.f. or joint c.d.f. of X and is denoted by FX . On the other hand, for any c.d.f. or joint c.d.f. F , there exists at least one random variable or vector (usually there are many) defined on some probability space for which FX = F . The following are some examples of random variables and their c.d.f.’s. More examples can be found in §1.3.1. Example 1.3 (Discrete c.d.f.’s). Let a1 < a2 < · · · be a sequence of real numbers P∞ and let pn , n = 1, 2, ..., be a sequence of positive numbers such that n=1 pn = 1. Define F (x) =

Pn

i=1

an ≤ x < an+1 , n = 1, 2, ... −∞ < x < a1 .

pi

0

(1.10)

Then F is a stepwise c.d.f. It has a jump of size pn at each an and is flat between an and an+1 , n = 1, 2, .... Such a c.d.f. is called a discrete c.d.f. and the corresponding random variable is called a discrete random variable. We can easily obtain a random variable having F in (1.10) as its c.d.f. For example, let Ω = {a1 , a2 , ...}, F be the collection of all subsets of Ω, X

P (A) =

i:ai ∈A

pi ,

A ∈ F,

(1.11)

and X(ω) = ω. One can show that P is a probability measure and the c.d.f. of X is F in (1.10). Example 1.4 (Continuous c.d.f.’s). Opposite to the class of discrete c.d.f.’s is the class of continuous c.d.f.’s. Without the concepts of integration and differentiation introduced in the next section, we can only provide a few examples of continuous c.d.f.’s. One such example is the uniform c.d.f. on the interval [a, b] defined as

F (x) =

0

x−a b−a

1

−∞ < x < a a≤x 0 and a∞ = R 0 if a = 0, the right-hand side of (1.12) is always well defined, although ϕdν = ∞ is possible. Note that different ai ’s and Ai ’s may produce the same function ϕ; for example, with Ω = R, 2I(0,1) (x) + I[1,2] (x) = I(0,2] (x) + I(0,1) (x). However, one can show Rthat different representations of ϕ in (1.8) produce the same value for ϕdν so that the integral of a nonnegative simple function is well defined. Next, we consider a nonnegative Borel function f . Definition 1.4(b). Let f be a nonnegative Borel function and let Sf be the collection of all nonnegative simple functions of the form (1.8) satisfying ϕ(ω) ≤ f (ω) for any ω ∈ Ω. The integral of f w.r.t. ν is defined as Z Z f dν = sup ϕdν : ϕ ∈ Sf . Hence, for any Borel function f ≥ 0, there exists a sequence Rof simple functions ϕ1 , ϕ2 , ... such that 0 ≤ ϕi ≤ f for all i and limn→∞ ϕn dν = R f dν.

11

1.2. Integration and Differentiation

Finally, for a Borel function f , we first define the positive part of f by f+ (ω) = max{f (ω), 0} and the negative part of f by f− (ω) = max{−f (ω), 0}. Note that f+ and f− are nonnegative Borel functions, f (ω) = f+ (ω) − f− (ω), and |f (ω)| = f+ (ω) + f− (ω).

R Definition 1.4(c). Let f Rbe a Borel function. We say that f dν exists if R and only if at least one of f+ dν and f− dν is finite, in which case Z Z Z (1.13) f dν = f+ dν − f− dν. R R When both f+ dν and f− dν are finite, we say that f is integrable. Let A be a measurable set and IA be its indicator function. The integral of f over A is defined as Z Z f dν = IA f dν. A

Note that a Borel function f is integrable if and only if |f | is integrable. It is convenient to define the integral of a measurable function f from ¯ B), ¯ where R ¯ = R ∪ {−∞, ∞}, B¯ = σ(B ∪ {{∞}, {−∞}}). (Ω, F , ν) to (R, R Let A+ = {f = ∞} and A f+ dν − R= {f = −∞}.R If ν(A+ ) = 0, we define R to be IAc+ f+ dν; otherwise f+ dν = ∞. f− dν is similarly defined. If at R R R least one of f+ dν and f− dν is finite, then f dν is defined by (1.13). The integral of f may be denoted differently whenever there is a need to indicateR the variable(s) to Rbe integrated andR the integration domain; for R example, Ω f dν, f (ω)dν,R f (ω)dν(ω), or f (ω)ν(dω), and so on. In probability and statistics, XdP is usually written as EX or E(X) and called the Rexpectation or expected value R of X. If F isR the c.d.f. of P on (Rk , B k ), f (x)dP is also denoted by f (x)dF (x) or f dF . Example 1.5. Let Ω be a countable set, F be all subsets of Ω, and ν be the counting measure given in Example 1.1. For any Borel function f , it can be shown (exercise) that Z X f dν = f (ω). (1.14) ω∈Ω

Example 1.6. If Ω = R and ν is the Lebesgue measure, then the Lebesgue R Rb integral of f over an interval [a, b] is written as [a,b] f (x)dx = a f (x)dx, which agrees with the Riemann integral in calculus when the latter is well

12

1. Probability Theory

defined. However, there are functions for which the Lebesgue integrals are defined but not the Riemann integrals. We now introduce some properties of integrals. The proof of the following result is left to the reader. Proposition 1.5 (Linearity of integrals). Let (Ω, F, ν) be a measure space and fRand g be Borel functions. R R (i) If f dν Rexists and aR ∈ R, then (afR)dν exists R and is equal to a f dν. R(ii) If both f dν and gdν existRand fRdν + gdν is well defined, then (f + g)dν exists and is equal to f dν + gdν.

If N is an event with ν(N ) = 0 and a statement holds for all ω in the complement N c , then the statement is said to hold a.e. (almost everywhere) ν (or simply a.e. if the measure ν is clear from the context). If ν is a probability measure, then a.e. may be replaced by a.s. (almost surely).

Proposition 1.6. LetR (Ω, F, ν) R be a measure space and f and g be Borel. (i) If f ≤ g a.e., thenR f dν ≤ gdν, provided that the integrals exist. (ii) If f ≥ 0 a.e. and f dν = 0, then f = 0 a.e. Proof. (i) The proof for part (i) is left to the reader. (ii) Let A = {f > 0} and An = {f ≥ n−1 }, n = 1, 2, .... Then An ⊂ A for any n and limn→∞ An = ∪An = A (why?). By Proposition 1.1(iii), limn→∞ ν(An ) = ν(A). Using part (i) and Proposition 1.5, we obtain that Z Z Z n−1 ν(An ) = n−1 IAn dν ≤ f IAn dν ≤ f dν = 0 for any n. Hence ν(A) = 0 and f = 0 a.e.

R R Some direct consequences of Proposition 1.6(i) are: R R | f dν| R ≤ |f |dν; if f ≥ 0 a.e., then f dν ≥ 0; and if f = g a.e., then f dν = gdν. It is sometimes required to know whether the following interchange of two operations is valid: Z Z lim fn dν = lim fn dν, (1.15) n→∞

n→∞

where {fn : n = 1, 2, ...} is a sequence of Borel functions. Note that we only require limn→∞ fn exists a.e. Also, limn→∞ fn is Borel (Proposition 1.4). The following example shows that (1.15) is not always true.

Example 1.7. Consider (R, B) and the Lebesgue measure. Define fn (x) = nI[0,n−1 ] (x), n = 1, 2, .... Then limn→∞ fn (x) = 0 for all x but x = 0. Since the Lebesgue measure Rof a single point set is 0 (see Example 1.2), lim R n→∞ fn (x) = 0 a.e. and limn→∞ fn (x)dx R = 0. On the other hand, fn (x)dx = 1 for any n and, hence, limn→∞ fn (x)dx = 1.

13

1.2. Integration and Differentiation

The following result gives sufficient conditions under which (1.15) holds. Theorem 1.1. Let f1 , f2 , ... be a sequence of Borel functions on (Ω, F , ν). (i) (Fatou’s lemma). If fn ≥ 0, then Z Z lim inf fn dν ≤ lim inf fn dν. n

n

(ii) (Dominated convergence theorem). If limn→∞ fn = f a.e. and there exists an integrable function g such that |fn | ≤ g a.e., then (1.15) holds. (iii) (Monotone convergence theorem). If 0 ≤ f1 ≤ f2 ≤ · · · and limn→∞ fn = f a.e., then (1.15) holds. Proof. The results in (i) and (iii) are equivalent (exercise). R Applying Fatou’s Rlemma to functions g + f and g − f , we obtain that (g + f )dν ≤ n n R R lim inf n R(g + fn )dν and (g − fR )dν ≤ lim inf n (g − fn )dν (which is the same as (f − g)dνR ≥ lim supn (fn − R g)dν). Since g is Rintegrable,R these results imply that f dν ≤ lim inf n fn dν ≤ lim supn fn dν ≤ f dν. Hence, the result in (i) implies the result in (ii). It remains to show part (iii). Let f, f1 , fR2 , ... be given R in part (iii). From Proposition 1.6(i), there exists limn→∞ fn dν ≤ f dν. Let ϕ be a simple function with 0 ≤ R ϕ ≤ f and let Aϕ = {ϕ > 0}. Suppose that ν(Aϕ ) = ∞. Then f dν = ∞. Let a = 2−1 minω∈Aϕ ϕ(ω) and An = {fn > a}. Then a > 0, A1 ⊂ A2 ⊂ · · ·, and Aϕ ⊂ ∪AnR (why?). By R Proposition 1.1, ν(An ) → ν(∪An ) ≥ ν(Aϕ ) = ∞ and, hence, fn dν ≥ An fn dν ≥ aν(An ) → ∞. Suppose now ν(Aϕ ) < ∞. By Egoroff’s theorem (Exercise 20 in §1.6), for any ǫ > 0, there is B ⊂ AϕR with ν(B) R < ǫ such that fn converges to f uniformly on Aϕ ∩ B c . Hence, fn dν ≥ Aϕ ∩B c fn dν → R R R R R = ϕdν − B ϕdν ≥ ϕdν − ǫ maxω ϕ(ω). Since Aϕ ∩B c f dν ≥ Aϕ ∩B c ϕdν R R ǫ is arbitrary, lim R n→∞ fRn dν ≥ ϕdν. Since ϕ is arbitrary, by Definition 1.4(b), limn→∞ fn dν ≥ f dν. This completes the proof.

Example 1.8 (Interchange of differentiation and integration). Let (Ω, F , ν) be a measure space and, for any fixed θ ∈ R, let f (ω, θ) be a Borel function on Ω. Suppose that ∂f (ω, θ)/∂θ exists a.e. for θ ∈ (a, b) ⊂ R and that |∂f (ω, θ)/∂θ| ≤ g(ω) a.e., where g is an integrable function on Ω. Then, for each θ ∈ (a, b), ∂f (ω, θ)/∂θ is integrable and, by Theorem 1.1(ii), Z Z d ∂f (ω, θ) f (ω, θ)dν = dν. dθ ∂θ

Theorem 1.2 (Change of variables). Let f be measurable from (Ω, F, ν) to (Λ, G) and g be Borel on (Λ, G). Then Z Z g ◦ f dν = gd(ν ◦ f −1 ), (1.16) Ω

Λ

i.e., if either integral exists, then so does the other, and the two are the same.

14

1. Probability Theory

The reader is encouraged to provide a proof. A complete proof is in Billingsley (1986, p. 219). This the change of variable formula R result extends R for Riemann integrals, i.e., g(y)dy = g(f (x))f ′ (x)dx, y = f (x). Result (1.16) is very important in probability and statistics. R Let X be a random variable on a probability space (Ω, F, P ). IfREX = Ω XdP exists, then usually it is much simpler to compute EX = R xdPX , where PX = P ◦ X −1 is the law of X. Let Y be a random vector from Ω to Rk and k Rg be Borel from RR to R. According to (1.16), Eg(Y ) can be computed as g(y)dP or Y Rk R xdPg(Y ) , depending on which of PY and Pg(Y ) is easier to handle. As a more specific example, consider k = 2, Y = (X1 , X2 ), and g(Y ) = X1 + X2 . Using Proposition 1.5(ii), E(X1 + X2 ) = EX1 + EX2 R R and, hence, E(X1 + X2 ) = R xdPX1 + R xdPX2 . Then we need to handle Rtwo integrals involving PX1 and PX2 . On the other hand, E(X1 + X2 ) = R xdPX1 +X2 , which involves one integral w.r.t. PX1 +X2 . Unless we have some knowledge about the joint c.d.f. of (X1 , X2 ), it is not easy to obtain PX1 +X2 . The following theorem states how to evaluate an integral w.r.t. a product measure via iterated integration. The reader is encouraged to prove this theorem. A complete proof can be found in Billingsley (1986, pp. 236-238). Theorem 1.3 (Fubini’s theorem). Let νi be a σ-finite measure on (Ωi , Fi ), Q2 i = 1, 2, and let f be a Borel function on i=1 (Ωi , Fi ). Suppose that either f ≥ 0 or f is integrable w.r.t. ν1 × ν2 . Then Z g(ω2 ) = f (ω1 , ω2 )dν1 Ω1

exists a.e. ν2 and defines a Borel function on Ω2 whose integral w.r.t. ν2 exists, and Z Z Z f (ω1 , ω2 )dν1 × ν2 = f (ω1 , ω2 )dν1 dν2 . Ω1 ×Ω2

Ω2

Ω1

This result Q can be naturally extended to the integral w.r.t. the product measure on ki=1 (Ωi , Fi ) for any finite positive integer k.

Example 1.9. Let Ω1 = Ω2 = {0, 1, 2, ...}, and ν1 = ν2 be the counting measure (Example 1.1). A function f on Ω1 ×Ω2 defines a double sequence. R If f ≥ 0 or |f |dν1 × ν2 < ∞, then Z ∞ ∞ ∞ X ∞ X X X f (i, j) = f (i, j) (1.17) f dν1 × ν2 = i=0 j=0

j=0 i=0

(by Theorem 1.3 and Example 1.5). Thus, a double series can be summed in either order, if it is summable or f ≥ 0.

15

1.2. Integration and Differentiation

1.2.2 Radon-Nikodym derivative Let (Ω, F, ν) be a measure space and f be a nonnegative Borel function. One can show that the set function Z λ(A) = f dν, A ∈ F , (1.18) A

is a measure on (Ω, F) (verify). Note that ν(A) = 0 implies

λ(A) = 0.

(1.19)

If (1.19) holds for two measures λ and ν defined on the same measurable space, then we say λ is absolutely continuous w.r.t. ν and write λ ≪ ν. Formula (1.18) gives us not only a way of constructing measures, but also a method of computing measures of measurable sets. Let ν be a wellknown measure (such as the Lebesgue measure or the counting measure) and λ a relatively unknown measure. If we can find a function f such that (1.18) holds, then computing λ(A) can be done through integration. A necessary condition for (1.18) is clearly λ ≪ ν. The following result shows that λ ≪ ν is also almost sufficient for (1.18). Theorem 1.4 (Radon-Nikodym theorem). Let ν and λ be two measures on (Ω, F) and ν be σ-finite. If λ ≪ ν, then there exists a nonnegative Borel function f on RΩ such that (1.18) holds. Furthermore, f is unique a.e. ν, i.e., if λ(A) = A gdν for any A ∈ F, then f = g a.e. ν.

The proof of this theorem can be found in Billingsley (1986, pp. 443444). If (1.18) holds, then the function f is called the Radon-Nikodym derivative or density of λ w.r.t. ν and is denoted by dλ/dν. R A useful consequence of Theorem 1.4 is that if f is Borel on (Ω, F ) and f dν = 0 for any A ∈ F, then f = 0 a.e. A R If f dν = 1 for an f ≥ 0 a.e. ν, then λ given by (1.18) is a probability measure and f is called its probability density function (p.d.f.) w.r.t. ν. For any probability measure P on (Rk , B k ) corresponding to a c.d.f. F or a random vector X, if P has a p.d.f. f w.r.t. a measure ν, then f is also called the p.d.f. of F or X w.r.t. ν.

Example 1.10 (p.d.f. of a discrete c.d.f.). Consider the discrete c.d.f. F in (1.10) of Example 1.3 with its probability measure given by (1.11). Let Ω = {a1 , a2 , ...} and ν be the counting measure on the power set of Ω. By Example 1.5, Z X P (A) = f dν = f (ai ), A ⊂ Ω, (1.20) A

ai ∈A

16

1. Probability Theory

where f (ai ) = pi , i = 1, 2, .... That is, f is the p.d.f. of P or F w.r.t. ν. Hence, any discrete c.d.f. has a p.d.f. w.r.t. counting measure. A p.d.f. w.r.t. counting measure is called a discrete p.d.f. Example 1.11. Let F be a c.d.f. Assume that F is differentiable in the usual sense in calculus. Let f be the derivative of F . From calculus, Z x F (x) = f (y)dy, x ∈ R. (1.21) −∞

Let P be theR probability measure corresponding to F . It can be shown that P (A) = A f dm for any A ∈ B, where m is the Lebesgue measure on R. Hence, f is the p.d.f. of P or F w.r.t. Lebesgue measure. In this case, the Radon-Nikodym derivative is the same as the usual derivative of F in calculus. A continuous c.d.f. may not have a p.d.f. w.r.t. Lebesgue measure. A necessary and sufficient condition for a c.d.f. F having a p.d.f. w.r.t. Lebesgue measure is that F is absolute continuous in the sense that for any ǫ > 0, there exists a δ > 0 such collection of disjoint P that for each finite P bounded open intervals (ai , bi ), (bi −ai ) < δ implies [F (bi )−F (ai )] < ǫ. Absolute continuity is weaker than differentiability, but is stronger than continuity. Thus, any discontinuous c.d.f. (such as a discrete c.d.f.) is not absolute continuous. Note that every c.d.f. is differentiable a.e. Lebesgue measure (Chung, 1974, Chapter 1). Hence, if f is the p.d.f. of F w.r.t. Lebesgue measure, then f is the usual derivative of F a.e. Lebesgue measure and (1.21) holds. In such a case probabilities can be computed through integration. It can be shown that the uniform and exponential c.d.f.’s in Example 1.4 are absolute continuous and their p.d.f.’s are, respectively, ( 1 a≤x0 (1 − p)x−1 p, x = 1, 2, ... pet /[1 − (1 − p)et ], t < − log(1 − p) 1/p (1 − p)/p2 p ∈ [0, 1] N m (nx ) r−x r x = 0, 1, ..., min{r, n}, r − x ≤ m No explicit form rn/N rnm(N − r)/[N 2 (N − 1)] r, n, m = 1, 2, ..., N = n + m r x−1 x−r , x = r, r + 1, ... r−1 p (1 − p) r rt p e /[1 − (1 − p)et ]r , t < − log(1 − p) r/p r(1 − p)/p2 p ∈ [0, 1], r = 1, 2, ... −(log p)−1 x−1 (1 − p)x , x = 1, 2, ... log[1 − (1 − p)et ]/ log p, t ∈ R −(1 − p)/(p log p) −(1 − p)[1 + (1 − p)/ log p]/(p2 log p) p ∈ (0, 1)

All p.d.f.’s are w.r.t. counting measure.

19

1.3. Distributions and Their Characteristics

forms, whereas many others do not and they have to be evaluated numerically or computed using tables or software. There are p.d.f.’s that are neither discrete nor Lebesgue. Example 1.12. Let X be a random variable on (Ω, F , P ) whose c.d.f. FX has a Lebesgue p.d.f. fX and FX (c) < 1, where c is a fixed constant. Let Y = min{X, c}, i.e., Y is the smaller of X and c. Note that Y −1 ((−∞, x]) = Ω if x ≥ c and Y −1 ((−∞, x]) = X −1 ((∞, x]) if x < c. Hence Y is a random variable and the c.d.f. of Y is 1 x≥c FY (x) = x < c. FX (x) This c.d.f. is discontinuous at c, since FX (c) < 1. Thus, it does not have a Lebesgue p.d.f. It is not discrete either. Does PY , the probability measure corresponding to FY , have a p.d.f. w.r.t. some measure? Define a probability measure on (R, B), called point mass at c, by 1 c∈A διc (A) = A∈B (1.22) 0 c 6∈ A, (which is a special case of the discrete uniform distribution in Table 1.1). Then PY ≪ m + διc , where m is the Lebesgue measure, and the p.d.f. of PY is x>c 0 dPY (x) = (1.23) 1 − FX (c) x=c d(m + διc ) x < c. fX (x)

A p.d.f. corresponding to a joint c.d.f. is called a joint p.d.f. The following is a joint Lebesgue p.d.f. on Rk that is important in statistics: f (x) = (2π)−k/2 [Det(Σ)]−1/2 e−(x−µ) k

τ

Σ−1 (x−µ)/2

,

x ∈ Rk ,

(1.24)

where µ ∈ R , Σ is a positive definite k × k matrix, Det(Σ) is the determinant of Σ and, when matrix algebra is involved, any k-vector c is treated as a k × 1 matrix (column vector) and cτ denotes its transpose (row vector). The p.d.f. in (1.24) and its c.d.f. are called the k-dimensional multivariate normal p.d.f. and c.d.f., and both are denoted by Nk (µ, Σ). Random vectors distributed as Nk (µ, Σ) are also denoted by Nk (µ, Σ) for convenience. The normal distribution N (µ, σ 2 ) in Table 1.2 is a special case of Nk (µ, Σ) with k = 1. In particular, N (0, 1) is called the standard normal distribution. When Σ is a nonnegative definite but singular matrix, we define X to be Nk (µ, Σ) if and only if cτ X is N (cτ µ, cτ Σc) for any c ∈ Rk (N (a, 0) is defined to be the c.d.f. of the point mass at a), which is an important property of Nk (µ, Σ) with a nonsingular Σ (Exercise 81). Another important joint p.d.f. will be introduced in Example 2.7.

20

1. Probability Theory

Table 1.2. Distributions on R with Lebesgue p.d.f.’s Uniform U (a, b)

Normal 2

N (µ, σ )

Exponential E(a, θ)

Chi-square χ2k

Gamma Γ(α, γ)

Beta B(α, β)

Cauchy C(µ, σ)

p.d.f. m.g.f. Expectation Variance Parameter p.d.f.

(b − a)−1 I(a,b) (x) (ebt − eat )/[(b − a)t], t ∈ R (a + b)/2 (b − a)2 /12 a, b ∈ R, a < b 2 2 √ 1 e−(x−µ) /2σ 2πσ

m.g.f.

eµt+σ

Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter

µ σ2 µ ∈ R, σ > 0 θ−1 e−(x−a)/θ I(a,∞) (x) eat (1 − θt)−1 , t < θ−1 θ+a θ2 θ > 0, a ∈ R 1 xk/2−1 e−x/2 I(0,∞) (x) Γ(k/2)2k/2 −k/2 , t < 1/2 (1 − 2t) k 2k k = 1, 2, ... 1 α−1 −x/γ e I(0,∞) (x) Γ(α)γ α x −α (1 − γt) , t < γ −1 αγ αγ 2 γ > 0, α > 0 Γ(α+β) α−1 (1 − x)β−1 I(0,1) (x) Γ(α)Γ(β) x No explicit form α/(α + β) αβ/[(α + β + 1)(α + β)2 ] α > 0, β > 0 h i−1 x−µ 2 1 πσ 1 + σ

p.d.f. ch.f. Expectation Variance Parameter

√

2 2

t /2

, t∈R

e −1µt−σ|t| Does not exist Does not exist µ ∈ R, σ > 0

21

1.3. Distributions and Their Characteristics

Table 1.2. (continued) t-distribution tn

F-distribution Fn,m

Log-normal LN (µ, σ 2 )

Weibull W (α, θ)

DE(µ, θ) Pareto P a(a, θ)

Logistic LG(µ, σ)

x2 n

−(n+1)/2

Γ[(n+1)/2] √ nπΓ(n/2)

ch.f. Expectation Variance Parameter

No explicit form 0, (n > 1) n/(n − 2), (n > 2) n = 1, 2, ...

p.d.f. ch.f. Expectation Variance Parameter p.d.f. ch.f. Expectation Variance Parameter p.d.f. ch.f. Expectation Variance

Double Exponential

p.d.f.

Parameter p.d.f. m.g.f. Expectation Variance Parameter p.d.f. ch.f. Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter

1+

nn/2 mm/2 Γ[(n+m)/2]xn/2−1 I (x) Γ(n/2)Γ(m/2)(m+nx)(n+m)/2 (0,∞)

No explicit form m/(m − 2), (m > 2) 2m2 (n + m − 2)/[n(m − 2)2 (m − 4)], (m > 4) n = 1, 2, ..., m = 1, 2, ... 2 2 √ 1 x−1 e−(log x−µ) /2σ I(0,∞) (x) 2πσ No explicit form 2 eµ+σ /2 2 2 e2µ+σ (eσ − 1) µ ∈ R, σ > 0 α α−1 −xα /θ e I(0,∞) (x) θx No explicit form −1 + 1) θ1/α Γ(α n 2 o 2/α Γ(2α−1 + 1) − Γ(α−1 + 1) θ

θ > 0, α > 0 1 −|x−µ|/θ 2θ e eµt /(1 − θ2 t2 ), |t| < θ−1 µ 2θ2 µ ∈ R, θ > 0 θaθ x−(θ+1) I(a,∞) (x) No explicit form θa/(θ − 1), (θ > 1) θa2 /[(θ − 1)2 (θ − 2)], (θ > 2) θ > 0, a > 0 σ −1 e−(x−µ)/σ /[1 + e−(x−µ)/σ ]2 eµt Γ(1 + σt)Γ(1 − σt), |t| < σ −1 µ σ 2 π 2 /3 µ ∈ R, σ > 0

22

1. Probability Theory

If a random k-vector (X1 , ..., Xk ) has a joint p.d.f. f w.r.t. a product measure ν1 × · · · × νk defined on B k , then Xi has the following marginal p.d.f. w.r.t. νi : Z f (x1 , ..., xi−1 , x, xi+1 , ..., xk )dν1 · · · dνi−1 dνi+1 · · · dνk . fi (x) = Rk−1

Let F be the joint c.d.f. of a random k-vector (X1 , ..., Xk ) and Fi be the marginal c.d.f. of Xi , i = 1, ..., k. If (1.7) holds, then random variables X1 , ..., Xk are said to be independent. From the discussion in the end of §1.1.1, this independence means that the probability measure corresponding to F is the product measure of the k probability measures corresponding to Fi ’s. The meaning of independence is further discussed in §1.4.2. If (X1 , ..., Xk ) has a joint p.d.f. f w.r.t. a product measure ν1 × · · · × νk defined on B k , then X1 , ..., Xk are independent if and only if f (x1 , ..., xk ) = f1 (x1 ) · · · fk (xk ),

(x1 , ..., xk ) ∈ Rk ,

(1.25)

where fi is the p.d.f. of Xi w.r.t. νi , i = 1, ..., k. For example, using (1.24), one can show (exercise) that the components of Nk (µ, Σ) are independent if and only if Σ is a diagonal matrix. The following lemma is useful in considering the independence of functions of independent random variables. Lemma 1.1. Let X1 , ..., Xn be independent random variables. Then random variables g(X1 , ..., Xk ) and h(Xk+1 , ..., Xn ) are independent, where g and h are Borel functions and k is an integer between 1 and n. Lemma 1.1 can be proved directly (exercise). But it is a simple consequence of an equivalent definition of independence introduced in §1.4.2. Let X1 , ..., Xk be random variables. If Xi and Xj are independent for every pair i 6= j, then X1 , ..., Xk are said to be pairwise independent. If X1 , ..., Xk are independent, then clearly they are pairwise independent. However, the converse is not true. The following is an example. Example 1.13. Let X1 and X2 be independent random variables each assuming the values 1 and −1 with probability 0.5, and X3 = X1 X2 . Let Ai = {Xi = 1}, i = 1, 2, 3. Then P (Ai ) = 0.5 for any i and P (A1 )P (A2 )P (A3 ) = 0.125. However, P (A1 ∩A2 ∩A3 ) = P (A1 ∩A2 ) = P (A1 )P (A2 ) = 0.25. This implies that (1.7) does not hold and, hence, X1 , X2 , X3 are not independent. We now show that X1 , X2 , X3 are pairwise independent. It is enough to show that X1 and X3 are independent. Let Bi = {Xi = −1}, i = 1, 2, 3. Note that A1 ∩ A3 = A1 ∩ A2 , A1 ∩ B3 = A1 ∩ B2 , B1 ∩ A3 = B1 ∩ B2 , and B1 ∩ B3 = B1 ∩ A2 . Then the result follows from the fact that P (Ai ) = P (Bi ) = 0.5 for any i and X1 and X2 are independent.

23

1.3. Distributions and Their Characteristics

The random variable Y in Example 1.12 is a transformation of the random variable X. Transformations of random variables or vectors are frequently used in statistics. For a random variable or vector X, g(X) is a random variable or vector as long as g is measurable (Proposition 1.4). How do we find the c.d.f. (or p.d.f.) of g(X) when the c.d.f. (or p.d.f.) of X is known? In many cases, the most effective method is direct computation. Example 1.12 is one example. The following is another one. Example 1.14. Let X be a random variable with c.d.f. FX and Lebesgue −1 p.d.f. fX , and let Y = X 2 . √Since √ Y ((−∞, x]) is empty if x < 0 and −1 −1 equals Y ([0, x]) = X ([− x, x ]) if x ≥ 0, the c.d.f. of Y is FY (x) = P ◦ Y −1 ((−∞, x]) √ √ = P ◦ X −1 ([− x, x ]) √ √ = FX ( x) − FX (− x) if x ≥ 0 and FY (x) = 0 if x < 0. Clearly, the Lebesgue p.d.f. of FY is √ √ 1 fY (x) = √ [fX ( x) + fX (− x)]I(0,∞) (x). 2 x

(1.26)

In particular, if

2 1 fX (x) = √ e−x /2 , (1.27) 2π which is the Lebesgue p.d.f. of the standard normal distribution N (0, 1) (Table 1.2), then 1 fY (x) = √ e−x/2 I(0,∞) (x), 2πx

which is the Lebesgue p.d.f. for the chi-square distribution χ21 (Table 1.2). This is actually an important result in statistics. In some cases, one may apply the following general result whose proof is left to the reader. Proposition 1.8. Let X be a random k-vector with a Lebesgue p.d.f. fX and let Y = g(X), where g is a Borel function from (Rk , B k ) to (Rk , B k ). Let A1 , ..., Am be disjoint sets in B k such that Rk − (A1 ∪ · · · ∪ Am ) has Lebesgue measure 0 and g on Aj is one-to-one with a nonvanishing Jacobian, i.e., the determinant Det(∂g(x)/∂x) 6= 0 on Aj , j = 1, ..., m. Then Y has the following Lebesgue p.d.f.: fY (x) =

m X Det (∂hj (x)/∂x) fX (hj (x)) , j=1

where hj is the inverse function of g on Aj , j = 1, ..., m.

24

1. Probability Theory

One may apply Proposition 1.8 to obtain result (1.26) in Example 1.14, √ using A1 √ = (−∞, 0), A2 = (0, ∞), and√g(x) = x2 . Note that h1 (x) = − x, h2 (x) = x, and |dhj (x)/dx| = 1/(2 x). Another immediate application of Proposition 1.8 is to show that Y = AX is Nk (Aµ, AΣAτ ) when X is Nk (µ, Σ), where Σ is positive definite, A is a k × k matrix of rank k, and Aτ denotes the transpose of A. Example 1.15. Let X = (X1 , X2 ) be a random 2-vector having a joint Lebesgue p.d.f. fX . Consider first the transformation g(x) = (x1 , x1 + x2 ). Using Proposition 1.8, one can show that the joint p.d.f. of g(X) is fg(X) (x1 , y) = fX (x1 , y − x1 ), where y = x1 + x2 (note that the Jacobian equals 1). The marginal p.d.f. of Y = X1 + X2 is then Z fY (y) = fX (x1 , y − x1 )dx1 . In particular, if X1 and X2 are independent, then Z fY (y) = fX1 (x1 )fX2 (y − x1 )dx1 .

(1.28)

Next, consider the transformation h(x1 , x2 ) = (x1 /x2 , x2 ), assuming that X2 6= 0 a.s. Using Proposition 1.8, one can show that the joint p.d.f. of h(X) is fh(X) (z, x2 ) = |x2 |fX (zx2 , x2 ),

where z = x1 /x2 . The marginal p.d.f. of Z = X1 /X2 is Z fZ (z) = |x2 |fX (zx2 , x2 )dx2 . In particular, if X1 and X2 are independent, then Z fZ (z) = |x2 |fX1 (zx2 )fX2 (x2 )dx2 .

(1.29)

A number of results can be derived from (1.28) and (1.29). For example, if X1 and X2 are independent and both have the standard normal p.d.f. given by (1.27), then, by (1.29), the Lebesgue p.d.f. of Z = X1 /X2 is Z 2 2 1 |x2 |e−(1+z )x2 /2 dx2 fZ (z) = 2π Z 1 ∞ −(1+z2 )x = e dx π 0 1 , = π(1 + z 2 )

25

1.3. Distributions and Their Characteristics

which is the p.d.f. of the Cauchy distribution C(0, 1) in Table 1.2. Another application of formula (1.29) leads to the following important result in statistics. Example 1.16 (t-distribution and F-distribution). Let X1 and X2 be independent random variables having the chi-square distributions χ2n1 and χ2n2 (Table 1.2), respectively. By (1.29), the p.d.f. of Z = X1 /X2 is fZ (z) =

z n1 /2−1 I(0,∞) (z) 2(n1 +n2 )/2 Γ(n1 /2)Γ(n2 /2)

Z

0 n1 /2−1

=

∞

(n1 +n2 )/2−1 −(1+z)x2 /2

x2

e

dx2

z Γ[(n1 + n2 )/2] I(0,∞) (z), Γ(n1 /2)Γ(n2 /2) (1 + z)(n1 +n2 )/2

where the last equality follows from the fact that 1 (n +n )/2−1 −x2 /2 x 1 2 e I(0,∞) (x2 ) 2(n1 +n2 )/2 Γ[(n1 + n2 )/2] 2 is the p.d.f. of the chi-square distribution χ2n1 +n2 . Using Proposition 1.8, one can show that the p.d.f. of Y = (X1 /n1 )/(X2 /n2 ) = (n2 /n1 )Z is the p.d.f. of the F-distribution Fn1 ,n2 given in Table 1.2. Let U1 be a random variable having the standard normal distribution N (0, 1) and U2 a random variable having the chi-square distribution χ2n . Using the same argument, one canp show that if U1 and U2 are independent, then the distribution of T = U1 / U2 /n is the t-distribution tn given in Table 1.2. This result can also be derived using the result given in this example as follows. Let X1 = U12 and X2 = U2 . Then X1 and X2 are independent (which can be shown directly but follows from Lemma 1.1). By Example 1.14, the distribution of X1 is χ21 . Then Y = X1 /(X2 /n) has the F-distribution F1,n and its Lebesgue p.d.f. is nn/2 Γ[(n + 1)/2]x−1/2 √ I(0,∞) (x). πΓ(n/2)(n + x)(n+1)/2 Note that

√ Y √ T = − Y

U1 ≥ 0 U1 < 0.

The result follows from Proposition 1.8 and the fact that P ◦ T −1 ((−∞, −t]) = P ◦ T −1 ([t, ∞)) ,

t > 0.

(1.30)

If a random variable T satisfies (1.30), i.e., T and −T have the same distribution, then T and its c.d.f. and p.d.f. (if it exists) are said to be

26

1. Probability Theory

symmetric about 0. If T has a Lebesgue p.d.f. fT , then T is symmetric about 0 if and only if fT (x) = fT (−x) for any x > 0. T and its c.d.f. and p.d.f. are said to be symmetric about a (or symmetric for simplicity) if and only if T − a is symmetric about 0 for a fixed a ∈ R. The c.d.f.’s of t-distributions are symmetric about 0 and the normal, Cauchy, and double exponential c.d.f.’s are symmetric. The chi-square, t-, and F-distributions in the previous examples are special cases of the following noncentral chi-square, t-, and F-distributions, which are useful in some statistical problems. Let X1 , ..., Xn be independent random variables and Xi = N (µi , σ 2 ), i = 1, ..., n. The distribution of the random variable Y = (X12 +· · ·+Xn2 )/σ 2 is called the noncentral chi-square distribution and denoted by χ2n (δ), where δ = (µ21 + · · · + µ2n )/σ 2 is the noncentrality parameter. The chi-square distribution χ2k in Table 1.2 is a special case of the noncentral chi-square distribution χ2k (δ) with δ = 0 and, therefore, is called a central chi-square distribution. It can be shown (exercise) that Y has the following Lebesgue p.d.f.: ∞ X (δ/2)j e−δ/2 f2j+n (x), (1.31) j! j=0 where fk (x) is the Lebesgue p.d.f. of the chi-square distribution χ2k . It follows from the definition of noncentral chi-square distributions that if Y1 , ..., Yk are independent random variables and Yi has the noncentral chisquare distribution χ2ni (δi ), i = 1, ..., k, then Y = Y1 + · · · + Yk has the noncentral chi-square distribution χ2n1 +···+nk (δ1 + · · · + δk ). The result for the t-distribution in Example 1.16 can be extended to the case where U1 has a nonzero expectation µ (U2 still has thepχ2n distribution and is independent of U1 ). The distribution of T = U1 / U2 /n is called the noncentral t-distribution and denoted by tn (δ), where δ = µ is the noncentrality parameter. Using the same argument as that in Example 1.15, one can show (exercise) that T has the following Lebesgue p.d.f.: Z ∞ √ 1 (n−1)/2 −[(x y/n−δ)2 +y]/2 √ y e dy. (1.32) 2(n+1)/2 Γ(n/2) πn 0 The t-distribution tn in Example 1.16 is called a central t-distribution, since it is a special case of the noncentral t-distribution tn (δ) with δ = 0. Similarly, the result for the F-distribution in Example 1.16 can be extended to the case where X1 has the noncentral chi-square distribution χ2n1 (δ), X2 has the central chi-square distribution χ2n2 , and X1 and X2 are independent. The distribution of Y = (X1 /n1 )/(X2 /n2 ) is called the noncentral F-distribution and denoted by Fn1 ,n2 (δ), where δ is the noncentrality parameter. The F-distribution Fn1 ,n2 in Example 1.16 is called a central

27

1.3. Distributions and Their Characteristics

F-distribution, since it is a special case of the noncentral F-distribution Fn1 ,n2 (δ) with δ = 0. It can be shown (exercise) that the noncentral Fdistribution Fn1 ,n2 (δ) has the following Lebesgue p.d.f.: ∞ X n1 (δ/2)j n1 x f2j+n1 ,n2 , (1.33) e−δ/2 j!(2j + n1 ) 2j + n1 j=0 where fk1 ,k2 (x) is the Lebesgue p.d.f. of the central F-distribution Fk1 ,k2 given in Table 1.2. Using some results from linear algebra, we can prove the following result useful in analysis of variance (Scheff´e, 1959; Searle, 1971). Theorem 1.5. (Cochran’s theorem). Suppose that X = Nn (µ, In ) and X τ X = X τ A1 X + · · · + X τ Ak X,

(1.34)

where In is the n × n identity matrix and Ai is an n × n symmetric matrix with rank ni , i = 1, ..., k. A necessary and sufficient condition that X τ Ai X has the noncentral chi-square distribution χ2ni (δi ), i = 1, ..., k, and X τ Ai X’s are independent is n = n1 + · · · + nk , in which case δi = µτ Ai µ and δ1 + · · · + δk = µτ µ. Proof. Suppose that X τ Ai X, i = 1, ..., k, are independent and X τ Ai X has the χ2ni (δi ) distribution. Then X τ X has the χ2n1 +···+nk (δ1 + · · · + δk ) distribution. By definition, X τ X has the noncentral chi-square distribution χ2n (µτ µ). By (1.34), n = n1 + · · · + nk and δ1 + · · · + δk = µτ µ. Suppose now that n = n1 + · · · + nk . From linear algebra, for each i there exists cij ∈ Rn , j = 1, ..., ni , such that X τ Ai X = ±(cτi1 X)2 ± · · · ± (cτini X)2 .

(1.35)

Let Ci be the n × ni matrix whose jth column is cij , and C τ = (C1 , ..., Ck ). By (1.34) and (1.35), X τ X = X τ C τ ∆CX with an n × n diagonal matrix ∆ whose diagonal elements are either 1 or −1. This implies C τ ∆C = In . Thus, C is of full rank and, hence, ∆ = (C τ )−1 C −1 , which is positive definite. This shows ∆ = In , which implies C τ C = In and n1 +···+ni−1 +ni

X τ Ai X =

X

Yj2 ,

(1.36)

j=n1 +···+ni−1 +1

where Yj is the jth component of Y = CX. Note that Y = Nn (Cµ, In ) (Exercise 43). Hence Yj ’s are independent and Yj = N (λj , 1), where λj is the jth component of Cµ. This shows that X τ Ai X has the χ2ni (δi ) distribution with δi = λ2n1 +···+ni−1 +1 + · · · + λ2n1 +···+ni−1 +ni . Letting X = µ in (1.36) and (1.34), we obtain that δi = µτ Ai µ and δ1 + · · · + δk = µτ C τ Cµ = µτ µ. Finally, from (1.36) and Lemma 1.1, we conclude that X τ Ai X, i = 1, ..., k, are independent.

28

1. Probability Theory

1.3.2 Moments and moment inequalities We have defined the expectation of a random variable in §1.2.1. It is an important characteristic of a random variable. In this section, we introduce moments, which are some other important characteristics of a random variable or vector. Let X be a random variable. If EX k is finite, where k is a positive integer, then EX k is called the kth moment of X or PX (the distribution of X). If E|X|a < ∞ for some real number a, then E|X|a is called the ath absolute moment of X or PX . If µ = EX and E(X − µ)k are finite for a positive integer k, then E(X − µ)k is called the kth central moment of X or PX . If E|X|a < ∞ for an a > 0, then E|X|t < ∞ for any positive t < a and EX k is finite for any positive integer k ≤ a (Exercise 54). The expectation and the second central moment (if they exist) are two important characteristics of a random variable (or its distribution) in statistics. They are listed in Tables 1.1 and 1.2 for those useful distributions. The expectation, also called the mean in statistics, is a measure of the central location of the distribution of a random variable. The second central moment, also called the variance in statistics, is a measure of dispersion or spread of a random variable. The variance of a random variable X is denoted by Var(X). The variance is always nonnegative. If the variance of X is 0, then X is equal to its mean a.s. (Proposition 1.6). The squared root of the variance is called the standard deviation, another important characteristic of a random variable in statistics. The concept of mean and variance can be extended to random vectors. The expectation of a random matrix M with (i, j)th element Mij is defined to be the matrix whose (i, j)th element is EMij . Thus, for a random kvector X = (X1 , ..., Xk ), its mean is EX = (EX1 , ..., EXk ). The extension of variance is the variance-covariance matrix of X defined as Var(X) = E(X − EX)(X − EX)τ , which is a k × k symmetric matrix whose diagonal elements are variances of Xi ’s. The (i, j)th element of Var(X), i 6= j, is E(Xi − EXi )(Xj − EXj ), which is called the covariance of Xi and Xj and is denoted by Cov(Xi , Xj ). Let c ∈ Rk and X = (X1 , ..., Xk ) be a random k-vector. Then Y = τ c X is a random variable and, by Proposition 1.5 (linearity of integrals), EY = cτ EX if EX exists. Also, when Var(X) is finite (i.e., all elements of Var(X) are finite), Var(Y ) = E(cτ X − cτ EX)2 = E[cτ (X − EX)(X − EX)τ c] = cτ [E(X − EX)(X − EX)τ ]c = cτ Var(X)c.

29

1.3. Distributions and Their Characteristics

Since Var(Y ) ≥ 0 for any c ∈ Rk , the matrix Var(X) is nonnegative definite. Consequently, [Cov(Xi , Xj )]2 ≤ Var(Xi )Var(Xj ),

i 6= j.

(1.37)

An important quantity in p statistics is the correlation coefficient defined to be ρXi ,Xj = Cov(Xi , Xj )/ Var(Xi )Var(Xj ), which, by inequality (1.37), is always between −1 and 1. It is a measure of relationship between Xi and Xj ; if ρXi ,Xj is positive (or negative), then Xi and Xj tend to be positively (or negatively) related; if ρXi ,Xj = ±1, then P (Xi = c1 ± c2 Xj ) = 1 with some constants c1 and c2 > 0; if ρXi ,Xj = 0 (i.e., Cov(Xi , Xj ) = 0), then Xi and Xj are said to be uncorrelated. If Xi and Xj are independent, then they are uncorrelated. This follows from the following more general result. If X1 , ..., Xn are independent random variables and E|X1 · · · Xn | < ∞, then, by Fubini’s theorem and the fact that the joint c.d.f. of (X1 , ..., Xn ) corresponds to a product measure, we obtain that E(X1 · · · Xn ) = EX1 · · · EXn .

(1.38)

In fact, pairwise independence of X1 , ..., Xn implies that Xi ’s are uncorrelated, since Cov(Xi , Xj ) involves only a pair of random variables. However, the converse is not necessarily true: uncorrelated random variables may not be pairwise independent. Examples can be found in Exercises 60-61. Let RM = {y ∈ Rk : y = M x with some x ∈ Rk } for any k × k symmetric matrix M . If a random k-vector X has a finite Var(X), then P (X − EX ∈ RVar(X) ) = 1. This means that if the rank of Var(X) is r < k, then X is in a subspace of Rk with dimension r. Consequently, if PX ≪ Lebesgue measure on Rk , then the rank of Var(X) is k. Example 1.17. Let X be a random k-vector having the Nk (µ, Σ) distribution. It can be shown (exercise) that EX = µ and Var(X) = Σ. Thus, µ and Σ in (1.24) are the mean vector and the variance-covariance matrix of X. If Σ is a diagonal matrix (i.e., all components of X are uncorrelated), then by (1.25), the components of X are independent. This shows an important property of random variables having normal distributions: they are independent if and only if they are uncorrelated. There are many useful inequalities related to moments. The inequality in (1.37) is in fact the well-known Cauchy-Schwartz inequality whose general form is [E(XY )]2 ≤ EX 2 EY 2 , (1.39)

where X and Y are random variables with a well-defined E(XY ). Inequality (1.39) is a special case of the following H¨ older’s inequality: E|XY | ≤ (E|X|p )1/p (E|Y |q )1/q ,

(1.40)

30

1. Probability Theory

where p and q are constants satisfying p > 1 and p−1 + q −1 = 1. To show inequality (1.40), we use the following inequality (Exercise 62): xt y 1−t ≤ tx + (1 − t)y,

(1.41)

where x and y are nonnegative real numbers and t ∈ (0, 1). If either E|X|p or E|Y |q is ∞, then (1.40) holds. Hence we can assume that both E|X|p and E|Y |q are finite. Let a = (E|X|p )1/p and b = (E|Y |q )1/q . If either a = 0 or b = 0, then the equality in (1.40) holds because of Proposition 1.6(ii). Assume now a 6= 0 and b 6= 0. Letting x = |X/a|p , y = |Y /b|q , and t = p−1 in (1.41), we obtain that XY |X|p |Y |q ≤ + . ab pap qbq Taking expectations on both sides of this expression, we obtain that E|XY | E|X|p E|Y |q 1 1 ≤ + = + = 1, ab pap qaq p q which is (1.40). In fact, the equality in (1.40) holds if and only if α|X|p = β|Y |q a.s. for some nonzero constants α and β (Exercise 62). Using H¨older’s inequality, we can prove Liapounov’s inequality (E|X|r )1/r ≤ (E|X|s )1/s ,

(1.42)

where r and s are constants satisfying 1 ≤ r ≤ s, and Minkowski’s inequality (E|X + Y |p )1/p ≤ (E|X|p )1/p + (E|Y |p )1/p , (1.43) where X and Y are random variables and p is a constant larger than or equal to 1 (Exercise 63). Minkowski’s inequality can be extended to the case of more than two random variables (Exercise 63). The following inequality is a tightened form of Minkowski’s inequality due to Esseen and von Bahr (1965). Let X1 , ..., Xn be independent random variables with mean 0 and E|Xi |p < ∞, i = 1, ..., n, where p is a constant in [1, 2]. Then X p n X n E Xi ≤ Cp E|Xi |p , i=1

(1.44)

i=1

where Cp is a constant depending only on p. When 1 < p < 2, inequality (1.44) can be proved (Exercise 63) using inequality |a + b|p ≤ |a|p + psgn(a)|a|p−1 b + Cp |b|p ,

a ∈ R, b ∈ R,

31

1.3. Distributions and Their Characteristics

where sgn(x) is 1 or −1 as x is positive or negative and Cp =

sup (|1 + x|p − 1 − px)/|x|p .

x∈R,x6=0

For p ≥ 2, there is a similar inequality due to Marcinkiewicz and Zygmund: n n X p Cp X Xi ≤ 1−p/2 E|Xi |p , E n i=1 i=1

(1.45)

where Cp is a constant depending only on p. A proof of inequality (1.45) can be found in Lo`eve (1977, p. 276). Recall from calculus that a subset A of Rk is convex if and only if x ∈ A and y ∈ A imply tx + (1 − t)y ∈ A for any t ∈ [0, 1]; a function f from a convex A ⊂ Rk to R is convex if and only if f (tx + (1 − t)y) ≤ tf (x) + (1 − t)f (y),

x ∈ A, y ∈ A, t ∈ [0, 1];

(1.46)

and f is strictly convex if and only if (1.46) holds with ≤ replaced by the strict inequality . By Proposition 1.6(ii), Ef (X) > f (EX) unless P (f (X) = Ef (X)) = 1. Example 1.18. A direct application of Jensen’s inequality (1.47) is that if X is a nonconstant positive random variable with finite mean, then (EX)−1 < E(X −1 )

and

E(log X) < log(EX),

32

1. Probability Theory

since t−1 and − log t are convex functions on (0, ∞). Another application is to prove the following inequality related to entropy. Let f and g be positive integrable functions on a measure space with a σ-finite measure ν. R R If f dν ≥ gdν > 0, then one can show (exercise) that Z f f log dν ≥ 0. (1.49) g The next inequality, Chebyshev’s inequality, is almost trivial but very useful and famous. Let X be a random variable and ϕ a nonnegative and nondecreasing function on [0, ∞) satisfying ϕ(−t) = ϕ(t). Then, for each constant t ≥ 0, Z ϕ(t)P (|X| ≥ t) ≤ ϕ(X)dP ≤ Eϕ(X), (1.50) {|X|≥t}

where both inequalities in (1.50) follow from Proposition 1.6(i) and the first inequality also uses the fact that on the set {|X| ≥ t}, ϕ(X) ≥ ϕ(t). The most familiar application of (1.50) is when ϕ(t) = |t|p for p ∈ (0, ∞), in which case inequality (1.50) is also called Markov’s inequality. Chebyshev’s inequality, sometimes together with one of the moment inequalities introduced in this section, can be used to yield a desired upper bound for the “tail” probability P (|X| ≥ t). For example, let Y be a random variable with mean µ and variance σ 2 . Then X = (Y − µ)/σ has mean 0 and variance 1 and, by (1.50) with ϕ(t) = t2 , P (|X| ≥ 2) ≤ 41 . This means that the probability that the random variable |Y − µ| exceeds twice its standard deviation is bounded by 14 . Similarly, we can also claim that the probability of |Y − µ| exceeding 3σ is bounded by 19 . These bounds are rough but they can be applied to any random variable with a finite variance. Other applications of Chebyshev’s inequality can be found in §1.5. In some cases, we need an improvement over inequality (1.50) when X is of some special form. Let Y1 , ..., Yn be independent random variables having finite variances. The following inequality is due to H´ajek and R`enyi: ! l n X 1 X 2 P max cl (Yi − EYi ) > t ≤ 2 c Var(Yi ), t > 0, (1.51) 1≤l≤n t i=1 i i=1 where ci ’s are positive constants satisfying c1 ≥ c2 ≥ · · · ≥ cn . If ci = 1 for all i, then inequality (1.51) reduces to the famous Kolmogorov’s inequality. A proof for (1.51) is given in Sen and Singer (1993, pp. 65-66).

1.3.3 Moment generating and characteristic functions Moments are important characteristics of a distribution, but they do not determine a distribution in the sense that two different distributions may

33

1.3. Distributions and Their Characteristics

have the same moments of all orders. Functions that determine a distribution are introduced in the following definition. Definition 1.5. Let X be a random k-vector. (i) The moment generating function (m.g.f.) of X or PX is defined as τ

ψX (t) = Eet

X

,

t ∈ Rk .

(ii) The characteristic function (ch.f.) of X or PX is defined as √ √ τ φX (t) = Ee −1t X = E[cos(tτ X)] + −1 E[sin(tτ X)], t ∈ Rk . Obviously ψX (0) = φX (0) = 1 for any random vector X. The ch.f. is complex-valued and always well defined. In fact, any ch.f. is bounded by 1 and is a uniformly continuous function on Rk (exercise). The m.g.f. is nonnegative but may be ∞ everywhere except at t = 0 (Example 1.19). If the m.g.f. is finite in a neighborhood of 0 ∈ Rk , then φX (t) can be obtained √ by replacing t in ψX (t) by −1t. Tables 1.1 and 1.2 contain the m.g.f. (or ch.f. when the m.g.f. is ∞ everywhere except at 0) for distributions useful in statistics. For a linear transformation Y = Aτ X + c, where A is a k × m matrix and c ∈ Rm , it follows from Definition 1.5 that τ

ψY (u) = ec u ψX (Au) and φY (u) = e

√ −1cτ u

u ∈ Rm . (1.52) For a random variable X, if its m.g.f. is finite at t and −t for a t 6= 0, then X has finite moments and absolute moments of any order. To compute moments of X using its m.g.f., a condition stronger than the finiteness of the m.g.f. at some t 6= 0 is needed. Consider a random k-vector X. If ψX is finite in a neighborhood of 0, then µr1 ,...,rk = E(X1r1 · · · Xkrk ) is finite for any nonnegative integers r1 , ..., rk , where Xj is the jth component of X, and ψX has the power series expansion ψX (t) =

X

(r1 ,...,rk )∈Z

φX (Au),

µr1 ,...,rk tr11 · · · trkk r1 ! · · · rk !

(1.53)

for t in the neighborhood of 0, where tj is the jth component of t and Z ⊂ Rk containing vectors whose components are nonnegative integers. Consequently, the components of X have finite moments of all orders and ∂ r1 +···+rk ψX (t) rk r1 E(X1 · · · Xk ) = , ∂tr11 · · · ∂trkk t=0 which are also called moments of X. In particular, ∂ψX (t) ∂ 2 ψX (t) = EX, = E(XX τ ), ∂t t=0 ∂t∂tτ t=0

(1.54)

34

1. Probability Theory (p)

and, when k = 1 and p is a positive integer, ψX (0) = EX p , where g (p) (t) denotes the pth order derivative of a function g(t). If 0 < ψX (t) < ∞, then κX (t) = log ψX (t) is called the cumulant generating function of X or PX . If 0 < ψX (t) < ∞ for t in a neighborhood of 0, then κX has a power series expansion similar to that in (1.53): X

κX (t) =

(r1 ,...,rk )∈Z

κr1 ,...,rk tr11 · · · trkk , r1 ! · · · rk !

(1.55)

where κr1 ,...,rk ’s are called cumulants of X. There is a one-to-one correspondence between the set of moments and the set of cumulants. An example for the case of k = 1 is given in Exercise 68. When ψX is not finite, finite moments of X can be obtained by differentiating its ch.f. φX . Suppose that E|X1r1 · · · Xkrk | < ∞ for some nonnegative integers r1 , ..., rk . Let r = r1 + · · · + rk and √

τ

√ τ ∂ r e −1t X r/2 r1 X1 · · · Xkrk e −1t X . g(t) = r1 rk = (−1) ∂t1 · · · ∂tk

Then |g(t)| ≤ |X1r1 · · · Xkrk |, which is integrable. Hence, from Example 1.8, √ τ ∂ r φX (t) r/2 E X1r1 · · · Xkrk e −1t X r1 rk = (−1) ∂t1 · · · ∂tk

and

In particular,

(1.56)

∂ r φX (t) = (−1)r/2 E(X1r1 · · · Xkrk ). ∂tr11 · · · ∂trkk t=0

√ ∂φX (t) = −1EX, ∂t t=0

∂ 2 φX (t) = −E(XX τ ), ∂t∂tτ t=0 (p)

and, if k = 1 and p is a positive integer, then φX (0) = (−1)p/2 EX p , provided that all moments involved are finite. In fact, when k = 1, if φX has a finite derivative of even order p at t = 0, then EX p < ∞ (see, e.g., Chung, 1974, pp. 166-168). 2 2

Example 1.19. Let X = N (µ, σ 2 ). From Table 1.2, ψX (t) = eµt+σ t /2 . A ′ ′′ direct calculation shows that EX = ψX (0) = µ, EX 2 = ψX (0) = σ 2 + µ2 , (3) (4) EX 3 = ψX (0) = 3σ 2 µ+µ3 , and EX 4 = ψX (0) = 3σ 4 +6σ 2 µ2 +µ4 . If µ = 0, then EX p = 0 when p is an odd integer and EX p = (p−1)(p−3) · · · 3·1σ p when p is an even integer (exercise). The cumulant generating function of X is κX (t) = log ψX (t) = µt + σ 2 t2 /2. Hence, κ1 = µ, κ2 = σ 2 , and κr = 0 for r = 3, 4, ....

35

1.3. Distributions and Their Characteristics

We now find a random variable having finite moments of all order but having an m.g.f. = ∞ except for t = 0. Let Pn be the probability meaP∞ sure for the N (0, σn2 ) distribution, n = 1, 2, .... Then P = n=1 2−n Pn is a probability measure (Exercise 35). Let X be a random variable hav2 2 ing distribution P . Since the m.g.f. of N (0, σn2 ) is eσn t /2 , it follows from P∞ −n σ2 t2 /2 . When Fubini’s theorem that the m.g.f. of X is ψX (t) = n=1 2 e n σn2 = n2 ,PψX (t) = ∞ for any t 6= 0 but EX k = 0 for any odd integer k and ∞ EX k = n=1 2−n (k − 1)(k − 3) · · · 1nk < ∞ for any even integer k. When √ 2 σn2 = n, ψX (t) = (2e−t /2 − 1)−1 for |t| < log 4 and, hence, the moments of X can be obtained by differentiating ψX . For example, EX = φ′X (0) = 0 and EX 2 = φ′′X (0) = 2. A fundamental fact about ch.f.’s is that there is a one-to-one correspondence between the set of all distributions on Rk and the set of all ch.f.’s defined on Rk . The same fact is true for m.g.f.’s, but we have to focus on distributions having m.g.f.’s finite in neighborhoods of 0. Theorem 1.6. (Uniqueness). Let X and Y be random k-vectors. (i) If φX (t) = φY (t) for all t ∈ Rk , then PX = PY . (ii) If ψX (t) = ψY (t) < ∞ for all t in a neighborhood of 0, then PX = PY . Proof. (i) The result follows from the following inversion formula whose proof can be found, for example, in Billingsley (1986, p. 395): for any a = (a1 , ..., ak ) ∈ Rk , b = (b1 , ..., bk ) ∈ Rk , and (a, b] = (a1 , b1 ] × · · · × (ak , bk ] satisfying PX (the boundary of (a, b]) = 0, √ √ Z c Z c k φX (t1 , ..., tk ) Y e− −1ti ai − e− −1ti bi ··· dti . PX (a, b] = lim k/2 (2π)k c→∞ −c ti −c (−1) i=1

(ii) First consider the case of k = 1. From es|x| ≤ esx + e−sx , we conclude that |X| has an m.g.f. that is finite in the neighborhood (−c, c) for some c >√0 and |X| has finite moments of all order. Using the inequality √ Pn √ |e −1tx [e −1ax − j=0 ( −1ax)j /j!]| ≤ |ax|n+1 /(n + 1)!, we obtain that n X √ |a|n+1 E|X|n+1 √ aj j −1tX φX (t + a) − E[( , −1X) e ] ≤ j! (n + 1)! j=0 which together with (1.53) and (1.56) imply that, for any t ∈ R, φX (t + a) =

∞ (j) X φ (t) X

j=0

j!

aj ,

|a| < c.

(1.57)

Similarly, (1.57) holds with φX replaced by φY . Under the assumption that ψX = ψY < ∞ in a neighborhood of 0, X and Y have the same moments of (j) (j) all order. By (1.56), φX (0) = φY (0) for all j = 1, 2, ..., which and (1.57)

36

1. Probability Theory

with t = 0 imply that φX and φY are the same on the interval (−c, c) and hence have identical derivatives there. Considering t = c − ǫ and −c + ǫ for an arbitrarily small ǫ > 0 in (1.57) shows that φX and φY also agree on (−2c + ǫ, 2c − ǫ) and hence on (−2c, 2c). By the same argument φX and φY are the same on (−3c, 3c) and so on. Hence, φX (t) = φY (t) for all t and, by part (i), PX = PY . Consider now the general case of k ≥ 2. If PX 6= PY , then by part (i) there exists t ∈ Rk such that φX (t) 6= φY (t). Then φtτ X (1) 6= φtτ Y (1), which implies that Ptτ X 6= Ptτ Y . But ψX = ψY < ∞ in a neighborhood of 0 ∈ Rk implies that ψtτ X = ψtτ Y < ∞ in a neighborhood of 0 ∈ R and, by the proved result for k = 1, Ptτ X = Ptτ Y . This contradiction shows that PX = PY . Applying result (1.38) and Lemma 1.1, we obtain that ψX+Y (t) = ψX (t)ψY (t)

and φX+Y (t) = φX (t)ψY (t),

t ∈ Rk , (1.58)

for independent random k-vectors X and Y . This result, together with Theorem 1.6, provides a useful tool to obtain distributions of sums of independent random vectors with known distributions. The following example is an illustration. Example 1.20. Let Xi , i = 1, ..., k, be independent random variables and Xi have the gamma distribution Γ(αi , γ) (Table 1.2), i = 1, ..., k. From Table 1.2, Xi has the m.g.f. ψXi (t) = (1 − γt)−αi , t < γ −1 , i = 1, ..., k. By result (1.58), the m.g.f. of Y = X1 + · · · + Xk is equal to ψY (t) = (1 − γt)−(α1 +···+αk ) , t < γ −1 . From Table 1.2, the gamma distribution Γ(α1 + · · · + αk , γ) has the m.g.f. ψY (t) and, hence, is the distribution of Y (by Theorem 1.6). Similarly, result (1.52) and Theorem 1.6 can be used to determine distributions of linear transformations of random vectors with known distributions. The following is another interesting application of Theorem 1.6. Note that a random variable X is symmetric about 0 (defined according to (1.30)) if and only if X and −X have the same distribution, which can then be used as the definition of a random vector X symmetric about 0. We now show that X is symmetric about 0 if and only if its ch.f. φX is realvalued. If X and −X have the same distribution, then by Theorem 1.6, φX (t) = φ−X (t). From (1.52), φ−X (t) = φX (−t). Then φX (t) = φX (−t). Since sin(−tτ X) = − sin(tτ X) and cos(tτ X) = cos(−tτ X), this proves E[sin(tτ X)] = 0 and, thus, φX is real-valued. Conversely, if φX is realvalued, then φX (t) = E[cos(tτ X)] and φ−X (t) = φX (−t) = φX (t). By Theorem 1.6, X and −X must have the same distribution. Other applications of ch.f.’s can be found in §1.5.

1.4. Conditional Expectations

37

1.4 Conditional Expectations In elementary probability the conditional probability of an event B given an event A is defined as P (B|A) = P (A ∩ B)/P (A), provided that P (A) > 0. In probability and statistics, however, we sometimes need a notion of “conditional probability” even for A’s with P (A) = 0; for example, A = {Y = c}, where c ∈ R and Y is a random variable having a continuous c.d.f. General definitions of conditional probability, expectation, and distribution are introduced in this section, and they are shown to agree with those defined in elementary probability in special cases.

1.4.1 Conditional expectations Definition 1.6. Let X be an integrable random variable on (Ω, F , P ). (i) Let A be a sub-σ-field of F. The conditional expectation of X given A, denoted by E(X|A), is the a.s.-unique random variable satisfying the following two conditions: (a) E(X|A) is measurable from (Ω, A) to (R, B); R R (b) A E(X|A)dP = A XdP for any A ∈ A. (Note that the existence of E(X|A) follows from Theorem 1.4.) (ii) Let B ∈ F . The conditional probability of B given A is defined to be P (B|A) = E(IB |A). (iii) Let Y be measurable from (Ω, F , P ) to (Λ, G). The conditional expectation of X given Y is defined to be E(X|Y ) = E[X|σ(Y )]. Essentially, the σ-field σ(Y ) contains “the information in Y ”. Hence, E(X|Y ) is the “expectation” of X given the information provided by σ(Y ). The following useful result shows that there is a Borel function h defined on the range of Y such that E(X|Y ) = h ◦ Y . Lemma 1.2. Let Y be measurable from (Ω, F ) to (Λ, G) and Z a function from (Ω, F ) to Rk . Then Z is measurable from (Ω, σ(Y )) to (Rk , B k ) if and only if there is a measurable function h from (Λ, G) to (Rk , B k ) such that Z = h ◦ Y . The function h in E(X|Y ) = h ◦ Y is a Borel function on (Λ, G). Let y ∈ Λ. We define E(X|Y = y) = h(y) to be the conditional expectation of X given Y = y. Note that h(y) is a function on Λ, whereas h ◦ Y = E(X|Y ) is a function on Ω. For a random vector X, E(X|A) is defined as the vector of conditional expectations of components of X.

38

1. Probability Theory

Example 1.21. Let X be an integrable random variable on (Ω, F , P ), A1 , A2 , ... be disjoint events on (Ω, F , P ) such that ∪Ai = Ω and P (Ai ) > 0 for all i, and let a1 , a2 , ... be distinct real numbers. Define Y = a1 IA1 + a2 IA2 + · · ·. We now show that R ∞ X Ai XdP IAi . E(X|Y ) = (1.59) P (Ai ) i=1 We need to verify (a) and (b) in Definition 1.6 with A = σ(Y ). Since σ(Y ) = σ({A1 , A2 , ...}), it is clear that the function on the right-hand side of (1.59) is measurable on (Ω, σ(Y )). For any B ∈ B, Y −1 (B) = ∪i:ai ∈B Ai . Using properties of integrals, we obtain that Z X Z XdP = XdP Y −1 (B)

=

i:ai ∈B Ai R ∞ X Ai XdP i=1

=

Z

P Ai ∩ Y −1 (B)

P (Ai ) "∞ R X A XdP i

Y −1 (B)

i=1

P (Ai )

#

IAi dP.

This verifies (b) and thus (1.59) holds. R Let h be a Borel function on R satisfying h(ai ) = Ai XdP/P (Ai ). Then, by (1.59), E(X|Y ) = h ◦ Y and E(X|Y = y) = h(y). Let A ∈ F and X = IA . Then P (A|Y ) = E(X|Y ) =

∞ X P (A ∩ Ai ) i=1

P (Ai )

IAi ,

which equals P (A ∩ Ai )/P (Ai ) = P (A|Ai ) if ω ∈ Ai . Hence, the definition of conditional probability in Definition 1.6 agrees with that in elementary probability. The next result generalizes the result in Example 1.21 to conditional expectations of random variables having p.d.f.’s. Proposition 1.9. Let X be a random n-vector and Y a random m-vector. Suppose that (X, Y ) has a joint p.d.f. f (x, y) w.r.t. ν × λ, where ν and λ are σ-finite measures on (Rn , B n ) and (Rm , B m ), respectively. Let g(x, y) be a Borel function on Rn+m for which E|g(X, Y )| < ∞. Then R g(x, Y )f (x, Y )dν(x) R E[g(X, Y )|Y ] = a.s. (1.60) f (x, Y )dν(x)

39

1.4. Conditional Expectations

Proof. Denote the right-hand side of (1.60) by h(Y ). By Fubini’s theorem, h is Borel. Then, by Lemma R 1.2, h(Y ) is Borel on (Ω, σ(Y )). Also, by Fubini’s theorem, fY (y) = f (x, y)dν(x) is the p.d.f. of Y w.r.t. λ. For B ∈ Bm, Z Z h(Y )dP = h(y)dPY Y −1 (B) B Z R g(x, y)f (x, y)dν(x) R fY (y)dλ(y) = f (x, y)dν(x) B Z = g(x, y)f (x, y)dν × λ Rn ×B Z = g(x, y)dP(X,Y ) Rn ×B Z = g(X, Y )dP, Y −1 (B)

where the first and the last equalities follow from Theorem 1.2, the second and the next to last equalities follow from the definition of h and p.d.f.’s, and the third equality follows from Theorem 1.3 (Fubini’s theorem). For a random vector (X, Y ) with a joint p.d.f. f (x, y) w.r.t. ν × λ, define the conditional p.d.f. of X given Y = y to be fX|Y (x|y) =

f (x, y) , fY (y)

(1.61)

R where fY (y) = f (x, y)dν(x) is the marginal p.d.f. of Y w.r.t. λ. One can easily check that for each fixed y with fY (y) > 0, fX|Y (x|y) in (1.61) is a p.d.f. w.r.t. ν. Then equation (1.60) can be rewritten as Z E[g(X, Y )|Y ] = g(x, Y )fX|Y (x|Y )dν(x). Again, this agrees with the conditional expectation defined in elementary probability (i.e., the conditional expectation of g(X, Y ) given Y is equal to the expectation of g(X, Y ) w.r.t. the conditional p.d.f. of X given Y ). Now we list some useful properties of conditional expectations. The proof is left to the reader. Proposition 1.10. Let X, Y , X1 , X2 , ... be integrable random variables on (Ω, F , P ) and A be a sub-σ-field of F. (i) If X = c a.s., c ∈ R, then E(X|A) = c a.s. (ii) If X ≤ Y a.s., then E(X|A) ≤ E(Y |A) a.s. (iii) If a ∈ R and b ∈ R, then E(aX + bY |A) = aE(X|A) + bE(Y |A) a.s.

40

1. Probability Theory

(iv) E[E(X|A)] = EX. (v) E[E(X|A)|A0 ] = E(X|A0 ) = E[E(X|A0 )|A] a.s., where A0 is a sub-σfield of A. (vi) If σ(Y ) ⊂ A and E|XY | < ∞, then E(XY |A) = Y E(X|A) a.s. (vii) If X and Y are independent and E|g(X, Y )| < ∞ for a Borel function g, then E[g(X, Y )|Y = y] = E[g(X, y)] a.s. PY . (viii) If EX 2 < ∞, then [E(X|A)]2 ≤ E(X 2 |A) a.s. (ix) (Fatou’s lemma). If Xn ≥ 0 for any n, then E lim inf n Xn A ≤ lim inf n E(Xn |A) a.s. (x) (Dominated convergence theorem). Suppose that |Xn | ≤ Y for any n and Xn →a.s. X. Then E(Xn |A) →a.s. E(X|A). Although part (vii) of Proposition 1.10 can be proved directly, it is a consequence of a more general result given in Theorem 1.7(i). Since E(X|A) is defined only for integrable X, a version of monotone convergence theorem (i.e., 0 ≤ X1 ≤ X2 ≤ · · · and Xn →a.s. X imply E(Xn |A) →a.s. E(X|A)) becomes a special case of Proposition 1.10(x). It can also be shown (exercise) that H¨ older’s inequality (1.40), Liapounov’s inequality (1.42), Minkowski’s inequality (1.43), and Jensen’s inequality (1.47) hold a.s. with the expectation E replaced by the conditional expectation E(·|A). As an application, we consider the following example. Example 1.22. Let X be a random variable on (Ω, F , P ) with EX 2 < ∞ and let Y be a measurable function from (Ω, F , P ) to (Λ, G). One may wish to predict the value of X based on an observed value of Y . Let g(Y ) be a predictor, i.e., g ∈ ℵ = {all Borel functions g with E[g(Y )]2 < ∞}. Each predictor is assessed by the “mean squared prediction error” E[X − g(Y )]2 . We now show that E(X|Y ) is the best predictor of X in the sense that E[X − E(X|Y )]2 = min E[X − g(Y )]2 . g∈ℵ

(1.62)

First, Proposition 1.10(viii) implies E(X|Y ) ∈ ℵ. Next, for any g ∈ ℵ, E[X − g(Y )]2 = E[X − E(X|Y ) + E(X|Y ) − g(Y )]2 = E[X − E(X|Y )]2 + E[E(X|Y ) − g(Y )]2

+ 2E{[X − E(X|Y )][E(X|Y ) − g(Y )]} = E[X − E(X|Y )]2 + E[E(X|Y ) − g(Y )]2 + 2E E{[X − E(X|Y )][E(X|Y ) − g(Y )]|Y } = E[X − E(X|Y )]2 + E[E(X|Y ) − g(Y )]2 + 2E{[E(X|Y ) − g(Y )]E[X − E(X|Y )|Y ]} = E[X − E(X|Y )]2 + E[E(X|Y ) − g(Y )]2 ≥ E[X − E(X|Y )]2 ,

1.4. Conditional Expectations

41

where the third equality follows from Proposition 1.10(iv), the fourth equality follows from Proposition 1.10(vi), and the last equality follows from Proposition 1.10(i), (iii), and (vi).

1.4.2 Independence Definition 1.7. Let (Ω, F, P ) be a probability space. (i) Let C be a collection of subsets in F . Events in C are said to be independent if and only if for any positive integer n and distinct events A1 ,...,An in C, P (A1 ∩ A2 ∩ · · · ∩ An ) = P (A1 )P (A2 ) · · · P (An ). (ii) Collections Ci ⊂ F, i ∈ I (an index set that can be uncountable), are said to be independent if and only if events in any collection of the form {Ai ∈ Ci : i ∈ I} are independent. (iii) Random elements Xi , i ∈ I, are said to be independent if and only if σ(Xi ), i ∈ I, are independent. The following result is useful for checking the independence of σ-fields. Lemma 1.3. Let Ci , i ∈ I, be independent collections of events. Suppose that each Ci has the property that if A ∈ Ci and B ∈ Ci , then A ∩ B ∈ Ci . Then σ(Ci ), i ∈ I, are independent. An immediate application of Lemma 1.3 is to show (exercise) that random variables Xi , i = 1, ..., k, are independent according to Definition 1.7 if and only if (1.7) holds with F being the joint c.d.f. of (X1 , ..., Xk ) and Fi being the marginal c.d.f. of Xi . Hence, Definition 1.7(iii) agrees with the concept of independence of random variables discussed in §1.3.1. It is easy to see from Definition 1.7 that if X and Y are independent random vectors, then so are g(X) and h(Y ) for Borel functions g and h. Since the independence in Definition 1.7 is equivalent to the independence discussed in §1.3.1, this provides a simple proof of Lemma 1.1. For two events A and B with P (A) > 0, A and B are independent if and only if P (B|A) = P (B). This means that A provides no information about the probability of the occurrence of B. The following result is a useful extension. Proposition 1.11. Let X be a random variable with E|X| < ∞ and let Yi be random ki -vectors, i = 1, 2. Suppose that (X, Y1 ) and Y2 are independent. Then E[X|(Y1 , Y2 )] = E(X|Y1 ) a.s.

42

1. Probability Theory

Proof. First, E(X|Y1 ) is Borel on (Ω, σ(Y1 , Y2 )), since σ(Y1 ) ⊂ σ(Y1 , Y2 ). Next, we need to show that for any Borel set B ∈ B k1 +k2 , Z Z XdP = E(X|Y1 )dP. (1.63) (Y1 ,Y2 )−1 (B)

(Y1 ,Y2 )−1 (B)

If B = B1 × B2 , where Bi ∈ B ki , then (Y1 , Y2 )−1 (B) = Y1−1 (B1 ) ∩ Y2−1 (B2 ) and Z Z E(X|Y1 )dP = IY −1 (B1 ) IY −1 (B2 ) E(X|Y1 )dP 1

Y1−1 (B1 )∩Y2−1 (B2 )

= = = =

Z

Z

Z

Z

2

Z

IY −1 (B1 ) E(X|Y1 )dP IY −1 (B2 ) dP 1 2 Z IY −1 (B1 ) XdP IY −1 (B2 ) dP 1

2

IY −1 (B1 ) IY −1 (B2 ) XdP 1

2

Y1−1 (B1 )∩Y2−1 (B2 )

XdP,

where the second and the next to last equalities follow from result (1.38) and the independence of (X, Y1 ) and Y2 , and the third equality follows from the fact that E(X|Y1 ) is the conditional expectation of X given Y1 . This shows that (1.63) holds for B = B1 × B2 . We can show that the collection H = {B ⊂ Rk1 +k2 : B satisfies (1.63)} is a σ-field. Since we have already shown that B k1 × B k2 ⊂ H, B k1 +k2 = σ(B k1 × B k2 ) ⊂ H and thus the result follows. Clearly, the result in Proposition 1.11 still holds if X is replaced by h(X) for any Borel h and, hence, P (A|Y1 , Y2 ) = P (A|Y1 ) a.s. for any A ∈ σ(X),

(1.64)

if (X, Y1 ) and Y2 are independent. If Y1 is a constant and Y = Y2 , (1.64) reduces to P (A|Y ) = P (A) a.s. for any A ∈ σ(X), if X and Y are independent, i.e., σ(Y ) does not provide any additional information about the stochastic behavior of X. This actually provides another equivalent but more intuitive definition of the independence of X and Y (or two σ-fields). With a nonconstant Y1 , we say that given Y1 , X and Y2 are conditionally independent if and only if (1.64) holds. Then the result in Proposition 1.11 can be stated as: if Y2 and (X, Y1 ) are independent, then given Y1 , X and Y2 are conditionally independent. It is important to know that the result in Proposition 1.11 may not be true if Y2 is independent of X but not (X, Y1 ) (Exercise 96).

43

1.4. Conditional Expectations

1.4.3 Conditional distributions The conditional p.d.f. was introduced in §1.4.1 for random variables having p.d.f.’s w.r.t. some measures. We now consider conditional distributions in general cases where we may not have any p.d.f. Let X and Y be two random vectors defined on a common probability space. It is reasonable to consider P [X −1 (B)|Y = y] as a candidate for the conditional distribution of X, given Y = y, where B is any Borel set. However, since conditional probability is defined almost surely, for any fixed y, P [X −1 (B)|Y = y] may not be a probability measure. The first part of the following theorem (whose proof can be found in Billingsley (1986, pp. 460-461)) shows that there exists a version of conditional probability such that P [X −1 (B)|Y = y] is a probability measure for any fixed y. Theorem 1.7. (i) (Existence of conditional distributions). Let X be a random n-vector on a probability space (Ω, F, P ) and A be a sub-σ-field of F . Then there exists a function P (B, ω) on B n × Ω such that (a) P (B, ω) = P [X −1 (B)|A] a.s. for any fixed B ∈ B n , and (b) P (·, ω) is a probability measure on (Rn , B n ) for any fixed ω ∈ Ω. Let Y be measurable from (Ω, F , P ) to (Λ, G). Then there exists PX|Y (B|y) such that (a) PX|Y (B|y) = P [X −1 (B)|Y = y] a.s. PY for any fixed B ∈ B n , and (b) PX|Y (·|y) is a probability measure on (Rn , B n ) for any fixed y ∈ Λ. Furthermore, if E|g(X, Y )| < ∞ with a Borel function g, then Z E[g(X, Y )|Y = y] = E[g(X, y)|Y = y] = g(x, y)dPX|Y (x|y) a.s. PY . Rn

(ii) Let (Λ, G, P1 ) be a probability space. Suppose that P2 is a function from B n × Λ to R and satisfies (a) P2 (·, y) is a probability measure on (Rn , B n ) for any y ∈ Λ, and (b) P2 (B, ·) is Borel for any B ∈ B n . Then there is a unique probability measure P on (Rn × Λ, σ(B n × G)) such that, for B ∈ B n and C ∈ G, Z P (B × C) = P2 (B, y)dP1 (y). (1.65) C

m

m

Furthermore, if (Λ, G) = (R , B ), and X(x, y) = x and Y (x, y) = y define the coordinate random vectors, then PY = P1 , PX|Y (·|y) = P2 (·, y), and the probability measure in (1.65) is the joint distribution of (X, Y ), which has the following joint c.d.f.: Z F (x, y) = PX|Y (−∞, x]|z dPY (z), x ∈ Rn , y ∈ Rm , (1.66) (−∞,y]

where (−∞, a] denotes (−∞, a1 ] × · · · × (−∞, ak ] for a = (a1 , ..., ak ).

44

1. Probability Theory

For a fixed y, PX|Y =y = PX|Y (·|y) is called the conditional distribution of X given Y = y. Under the conditions in Theorem 1.7(i), if Y is a random m-vector and (X, Y ) has a p.d.f. w.r.t. ν × λ (ν and λ are σ-finite measures on (Rn , B n ) and (Rm , B m ), respectively), then fX|Y (x|y) defined in (1.61) is the p.d.f. of PX|Y =y w.r.t. ν for any fixed y. The second part of Theorem 1.7 states that given a distribution on one space and a collection of conditional distributions (which are conditioned on values of the first space) on another space, we can construct a joint distribution in the product space. It is sometimes called the “two-stage experiment theorem” for the following reason. If Y ∈ Rm is selected in stage 1 of an experiment according to its marginal distribution PY = P1 , and X is chosen afterward according to a distribution P2 (·, y), then the combined two-stage experiment produces a jointly distributed pair (X, Y ) with distribution P(X,Y ) given by (1.65) and PX|Y =y = P2 (·, y). This provides a way of generating dependent random variables. The following is an example. Example 1.23. A market survey is conducted to study whether a new product is preferred over the product currently available in the market (old product). The survey is conducted by mail. Questionnaires are sent along with the sample products (both new and old) to N customers randomly selected from a population, where N is a positive integer. Each customer is asked to fill out the questionnaire and return it. Responses from customers are either 1 (new is better than old) or 0 (otherwise). Some customers, however, do not return the questionnaires. Let X be the number of ones in the returned questionnaires. What is the distribution of X? If every customer returns the questionnaire, then (from elementary probability) X has the binomial distribution Bi(p, N ) in Table 1.1 (assuming that the population is large enough so that customers respond independently), where p ∈ (0, 1) is the overall rate of customers who prefer the new product. Now, let Y be the number of customers who respond. Then Y is random. Suppose that customers respond independently with the same probability π ∈ (0, 1). Then PY is the binomial distribution Bi(π, N ). Given Y = y (an integer between 0 and N ), PX|Y =y is the binomial distribution Bi(p, y) if y ≥ 1 and the point mass at 0 (see (1.22)) if y = 0. Using (1.66) and the fact that binomial distributions have p.d.f.’s w.r.t. counting measure, we obtain that the joint c.d.f. of (X, Y ) is F (x, y) =

y X

k=0

=

N k PX|Y =k (−∞, x] π (1 − π)N −k k

y min{x,k} X X k N k pj (1 − p)k−j π (1 − π)N −k j k j=0

k=0

45

1.4. Conditional Expectations

for x = 0, 1, ..., y, y = 0, 1, ..., N . The marginal c.d.f. FX (x) = F (x, ∞) = F (x, N ). The p.d.f. of X w.r.t. counting measure is N X k

N k fX (x) = p (1 − p) π (1 − π)N −k x k k=x k−x N −k N X N − x π − πp 1−π N x N −x = (πp) (1 − πp) 1 − πp 1 − πp k−x x k=x N = (πp)x (1 − πp)N −x x x

k−x

for x = 0, 1, ..., N . It turns out that the marginal distribution of X is the binomial distribution Bi(πp, N ).

1.4.4 Markov chains and martingales As applications of conditional expectations, we introduce here two important types of dependent sequences of random variables. Markov chains A sequence of random vectors {Xn : n = 1, 2, ...} is said to be a Markov chain or Markov process if and only if P (B|X1 , ..., Xn ) = P (B|Xn ) a.s., B ∈ σ(Xn+1 ), n = 2, 3, ....

(1.67)

Comparing (1.67) with (1.64), we conclude that (1.67) implies that Xn+1 (tomorrow) is conditionally independent of (X1 , ..., Xn−1 ) (the past), given Xn (today). But (X1 , ..., Xn−1 ) is not necessarily independent of (Xn , Xn+1 ). Clearly, a sequence of independent random vectors forms a Markov chain since, by Proposition 1.11, both quantities on two sides of (1.67) are equal to P (B) for independent Xi ’s. The following example describes some Markov processes of dependent random variables. Example 1.24 (First-order autoregressive processes). Let ε1 , ε2 , ... be independent random variables defined on a probability space, X1 = ε1 , and Xn+1 = ρXn + εn+1 , n = 1, 2, ..., where ρ is a constant in R. Then {Xn } is called a first-order autoregressive process. We now show that for any B ∈ B and n = 1, 2, ..., P (Xn+1 ∈ B|X1 , ..., Xn ) = Pεn+1 (B − ρXn ) = P (Xn+1 ∈ B|Xn ) a.s.,

46

1. Probability Theory

where B − y = {x ∈ R : x + y ∈ B}, which implies that {Xn } is a Markov chain. For any y ∈ R, Z Pεn+1 (B − y) = P (εn+1 + y ∈ B) = IB (x + y)dPεn+1 (x) and, by Fubini’s theorem, Pεn+1 (B − y) is Borel. Hence, Pεn+1 (B − ρXn ) is Borel w.r.t. σ(Xn ) and, thus, is Borel w.r.t. σ(X1 , ..., Xn ). Let Bj ∈ B, j = 1, ..., n, and A = ∩nj=1 Xj−1 (Bj ). Since εn+1 + ρXn = Xn+1 and εn+1 is independent of (X1 , ..., Xn ), it follows from Theorem 1.2 and Fubini’s theorem that Z Z Z Pεn+1 (B − ρXn )dP = dPεn+1 (t)dPX (x) A

xj ∈Bj ,j=1,...,n

= =

Z

t∈B−ρxn

xj ∈Bj ,j=1,...,n,xn+1 ∈B −1 (B) , P A ∩ Xn+1

dP(X,εn+1 ) (x, t)

where X and x denote (X1 , ..., Xn ) and (x1 , ..., xn ), respectively, and xn+1 denotes ρxn + t. Using this and the argument in the end of the proof for Proposition 1.11, we obtain P (Xn+1 ∈ B|X1 , ..., Xn ) = Pεn+1 (B − ρXn ) a.s. The proof for Pεn+1 (B − ρXn ) = P (Xn+1 ∈ B|Xn ) a.s. is similar and simpler. The following result provides some characterizations of Markov chains. Proposition 1.12. A sequence of random vectors {Xn } is a Markov chain if and only if one of the following three conditions holds. (a) For any n = 2, 3, ... and any integrable h(Xn+1 ) with a Borel function h, E[h(Xn+1 )|X1 , ..., Xn ] = E[h(Xn+1 )|Xn ] a.s. (b) For any n = 1, 2, ... and B ∈ σ(Xn+1 , Xn+2 , ...), P (B|X1 , ..., Xn ) = P (B|Xn ) a.s. (c) For any n = 2, 3, ..., A ∈ σ(X1 , ..., Xn ), and B ∈ σ(Xn+1 , Xn+2 , ...), P (A ∩ B|Xn ) = P (A|Xn )P (B|Xn ) a.s. Proof. (i) It is clear that (a) implies (1.67). If h is a simple function, then (1.67) and Proposition 1.10(iii) imply (a). If h is nonnegative, then by Exercise 17 there are nonnegative simple functions h1 ≤ h2 ≤ · · · ≤ h such that hj → h. Then (1.67) together with Proposition 1.10(iii) and (x) imply (a). Since h = h+ − h− , we conclude that (1.67) implies (a). (ii) It is also clear that (b) implies (1.67). We now show that (1.67) implies (b). Note that σ(Xn+1 , Xn+2 , ...) = σ ∪∞ j=1 σ(Xn+1 , ..., Xn+j ) (Exercise 19). Hence, it suffices to show that P (B|X1 , ..., Xn ) = P (B|Xn ) a.s. for B ∈ σ(Xn+1 , ..., Xn+j ) for any j = 1, 2, .... We use induction. The result for j = 1 follows from (1.67). Suppose that the result holds for any B ∈

47

1.4. Conditional Expectations

σ(Xn+1 , ..., Xn+j ). To show the result for any B ∈ σ(Xn+1 , ..., Xn+j+1 ), it is enough (why?) to show that for any B1 ∈ σ(Xn+j+1 ) and any B2 ∈ σ(Xn+1 , ..., Xn+j ), P (B1 ∩ B2 |X1 , ..., Xn ) = P (B1 ∩ B2 |Xn ) a.s. From the proof in (i), the induction assumption implies E[h(Xn+1 , ..., Xn+j )|X1 , ..., Xn ] = E[h(Xn+1 , ..., Xn+j )|Xn ]

(1.68)

for any Borel function h. The result follows from E(IB1 IB2 |X1 , ..., Xn ) = E[E(IB1 IB2 |X1 , ..., Xn+j )|X1 , ..., Xn ] = E[IB2 E(IB1 |X1 , ..., Xn+j )|X1 , ..., Xn ] = E[IB2 E(IB1 |Xn+j )|X1 , ..., Xn ]

= E[IB2 E(IB1 |Xn+j )|Xn ] = E[IB2 E(IB1 |Xn , ..., Xn+j )|Xn ] = E[E(IB1 IB2 |Xn , ..., Xn+j )|Xn ] = E(IB1 IB2 |Xn ) a.s.,

where the first and last equalities follow from Proposition 1.10(v), the second and sixth equalities follow from Proposition 1.10(vi), the third and fifth equalities follow from (1.67), and the fourth equality follows from (1.68). (iii) Let A ∈ σ(X1 , ..., Xn ) and B ∈ σ(Xn+1 , Xn+2 , ...). If (b) holds, then E(IA IB |Xn ) = E[E(IA IB |X1 , ..., Xn )|Xn ] = E[IA E(IB |X1 , ..., Xn )|Xn ] = E[IA E(IB |Xn )|Xn ] = E(IA |Xn )E(IB |Xn ), which is (c). Assume that (c) holds. Let A1 ∈ σ(Xn ), A2 ∈ σ(X1 , ..., Xn−1 ), and B ∈ σ(Xn+1 , Xn+2 , ...). Then Z Z E(IB |Xn )dP = IA2 E(IB |Xn )dP A1 ∩A2 A1 Z = E[IA2 E(IB |Xn )|Xn ]dP A1 Z = E(IA2 |Xn )E(IB |Xn )dP A1 Z = E(IA2 IB |Xn )dP A1

= P (A1 ∩ A2 ∩ B).

Since disjoint unions of events of the form A1 ∩ A2 as specified above generate σ(X1 , ..., Xn ), this shows that E(IB |Xn ) = E(IB |X1 , ..., Xn ) a.s., which is (b). Note that condition (b) in Proposition 1.12 can be stated as “the past and the future are conditionally independent given the present”, which is a property of any Markov chain. More discussions and applications of Markov chains can be found in §4.1.4.

48

1. Probability Theory

Martingales Let {Xn } be a sequence of integrable random variables on a probability space (Ω, F , P ) and F1 ⊂ F2 ⊂ · · · ⊂ F be a sequence of σ-fields such that σ(Xn ) ⊂ Fn , n = 1, 2, .... The sequence {Xn , Fn : n = 1, 2, ...} is said to be a martingale if and only if E(Xn+1 |Fn ) = Xn a.s., n = 1, 2, ...,

(1.69)

a submartingale if and only if (1.69) holds with = replaced by ≥, and a supermartingale if and only if (1.69) holds with = replaced by ≤. {Xn } is said to be a martingale (submartingale or supermartingale) if and only if {Xn , σ(X1 , ..., Xn )} is a martingale (submartingale or supermartingale). From Proposition 1.10(v), if {Xn , Fn } is a martingale (submartingale or supermartingale), then so is {Xn }. A simple property of a martingale (or a submartingale) {Xn , Fn } is that E(Xn+j |Fn ) = Xn a.s. (or E(Xn+j |Fn ) ≥ Xn a.s.) and EX1 = EXj (or EX1 ≤ EX2 ≤ · · ·) for any j = 1, 2, ... (exercise). For any probability space (Ω, F, P ) and σ-fields F1 ⊂ F2 ⊂ · · · ⊂ F, we can always construct a martingale {E(Y |Fn )} by using an integrable random variable Y . Another way to construct a martingale is to use a sequence of independent integrable random variables {εn } by letting Xn = ε1 + · · · + εn , n = 1, 2, .... Since E(Xn+1 |X1 , ..., Xn ) = E(Xn + εn+1 |X1 , ..., Xn ) = Xn + Eεn+1 a.s., {Xn } is a martingale if Eεn = 0 for all n, a submartingale if Eεn ≥ 0 for all n, and a supermartingale if Eεn ≤ 0 for all n. Note that in Example 1.24 with ρ = 1, {Xn } is shown to be a Markov chain. The next example provides another example of martingales. Example 1.25 (Likelihood ratio). Let (Ω, F , P ) be a probability space, Q be a probability measure on F, and F1 ⊂ F2 ⊂ · · · ⊂ F be a sequence of σ-fields. Let Pn and Qn be P and Q restricted to Fn , respectively, n = 1, 2, .... Suppose that Qn ≪ Pn for each n. Then {Xn , Fn } is a martingale, where Xn = dQn /dPn (the Radon-Nikodym derivative of Qn w.r.t. Pn ), n = 1, 2, ... (exercise). Suppose now that {Yn } is a sequence of random variables on (Ω, F , P ), Fn = σ(Y1 , ..., Yn ) and that there exists a σfinite measure νn on Fn such that Pn ≪ νn and νn ≪ Pn , n = 1, 2, .... Let pn (Y1 , ..., Yn ) = dPn /dνn and qn (Y1 , ..., Yn ) = dQn /dνn . By Proposition 1.7(iii), Xn = qn (Y1 , ..., Yn )/pn (Y1 , ..., Yn ), which is called a likelihood ratio in statistical terms. The following results contain some useful properties of martingales and submartingales.

1.5. Asymptotic Theory

49

Proposition 1.13. Let ϕ be a convex function on R. (i) If {Xn , Fn } is a martingale and ϕ(Xn ) is integrable for all n, then {ϕ(Xn ), Fn } is a submartingale. (ii) If {Xn , Fn } is a submartingale, ϕ(Xn ) is integrable for all n, and ϕ is nondecreasing, then {ϕ(Xn ), Fn } is a submartingale. Proof. (i) Note that ϕ(Xn ) = ϕ(E(Xn+1 |Fn )) ≤ E[ϕ(Xn+1 |Fn )] a.s. by Jensen’s inequality for conditional expectations (Exercise 89(c)). (ii) Since ϕ is nondecreasing and {Xn , Fn } is a submartingale, ϕ(Xn ) ≤ ϕ(E(Xn+1 |Fn )) ≤ E[ϕ(Xn+1 |Fn )] a.s. An application of Proposition 1.13 shows that if {Xn , Fn } is a submartingale, then so is {(Xn )+ , Fn }; if {Xn , Fn } is a martingale, then {|Xn |, Fn } is a submartingale and so are {|Xn |p , Fn }, where p > 1 is a constant, and {|Xn |(log |Xn |)+ , Fn }, provided that |Xn |p and |Xn |(log |Xn |)+ are integrable for all n. Proposition 1.14 (Doob’s decomposition). Let {Xn , Fn } be a submartingale. Then Xn = Yn + Zn , n = 1, 2, ..., where {Yn , Fn } is a martingale, 0 = Z1 ≤ Z2 ≤ · · ·, and EZn < ∞ for all n. Furthermore, if supn E|Xn | < ∞, then supn E|Yn | < ∞ and supn EZn < ∞. Proof. Define η1 = ξ1 , ζ1 = 0, ηn = Xn −Xn−1 −E(X P n −Xn−1 |Fn−1 P), and ζn = E(Xn − Xn−1 |Fn−1 ) for n ≥ 2. Then Yn = ni=1 ηi and Zn = ni=1 ζi satisfy Xn = Yn + Zn and the required conditions (exercise). Assume now that supn E|Xn | < ∞. Since EY1 = EYn for any n and Zn ≤ |Xn | − Yn , EZn ≤ E|Xn | − EY1 . Hence supn EZn < ∞. Also, |Yn | ≤ |Xn | + Zn . Hence supn E|Yn | < ∞. The following martingale convergence theorem, due to Doob, has many applications (see, e.g., Example 1.27 in §1.5.1). Its proof can be found, for example, in Billingsley (1986, pp. 490-491). Proposition 1.15. Let {Xn , Fn } be a submartingale. If c = supn E|Xn | < ∞, then limn→∞ Xn = X a.s., where X is a random variable satisfying E|X| ≤ c.

1.5 Asymptotic Theory Asymptotic theory studies limiting behavior of random variables (vectors) and their distributions. It is an important tool for statistical analysis. A more complete coverage of asymptotic theory in statistical analysis can be found in Serfling (1980), Shorack and Wellner (1986), Sen and Singer (1993), Barndorff-Nielsen and Cox (1994), and van der Vaart (1998).

50

1. Probability Theory

1.5.1 Convergence modes and stochastic orders There are several convergence modes for random variables/vectors. Let r > 0 be a constant. For any c = (c1 , ..., ck ) ∈ Rk , we define kckr = Pk ( j=1 |cj |r )1/r . If r ≥ 1, then kckr is the Lr -distance between 0 and c. √ When r = 2, the subscript r is omitted and kck = kck2 = cτ c. Definition 1.8. Let X, X1 , X2 , . . . be random k-vectors defined on a probability space. (i) We say that the sequence {Xn } converges to X almost surely (a.s.) and write Xn →a.s. X if and only if limn→∞ Xn = X a.s. (ii) We say that {Xn } converges to X in probability and write Xn →p X if and only if, for every fixed ǫ > 0, lim P (kXn − Xk > ǫ) = 0.

n→∞

(1.70)

(iii) We say that {Xn } converges to X in Lr (or in rth moment) and write Xn →Lr X if and only if lim EkXn − Xkrr = 0,

n→∞

where r > 0 is a fixed constant. (iv) Let F , Fn , n = 1, 2, ..., be c.d.f.’s on Rk and P , Pn , n = 1, ..., be their corresponding probability measures. We say that {Fn } converges to F weakly (or {Pn } converges to P weakly) and write Fn →w F (or Pn →w P ) if and only if, for each continuity point x of F , lim Fn (x) = F (x).

n→∞

We say that {Xn } converges to X in distribution (or in law) and write Xn →d X if and only if FXn →w FX . The a.s. convergence has already been considered in previous sections. The concept of convergence in probability, convergence in Lr , or a.s. convergence represents a sense in which, for n sufficiently large, Xn and X approximate each other as functions on the original probability space. The concept of convergence in distribution in Definition 1.8(iv), however, depends only on the distributions FXn and FX (or probability measures PXn and PX ) and does not necessitate that Xn and X are close in any sense; in fact, Definition 1.8(iv) still makes sense even if X and Xn ’s are not defined on the same probability space. In Definition 1.8(iv), it is not required that limn→∞ Fn (x) = F (x) for every x. However, if F is a continuous function, then we have the following stronger result.

51

1.5. Asymptotic Theory

Proposition 1.16 (P´ olya’s theorem). If Fn →w F and F is continuous on Rk , then lim sup |Fn (x) − F (x)| = 0. n→∞ x∈Rk

A useful characterization of a.s. convergence is given in the following lemma. Lemma 1.4. For random k-vectors X, X1 , X2 , . . . on a probability space, Xn →a.s. X if and only if for every ǫ > 0, ! ∞ [ {kXm − Xk > ǫ} = 0. (1.71) lim P n→∞

∪∞ n=1

m=n

∩∞ m=n

Proof. Let Aj = {kXm − Xk ≤ j −1 }, j = 1, 2, .... By Proposition 1.1(iii) and DeMorgan’s law, (1.71) holds for every ǫ > 0 if and only if P (Aj ) = 1 for every j, which is equivalent to P (∩∞ j=1 Aj ) = 1. The result follows from ∩∞ A = {ω : lim X (ω) = X(ω)} (exercise). n→∞ n j=1 j The following result describes the relationship among the four convergence modes in Definition 1.8. Theorem 1.8. Let X, X1 , X2 , . . . be random k-vectors. (i) If Xn →a.s. X, then Xn →p X. (ii) If Xn →Lr X for an r > 0, then Xn →p X. (iii) If Xn →p X, then Xn →d X. (iv) (Skorohod’s theorem). If Xn →d X, then there are random vectors Y, Y1 , Y2 , ... defined on a common probability space such that PY = PX , PYn = PXn , n = 1, 2,...,Pand Yn →a.s. Y . ∞ (v) If, for every ǫ > 0, n=1 P (kXn − Xk ≥ ǫ) < ∞, then Xn →a.s. X. (vi) If Xn →p X, then there is a subsequence {Xnj , j = 1, 2, ...} such that Xnj →a.s. X as j → ∞. (vii) If Xn →d X and P (X = c) = 1, where c ∈ Rk is a constant vector, then Xn →p c. (viii) Suppose that Xn →d X. Then, for any r > 0, lim EkXn krr = EkXkrr < ∞

n→∞

if and only if {kXn krr } is uniformly integrable in the sense that lim sup E kXn krr I{kXn kr >t} = 0. t→∞ n

(1.72)

(1.73)

The proof of Theorem 1.8 is given after the following discussion and example.

52

1. Probability Theory

The converse of Theorem 1.8(i), (ii), or (iii) is generally not true (see Example 1.26 and Exercise 116). Note that part (iv) of Theorem 1.8 (Skorohod’s theorem) is not a converse of part (i), but it is an important result in probability theory. It is useful when we study convergence of quantities related to FXn and FX when Xn →d X (see, e.g., the proofs of Theorems 1.8 and 1.9). Part (v) of Theorem 1.8 indicates that the converse of part (i) is true under the additional condition that P (kXn − Xk ≥ ǫ) tends to 0 fast enough. Part (vi) provides a partial converse of part (i) whereas part (vii) is a partial converse of part (iii). A consequence of Theorem 1.8(viii) is that if Xn →p X and {kXn − Xkrr } is uniformly integrable, then Xn →Lr X; i.e., the converse of Theorem 1.8(ii) is true under the additional condition of uniform integrability. A useful sufficient condition for uniform integrability of {kXn krr } is that sup EkXn kr+δ 0. Some other sufficient conditions are given in Exercises 117-120. Example 1.26. Let θn = 1 + n−1 and Xn be a random variable having the exponential distribution E(0, θn ) (Table 1.2), n = 1, 2, .... Let X be a random variable having the exponential distribution E(0, 1). For any x > 0, FXn (x) = 1 − e−x/θn → 1 − e−x = FX (x)

as n → ∞. Since FXn (x) ≡ 0 ≡ FX (x) for x ≤ 0, we have shown that Xn →d X. Is it true that Xn →p X? This question cannot be answered without any further information about the random variables X and Xn . We consider two cases in which different answers can be obtained. First, suppose that Xn ≡ θn X (then Xn has the given c.d.f.). Note that Xn − X = (θn − 1)X = n−1 X, which has the c.d.f. (1 − e−nx )I[0,∞) (x). Hence P (|Xn − X| ≥ ǫ) = e−nǫ → 0 for any ǫ > 0. In fact, by Theorem 1.8(v), Xn →a.s. X; since E|Xn − X|p = n−p EX p < ∞ for any p > 0, Xn →Lp X for any p > 0. Next, suppose that Xn and X are independent random variables. Using result (1.28) and the fact that the p.d.f.’s for Xn and −X are θn−1 e−x/θn I(0,∞) (x) and ex I(−∞,0) (x), respectively, we obtain that Z ǫZ θn−1 e−x/θn ey−x I(0,∞) (x)I(−∞,x) (y)dxdy, P (|Xn − X| ≤ ǫ) = −ǫ

which converges to (by the dominated convergence theorem) Z ǫ Z e−x ey−x I(0,∞) (x)I(−∞,x) (y)dxdy = 1 − e−ǫ . −ǫ

53

1.5. Asymptotic Theory

Thus, P (|Xn − X| ≥ ǫ) → e−ǫ > 0 for any ǫ > 0 and, therefore, {Xn } does not converge to X in probability. The previous discussion, however, indicates how to construct the random variables Yn and Y in Theorem 1.8(iv) for this example. The following famous result is used in the proof of Theorem 1.8(v). Its proof is left to the reader. Lemma 1.5. (Borel-Cantelli lemma). Let An be a sequence of events in a ∞ ∞ probability P∞ space and lim supn An = ∩n=1 ∪m=n Am . (i) If n=1 P (An ) < ∞, then P (lim supn An ) = 0. P∞ (ii) If A1 , A2 , ... are pairwise independent and n=1 P (An ) = ∞, then P (lim supn An ) = 1. Proof of Theorem 1.8. (i) The result follows from Lemma 1.4, since (1.71) implies (1.70). (ii) The result follows from Chebyshev’s inequality with ϕ(t) = |t|r . (iii) For any c = (c1 , ..., ck ) ∈ Rk , define (−∞, c] = (−∞, c1 ]×· · ·×(−∞, ck ]. Let x be a continuity point of FX , ǫ > 0 be given, and Jk be the k-vector of ones. Then {X ∈ (−∞, x − ǫJk ], Xn 6∈ (−∞, x]} ⊂ {kXn − Xk > ǫ} and FX (x − ǫJk ) = P X ∈ (−∞, x − ǫJk ] ≤ P Xn ∈ (−∞, x] + P X ∈ (−∞, x − ǫJk ], Xn 6∈ (−∞, x] ≤ FXn (x) + P (kXn − Xk > ǫ) . Letting n → ∞, we obtain that FX (x − ǫJk ) ≤ lim inf n FXn (x). Similarly, we can show that FX (x + ǫJk ) ≥ lim supn FXn (x). Since ǫ is arbitrary and FX is continuous at x, FX (x) = limn→∞ FXn (x). (iv) The proof of this part can be found in Billingsley (1986, pp. 399-402). (v) Let An = {kXn − Xk ≥ ǫ}. The result follows from Lemma 1.4, Lemma 1.5(i), and Proposition 1.1(iii). (vi) From (1.70), for every j = 1, 2, ..., there is a positive integer nj such that P (kXnj − Xk > 2−j ) < 2−j . For any ǫ > 0, there is aP kǫ such that for −j j ≥ kǫ , P (kXnj − Xk > ǫ) < P (kXnj − Xk > 2−j ). Since ∞ = 1, it j=1 2 follows from the result in (v) that Xnj →a.s. X as j → ∞. (vii) The proof for this part is left as an exercise. (viii) First, by part (iv), we may assume that Xn →a.s. X (why?). Assume that {kXn krr } is uniformly integrable. Then supn EkXn krr < ∞ (why?) and by Fatou’s lemma (Theorem 1.1(i)), EkXkrr ≤ lim inf n EkXn krr < ∞. Hence, (1.72) follows if we can show that lim sup EkXn krr ≤ EkXkrr .

(1.75)

n

For any ǫ > 0 and t > 0, let An = {kXn −Xkr ≤ ǫ} and Bn = {kXn kr > t}.

54

1. Probability Theory

Then EkXn krr = E(kXn krr IAcn ∩Bn ) + E(kXn krr IAcn ∩Bnc ) + E(kXn krr IAn ) ≤ E(kXn krr IBn ) + tr P (Acn ) + EkXn IAn krr .

For r ≤ 1, kXn IAn krr ≤ (kXn − Xkrr + kXkrr )IAn and EkXn IAn krr ≤ E[(kXn − Xkrr + kXkrr )IAn ] ≤ ǫr + EkXkrr . For r > 1, an application of Minkowski’s inequality leads to EkXn IAn krr = Ek(Xn − X)IAn + XIAn krr r ≤ E [k(Xn − X)IAn kr + kXIAn kr ] or n 1/r 1/r ≤ [Ek(Xn − X)IAn krr ] + [EkXIAn krr ] or n ≤ ǫ + [EkXkrr ]1/r .

In any case, since ǫ is arbitrary, lim supn EkXn IAn krr ≤ EkXkrr . This result and the previously established inequality imply that lim sup EkXn krr ≤ lim sup E(kXn krr IBn ) + tr lim P (Acn ) n

n→∞

n

≤

+ lim sup EkXn IAn krr n sup E(kXn krr I{kXn kr >t} ) n

+ EkXkrr ,

since P (Acn ) → 0. Since {kXn krr } is uniformly integrable, letting t → ∞ we obtain (1.75). Suppose now that (1.72) holds. Let ξn = kXn krr IBnc − kXkrr IBnc . Then ξn →a.s. 0 and |ξn | ≤ tr + kXkrr , which is integrable. By the dominated convergence theorem, Eξn → 0; this and (1.72) imply that E(kXn krr IBn ) − E(kXkrr IBn ) → 0. From the definition of Bn , Bn ⊂ {kXn − Xkr > t/2} ∪ {kXkr > t/2}. Since EkXkrr < ∞, it follows from the dominated convergence theorem that E(kXkrr I{kXn −Xkr >t/2} ) → 0 as n → ∞. Hence, lim sup E(kXn krr IBn ) ≤ lim sup E(kXkrr IBn ) ≤ E(kXkrr I{kXkr >t/2} ). n

n

Letting t → ∞, it follows from the dominated convergence theorem that lim lim sup E(kXn krr IBn ) ≤ lim E(kXkrr I{kXkr >t/2} ) = 0.

t→∞

n

This proves (1.73).

t→∞

1.5. Asymptotic Theory

55

Example 1.27. As an application of Theorem 1.8(viii) and Proposition 1.15, we consider again the prediction problem in Example 1.22. Suppose that we predict a random variable X by a random n-vector Y = (Y1 , ..., Yn ). It is shown in Example 1.22 that Xn = E(X|Y1 , ..., Yn ) is the best predictor in terms of the mean squared prediction error, when EX 2 < ∞. We now show that Xn →a.s. X when n → ∞ under the assumption that σ(X) ⊂ F∞ = σ(Y1 , Y2 , ...) (i.e., X provides no more information than Y1 , Y2 , ...). From the discussion in §1.4.4, {Xn } is a martingale. Also, supn E|Xn | ≤ supn E[E(|X||Y1 , ..., Yn )] = E|X| < ∞. Hence, by Proposition 1.15, Xn →a.s. Z for some random variable Z. We now need to show Z = X a.s. Since σ(X) ⊂ F∞ , X = E(X|F∞ ) a.s. Hence, it suffices to show that Z = E(X|F∞ ) a.s. Since EXn2 ≤ EX 2 < ∞ (why?), condition (1.74) holds for sequence {|Xn |} and, hence, {|Xn |} is uniformly integrable. By Theorem R R 1.8(viii), E|Xn − Z| → 0, which implies A Xn dP → A ZdP for any event A. for n ≥ m and R Note thatR if A ∈ σ(Y1 , ..., Ym ), then A ∈ σ(Y1 , ..., Yn ) ∞ X dP = XdP . This implies that for any A ∈ ∪ n j=1 σ(Y1 , ..., Yj ), R A RA ∞ XdP = ZdP . Since ∪ σ(Y , ..., Y ) generates F 1 j ∞ , we conclude j=1 A A R R that A XdP = A ZdP for any A ∈ F∞ and thus Z = E(X|F∞ ) a.s. In the proof above, the condition EX 2 < ∞ is used only for showing the uniform integrability of {|Xn |}. But by Exercise 120, {|Xn |} is uniformly integrable as long as E|X| < ∞. Hence Xn →a.s. X is still true if the condition EX 2 < ∞ is replaced by E|X| < ∞. We now introduce the notion of O( · ), o( · ), and stochastic O( · ) and o( · ). In calculus, two sequences of real numbers, {an } and {bn }, satisfy an = O(bn ) if and only if |an | ≤ c|bn | for all n and a constant c; and an = o(bn ) if and only if an /bn → 0 as n → ∞. Definition 1.9. Let X1 , X2 , ... be random vectors and Y1 , Y2 , ... be random variables defined on a common probability space. (i) Xn = O(Yn ) a.s. if and only if P (kXn k = O(|Yn |)) = 1. (ii) Xn = o(Yn ) a.s. if and only if Xn /Yn →a.s. 0. (iii) Xn = Op (Yn ) if and only if, for any ǫ > 0, there is a constant Cǫ > 0 such that supn P (kXn k ≥ Cǫ |Yn |) < ǫ. (iv) Xn = op (Yn ) if and only if Xn /Yn →p 0. Note that Xn = op (Yn ) implies Xn = Op (Yn ); Xn = Op (Yn ) and Yn = Op (Zn ) implies Xn = Op (Zn ); but Xn = Op (Yn ) does not imply Yn = Op (Xn ). The same conclusion can be obtained if Op ( · ) and op ( · ) are replaced by O( · ) a.s. and o( · ) a.s., respectively. Some results related to Op are given in Exercise 127. For example, if Xn →d X for a random variable X, then Xn = Op (1). Since an = O(1) means that {an } is bounded, {Xn } is said to be bounded in probability if Xn = Op (1).

56

1. Probability Theory

1.5.2 Weak convergence We now discuss more about convergence in distribution or weak convergence of probability measures. A sequence {Pn } of probability measures on (Rk , B k ) is tight if for every ǫ > 0, there is a compact set C ⊂ Rk such that inf n Pn (C) > 1 − ǫ. If {Xn } is a sequence of random k-vectors, then the tightness of {PXn } is the same as the boundedness of {kXn k} in probability. The proof of the following result can be found in Billingsley (1986, pp. 392-395). Proposition 1.17. Let {Pn } be a sequence of probability measures on (Rk , B k ). (i) Tightness of {Pn } is a necessary and sufficient condition that for every subsequence {Pni } there exists a further subsequence {Pnj } ⊂ {Pni } and a probability measure P on (Rk , B k ) such that Pnj →w P as j → ∞. (ii) If {Pn } is tight and if each subsequence that converges weakly at all converges to the same probability measure P , then Pn →w P . The following result gives some useful sufficient and necessary conditions for convergence in distribution. Theorem 1.9. Let X, X1 , X2 , . . . be random k-vectors. (i) Xn →d X is equivalent to any one of the following conditions: (a) E[h(Xn )] → E[h(X)] for every bounded continuous function h; (b) lim supn PXn (C) ≤ PX (C) for any closed set C ⊂ Rk ; (c) lim inf n PXn (O) ≥ PX (O) for any open set O ⊂ Rk . (ii) (L´evy-Cram´er continuity theorem). Let φX , φX1 , φX2 , ... be the ch.f.’s of X, X1 , X2 , ..., respectively. Xn →d X if and only if limn→∞ φXn (t) = φX (t) for all t ∈ Rk . (iii) (Cram´er-Wold device). Xn →d X if and only if cτ Xn →d cτ X for every c ∈ Rk . Proof. (i) First, we show Xn →d X implies (a). By Theorem 1.8(iv) (Skorohod’s theorem), there exists a sequence of random vectors {Yn } and a random vector Y such that PYn = PXn for all n, PY = PX and Yn →a.s. Y . For bounded continuous h, h(Yn ) →a.s. h(Y ) and, by the dominated convergence theorem, E[h(Yn )] → E[h(Y )]. Then (a) follows from E[h(Xn )] = E[h(Yn )] for all n and E[h(X)] = E[h(Y )]. Next, we show (a) implies (b). Let C be a closed set and fC (x) = inf{kx − yk : y ∈ C}. Then fC is continuous. For j = 1, 2, ..., define ϕj (t) = I(−∞,0] + (1 − jt)I(0,j −1 ] . Then hj (x) = ϕj (fC (x)) is continuous and bounded, hj ≥ hj+1 , j = 1, 2, ..., and hj (x) → IC (x) as j → ∞. Hence lim supn PXn (C) ≤ limn→∞ E[hj (Xn )] = E[hj (X)] for each j (by (a)). By the dominated convergence theorem, E[hj (X)] → E[IC (X)] = PX (C).

1.5. Asymptotic Theory

57

This proves (b). For any open set O, Oc is closed. Hence, (b) is equivalent to (c). Now, we show (b) and (c) imply Xn →d X. For x = (x1 , ..., xk ) ∈ Rk , let (−∞, x] = (−∞, x1 ] × · · · × (−∞, xk ] and (−∞, x) = (−∞, x1 ) × · · · × (−∞, xk ). From (b) and (c), PX (−∞, x) ≤ lim inf n PXn (−∞, x) ≤ liminf n FXn (x) ≤ lim supn FXn (x) = lim supn PXn (−∞, x] ≤ PX (−∞, x] = FX (x). If x is a continuity point of FX , then PX (−∞, x) = FX (x). This proves Xn →d X and completes the proof of (i). √ τ (ii) From (a) of part (i), Xn →d X implies φXn (t) → φX (t), since e −1t x = √ cos(tτ x) + −1 sin(tτ x) and cos(tτ x) and sin(tτ x) are bounded continuous functions for any fixed t. Suppose now that k = 1 and that φXn (t) → φX (t) for every t ∈ R. By Fubini’s theorem, Z Z ∞ Z u √ 1 u 1 −1tx [1 − φXn (t)]dt = (1 − e )dt dPXn (x) u −u −∞ u −u Z ∞ sin ux dPXn (x) 1− =2 ux −∞ Z 1 dPXn (x) 1− ≥2 |ux| {|x|>2u−1 } ≥ PXn (−∞, −2u−1) ∪ (2u−1 , ∞) for any u > 0. Since φX is continuous at 0 and φX (0) = 1, for any ǫ > 0 Ru there is a u > 0 such that u−1 −u [1 − φX (t)]dt < ǫ/2. Since φXn → φX , Ru by the dominated convergence theorem, supn {u−1 −u [1 − φXn (t)]dt} < ǫ. Hence, Z u 1 inf PXn [−2u−1 , 2u−1 ] ≥ 1 − sup [1 − φXn (t)]dt ≥ 1 − ǫ, n u −u n

i.e., {PXn } is tight. Let {PXnj } be any subsequence that converges to a probability measure P . By the first part of the proof, φXnj → φ, which is the ch.f. of P . By the convergence of φXn , φ = φX . By Theorem 1.6(i), P = PX . By Proposition 1.17(ii), Xn →d X. Consider now the case where k ≥ 2 and φXn → φX . Let Ynj be the jth component of Xn and Yj be the jth component of X. Then φYnj → φYj for each j. By the proof for the case of k = 1, Ynj →d Yj . By Proposition 1.17(i), {PYnj } is tight, j = 1, ..., k. This implies that {PXn } is tight (why?). Then the proof for Xn →d X is the same as that for the case of k = 1. (iii) From (1.52), φcτ Xn (u) = φXn (uc) and φcτ X (u) = φX (uc) for any u ∈ R and any c ∈ Rk . Hence, convergence of φXn to φX is equivalent to convergence of φcτ Xn to φcτ X for every c ∈ Rk . Then the result follows from part (ii).

58

1. Probability Theory

Example 1.28. Let X1 , ..., Xn be independent random variables having a common c.d.f. and Tn = X1 + · · · + Xn , n = 1, 2, .... Suppose that E|X1 | < ∞. It follows from (1.56) and a result in calculus that the ch.f. of X1 satisfies √ φX1 (t) = φX1 (0) + −1µt + o(|t|) as |t| → 0, where µ = EX1 . From (1.52) and (1.58), the ch.f. of Tn /n is √ n n −1µt t t +o φTn /n (t) = φX1 = 1+ n n n for any t ∈ R, as n → ∞. Since (1 + cn /n)n → ec for any complex sequence √ {cn } satisfying cn → c, we obtain that φTn /n (t) → e −1µt , which is the ch.f. of the distribution degenerated at µ (i.e., the point mass probability measure at µ; see (1.22)). By Theorem 1.9(ii), Tn /n →d µ. From Theorem 1.8(vii), this also shows that Tn /n →p µ. Similarly, µ = 0 and σ 2 = Var(X1 ) < ∞ imply 2 n t σ 2 t2 +o φTn /√n (t) = 1 − 2n n 2 2

for any t ∈ R, which implies that φTn /√n (t) → e−σ t /2 , the ch.f. of √ N (0, σ 2 ). Hence Tn / n →d N (0, σ 2 ). (Recall that N (µ, σ 2 ) denotes a random variable having the N (µ, σ 2 ) distribution.) If µ 6= 0, a transforma√ tion of Yi = Xi − µ leads to (Tn − nµ)/ n →d N (0, σ 2 ). Suppose now that X1 , ..., Xn are random k-vectors and µ = EX1 and Σ = Var(X1 ) are finite. For any√fixed c ∈ Rk , it follows from the previous discussion that (cτ Tn − ncτ µ)/ n →d N (0, cτ Σc). From Theorem 1.9(iii) and a property √ of the normal distribution (Exercise 81), we conclude that (Tn − nµ)/ n →d Nk (0, Σ). Example 1.28 shows that Theorem 1.9(ii) together with some properties of ch.f.’s can be applied to show convergence in distribution for sums of independent random variables (vectors). The following is another example. Example 1.29. Let X1 , ..., Xn be independent random variables having a common Lebesgue p.d.f. f (x) = (1 − cos x)/(πx2 ). Then the ch.f. of X1 is max{1 − |t|, 0} (Exercise 73) and the ch.f. of Tn /n = (X1 + · · · + Xn )/n is n |t| max 1 − , 0 → e−|t| , t ∈ R. n Since e−|t| is the ch.f. of the Cauchy distribution C(0, 1) (Table 1.2), we conclude that Tn /n →d X, where X has the Cauchy distribution C(0, 1). Does this result contradict the first result in Example 1.28?

1.5. Asymptotic Theory

59

Other examples of applications of Theorem 1.9 are given in Exercises 135-140 in §1.6. The following result can be used to check whether Xn →d X when X has a p.d.f. f and Xn has a p.d.f. fn . Proposition 1.18 (Scheff´e’s theorem). Let {fn } be a sequence of p.d.f.’s on Rk w.r.t. a measure ν. Suppose that R limn→∞ fn (x) = f (x) a.e. ν and f (x) is a p.d.f. w.r.t. ν. Then limn→∞ |fn (x) − f (x)|dν = 0. Proof. Let gn (x) = [f (x) − fn (x)]I{f ≥fn } (x), n = 1, 2,.... Then Z Z |fn (x) − f (x)|dν = 2 gn (x)dν. Since 0 ≤ gn (x) ≤ f (x) for all x and gn → 0 a.e. ν, the result follows from the dominated convergence theorem. As an example, consider the Lebesgue p.d.f. fn of the t-distribution tn (Table 1.2), n = 1, 2,.... One can show (exercise) that fn → f , where f is the standard normal p.d.f. This is an important result in statistics.

1.5.3 Convergence of transformations Transformation is an important tool in statistics. For random vectors Xn converging to X in some sense, we often want to know whether g(Xn ) converges to g(X) in the same sense. The following result provides an answer to this question in many problems. Its proof is left to the reader. Theorem 1.10. Let X, X1 , X2 , ... be random k-vectors defined on a probability space and g be a measurable function from (Rk , B k ) to (Rl , B l ). Suppose that g is continuous a.s. PX . Then (i) Xn →a.s. X implies g(Xn ) →a.s. g(X); (ii) Xn →p X implies g(Xn ) →p g(X); (iii) Xn →d X implies g(Xn ) →d g(X). Example 1.30. (i) Let X1 , X2 , ... be random variables. If Xn →d X, where X has the N (0, 1) distribution, then Xn2 →d Y , where Y has the chi-square distribution χ21 (Example 1.14). (ii) Let (Xn , Yn ) be random 2-vectors satisfying (Xn , Yn ) →d (X, Y ), where X and Y are independent random variables having the N (0, 1) distribution, then Xn /Yn →d X/Y , which has the Cauchy distribution C(0, 1) (§1.3.1). (iii) Under the conditions in part (ii), max{Xn , Yn } →d max{X, Y }, which has the c.d.f. [Φ(x)]2 (Φ(x) is the c.d.f. of N (0, 1)). In Example 1.30(ii) and (iii), the condition that (Xn , Yn ) →d (X, Y ) cannot be relaxed to Xn →d X and Yn →d Y (exercise); i.e., we need the

60

1. Probability Theory

convergence of the joint c.d.f. of (Xn , Yn ). This is different when →d is replaced by →p or →a.s. . The following result, which plays an important role in probability and statistics, establishes the convergence in distribution of Xn + Yn or Xn Yn when no information regarding the joint c.d.f. of (Xn , Yn ) is provided. Theorem 1.11 (Slutsky’s theorem). Let X, X1 , X2 , ..., Y1 , Y2 , ... be random variables on a probability space. Suppose that Xn →d X and Yn →p c, where c is a fixed real number. Then (i) Xn + Yn →d X + c; (ii) Yn Xn →d cX; (iii) Xn /Yn →d X/c if c 6= 0. Proof. We prove (i) only. The proofs of (ii) and (iii) are left as exercises. Let t ∈ R and ǫ > 0 be fixed constants. Then FXn +Yn (t) = P (Xn + Yn ≤ t)

≤ P ({Xn + Yn ≤ t} ∩ {|Yn − c| < ǫ}) + P (|Yn − c| ≥ ǫ) ≤ P (Xn ≤ t − c + ǫ) + P (|Yn − c| ≥ ǫ)

and, similarly, FXn +Yn (t) ≥ P (Xn ≤ t − c − ǫ) − P (|Yn − c| ≥ ǫ). If t − c, t − c + ǫ, and t − c − ǫ are continuity points of FX , then it follows from the previous two inequalities and the hypotheses of the theorem that FX (t − c − ǫ) ≤ lim inf FXn +Yn (t) ≤ lim sup FXn +Yn (t) ≤ FX (t − c + ǫ). n

n

Since ǫ can be arbitrary (why?), lim FXn +Yn (t) = FX (t − c).

n→∞

The result follows from FX+c (t) = FX (t − c). An application of Theorem 1.11 is given in the proof of the following important result. Theorem 1.12. Let X1 , X2 , ... and Y be random k-vectors satisfying an (Xn − c) →d Y,

(1.76)

where c ∈ Rk and {an } is a sequence of positive numbers with limn→∞ an = ∞. Let g be a function from Rk to R. (i) If g is differentiable at c, then an [g(Xn ) − g(c)] →d [∇g(c)]τ Y,

(1.77)

61

1.5. Asymptotic Theory

where ∇g(x) denotes the k-vector of partial derivatives of g at x. (ii) Suppose that g has continuous partial derivatives of order m > 1 in a neighborhood of c, with all the partial derivatives of order j, 1 ≤ j ≤ m− 1, vanishing at c, but with the mth-order partial derivatives not all vanishing at c. Then k k X 1 X ∂mg m Yi1 · · · Yim , (1.78) an [g(Xn ) − g(c)] →d ··· m! i =1 ∂x i 1 · · · ∂xim x=c i =1 1

m

where Yj is the jth component of Y . Proof. We prove (i) only. The proof of (ii) is similar. Let

Zn = an [g(Xn ) − g(c)] − an [∇g(c)]τ (Xn − c). If we can show that Zn = op (1), then by (1.76), Theorem 1.9(iii), and Theorem 1.11(i), result (1.77) holds. The differentiability of g at c implies that for any ǫ > 0, there is a δǫ > 0 such that |g(x) − g(c) − [∇g(c)]τ (x − c)| ≤ ǫkx − ck (1.79) whenever kx − ck < δǫ . Let η > 0 be fixed. By (1.79),

P (|Zn | ≥ η) ≤ P (kXn − ck ≥ δǫ ) + P (an kXn − ck ≥ η/ǫ). Since an → ∞, (1.76) and Theorem 1.11(ii) imply Xn →p c. By Theorem 1.10(iii), (1.76) implies an kXn − ck →d kY k. Without loss of generality, we can assume that η/ǫ is a continuity point of FkY k . Then lim sup P (|Zn | ≥ η) ≤ lim P (kXn − ck ≥ δǫ ) n

n→∞

+ lim P (an kXn − ck ≥ η/ǫ) n→∞

= P (kY k ≥ η/ǫ). The proof is complete since ǫ can be arbitrary. In statistics, we often need a nondegenerated limiting distribution of an [g(Xn ) − g(c)] so that probabilities involving an [g(Xn ) − g(c)] can be approximated by the c.d.f. of [∇g(c)]τ Y , if (1.77) holds. Hence, result (1.77) is not useful for this purpose if ∇g(c) = 0, and in such cases result (1.78) may be applied. A useful method in statistics, called the delta-method, is based on the following corollary of Theorem 1.12. Corollary 1.1. Assume the conditions of Theorem 1.12. If Y has the Nk (0, Σ) distribution, then an [g(Xn ) − g(c)] →d N (0, [∇g(c)]τ Σ∇g(c)) .

62

1. Probability Theory

Example 1.31. Let {Xn } be a sequence of random variables satisfying √ n(Xn − c) →d N (0, 1). Consider the function g(x) = x2 . If c 6= 0, then an √ 2 application of Corollary 1.1 gives that n(Xn − c2 ) →d N (0, 4c2 ). If c = 0, the first-order derivative of g at 0 is 0 but the second-order derivative of g ≡ 2. Hence, an application of result (1.78) gives that nXn2 →d [N (0, 1)]2 , which has the chi-square distribution χ21 (Example 1.14). The last result can also be obtained by applying Theorem 1.10(iii).

1.5.4 The law of large numbers The law of large numbers concerns the limiting behavior of sums of independent random variables. The weak law of large numbers (WLLN) refers to convergence in probability, whereas the strong law of large numbers (SLLN) refers to a.s. convergence. The following lemma is useful in establishing the SLLN. Its proof is left as an exercise. Lemma 1.6. (Kronecker’s lemma). Let xnP∈ R, an ∈ R, 0 < an ≤ ∞ an+1P , n = 1, 2, ..., and an → ∞. If the series n=1 xn /an converges, then n a−1 n i=1 xi → 0.

Our first result gives the WLLN and SLLN for a sequence of independent and identically distributed (i.i.d.) random variables.

Theorem 1.13. Let X1 , X2 , ... be i.i.d. random variables. (i) (The WLLN). A necessary and sufficient condition for the existence of a sequence of real numbers {an } for which n

1X Xi − an →p 0 n i=1

(1.80)

is that nP (|X1 | > n) → 0, in which case we may take an = E(X1 I{|X1 |≤n} ). (ii) (The SLLN). A necessary and sufficient condition for the existence of a constant c for which n 1X Xi →a.s. c (1.81) n i=1 is that E|X1 | < ∞, in which case c = EX1 and n

1X ci (Xi − EX1 ) →a.s. 0 n i=1 for any bounded sequence of real numbers {ci }.

(1.82)

63

1.5. Asymptotic Theory

Proof. (i) We prove the sufficiency. The proof of necessity can be found in Petrov (1975). Consider a sequence of random variables obtained by truncating Xj ’s at n: Ynj = Xj I{|Xj |≤n} . Let Tn = X1 + · · · + Xn and Zn = Yn1 + · · · + Ynn . Then P (Tn 6= Zn ) ≤

n X j=1

P (Ynj 6= Xj ) = nP (|X1 | > n) → 0.

(1.83)

For any ǫ > 0, it follows from Chebyshev’s inequality that 2 Zn − EZn Var(Zn ) EYn1 Var(Yn1 ) >ǫ ≤ ≤ , = P n ǫ 2 n2 ǫ2 n ǫ2 n

where the last equality follows from the fact that Ynj , j = 1, ..., n, are i.i.d. From integration by parts, we obtain that Z Z 2 EYn1 1 2 n = x2 dF|X1 | (x) = xP (|X1 | > x)dx − nP (|X1 | > n), n n [0,n] n 0

which converges to 0 since nP (|X1 | > n) → 0 (why?). This proves that (Zn − EZn )/n →p 0, which together with (1.83) and the fact that EYnj = E(X1 I{|X1 |≤n} ) imply the result. (ii) For the sufficiency, let Yn = Xn I{|Xn |≤n} , n = 1, 2, .... Let m > 0 be an integer smaller than n. If we define ci = i−1 for i ≥ m, Z1 = · · · = Zm−1 = 0, Zm = Y1 + · · · + Ym , Zi = Yi , i = m + 1, ..., n, and apply the H´ajek-R`enyi inequality (1.51) to Zi ’s, then we obtain that for any ǫ > 0, m n 1 X 1 X Var(Yi ) P max |ξl | > ǫ ≤ 2 2 Var(Yi ) + 2 , (1.84) m≤l≤n ǫ m i=1 ǫ i=m+1 i2 where ξn = n−1 that

Pn

i=1 (Zi

− EZi ) (= n−1

Pn

i=1 (Yi

− EYi ) if l ≥ m). Note

n ∞ ∞ X X X E(X12 I{j−1 n) =

∞ X

n=1

P (|X1 | > n) < ∞

∞ (Exercise 54) and Lemma 1.5(i) that P (∩∞ n=1 ∪m=n {Xm 6= Ym }) = 0, i.e., there is an event A with P (A) = 1 such that if ω ∈ A, then Xn (ω) = Yn (ω) for sufficiently large n. This implies n

n

1X 1X Xi − Yi →a.s. 0, n i=1 n i=1

(1.85)

which proves the sufficiency. The proof of (1.82) is left as an exercise. We now prove the necessity. Suppose that (1.81) holds for some c ∈ R. Then Xn Tn n − 1 Tn−1 c = −c− − c + →a.s. 0. n n n n−1 n

From Exercise 114, Xn /n →a.s. 0 and the i.i.d. assumption on Xn ’s imply ∞ X

n=1

P (|Xn | ≥ n) =

∞ X

n=1

P (|X1 | ≥ n) < ∞,

which implies E|X1 | < ∞ (Exercise 54). From the proved sufficiency, c = EX1 . If E|X1 | < ∞, then an in (1.80) converges to EX1 and result (1.80) is actually established in Example 1.28 in a much simpler way. On the other hand, if E|X1 | < ∞, then the stronger result (1.81) can be obtained. Some results for the case of E|X1 | = ∞ can be found in Exercise 148 in §1.6 and Theorem 5.4.3 in Chung (1974). The next result is for sequences of independent but not necessarily identically distributed random variables.

65

1.5. Asymptotic Theory

Theorem 1.14. Let X1 , X2 , ... be independent random variables with finite expectations. (i) (The SLLN). If there is a constant p ∈ [1, 2] such that ∞ X E|Xi |p i=1

then

ip

< ∞,

(1.86)

n

1X (Xi − EXi ) →a.s. 0. n i=1

(1.87)

(ii) (The WLLN). If there is a constant p ∈ [1, 2] such that n 1 X E|Xi |p = 0, n→∞ np i=1

(1.88)

lim

then

n

1X (Xi − EXi ) →p 0. n i=1

(1.89)

Proof. (i) Consider again the truncated Xn : Yn = Xn I{|Xn |≤n} , n = 1, 2, .... Since Xn2 I{|Xn |≤n} ≤ n2−p |Xn |p , ∞ ∞ ∞ X X E(Xn2 I{|Xn |≤n} ) X E|Xn |p EYn2 = ≤ < ∞. n2 n2 np n=1 n=1 n=1

It follows from the proof of Theorem 1.13(ii) that n−1 0. Also, ∞ X

n=1

P (Xn 6= Yn ) =

∞ X

n=1

P (|Xn | > n) ≤

Pn

i=1 (Yi −EYi )

→a.s.

∞ X E|Xn |p < ∞. np n=1

Hence, it follows from the proof of Theorem 1.13(ii) that (1.85) holds. Finally, ∞ ∞ ∞ X |E(Xn − Yn )| X E(|Xn |I{|Xn |>n} ) X E|Xn |p = ≤ < ∞, n n np n=1 n=1 n=1

P which together with Lemma 1.6 imply that n−1 ni=1 |E(Xi − Yi )| → 0 and thus (1.87) holds. (ii) For any ǫ > 0, an application of Chebyshev’s inequality and inequality (1.44) leads to ! n n X 1 X > ǫ ≤ Cp P (X − EX ) E|Xi − EXi |p , i i n i=1 ǫp np i=1 which converges to 0 under (1.88). This proves (1.89).

66

1. Probability Theory

Note that (1.86) implies (1.88) (Lemma 1.6). The result in Theorem 1.14(i) is called Kolmogorov’s SLLN when p = 2 and is due to Marcinkiewicz and Zygmund when 1 ≤ p < 2. An obvious sufficient condition for (1.86) with p ∈ (1, 2] is supn E|Xn |p < ∞. For dependent random variables, a result for Markov chains introduced in §1.4.4 is discussed in §4.1.4. We now consider martingales studied in §1.4.4. First, consider the WLLN. Inequality (1.44) still holds if the independence assumption of Xi ’s is replaced by the martingale assumption on Pn the sequence { i=1 (Xi −EXi )} (why?). Hence, from the proof of Theorem 1.14(ii) we conclude that (1.89) still holds if P the independence assumption of Xi ’s in Theorem 1.14 is replaced by that { ni=1 (Xi − EXi )} is a martingale. A result similar to the SLLN in Theorem 1.14(i) can be established ifP the independence assumption of Xi ’s is replaced by that the sequence n { i=1 (Xi − EXi )} is a martingale and if condition (1.86) is replaced by ∞ X E(|Xn |p |X1 , ..., Xn−1 ) < ∞ a.s., np n=2

which is the same as (1.86) if Xi ’s are independent. The proof of this martingale SLLN and many other versions of WLLN and SLLN can be found in standard probability textbooks, for example, Chung (1974) and Lo`eve (1977). The WLLN and SLLN have many applications in probability and statistics. The following is an example. Other examples can be found in later chapters. Example 1.32. Let f and g be continuous functions on [0, 1] satisfying 0 ≤ f (x) ≤ Cg(x) for all x, where C > 0 is a constant. We now show that lim

n→∞

Z

0

1

Z

0

1

···

Z

0

1

R1 Pn f (x)dx f (xi ) i=1 Pn dx1 dx2 · · · dxn = R01 g(x ) i i=1 0 g(x)dx

(1.90)

R1 (assuming that 0 g(x)dx 6= 0). Let X1 , X2 , ... be i.i.d. random variables having the uniform distribution on [0, 1]. By Theorem 1.2, E[f (X1 )] = R1 R1 0 f (x)dx < ∞ and E[g(X1 )] = 0 g(x)dx < ∞. By the SLLN (Theorem 1.13(ii)), n 1X f (Xi ) →a.s. E[f (X1 )], n i=1

and the same result holds when f is replaced by g. By Theorem 1.10(i), Pn f (Xi ) E[f (X1 )] Pi=1 . (1.91) →a.s. n E[g(X1 )] g(X ) i i=1

67

1.5. Asymptotic Theory

Since the random variable on the left-hand side of (1.91) is bounded by C, result (1.90) follows from the dominated convergence theorem and the fact that the left-hand side of (1.90) is the expectation of the random variable on the left-hand side of (1.91). Moment inequalities introduced in §1.3.2 play important roles in proving convergence theorems. They can also be P used to obtain convergence n rates of tail probabilities of the form P |n−1 i=1 (Xi − EXi )| > t . For example, an application of the Esseen-von Bahr, Marcinkiewicz-Zygmund, and Chebyshev inequalities produces ! n 1 X if 1 < p < 2 O(n1−p ) P (Xi − EXi ) > t ≤ if p ≥ 2 O(n−p/2 ) n i=1

for independent random variables X1 , ..., Xn with supn E|Xn |p < ∞.

1.5.5 The central limit theorem The WLLN and SLLN may not be useful in approximating the distributions of (normalized) sums of independent random variables. We need to use the central limit theorem (CLT), which plays a fundamental role in statistical asymptotic theory. Theorem 1.15 (Lindeberg’s CLT). Let {Xnj , j = 1, ..., kn } be independent P n random variables with 0 < σn2 = Var( kj=1 Xnj ) < ∞, n = 1, 2,..., and kn → ∞ as n → ∞. If kn X j=1

then

E (Xnj − EXnj )2 I{|Xnj −EXnj |>ǫσn } = o(σn2 ) for any ǫ > 0, (1.92) kn 1 X (Xnj − EXnj ) →d N (0, 1). σn j=1

(1.93)

Proof. Considering (Xnj − EXnj )/σn , without loss of generality we may assume EXnj = 0 and σn2 = 1 in this proof. Let t ∈ R be given. From the √ √ −1tx inequality |e − (1 + −1tx − t2 x2 /2)| ≤ min{|tx|2 , |tx|3 }, the ch.f. of Xnj satisfies 2 φXnj (t) − 1 − t2 σnj /2 ≤ E min{|tXnj |2 , |tXnj |3 } , (1.94) 2 = Var(Xnj ). For any ǫ > 0, the right-hand side of (1.94) is where σnj bounded by E(|tXnj |3 I{|Xnj |ǫ} ) → ǫ2 for arbitrary ǫ > 0. Hence

lim max

n→∞ j≤kn

2 σnj = 0. σn2

(1.96)

2 (Note that σn2 = 1 is assumed for convenience.) This implies that 1 − t2 σnj are all between 0 and 1 for large enough n. Using the inequality

|a1 · · · am − b1 · · · bm | ≤

m X j=1

|aj − bj |

for any complex numbers aj ’s and bj ’s with |aj | ≤ 1 and |bj | ≤ 1, j = 1, ..., m, we obtain that kn kn kn Y −t2 σ2 /2 Y −t2 σ2 /2 X 2 2 2 2 nj nj 1 − t σnj /2 ≤ e − − 1 − t σnj /2 , e j=1

j=1

j=1

Pkn 4

4 4 2 x which is bounded by t j=1 σnj ≤ t maxj≤kn σnj → 0, since |e −1−x| ≤ P n 2 σnj = σn2 = 1. Also, x2 /2 if |x| ≤ 12 and kj=1

kn kn Y Y 2 2 1 − t φ (t) − σ /2 Xnj nj j=1

j=1

is bounded by the quantity on the left-hand side of (1.95) and, hence, converges to 0 by (1.95). Thus, kn Y

j=1

φXnj (t) =

kn Y

2

e−t

2 σnj /2

2

+ o(1) = e−t

/2

+ o(1).

j=1

P n Xnj converges to the ch.f. of N (0, 1) for This shows that the ch.f. of kj=1 every t. By Theorem 1.9(ii), the result follows. Condition (1.92) is called Lindeberg’s condition. From the proof of Theorem 1.15, Lindeberg’s condition implies (1.96), which is called Feller’s condition. Feller’s condition (1.96) means that all terms in the sum σn2 =

69

1.5. Asymptotic Theory Pkn

2 σnj are uniformly negligible as n → ∞. If Feller’s condition is assumed, then Lindeberg’s condition is not only sufficient but also necessary for result (1.93), which is the well-known Lindeberg-Feller CLT. A proof can be found in Billingsley (1986, pp. 373-375). Note that neither Lindeberg’s condition nor Feller’s condition is necessary for result (1.93) (Exercise 158). A sufficient condition for Lindeberg’s condition is the following Liapounov’s condition, which is somewhat easier to verify: j=1

kn X j=1

E|Xnj − EXnj |2+δ = o(σn2+δ ) for some δ > 0.

(1.97)

Example 1.33. Let X1 , X2 , ... be independent random variables. Suppose that distribution Bi(pi , 1), i = 1, 2,..., and that σn2 = Pn Pn Xi has the binomial i=1 Var(Xi ) = i=1 pi (1 − pi ) → ∞ as n → ∞. For each i, EXi = p and E|X − EX |3 = (1 − pi )3 pi + p3i (1 − pi ) ≤ 2pi (1 − pi ). Hence i i i Pn 3 2 i=1 E|Xi − EXi | ≤ 2σn , i.e., Liapounov’s condition (1.97) holds with δ = 1. Thus, by Theorem 1.15, n 1 X (Xi − pi ) →d N (0, 1). σn i=1

(1.98)

It can be shown (exercise) that the condition σn → ∞ is also necessary for result (1.98). The following are useful corollaries of Theorem 1.15 (and Theorem 1.9(iii)). Corollary 1.2 is in fact proved in Example 1.28. The proof of Corollary 1.3 is left as an exercise. Corollary 1.2 (Multivariate CLT). Let X1 , ..., Xn be i.i.d. random kvectors with a finite Σ = Var(X1 ). Then n

1 X √ (Xi − EX1 ) →d Nk (0, Σ). n i=1 Corollary 1.3. Let Xni ∈ Rmi , i = 1, ..., kn , be independent random vectors with mi ≤ m (a fixed integer), n = 1, 2,..., kn → ∞ as n → ∞, and inf i,n λ− [Var(Xni )] > 0, where λ− [A] is the smallest eigenvalue of A. Let cni ∈ Rmi be vectors such that ! X kn max kcni k2 kcni k2 = 0. lim n→∞

1≤i≤kn

i=1

70

1. Probability Theory

(i) Suppose that supi,n EkXni k2+δ < ∞ for some δ > 0. Then kn X i=1

cτni (Xni

#1/2 "X kn τ − EXni ) Var(cni Xni ) →d N (0, 1).

(1.99)

i=1

(ii) Suppose that whenever mi = mj , 1 ≤ i < j ≤ kn , n = 1, 2, ..., Xni and Xnj have the same distribution with EkXni k2 < ∞. Then (1.99) holds. Applications of these corollaries can be found in later chapters. An extension of Lindeberg’s CLT is the so-called martingale CLT. In Theorem 1.15, if the independence assumption of Xnj , j = 1, ..., kn , is replaced by that {Yn } is a martingale and kn 1 X E[(Xnj − EXnj )2 |Xn1 , ..., Xn(j−1) ] →p 1, σn2 j=1

P n where Yn = kj=1 (Xnj − EXnj ) when n ≤ kn , Yn = Ykn when n > kn , and Xn0 is defined to be 0, then result (1.93) still holds (see, e.g., Billingsley, 1986, p. 498 and Sen and Singer 1993, p. 120). More results on the CLT can be found, for example, in Serfling (1980) and Shorack and Wellner (1986). Let Yn be a sequence of random variables, {µn } and {σn } be sequences of real numbers such that σn > 0 for all n, and (Yn − µn )/σn →d N (0, 1). Then, by Proposition 1.16, lim sup |F(Yn −µn )/σn (x) − Φ(x)| = 0,

n→∞ x

(1.100)

where Φ is the c.d.f. of N (0, 1). This implies that for any sequence of real n numbers {cn }, limn→∞ |P (Yn ≤ cn ) − Φ cnσ−µ | = 0, i.e., P (Yn ≤ cn ) can n n be approximated by Φ cnσ−µ , regardless of whether {cn } has a limit. Since n t−µn 2 Φ σn is the c.d.f. of N (µn , σn ), Yn is said to be asymptotically distributed Pkn τ cni Xni as N (µn , σn2 ) or simply asymptotically normal. For example, i=1 in Corollary 1.3 is asymptotically normal. This can be extended to ranPn dom vectors. For example, i=1 Xi in Corollary 1.2 is asymptotically distributed as Nk (nEX1 , nΣ).

1.5.6 Edgeworth and Cornish-Fisher expansions Let {Yn } be a sequence of random variables satisfying (1.100) and Wn = (Yn − µn )/σn . The convergence speed of (1.100) can be used to assess whether Φ provides a good approximation to the c.d.f. FWn . Also, sometimes we would like to find an approximation to FWn that is better than

71

1.5. Asymptotic Theory

Φ in terms of convergence speed. The Edgeworth expansion is a useful tool for these purposes. P To illustrate the idea, let Wn = n−1/2 ni=1 (Xi −µ)/σ, where X1 , X2 , ... are i.i.d. random variables with EX1 = µ and Var(X1 ) = σ 2 . Assume that the m.g.f. of Z = (X1 − µ)/σ is finite and positive in a neighborhood of 0. From (1.55), the cumulant generating function of Z has the expansion κ(t) =

∞ X κj j=1

j!

tj ,

where κj , j = 1, 2, ..., are cumulants of Z (e.g., κ1 = 0, κ2 = 1, κ3 = EZ 3 , and κ4 = EZ 4 − 3), and the m.g.f. of Wn is equal to 2 X ∞ √ n κ j tj t + , ψn (t) = exp{κ(t/ n)} = exp 2 j!n(j−2)/2 j=3 where exp{x} denotes the exponential function ex . Using the series expan2 sion for et /2 , we obtain that 2

ψn (t) = et

/2

2

+ n−1/2 r1 (t)et

/2

2

+ · · · + n−j/2 rj (t)et

/2

+ ···,

(1.101)

where rj is a polynomial of degree 3j depending on κ3 , ..., κj+2 but not on n, j = 1, 2, .... For example, it can be shown (exercise) that r1 (t) = 16 κ3 t3

and r2 (t) =

1 4 24 κ4 t

+

1 2 6 72 κ3 t .

(1.102)

R R 2 Since ψn (t) = etx dFWn (x) and et /2 = etx dΦ(x), expansion (1.101) suggests the inverse expansion FWn (x) = Φ(x) + n−1/2 R1 (x) + · · · + n−j/2 Rj (x) + · · · , R 2 where Rj (x) is a function satisfying etx dRj (x) = rj (t)et /2 , j = 1, 2, .... j d 1 Let ∇j = dx j be the differential operator and ∇ = ∇ . Then Rj (x) = rj (−∇)Φ(x), j = 1, 2, ..., where rj (−∇) is interpreted as a differential operator. Thus, Rj ’s can be obtained once rj ’s are derived. It follows from (1.102) (exercise) that R1 (x) = − 16 κ3 (x2 − 1)Φ′ (x)

(1.103)

and 1 κ4 x(x2 − 3) + R2 (x) = −[ 24

1 2 4 72 κ3 x(x

− 10x2 + 15)]Φ′ (x).

(1.104)

A rigorous statement of the Edgeworth expansion for a more general Wn is given in the following theorem whose proof can be found in Hall (1992).

72

1. Probability Theory

Theorem 1.16 (Edgeworth expansions). Let m be a positive integer and X1 , X2 , √ ... be i.i.d. random k-vectors having Pn finite m+2 moments. Consider −1 ¯ ¯ Wn = nh(X)/σ h , where X = n i=1 Xi , h is a Borel function on Rk that is m + 2 times continuously differentiable in a neighborhood of µ = EX1 , h(µ) = 0, and σh2 = [∇h(µ)]τ Var(X1 )∇h(µ) > 0. Assume that lim sup |φX1 (t)| < 1,

(1.105)

ktk→∞

where φX1 is the ch.f. of X1 . Then, FWn admits the Edgeworth expansion m X pj (x)Φ′ (x) 1 sup FWn (x) − Φ(x) − = o nm/2 , nj/2 x

(1.106)

j=1

where pj (x) is a polynomial of degree at most 3j − 1, odd for even j and even for odd j, with coefficients depending on the first m + 2 moments of X1 , j = 1, ..., m. In particular, (1.107) p1 (x) = −c1 σh−1 + 6−1 c2 σh−3 (x2 − 1) P P P P P k k k with c1 = 2−1 ki=1 kj=1 aij µij and c2 = i=1 j=1 l=1 ai aj al µijl + Pk Pk Pk Pk 3 i=1 j=1 l=1 h=1 ai aj alh µil µjh , where ai is the ith component of ∇h(µ), aij is the (i, j)th element of the Hessian matrix ∇2 h(µ), µij = E(Yi Yj ), µijl = E(Yi Yj Yl ), and Yi is the ith component of X1 − µ. Condition (1.105) is Cram´er’s continuity condition. It is satisfied if one component of X1 has a Lebesgue p.d.f. The polynomial pj with j ≥ 2 may be derived using the method in deriving (1.103) and (1.104), but the derivation is usually complicated (see Hall (1992)). Under the conditions of Theorem 1.16, the convergence speed of (1.100) Pm is O(n−1/2 ) and, as an approximation to FWn , Φ+ j=1 n−j/2 pj Φ′ is better than Φ, since its convergence speed is o(n−m/2 ). The results in Theorem 1.16 can be applied to many cases, as the following example indicates. ¯ = n−1 Pn Xi with i.i.d. random variables X1 , X2 , Example 1.34. Let X i=1 ... satisfying condition (1.105). First, consider the normalized random √ ¯ variable Wn = n(X − µ)/σ, where µ = EX1 and σ 2 = Var(X1 ). Then, Theorem 1.16 can be applied with h(x) = x − µ and σh2 = σ 2 , and the Edgeworth expansion in (1.106) holds if E|X1 |m+2 < ∞. In this case, results (1.103) and (1.104) imply that pj (x) = Rj (x)/Φ′ (x), j = 1, 2. √ ¯ Next, consider the studentized random variable Wn = n(X − µ)/ˆ σ, P n ¯ 2 . Assuming that EX 2m+4 < ∞ and applywhere σ ˆ 2 = n−1 i=1 (Xi − X) 1 ing Theorem 1.16 to random vectors (Xi , Xi2 ), i = 1, 2, ..., and h(x, y) =

73

1.5. Asymptotic Theory

(x−µ)/

p

(y − x2 ), we obtain the Edgeworth expansion (1.106) with σh = 1, p1 (x) = 16 κ3 (2x2 + 1)

(exercise). Furthermore, it can be found in Hall (1992, p. 73) that p2 (x) =

1 2 12 κ4 x(x

− 3) −

1 2 4 18 κ3 x(x

+ 2x2 − 3) − 14 x(x2 + 3).

√ σ 2 − σ 2 ). Theorem 1.16 can be Consider now the random variable n(ˆ 2 applied to random vectors (Xi , Xi ), i = 1, 2, ..., and h(x, y) = (y − x2 − σ 2 ). Assume that EX12m+4 < ∞. It can be √ shown (exercise) that the Edgeworth expansion in (1.106) holds with Wn = n(ˆ σ 2 − σ 2 )/σh , σh2 = E(X1 − µ)4 − 4 σ , and p1 (x) = (ν4 − 1)−1/2 [1 − 16 (ν4 − 1)−1 (ν6 − 3ν4 − 6ν32 + 2)(x2 − 1)], where νj = σ −j E(X1 − µ)j , j = 3, ..., 6. √ σ 2 − σ 2 )/ˆ τ, Finally, consider random variable Wn = n(ˆ Pnthe studentized 2 −1 4 4 ¯ where τˆ = n (X − X) − σ ˆ . Theorem 1.16 can be applied to i i=1 random vectors (Xi , Xi2 , Xi3 , Xi4 ), i = 1, 2, ..., and h(x, y, z, w) = (y − x2 − σ 2 )[w − y 2 − 4xz + 8x2 y − 4x4 ]−1/2 . Assume that EX14m+8 < ∞. It can be shown (exercise) that the Edgeworth expansion in (1.106) holds with σh2 = 1 and p1 (x) = −(ν4 −1)−3/2 [ 21 (4ν32 +ν4 −ν6 ) + 13 (3ν32 +3ν4 −ν6 −2)(x2 −1)]. An inverse Edgeworth expansion is referred to as a Cornish-Fisher expansion, which is useful in statistics (see §7.4). For α ∈ (0, 1), let zα = Φ−1 (α). Since the c.d.f. FWn may not be strictly increasing and continuous, we define wnα = inf{x : FWn (x) ≥ α}. The following result can be found in Hall (1992). Theorem 1.17 (Cornish-Fisher expansions). Under the conditions of Theorem 1.16, wnα admits the Cornish-Fisher expansion m X qj (zα ) 1 =o , (1.108) sup wnα − zα − nj/2 nm/2 ǫ 0, there is a δǫ such that ν(A) < δǫ and A ∈ F imply A |f |dν < ǫ.

25. Prove that part (i) and part (iii) of Theorem 1.1 are equivalent. 26. Prove Theorem 1.2.

27. Prove Theorem 1.3. (Hint: first consider simple nonnegative f .) 28. Consider Example 1.9. Show that (1.17) does not hold for i=j 1 f (i, j) = −1 i=j−1 0 otherwise. Does this contradict Fubini’s theorem?

29. Let f be a nonnegative Borel function on (Ω, F , ν) with a σ-finite ν, A = {(ω, x) ∈ Ω × R : 0 ≤ x ≤ f (ω)}, and Rm be the Lebesgue measure on (R, B). Show that A ∈ σ(F × B) and Ω f dν = ν × m(A).

1.6. Exercises

77

R 30. For any c.d.f. F and any a ≥ 0, show that [F (x + a) − F (x)]dx = a.

31. (Integration by parts). Let F and G be two c.d.f.’s on R. Show that if F andR G have no common points of discontinuity Rin the interval (a, b], then (a,b] G(x)dF (x) = F (b)G(b) − F (a)G(a) − (a,b] F (x)dG(x).

32. Let f be a Borel function on R2 such that f (x, y) = 0 for each x ∈ R and y 6∈ Cx , where m(Cx ) = 0 for each x and m is the Lebesgue measure. Show that f (x, y) = 0 for each y 6∈ C and x 6∈ By , where m(C) = 0 and m(By ) = 0 for each y 6∈ C. 33. RConsider Example 1.11. Show that if (1.21) holds, then R P (A) = f (x)dx for any Borel set A. (Hint: A = {A : P (A) = A A f (x)dx} is a σ-field containing all sets of the form (−∞, x].) 34. Prove Proposition 1.7. P∞ 35. Let {an } be a sequence of positive numbers satisfying n=1 an = 1 and let {Pn } be a sequence ofPprobability measures on a common ∞ measurable space. Define P = n=1 an Pn . (a) Show that P is a probability measure. (b) Show that Pn ≪ ν for all n and a measure only if P ≪ ν P∞ ν if and dPn and, when P ≪ ν and ν is σ-finite, dP = a . n n=1 dν dν (c) Derive the Lebesgue p.d.f. of P when Pn is the gamma distribution Γ(α, n−1 ) (Table 1.2) with α > 1 and an is proportional to n−α . 36. Let Fi be a c.d.f. having a Lebesgue p.d.f. fi , i = 1, 2. Assume that there is a c ∈ R such that F1 (c) < F2 (c). Define −∞ < x < c F1 (x) F (x) = c ≤ x < ∞. F2 (x) Show that the probability measure P corresponding to F satisfies P ≪ m + διc and find dP/d(m + διc ), where m + διc is given in (1.23). 37. Let (X, Y ) be a random 2-vector with the following Lebesgue p.d.f.: 8xy 0≤x≤y≤1 f (x, y) = 0 otherwise. Find the marginal p.d.f.’s of X and Y . Are X and Y independent? 38. Let (X, Y, Z) be a random 3-vector with the following Lebesgue p.d.f.: 1−sin x sin y sin z 0 ≤ x ≤ 2π, 0 ≤ y ≤ 2π, 0 ≤ z ≤ 2π 8π 3 f (x, y, z) = 0 otherwise. Show that X, Y , and Z are not independent, but are pairwise independent.

78

1. Probability Theory

39. Prove Lemma 1.1 without using Definition 1.7 for independence. 40. Let X be a random variable having a continuous c.d.f. F . Show that Y = F (X) has the uniform distribution U (0, 1) (Table 1.2). 41. Let U be a random variable having the uniform distribution U (0, 1) and let F be a c.d.f. Show that the c.d.f. of Y = F −1 (U ) is F , where F −1 (t) = inf{x ∈ R : F (x) ≥ t}. 42. Prove Proposition 1.8. 43. Let X = Nk (µ, Σ) with a positive definite Σ. (a) Let Y = AX + c, where A is an l × k matrix of rank l ≤ k and c ∈ Rl . Show that Y has the Nl (Aµ + c, AΣAτ ) distribution. (b) Show that the components of X are independent if and only if Σ is a diagonal matrix. (c) Let Λ be positive definite and Y = Nm (η, Λ) be independent of X. Show that (X, Y ) has the Nk+m ((µ, η), D) distribution, where D is a block diagonal matrix whose two diagonal blocks are Σ and Λ. 44. Let X be a random variable having the Lebesgue p.d.f. Derive the p.d.f. of Y = sin X.

2x π 2 I(0,π) (x).

45. Let Xi , i = 1, 2, 3, be independent random variables having the same Lebesgue p.d.f. f (x) = e−x I(0,∞) (x). Obtain the joint Lebesgue p.d.f. of (Y1 , Y2 , Y3 ), where Y1 = X1 + X2 + X3 , Y2 = X1 /(X1 + X2 ), and Y3 = (X1 + X2 )/(X1 + X2 + X3 ). Are Yi ’s independent? 46. Let X1 and X2 be independent random variables having the standard normalp distribution. Obtain the joint Lebesgue p.d.f. of (Y1 , Y2 ), where Y1 = X12 + X22 and Y2 = X1 /X2 . Are Yi ’s independent?

random variables and Y = X1 + X2 . 47. Let X1 and X2 be independent R Show that FY (y) = FX2 (y − x)dFX1 (x).

48. Show that the Lebesgue p.d.f.’s given by (1.31) and (1.33) are the p.d.f.’s of the χ2n (δ) and Fn1 ,n2 (δ) distributions, respectively. 49. Show that the Lebesgue p.d.f. given by (1.32) is the p.d.f. of the tn (δ) distribution.

50. Let X = Nn (µ, In ) and A be an n × n symmetric matrix. Show that if X τ AX has the χ2r (δ) distribution, then A2 = A, r is the rank of A, and δ = µτ Aµ. 51. Let X = Nn (µ, In ). Apply Cochran’s theorem (Theorem 1.5) to show that if A2 = A, then X τ AX has the noncentral chi-square distribution χ2r (δ), where A is an n × n symmetric matrix, r is the rank of A, and δ = µτ Aµ.

79

1.6. Exercises

and Xi =PN (0, σi2 ), i = 1, ..., n. Let 52. Let XP 1 , ..., Xn be independent Pn n n −2 −2 ˜ 2 . Apply ˜ X = i=1 σi Xi / i=1 σi and S˜2 = i=1 σi−2 (Xi − X) 2 2 ˜ ˜ Cochran’s theorem to show that X and S are independent and that S˜2 has the chi-square distribution χ2n−1 . 53. Let X = Nn (µ, In ) and Ai be an n × n symmetric matrix satisfying A2i = Ai , i = 1, 2. Show that a necessary and sufficient condition that X τ A1 X and X τ A2 X are independent is A1 A2 = 0. 54. Let X be a random variable and a > 0. Show that E|X|a < ∞ if and P∞ only if n=1 na−1 P (|X| ≥ n) < ∞.

55. Let X be a random variable. RShow that R0 ∞ (a) if EX exists, then EX = 0 P (X > x)dx − −∞ P (X ≤ x)dx; P∞ (b) if X has range {0, 1, 2, ...}, then EX = n=1 P (X ≥ n).

56. Let T be a random variable having the noncentral t-distribution tn (δ). Show that p (a) E(T ) = δΓ((n − 1)/2) n/2/Γ(n/2) when n > 1; h i2 2 ) δ 2 n Γ((n−1)/2) − when n > 2. (b) Var(T ) = n(1+δ n−2 2 Γ(n/2)

57. Let F be a random variable having the noncentral F-distribution Fn1 ,n2 (δ). Show that (n1 +δ) (a) E(F) = nn12 (n when n2 > 2; 2 −2) (b) Var(F) =

2n22 [(n1 +δ)2 +(n2 −2)(n1 +2δ)] n21 (n2 −2)2 (n2 −4)

when n2 > 4.

58. Let X = Nk (µ, Σ) with a positive definite Σ. (a) Show that EX = µ and Var(X) = Σ. (b) Let A be an l × k matrix and B be an m × k matrix. Show that AX and BX are independent if and only if AΣB τ = 0. (c) Suppose that k = 2, X = (X1 , X2 ), µ = 0, Var(X1 ) =pVar(X2 ) = 1, and Cov(X1 , X2 ) = ρ. Show that E(max{X1 , X2 }) = (1 − ρ)/π.

59. Let X be a random variable and g and h be nondecreasing functions on R. Show that Cov(g(X), h(X)) ≥ 0 when E|g(X)h(X)| < ∞.

60. Let X be a random variable with EX 2 < ∞ and let Y = |X|. Suppose that X has a Lebesgue p.d.f. symmetric about 0. Show that X and Y are uncorrelated, but they are not independent. 61. Let (X, Y ) be a random 2-vector with the following Lebesgue p.d.f.: −1 π x2 + y 2 ≤ 1 f (x, y) = 0 x2 + y 2 > 1. Show that X and Y are uncorrelated, but are not independent.

80

1. Probability Theory

62. Show that inequality (1.41) holds and that when 0 < E|X|p < ∞ and 0 < E|Y |q < ∞, the equality in (1.40) holds if and only if α|X|p = β|Y |q a.s. for some nonzero constants α and β. 63. Prove the following inequalities. (a) Liapounov’s inequality (1.42). (b) Minkowski’s inequality (1.43). (Hint: apply H¨older’s inequality to random variables |X + Y |p−1 and |X|.) (c) (Cr -inequality). E|X + Y |r ≤ Cr (E|X|r + E|Y |r ), where X and Y are random variables, r is a positive constant, and Cr = 1 if 0 < r ≤ 1 and Cr = 2r−1 if r > 1. (d) Let Xi be a random variable with E|Xi |p < ∞, i = 1, ..., n, where p is a constant larger than 1. Show that #p ) ( n " n n 1 X p 1X 1X p p 1/p E . Xi ≤ min E|Xi | , (E|Xi | ) n i=1 n i=1 n i=1 (e) Inequality (1.44). (Hint: prove the case of n = 2 first and then use induction.) (f) Inequality (1.49).

64. Show that the following functions of x are convex and discuss whether they are strictly convex. (a) |x − a|p , where p ≥ 1 and a ∈ R. (b) x−p , x ∈ (0, ∞), where p > 0. (c) ecx , where c ∈ R. (d) x log x, x ∈ (0, ∞). (e) g(ϕ(x)), x ∈ (a, b), where −∞ ≤ a < b ≤ ∞, ϕ is convex on (a, b), and g is convex Pk and nondecreasing on the range Qkof ϕ. (f) ϕ(x) = i=1 ci ϕi (xi ), x = (x1 , ..., xk ) ∈ i=1 Xi , where ci is a positive constant and ϕi is convex on Xi , i = 1, ..., k. 65. Let X = Nk (µ, Σ) with a positive definite Σ. τ τ (a) Show that the m.g.f. of X is et µ+t Σt/2 . (b) Show that EX = µ and Var(X) = Σ by applying (1.54). ′ (c) When k = 1 (Σ = σ 2 ), show that EX = ψX (0) = µ, EX 2 = (3) (4) ′′ ψX (0) = σ 2 + µ2 , EX 3 = ψX (0) = 3σ 2 µ + µ3 , and EX 4 = ψX (0) = 3σ 4 + 6σ 2 µ2 + µ4 . (d) In part (c), show that if µ = 0, then EX p = 0 when p is an odd integer and EX p = (p − 1)(p − 3) · · · 3 · 1σ p when p is an even integer. 66. Let X be a random variable having the gamma distribution Γ(α, γ). Find moments EX p , p = 1, 2, ..., by differentiating the m.g.f. of X. 67. Let X be a random variable with finite EetX and Ee−tX for a t 6= 0. Show that E|X|a < ∞ for any a > 0.

1.6. Exercises

81

68. Let X be a random variable having ψX (t) < ∞ for t in a neighborhood of 0. Show that the moments and cumulants of X satisfy the following equations: µ1 = κ1 , µ2 = κ2 + κ21 , µ3 = κ3 + 3κ1 κ2 + κ31 , and µ4 = κ4 + 3κ22 + 4κ1 κ3 + 6κ21 κ2 + κ41 , where µi and κi are the ith moment and cumulant of X, respectively. 69. Let X be a discrete random variable taking values 0,1,2.... The probability generating function of X is defined to be ρX (t) = E(tX ). Show that (a) ρX (t) = ψX (log t), where ψX is the m.g.f. of X; p (b) d ρdtXp(t) t=1 = E[X(X − 1) · · · (X − p + 1)] for any positive integer p, if ρX is finite in a neighborhood of 1. 70. Let Y be a random variable having the noncentral chi-square distribution χ2k (δ). Show that √ √ √ (a) the ch.f. of Y is (1 − 2 −1t)−k/2 e −1δt/(1−2 −1t) ; (b) E(Y ) = k + δ and Var(Y ) = 2k + 4δ.

71. Let φ be a ch.f. on Rk . Show that |φ| ≤ 1 and φ is uniformly continuous on Rk . √ and b are real numbers, 72. For a complex number√z = a+ −1b, where Pn aP n z¯ is defined to be a − −1b. Show that i=1 j=1 φ(ti − tj )zi z¯j ≥ 0, where φ is a ch.f. on Rk , t1 , ..., tn are k-vectors, and z1 , ..., zn are complex numbers. 73. Show that the following functions of t ∈ R are ch.f.’s, where a > 0 and b > 0 are constants: (a) a2 /(a2 + t2 ); √ (b) (1 + ab − abe −1t )−1/b ; (c) max{1 − |t|/a, 0}; (d) 2(1 − cos at)/(a2 t2 ); a (e) e−|t| , where 0 < a ≤ 2; 2 (f) |φ| R , where φ is a ch.f. on R; (g) φ(ut)dG(u), where φ is a ch.f. on R and G is a c.d.f. on R.

74. Let φn be the ch.f. of a probability measure P Pn , n = 1, 2,.... Let {an } ∞ be a sequence of nonnegative numbers with n=1 an = 1. Show that P∞ n=1 an φn is a ch.f. and find its corresponding probability measure. R 75. Let X be a random variable whose ch.f. φX satisfies |φX (t)|dt < ∞. √ R Show that (2π)−1 e− −1xt φX (t)dt is the Lebesgue p.d.f. of X. 76. A random variable PX or its distribution is of the lattice type if and only if FX (x) = ∞ j=−∞ pj I{a+jd} (x), x ∈ R, where a, d, pj ’s are

82

1. Probability Theory P constants, d > 0, pj ≥ 0, and ∞ j=−∞ pj = 1. Show that X is of the lattice type if and only if its ch.f. satisfies |φX (t)| = 1 for some t 6= 0.

77. Let φ be a ch.f. on R. Show that (a) if |φ(t )| = |φ(t2 )| = 1 and t1 /t2 is an irrational number, then √1 φ(t) = e −1at for some constant a; (b) if tn → 0, tn 6= 0, and |φ(tn )| = 1, then the result in (a) holds; (c) | cos t| is not a ch.f., although cos t is a ch.f. 78. Let X1 , ..., Xk be independent random variables and Y = X1 +· · ·+Xk . Prove the following statements, using Theorem 1.6 and result (1.58). (a) If Xi has the binomial distribution Bi(p, ni ), i = 1, ..., k, then Y has the binomial distribution Bi(p, n1 + · · · + nk ). (b) If Xi has the Poisson distribution P (θi ), i = 1, ..., k, then Y has the Poisson distribution P (θ1 + · · · + θk ). (c) If Xi has the negative binomial distribution N B(p, ri ), i = 1, ..., k, then Y has the negative binomial distribution N B(p, r1 + · · · + rk ). (d) If Xi has the exponential distribution E(0, θ), i = 1, ..., k, then Y has the gamma distribution Γ(k, θ). (e) If Xi has the Cauchy distribution C(0, 1), i = 1, ..., k, then Y /k has the same distribution as X1 . 79. Find an example of two random variables X and Y such that X and Y are not independent but their ch.f.’s satisfy φX (t)φY (t) = φX+Y (t) for all t ∈ R. 80. Let X1 , X2 , ... be independent random variables having the exponential distribution E(0, θ). For given t > 0, let Y be the maximum of n such that Tn ≤ t, where T0 = 0 and Tn = X1 + · · · + Xn , n = 1, 2, .... Show that Y has the Poisson distribution P (t/θ). 81. Let Σ be a k × k nonnegative definite matrix. (a) For a nonsingular Σ, show that X is Nk (µ, Σ) if and only if cτ X is N (cτ µ, cτ Σc) for any c ∈ Rk . (b) For a singular Σ, we define X to be Nk (µ, Σ) if and only if cτ X is N (cτ µ, cτ Σc) for any c ∈ Rk (N (a, 0) is the c.d.f. of the point mass at a). Show that the results in Exercise 43(a)-(c), Exercise 58(a)-(b), and Exercise 65(a) still hold for X = Nk (µ, Σ) with a singular Σ. 82. Let (X1 , X2 ) be Nk (µ, Σ) with a k × k positive definite Σ11 Σ12 , Σ= Σ21 Σ22 where X1 is a random l-vector and Σ11 is an l × l matrix. Show that the conditional Lebesgue p.d.f. of X2 given X1 = x1 is −1 Nk−l µ2 + Σ21 Σ−1 11 (x1 − µ1 ), Σ22 − Σ21 Σ11 Σ12 ,

1.6. Exercises

83

where µi = EXi , i = 1, 2. (Hint: consider X2 − µ2 − Σ21 Σ−1 11 (X1 − µ1 ) and X1 − µ1 .) 83. Let X be an integrable random variable with a Lebesgue p.d.f. fX and let Y = g(X), where g is a function with positive derivative on (0, ∞) and g(x) = g(−x). Find an expression for E(X|Y ) and verify that it is indeed the conditional expectation. 84. Prove Lemma 1.2. (Hint: first consider simple functions.) 85. Prove Proposition 1.10. (Hint for proving (ix): first show that 0 ≤ X1 ≤ X2 ≤ · · · and Xn →a.s. X imply E(Xn |A) →a.s. E(X|A).) 86. Let X and Y be integrable random variables on (Ω, F, P ) and A ⊂ F be a σ-field. Show that E[Y E(X|A)] = E[XE(Y |A)], assuming that both integrals exist. 87. Let X, X1 , X2 , ... be a sequence of random variables on (Ω, F, P ) and A ⊂ F be a σ-field. Suppose that E(Xn Y ) → E(XY ) for every integrable (or bounded) random variable Y . Show that E[E(Xn |A)Y ] → E[E(X|A)Y ] for every integrable (or bounded) random variable Y . 88. Let X be a nonnegative integrable randomRvariable on (Ω,F , P ) and ∞ A ⊂ F be a σ-field. Show that E(X|A) = 0 P X > t|A dt a.s.

89. Let X and Y be random variables on (Ω, F , P ) and A ⊂ F be a σfield. Prove the following inequalities for conditional expectations. (a) If E|X|p < ∞ and E|Y |q < ∞ for constants p and q with p > 1 and p−1 + q −1 = 1, then E(|XY ||A) ≤ [E(|X|p |A)]1/p [E(|Y |q |A)]1/q a.s. (b) If E|X|p < ∞ and E|Y |p < ∞ for a constant p ≥ 1, then [E(|X + Y |p |A)]1/p ≤ [E(|X|p |A)]1/p + [E(|Y |p |A)]1/p a.s. (c) If f is a convex function on R, then f (E(X|A)) ≤ E[f (X)|A] a.s.

90. Let X and Y be random variables on a probability space with Y = E(X|Y ) a.s. and let ϕ be a nondecreasing convex function on [0, ∞). (a) Show that if Eϕ(|X|) < ∞, then Eϕ(|Y |) < ∞. (b) Find an example in which Eϕ(|Y |) < ∞ but Eϕ(|X|) = ∞. (c) Suppose that Eϕ(|X|) = Eϕ(|Y |) < ∞ and ϕ is strictly convex and strictly increasing. Show that X = Y a.s. 91. Let X, Y , and Z be random variables on a probability space. Suppose that E|X| < ∞ and Y = h(Z) with a Borel h. Show that (a) if X and Z are independent and E|Z| < ∞, then E(XZ|Y ) = E(X)E(Z|Y ) a.s.; (b) if E[f (X)|Z] = f (Y ) for all bounded continuous functions f on R, then X = Y a.s.; (c) if E[f (X)|Z] ≥ f (Y ) for all bounded, continuous, nondecreasing functions f on R, then X ≥ Y a.s.

84

1. Probability Theory

92. Prove Lemma 1.3. 93. Show that random variables Xi , i = 1, ..., n, are independent according to Definition 1.7 if and only if (1.7) holds with F being the joint c.d.f. of Xi ’s and Fi being the marginal c.d.f. of Xi . 94. Show that a random variable X is independent of itself if and only if X is constant a.s. Can X and f (X) be independent for a Borel f ? 95. Let X, Y , and Z be independent random variables on a probability space and let U = X + Z and V = Y + Z. Show that given Z, U and V are conditionally independent. 96. Show that the result in Proposition 1.11 may not be true if Y2 is independent of X but not (X, Y1 ). 97. Let X and Y be independent random variables on a probability space. Show that if E|X|a < ∞ for some a ≥ 1 and E|Y | < ∞, then E|X + Y |a ≥ E|X + EY |a . 98. Let PY be a discrete distribution on {0, 1, 2, ...} and PX|Y =y be the binomial distribution Bi(p, y). Let (X, Y ) be the random vector having the joint c.d.f. given by (1.66). Show that (a) if Y has the Poisson distribution P (θ), then the marginal distribution of X is the Poisson distribution P (pθ); (b) if Y + r has the negative binomial distribution N B(π, r), then the marginal distribution of X + r is the negative binomial distribution N B(π/[1 − (1 − p)(1 − π)], r). 99. Let X1 , X2 , ... be i.i.d. random variables and Y be a discrete random variable taking positivePinteger values. Assume that Y and Xi ’s are Y independent. Let Z = i=1 Xi . (a) Obtain the ch.f. of Z. (b) Show that EZ = EY EX1 . (c) Show that Var(Z) = EY Var(X1 ) + Var(Y )(EX1 )2 . 100. Let X, Y , and Z be random variables having a positive joint Lebesgue p.d.f. Let fX|Y (x|y) and fX|Y,Z (x|y, z) be respectively the conditional p.d.f. of X given Y and the conditional p.d.f. of X given (Y, Z), as defined by (1.61). Show that Var(1/fX|Y (X|Y )|X) ≤ Var(1/fX|Y,Z (X|Y, Z)|X) a.s., where Var(ξ|ζ) = E{[ξ − E(ξ|ζ)]2 |ζ} for any random variables ξ and ζ with Eξ 2 < ∞. 101. Let {Xn } be a Markov chain. Show that if g is a one-to-one Borel function, then {g(Xn )} is also a Markov chain. Give an example to show that {g(Xn )} may not be a Markov chain in general.

1.6. Exercises

85

102. A sequence of random vectors {Xn } is said to be a Markov chain of order r for a positive integer r if P (B|X1 , ..., Xn ) = P (B|Xn−r+1 , ..., Xn ) a.s. for any B ∈ σ(Xn+1 ) and n = r, r + 1, .... (a) Let s > r be two positive integers. Show that if {Xn } is a Markov chain of order r, then it is a Markov chain of order s. (b) Let {Xn } be a sequence of random variables, r be a positive integer, and Yn = (Xn , Xn+1 , ..., Xn+r−1 ). Show that {Yn } is a Markov chain if and only if {Xn } is a Markov chain of order r. (c) (Autoregressive process of order r). Let {εn } be a sequence of independent random variables and r be a positive P integer. Show that r {Xn } is a Markov chain of order r, where Xn = j=1 ρj Xn−j + εn and ρj ’s are constants. 103. Show that if {Xn , Fn } is a martingale (or a submartingale), then E(Xn+j |Fn ) = Xn a.s. (or E(Xn+j |Fn ) ≥ Xn a.s.) and EX1 = EXj (or EX1 ≤ EX2 ≤ · · ·) for any j = 1, 2, .... 104. Show that {Xn } in Example 1.25 is a martingale. 105. Let {Xj } and {Zj } be sequences of random variables and let fn and gn denote the Lebesgue p.d.f.’s of Yn = (X1 , ..., Xn ) and (Z1 , ..., Zn ), respectively, n = 1, 2, .... Define λn = −gn (Yn )/fn (Yn )I{fn (Yn )>0} , n = 1, 2, .... Show that {λn } is a submartingale. 106. Let {Yn } be a sequence of independent random variables. (a) Suppose that EYn = 0 for all n. Let X1 = Y1 and Xn+1 = Xn + Yn+1 hn (X1 , ..., Xn ), n ≥ 2, where {hn } is a sequence of Borel functions. Show that {Xn } is a martingale. (b) that EYn = 0 and Var(Yn ) = σ 2 for all n. Let Xn = PnSuppose 2 ( j=1 Yj ) − nσ 2 . Show that {Xn } is a martingale. (c) Suppose that Yn > 0 and EYn = 1 for all n. Let Xn = Y1 · · · Yn . Show that {Xn } is a martingale. 107. Prove the claims in the proof of Proposition 1.14. 108. Show that every sequence of integrable random variables is the sum of a supermartingale and a submartingale. 109. Let {Xn } be a martingale. Show that if {Xn } is bounded either above or below, then supn E|Xn | < ∞. 2 110. Let {Xn } be a martingale satisfying PmEX1 = 0 and EXn 0, n=1 P (|Xn | ≥ ǫ) < ∞.

115. Let X1 , X2 , ... be a sequence of identically distributed random variables with a finite E|X1 | and let Yn = n−1 maxi≤n |Xi |. Show that (a) Yn →L1 0; (b) Yn →a.s. 0. 116. Let X, X1 , X2 , ... be random variables. Find an example for each of the following cases: (a) Xn →p X, but {Xn } does not converge to X a.s. (b) Xn →p X, but {Xn } does not converge to X in Lp for any p > 0. (c) Xn →d X, but {Xn } does not converge to X in probability (do not use Example 1.26). (d) Xn →p X, but {g(Xn )} does not converge to g(X) in probability for some function g.

117. Let X1 , X2 , ... be random variables. Show that (a) {|Xn |} is uniformly integrable if and only if supn E|Xn | < ∞ and, for any ǫ > 0, there is a δǫ > 0 such that supn E(|Xn |IA ) < ǫ for any event A with P (A) < δǫ ; (b) supn E|Xn |1+δ < ∞ for a δ > 0 implies that {|Xn |} is uniformly integrable. 118. Let X, X1 , X2 , ... be random variables satisfying P (|Xn | ≥ c) ≤ P (|X| ≥ c) for all n and c > 0. Show that if E|X| < ∞, then {|Xn |} is uniformly integrable. 119. Let X1 , X2 , ... and Y1 , Y2 , ... be random variables. Show that (a) if {|Xn |} and {|Yn |} are uniformly integrable, then {|Xn + Yn |} is uniformly integrable; Pn (b) if {|Xn |} is uniformly integrable, then {|n−1 i=1 Xi |} is uniformly integrable. 120. Let Y be an integrable random variable and {Fn } be a sequence of σ-fields. Show that {|E(Y |Fn )|} is uniformly integrable. 121. Let X, Y, X1 , X2 , ... be random variables satisfying Xn →p X and P (|Xn | ≤ |Y |) = 1 for all n. Show that if E|Y |r < ∞ for some r > 0, then Xn →Lr X. 122. Let X1 , X2 , ... be a sequence of random k-vectors. Show that Xn →p 0 if and only if E[kXn k/(1 + kXn k)] → 0.

1.6. Exercises

87

123. Let X, X1 , X2 , ... be random variables. Show that Xn →p X if and only if, for any subsequence {nk } of integers, there is a further subsequence {nj } ⊂ {nk } such that Xnj →a.s. X as j → ∞. 124. Let X1 , X2 , ... be a sequence of random variables satisfying |Xn | ≤ C1 and Var(Xn ) ≥ C2 for all n, where Ci ’s are positive constants. Show that Xn →p 0 does not hold. 125. Prove Lemma 1.5. P∞(Hint for part (ii): use Chebyshev’s inequality to show that P ( n=1 IAn = ∞) = 1, which can be shown to be equivalent to the result in (ii).) 126. Prove part (vii) of Theorem 1.8. 127. Let X, X1 , X2 , ..., Y1 , Y2 , ..., Z1 , Z2 , ... be random variables. Prove the following statements. (a) If Xn →d X, then Xn = Op (1). (b) If Xn = Op (Zn ) and P (Yn = 0) = 0, then Xn Yn = Op (Yn Zn ). (c) If Xn = Op (Zn ) and Yn = Op (Zn ), then Xn + Yn = Op (Zn ). (d) If E|Xn | = O(an ), then Xn = Op (an ), where an ∈ (0, ∞). (e) If Xn →a.s. X, then supn |Xn | = Op (1). 128. Let {Xn } and {Yn } be two sequences of random variables such that Xn = Op (1) and P (Xn ≤ t, Yn ≥ t + ǫ) + P (Xn ≥ t + ǫ, Yn ≤ t) = o(1) for any fixed t ∈ R and ǫ > 0. Show that Xn − Yn = op (1). 129. Let {Fn } be a sequence of c.d.f.’s on R, Gn (x) = Fn (an x + cn ), and Hn (x) = Fn (bn x+ dn ), where {an } and {bn } are sequences of positive numbers and {cn } and {dn } are sequences of real numbers. Suppose that Gn →w G and Hn →w H, where G and H are nondegenerate c.d.f.’s. Show that an /bn → a > 0, (cn − dn )/an → b ∈ R, and H(ax + b) = G(x) for all x ∈ R. 130. Let {Pn } be a sequence of probability measures on (R, B) and f be a R nonnegative Borel function such that supn f dPn < ∞ and f (x) → 0 as |x| → ∞. Show that {Pn } is tight. 131. Let P, P1 , P2 , ... be probability measures on (Rk , B k ). Show that if Pn (O) → P (O) for every open subset of R, then Pn (B) → P (B) for every B ∈ B k . 132. Let P, P1 , P2 , ... be probability measures on (R, B). Show that Pn →w P if and only if there exists a dense subset D of R such that limn→∞ Pn ((a, b]) = P ((a, b]) for any a < b, a ∈ D and b ∈ D. 133. Let Fn , n = 0, 1, 2, ..., be c.d.f.’s such that Fn →w F0 . Let Gn (U ) = sup{x : Fn (x) ≤ U }, n = 0, 1, 2, ..., where U is a random variable having the uniform U (0, 1) distribution. Show that Gn (U ) →p G0 (U ).

88

1. Probability Theory

134. Let P, P1 , P2 , ... be probability measures on (R, B). Suppose that Pn →w P and {gn } is a sequence of bounded R continuous R functions on R converging uniformly to g. Show that gn dPn → gdP .

135. Let X, X1 , X2 , ... be random k-vectors and Y, Y1 , Y2 , ... be random lvectors. Suppose that Xn →d X, Yn →d Y , and Xn and Yn are independent for each n. Show that (Xn , Yn ) converges in distribution to a random (k + l)-vector. 136. Let X1 , X2 , ... be independent variables with P (Xn = ±2−n ) Prandom n 1 = 2 , n = 1, 2, .... Show that i=1 Xi →d U , where U has the uniform distribution U (−1, 1). 137. Let {Xn } and {Yn } be two sequences of random variables. Suppose that Xn →d X and that PYn |Xn =xn →w PY almost surely for every sequence of numbers {xn }, where X and Y are independent random variables. Show that Xn + Yn →d X + Y . 138. Let X1 , X2 , ... be i.i.d. random variables having the ch.f. of the form 1 − c|t|a + o(|t|a )Pas t → 0, where 0 < a ≤ 2. Determine the constants b and u so that ni=1 Xi /(bnu ) converges in distribution to a random a variable having ch.f. e−|t| . 139. Let X, X1 , X2 , ... be random k-vectors and A1 , A2 , ... be events. Suppose that Xn →d X. Show that Xn IAn →d X if and only if P (An ) → 1. 140. Let Xn be a random variable having the N (µn , σn2 ) distribution, n = 1, 2,..., and X be a random variable having the N (µ, σ 2 ) distribution. Show that Xn →d X if and only if µn → µ and σn → σ. 141. Suppose that Xn is a random variable having the binomial distribution Bi(pn , n). Show that if npn → θ > 0, then Xn →d X, where X has the Poisson distribution P (θ). 142. Let fn be the Lebesgue p.d.f. of the t-distribution tn , n = 1, 2,.... Show that fn (x) → f (x) for any x ∈ R, where f is the Lebesgue p.d.f. of the standard normal distribution. 143. Prove Theorem 1.10. 144. Show by example that Xn →d X and Yn →d Y does not necessarily imply that g(Xn , Yn ) →d g(X, Y ), where g is a continuous function. 145. Prove Theorem 1.11(ii)-(iii) and Theorem 1.12(ii). Extend Theorem 1.12(i) to the case where g is a function from Rp to Rq with 2 ≤ q ≤ p. 146. Let U1 , U2 , ... be i.i.d. random variables having the uniform distribuQn √ −1/n tion on [0, 1] and Yn = ( i=1 Ui ) . Show that n(Yn − e) →d N (0, e2 ).

89

1.6. Exercises −1 147. Prove Lemma Pn1.6. (Hint: an where bn = i=1 xi /ai .)

Pn

i=1

xi = bn − a−1 n

Pn−1 i=0

bi (ai+1 − ai ),

148. In Theorem 1.13, (a) prove (1.82) for bounded ci ’s whenPE|X1 | < ∞; n (b) show that if EX1 = ∞, then n−1 i=1 Xi →P a.s. ∞; n (c) show that if E|X1 | = ∞, then P (lim supn {| i=1 Xi | > cn}) = P (lim supn {|XP n | > cn}) = 1 for any fixed positive constant c, and n lim supn |n−1 i=1 Xi | = ∞ a.s.

that for x = 3, 4, ..., 149. Let X1 , ..., Xn be i.i.d. random variables such P∞ −2 P (X1 = ±x) = (2cx2 logP x)−1 , where c = / log x. Show x=3 x n −1 that E|X1 | = ∞ but n X → 0, using Theorem 1.13(i). i p i=1

150. Let X1 , X2 , ... be i.i.d. random variables satisfying P (X1 = 2j ) = 2−j , j = 1, 2, .... Show that the WLLN does not hold for {Xn }, i.e., (1.80) does not hold for any {an }. Suppose that, as 151. Let X1 , XP 2 , ... be independent random variables. Pn n −2 2 n → ∞, P (|X | > n) → 0 and n E(X i i I{|Xi |≤n} ) → i=1 i=1P n 0. Show that (Tn − bn )/n →p 0, where Tn = i=1 Xi and bn = P n E(X I ). i {|Xi |≤n} i=1 Pn 152. Let Tn = i=1 Xi , where Xn ’s are independent random variables satisfying P (Xn = ±nθ ) = 0.5 and θ > 0 is a constant. Show that (a) when θ < 0.5, Tn /n →a.s. 0; (b) when θ ≥ 1, Tn /n →p 0 does not hold. 153. Let X2 , X3 , ... bepa sequence of independent random variables satisfying P (Xn = ± n/ log n) = 0.5. Show that (1.86) does not hold for p ∈ [1, 2] but (1.88) is satisfied for p = 2 and, thus, (1.89) holds. 154. Let X1 , ..., Xn beP i.i.d. random variables with Var(X1 ) < ∞. Show n that [n(n + 1)]−1 j=1 jXj →p EX1 . P ¯ = n Xi /n. 155. Let {Xn } be a sequence of random variables and let X i=1 ¯ →a.s. 0. (a) Show that if Xn →a.s. 0, then X ¯ →Lr 0, where r ≥ 1 is a constant. (b) Show that if Xn →Lr 0, then X (c) Show that the result in (b) may not be true for r ∈ (0, 1). ¯ →p 0. (d) Show that Xn →p 0 may not imply X 156. Let X1 , ..., Xn be random variables and {µn }, {σn }, {an }, and {bn } be sequences of real numbers with σn ≥ 0 and an ≥ 0. Suppose that Xn is asymptotically distributed as N (µn , σn2 ). Show that an Xn + bn is asymptotically distributed as N (µn , σn2 ) if and only if an → 1 and [µn (an − 1) + bn ]/σn → 0.

90

1. Probability Theory

157. Show that Liapounov’s condition (1.97) implies Lindeberg’s condition (1.92). 158. Let X1 , XP 2 , ... be a sequence of independent random variables and σn2 = Var( nj=1 Xj ). −n (a) Show that if Xn = N (0, Pn2 ), n = 1, 2, ..., then Feller’s condition (1.96) does not hold but j=1 (Xj − EXj )/σn →d N (0, 1). (b) Show that the result in (a) is still true if X1 has the uniform distribution U (−1, 1) and Xn = N (0, 2n−1 ), n = 2, 3, .... 159. In Example 1.33, show that (a) the condition σn2 → ∞ is also necessary for (1.98); Pn −1 (b) n P i=1 (Xi − pi ) →Lr 0 for any constant r > 0; (c) n−1 ni=1 (Xi − pi ) →a.s. 0.

160. Prove Corollary 1.3.

161. Suppose that Xn is a random variable having the binomial distribution Bi(θ, n), where 0 < θ < 1, n = 1, 2,.... Define Yn = log(Xn /n) when√Xn ≥ 1 and Yn = 1 when Xn = 0. Show that Yn →a.s. log θ and n(Yn − log θ) →d N 0, 1−θ . Establish similar results when θ Xn has the Poisson distribution P (nθ). 162. Let X1 , X2 , ... be independent random variables such that Xj has the uniform distribution on [−j, j], j = 1, 2,.... Show that Lindeberg’s condition is satisfied and state the resulting CLT. 163. Let X1 , X2 , ... be independent random variables such that for j = 1, 2,..., P (Xj = ±j a ) = 6−1 j −2(a−1) and P (Xj = 0) = 1 − 3−1 j −2(a−1) , where a > 1 is a constant. Show that Lindeberg’s condition is satisfied if and only if a < 1.5. 164. Let X1 , X2 , ... be independent random variables with P (Xj = ±j a ) = P (Xj = 0) = 1/3, where a > 0, j = 1, 2,.... Can we apply Theorem 1.15 to {Xj } by checking Liapounov’s condition (1.97)? 165. Let {X Pn } be a sequence of independent random variables.PSuppose that nj=1 (Xj − EXj )/σn →d N (0, 1), where σn2 = Var( nj=1 Xj ). Pn Show that n−1 j=1 (Xj − EXj ) →p 0 if and only if σn = o(n). p 166. Consider Exercise 152. Show that Tn / Var(Tn ) →d N (0, 1) and, when 0.5 ≤ θ < 1, Tn /n →p 0 does not hold. 167. Prove (1.102)-(1.104). √ ¯ √ 168. In Example 1.34, prove σh2 = 1 for n(X − µ)/ˆ σ and n(ˆ σ 2 − σ 2 )/ˆ τ and derive the expressions for p1 (x) in all four cases.

Chapter 2

Fundamentals of Statistics This chapter discusses some fundamental concepts of mathematical statistics. These concepts are essential for the material in later chapters.

2.1 Populations, Samples, and Models A typical statistical problem can be described as follows. One or a series of random experiments is performed; some data from the experiment(s) are collected; and our task is to extract information from the data, interpret the results, and draw some conclusions. In this book we do not consider the problem of planning experiments and collecting data, but concentrate on statistical analysis of the data, assuming that the data are given. A descriptive data analysis can be performed to obtain some summary measures of the data, such as the mean, median, range, standard deviation, etc., and some graphical displays, such as the histogram and boxand-whisker diagram, etc. (see, e.g., Hogg and Tanis (1993)). Although this kind of analysis is simple and requires almost no assumptions, it may not allow us to gain enough insight into the problem. We focus on more sophisticated methods of analyzing data: statistical inference and decision theory.

2.1.1 Populations and samples In statistical inference and decision theory, the data set is viewed as a realization or observation of a random element defined on a probability space (Ω, F, P ) related to the random experiment. The probability measure P is called the population. The data set or the random element that produces 91

92

2. Fundamentals of Statistics

the data is called a sample from P . The size of the data set is called the sample size. A population P is known if and only if P (A) is a known value for every event A ∈ F. In a statistical problem, the population P is at least partially unknown and we would like to deduce some properties of P based on the available sample. Example 2.1 (Measurement problems). To measure an unknown quantity θ (for example, a distance, weight, or temperature), n measurements, x1 , ..., xn , are taken in an experiment of measuring θ. If θ can be measured without errors, then xi = θ for all i; otherwise, each xi has a possible measurement error. In descriptive data analysis, a few summary measures may be calculated, for example, the sample mean n

x ¯= and the sample variance

1X xi n i=1 n

1 X 2 s = (xi − x ¯) . n − 1 i=1 2

However, what is the relationship between x ¯ and θ? Are they close (if not equal) in some sense? The sample variance s2 is clearly an average of squared deviations of xi ’s from their mean. But, what kind of information does s2 provide? Finally, is it enough to just look at x¯ and s2 for the purpose of measuring θ? These questions cannot be answered in descriptive data analysis. In statistical inference and decision theory, the data set, (x1 , ..., xn ), is viewed as an outcome of the experiment whose sample space is Ω = Rn . We usually assume that the n measurements are obtained in n independent trials of the experiment. Hence, we can define a random n-vector Qn X = (X1 , ..., Xn ) on i=1 (R, B, P ) whose realization is (x1 , ..., xn ). The population in this problem is P (note that the product probability measure is determined by P ) and is at least partially unknown. The random vector X is a sample and n is the sample size. Define n

X ¯= 1 X Xi n i=1

and

(2.1)

n

S2 =

1 X ¯ 2. Xi − X n − 1 i=1

(2.2)

¯ and S 2 are random variables that produce x Then X ¯ and s2 , respectively. Questions raised previously can be answered if some assumptions are imposed on the population P , which are discussed later.

93

2.1. Populations, Samples, and Models

When the sample (X1 , ..., Xn ) has i.i.d. components, which is often the case in applications, the population is determined by the marginal distribution of Xi . Example 2.2 (Life-time testing problems). Let x1 , ..., xn be observed lifetimes of some electronic components. Again, in statistical inference and decision theory, x1 , ..., xn are viewed as realizations of independent random variables X1 , ..., Xn . Suppose that the components are of the same type so that it is reasonable to assume that X1 , ..., Xn have a common marginal c.d.f. F . Then the population is F , which is often unknown. A quantity of interest in this problem is 1 − F (t) with a t > 0, which is the probability that a component does not fail at time t. It is possible that all xi ’s are smaller (or larger) than t. Conclusions about 1 − F (t) can be drawn based on data x1 , ..., xn when certain assumptions on F are imposed. Example 2.3 (Survey problems). A survey is often conducted when one is not able to evaluate all elements in a collection P = {y1 , ..., yN } containing N values in Rk , where k and N are finite positive integers but N may be very large. PN Suppose that the quantity of interest is the population total Y = i=1 yi . In a survey, a subset s of n elements are selected from {1, ..., N } and values yi , i ∈ s, are obtained. Can we draw some conclusion about Y based on data yi , i ∈ s? How do we define some random variables that produce the survey data? First, we need to specify how s is selected. A commonly used probability sampling plan can be described as follows. Assume that every element in {1, ..., N } can be selected at most once, i.e., we consider sampling without replacement. Let S be the collection of all subsets of n distinct elements from {1, ..., N }, Fs be the collection of all subsets of S, and p be a probability measure on (S, Fs ). Any s ∈ S is selected with probability p(s). Note that p(s) is a known value whenever s is given. Let X1 , ..., Xn be random variables such that P (X1 = yi1 , ..., Xn = yin ) =

p(s) , n!

s = {i1 , ..., in } ∈ S.

(2.3)

Then (yi , i ∈ s) can be viewed as a realization of the sample (X1 , ..., Xn ). If p(s) is constant, then the sampling plan is called the simple random sampling (without replacement) and (X1 , ..., Xn ) is called a simple random sample. Although X1 , ..., Xn are identically distributed, they are not necessarily independent. Thus, unlike in the previous two examples, the population in this problem may not be specified by the marginal distributions of Xi ’s. The population is determined by P and the known selection probability measure p. For this reason, P is often treated as the population. Conclusions about Y and other characteristics of P can be drawn based on data yi , i ∈ s, which are discussed later.

94

2. Fundamentals of Statistics

2.1.2 Parametric and nonparametric models A statistical model (a set of assumptions) on the population P in a given problem is often postulated to make the analysis possible or easy. Although testing the correctness of postulated models is part of statistical inference and decision theory, postulated models are often based on knowledge of the problem under consideration. Definition 2.1. A set of probability measures Pθ on (Ω, F ) indexed by a parameter θ ∈ Θ is said to be a parametric family if and only if Θ ⊂ Rd for some fixed positive integer d and each Pθ is a known probability measure when θ is known. The set Θ is called the parameter space and d is called its dimension. A parametric model refers to the assumption that the population P is in a given parametric family. A parametric family {Pθ : θ ∈ Θ} is said to be identifiable if and only if θ1 6= θ2 and θi ∈ Θ imply Pθ1 6= Pθ2 . In most cases an identifiable parametric family can be obtained through reparameterization. Hence, we assume in what follows that every parametric family is identifiable unless otherwise stated. Let P be a family of populations and ν be a σ-finite measure on (Ω, F ). If P ≪ ν for all P ∈ P, then P is said to be dominated by ν, in which case P dPθ can be identified by the family of densities { dP dν : P ∈ P} (or { dν : θ ∈ Θ} for a parametric family). Many examples of parametric families can be obtained from Tables 1.1 and 1.2 in §1.3.1. All parametric families from Tables 1.1 and 1.2 are dominated by the counting measure or the Lebesgue measure on R. Example 2.4 (The k-dimensional normal family). Consider the normal distribution Nk (µ, Σ) given by (1.24) for a fixed positive integer k. An important parametric family in statistics is the family of normal distributions P = {Nk (µ, Σ) : µ ∈ Rk , Σ ∈ Mk }, where Mk is a collection of k ×k symmetric positive definite matrices. This family is dominated by the Lebesgue measure on Rk . In the measurement problem described in Example 2.1, Xi ’s are often i.i.d. from the N (µ, σ 2 ) distribution. Hence, we can impose a parametric model on the population, i.e., P ∈ P = {N (µ, σ 2 ) : µ ∈ R, σ 2 > 0}. The normal parametric model is perhaps not a good model for the lifetime testing problem described in Example 2.2, since clearly Xi ≥ 0 for all i. In practice, the normal family {N (µ, σ 2 ) : µ ∈ R, σ 2 > 0} can be used for a life-time testing problem if one puts some restrictions on µ and σ so that P (Xi < 0) is negligible. Common parametric models for

2.1. Populations, Samples, and Models

95

life-time testing problems are the exponential model (containing the exponential distributions E(0, θ) with an unknown parameter θ; see Table 1.2 in §1.3.1), the gamma model (containing the gamma distributions Γ(α, γ) with unknown parameters α and γ), the log-normal model (containing the log-normal distributions LN (µ, σ2 ) with unknown parameters µ and σ), the Weibull model (containing the Weibull distributions W (α, θ) with unknown parameters α and θ), and any subfamilies of these parametric families (e.g., a family containing the gamma distributions with one known parameter and one unknown parameter). The normal family is often not a good choice for the survey problem discussed in Example 2.3. In a given problem, a parametric model is not useful if the dimension of Θ is very high. For example, the survey problem described in Example 2.3 has a natural parametric model, since the population P can be indexed by the parameter θ = (y1 , ..., yN ). If there is no restriction on the y-values, however, the dimension of the parameter space is kN , which is usually much larger than the sample size n. If there are some restrictions on the y-values (for example, yi ’s are nonnegative integers no larger than a fixed integer m), then the dimension of the parameter space is at most m + 1 and the parametric model becomes useful. A family of probability measures is said to be nonparametric if it is not parametric according to Definition 2.1. A nonparametric model refers to the assumption that the population P is in a given nonparametric family. There may be almost no assumption on a nonparametric family, for example, the family of all probability measures on (Rk , B k ). But in many applications, we may use one or a combination of the following assumptions to form a nonparametric family on (Rk , B k ): (1) The joint c.d.f.’s are continuous. (2) The joint c.d.f.’s have finite moments of order ≤ a fixed integer. (3) The joint c.d.f.’s have p.d.f.’s (e.g., Lebesgue p.d.f.’s). (4) k = 1 and the c.d.f.’s are symmetric. For instance, in Example 2.1, we may assume a nonparametric model with symmetric and continuous c.d.f.’s. The symmetry assumption may not be suitable for the population in Example 2.2, but the continuity assumption seems to be reasonable. In statistical inference and decision theory, methods designed for parametric models are called parametric methods, whereas methods designed for nonparametric models are called nonparametric methods. However, nonparametric methods are used in a parametric model when parametric methods are not effective, such as when the dimension of the parameter

96

2. Fundamentals of Statistics

space is too high (Example 2.3). On the other hand, parametric methods may be applied to a semi-parametric model, which is a nonparametric model having a parametric component. Some examples are provided in §5.1.4.

2.1.3 Exponential and location-scale families In this section, we discuss two types of parametric families that are of special importance in statistical inference and decision theory. Definition 2.2 (Exponential families). A parametric family {Pθ : θ ∈ Θ} dominated by a σ-finite measure ν on (Ω, F ) is called an exponential family if and only if dPθ (ω) = exp [η(θ)]τ T (ω) − ξ(θ) h(ω), dν

ω ∈ Ω,

(2.4)

where exp{x} = ex , T is a random p-vector with a fixed positive integer p, η is a function R from Θ to Rp , h is a nonnegative Borel function on (Ω, F), and ξ(θ) = log Ω exp{[η(θ)]τ T (ω)}h(ω)dν(ω) .

In Definition 2.2, T and h are functions of ω only, whereas η and ξ are functions of θ only. Ω is usually Rk . The representation (2.4) of an exponential family is not unique. In fact, any transformation η˜(θ) = Dη(θ) with a p × p nonsingular matrix D gives another representation (with T replaced by T˜ = (Dτ )−1 T ). A change of the measure that dominates the Rfamily also changes the representation. For example, if we define λ(A) = hdν for any A ∈ F, then we obtain an exponential family with densities A dPθ (ω) = exp [η(θ)]τ T (ω) − ξ(θ) . dλ

(2.5)

In an exponential family, consider the reparameterization η = η(θ) and fη (ω) = exp η τ T (ω) − ζ(η) h(ω), ω ∈ Ω, (2.6) R where ζ(η) = log Ω exp{η τ T (ω)}h(ω)dν(ω) . This is the canonical form for the family, which is not unique for the reasons discussed previously. The new parameter η is called the natural parameter. The new parameter space Ξ = {η(θ) : θ ∈ Θ}, a subset of Rp , is called the natural parameter space. An exponential family in canonical form is called a natural exponential family. If there is an open set contained in the natural parameter space of an exponential family, then the family is said to be of full rank. Example 2.5. Let Pθ be the binomial distribution Bi(θ, n) with parameter θ, where n is a fixed positive integer. Then {Pθ : θ ∈ (0, 1)} is an

2.1. Populations, Samples, and Models

97

exponential family, since the p.d.f. of Pθ w.r.t. the counting measure is n o n θ + n log(1 − θ) I{0,1,...,n} (x) fθ (x) = exp x log 1−θ x θ , ξ(θ) = −n log(1 − θ), and h(x) = nx I{0,1,...,n} (x)). (T (x) = x, η(θ) = log 1−θ θ If we let η = log 1−θ , then Ξ = R and the family with p.d.f.’s n I{0,1,...,n} (x) fη (x) = exp {xη − n log(1 + eη )} x is a natural exponential family of full rank. Example 2.6. The normal family {N (µ, σ 2 ) : µ ∈ R, σ > 0} is an exponential family, since the Lebesgue p.d.f. of N (µ, σ 2 ) can be written as 1 µ 1 2 µ2 √ exp x − x − − log σ . σ2 2σ 2 2σ 2 2π

µ2 Hence, T (x) = (x, −x2 ), η(θ) = σµ2 , 2σ1 2 , θ = (µ, σ 2 ), ξ(θ) = 2σ 2 + log σ, √ and h(x) = 1/ 2π. Let η = (η1 , η2 ) = σµ2 , 2σ1 2 . Then Ξ = R × (0, ∞) and we can obtain √ a natural exponential family of full rank with ζ(η) = η12 /(4η2 ) + log(1/ 2η2 ). A subfamily of the previous normal family, {N (µ, µ2 ) : µ ∈ R, µ 6= 0}, is also an exponential family with the natural parameter η = µ1 , 2µ1 2 and natural parameter space Ξ = {(x, y) : y = 2x2 , x ∈ R, y > 0}. This exponential family is not of full rank. For an exponential family, (2.5) implies that there is a nonzero measure λ such that dPθ (ω) > 0 for all ω and θ. (2.7) dλ We can use this fact to show that a family of distributions is not an exponential family. For example, consider the family of uniform distributions, i.e., Pθ is U (0, θ) with an unknown θ ∈ (0, ∞). If {Pθ : θ ∈ (0, ∞)} is an exponential family, then from the previous discussion we have a nonzero measure λ such that (2.7) holds. For any t > 0, there is a θ < t such that Pθ ([t, ∞)) = 0, which with (2.7) implies that λ([t, ∞)) = 0. Also, for any t ≤ 0, Pθ ((−∞, t]) = 0, which with (2.7) implies that λ((−∞, t]) = 0. Since t is arbitrary, λ ≡ 0. This contradiction implies that {Pθ : θ ∈ (0, ∞)} cannot be an exponential family. The reader may verify which of the parametric families from Tables 1.1 and 1.2 are exponential families. As another example, we consider an important exponential family containing multivariate discrete distributions.

98

2. Fundamentals of Statistics

Example 2.7 (The multinomial family). Consider an experiment having k + 1 possiblePoutcomes with pi as the probability for the ith outcome, k i = 0, 1, ..., k, i=0 pi = 1. In n independent trials of this experiment, let Xi be the number of trials resulting in the ith outcome, i = 0, 1, ..., k. Then the joint p.d.f. (w.r.t. counting measure) of (X0 , X1 , ..., Xk ) is n! px0 px1 · · · pxk k IB (x0 , x1 , ..., xk ), x0 !x1 ! · · · xk ! 0 1 P where B = {(x0 , x1 , ..., xk ) : xi ’s are integers ≥ 0, ki=0 xi = n} and θ = (p0 , p1 , ..., pk ). The distribution of (X0 , X1 , ..., Xk ) is called the multinomial distribution, which is an extension of the binomial distribution. In fact, the marginal c.d.f. of each Xi is the binomial distribution Bi(pi , n). Let P Θ = {θ ∈ Rk+1 : 0 < pi < 1, ki=0 pi = 1}. The parametric family {fθ : θ ∈ Θ} is called the multinomial family. Let x = (x0 , x1 , ..., xk ), η = (log p0 , log p1 , ..., log pk ), and h(x) = [n!/(x0 !x1 ! · · · xk !)]IB (x). Then fθ (x0 , x1 , ..., xk ) =

fθ (x0 , x1 , ..., xk ) = exp {η τ x} h(x),

x ∈ Rk+1 .

(2.8)

Hence, the multinomial family is a natural exponential family with natural parameter η. However, representation (2.8) does not provide an exponential family of full rank, since there is no open set of Rk+1 contained in the natural parameter space. A reparameterization leads to anPexponential P family with full rank. Using the fact that ki=0 Xi = n and ki=0 pi = 1, we obtain that fθ (x0 , x1 , ..., xk ) = exp {η∗τ x∗ − ζ(η∗ )} h(x),

x ∈ Rk+1 ,

(2.9)

where x∗ = (x1 , ..., xk ), η∗ = (log(p1 /p0 ), ..., log(pk /p0 )), and ζ(η∗ ) = −n log p0 . The η∗ -parameter space is Rk . Hence, the family of densities given by (2.9) is a natural exponential family of full rank. If X1 , ..., Xm are independent random vectors with p.d.f.’s in exponential families, then the p.d.f. of (X1 , ..., Xm ) is again in an exponential family. The following result summarizes some other useful properties of exponential families. Its proof can be found in Lehmann (1986). Theorem 2.1. Let P be a natural exponential family given by (2.6). (i) Let T = (Y, U ) and η = (ϑ, ϕ), where Y and ϑ have the same dimension. Then, Y has the p.d.f. fη (y) = exp{ϑτ y − ζ(η)} w.r.t. a σ-finite measure depending on ϕ. In particular, T has a p.d.f. in a natural exponential family. Furthermore, the conditional distribution of Y given U = u has the p.d.f. (w.r.t. a σ-finite measure depending on u) fϑ,u (y) = exp{ϑτ y − ζu (ϑ)},

99

2.1. Populations, Samples, and Models

which is in a natural exponential family indexed by ϑ. (ii) If η0 is an interior point of the natural parameter space, then the m.g.f. ψη0 of Pη0 ◦ T −1 is finite in a neighborhood of 0 and is given by ψη0 (t) = exp{ζ(η0 + t) − ζ(η0 )}. R Furthermore, if f is a Borel function satisfying |f |dPη0 < ∞, then the function Z f (ω) exp{η τ T (ω)}h(ω)dν(ω)

is infinitely often differentiable in a neighborhood of η0 , and the derivatives may be computed by differentiation under the integral sign. Using Theorem 2.1(ii) and the result in Example 2.5, we obtain that the m.g.f. of the binomial distribution Bi(p, n) is ψη (t) = exp{n log(1 + eη+t ) − n log(1 + eη )} n 1 + eη et = 1 + eη = (1 − p + pet )n ,

since p = eη /(1 + eη ). Definition 2.3 (Location-scale families). Let P be a known probability measure on (Rk , B k ), V ⊂ Rk , and Mk be a collection of k × k symmetric positive definite matrices. The family {P(µ,Σ) : µ ∈ V, Σ ∈ Mk } is called a location-scale family (on Rk ), where P(µ,Σ) (B) = P Σ−1/2 (B − µ) ,

(2.10)

B ∈ Bk ,

Σ−1/2 (B − µ) = {Σ−1/2 (x − µ) : x ∈ B} ⊂ Rk , and Σ−1/2 is the inverse of the “square root” matrix Σ1/2 satisfying Σ1/2 Σ1/2 = Σ. The parameters µ and Σ1/2 are called the location and scale parameters, respectively.

The following are some important examples of location-scale families. The family {P(µ,Ik ) : µ ∈ Rk } is called a location family, where Ik is the k × k identity matrix. The family {P(0,Σ) : Σ ∈ Mk } is called a scale family. In some cases, we consider a location-scale family of the form {P(µ,σ2 Ik ) : µ ∈ Rk , σ > 0}. If X1 , ..., Xk are i.i.d. with a common distribution in the location-scale family {P(µ,σ2 ) : µ ∈ R, σ > 0}, then the joint distribution of the vector (X1 , ..., Xk ) is in the location-scale family {P(µ,σ2 Ik ) : µ ∈ V, σ > 0} with V = {(x, ..., x) ∈ Rk : x ∈ R}.

100

2. Fundamentals of Statistics

A location-scale family can be generated as follows. Let X be a random k-vector having a distribution P . Then the distribution of Σ1/2 X + µ is P(µ,Σ) . On the other hand, if X is a random k-vector whose distribution is in the location-scale family (2.10), then the distribution DX + c is also in the same family, provided that Dµ + c ∈ V and DΣDτ ∈ Mk . Let F be the c.d.f. of P . Then the c.d.f. of P(µ,Σ) is F Σ−1/2 (x − µ) , x ∈ Rk . If F has a Lebesgue p.d.f. f , then the Lebesgue p.d.f. of P(µ,Σ) is Det(Σ−1/2 )f Σ−1/2 (x − µ) , x ∈ Rk (Proposition 1.8). Many families of distributions in Table 1.2 (§1.3.1) are location, scale, or location-scale families. For example, the family of exponential distributions E(a, θ) is a location-scale family on R with location parameter a and scale parameter θ; the family of uniform distributions U (0, θ) is a scale family on R with a scale parameter θ. The k-dimensional normal family discussed in Example 2.4 is a location-scale family on Rk .

2.2 Statistics, Sufficiency, and Completeness Let us assume now that our data set is a realization of a sample X (a random vector) from an unknown population P on a probability space.

2.2.1 Statistics and their distributions A measurable function of X, T (X), is called a statistic if T (X) is a known value whenever X is known, i.e., the function T is a known function. Statistical analyses are based on various statistics, for various purposes. Of course, X itself is a statistic, but it is a trivial statistic. The range of a nontrivial statistic T (X) is usually simpler than that of X. For example, X may be a random n-vector and T (X) may be a random p-vector with a p much smaller than n. This is desired since T (X) simplifies the original data. From a probabilistic point of view, the “information” within the statistic T (X) concerning the unknown distribution of X is contained in the σfield σ(T (X)). To see this, assume that S is any other statistic for which σ(S(X)) = σ(T (X)). Then, by Lemma 1.2, S is a measurable function of T , and T is a measurable function of S. Thus, once the value of S (or T ) is known, so is the value of T (or S). That is, it is not the particular values of a statistic that contain the information, but the generated σ-field of the statistic. Values of a statistic may be important for other reasons. Note that σ(T (X)) ⊂ σ(X) and the two σ-fields are the same if and only if T is one-to-one. Usually σ(T (X)) simplifies σ(X), i.e., a statistic provides a “reduction” of the σ-field.

2.2. Statistics, Sufficiency, and Completeness

101

Any T (X) is a random element. If the distribution of X is unknown, then the distribution of T may also be unknown, although T is a known function. Finding the form of the distribution of T is one of the major problems in statistical inference and decision theory. Since T is a transformation of X, tools we learn in Chapter 1 for transformations may be useful in finding the distribution or an approximation to the distribution of T (X). Example 2.8. Let X1 , ..., Xn be i.i.d. random variables having a common ¯ and sample distribution P and X = (X1 , ..., Xn ). The sample mean X variance S 2 defined in (2.1) and (2.2), respectively, are two commonly used ¯ and S 2 ? statistics. Can we find the joint or the marginal distributions of X It depends on how much we know about P . ¯ and S 2 . Assume that P has a First, let us consider the moments of X finite mean denoted by µ. Then ¯ = µ. EX R ¯ = xdPθ = µ(θ) If P is in a parametric family {Pθ : θ ∈ Θ}, then E X for some function µ(·). Even if the form of µ is known, µ(θ) may still be unknown when θ is unknown. Assume now that P has a finite variance denoted by σ 2 . Then ¯ = σ 2 /n, Var(X) which equals σ 2 (θ)/n for some function σ 2 (·) if P is in a parametric family. With a finite σ 2 = Var(X1 ), we can also obtain that ES 2 = σ 2 . ¯ 3 and Cov(X, ¯ S 2 ), and with a With a finite E|X1 |3 , we can obtain E(X) 4 2 finite E|X1 | , we can obtain Var(S ) (exercise). ¯ If P is in a parametric family, we Next, consider the distribution of X. ¯ See Example 1.20 and some exercises can often find the distribution of X. ¯ is N (µ, σ 2 /n) if P is N (µ, σ 2 ); nX ¯ has the gamma in §1.6. For example, X distribution Γ(n, θ) if P is the exponential distribution E(0, θ). If P is not in a parametric family, then it is usually hard to find the exact form of the ¯ One can, however, use the CLT (§1.5.4) to obtain an distribution of X. ¯ Applying Corollary 1.2 (for the approximation to the distribution of X. case of k = 1), we obtain that √ ¯ − µ) →d N (0, σ 2 ) n(X

¯ can be approximated by N (µ, σ 2 /n), and, by (1.100), the distribution of X 2 where µ and σ are the mean and variance of P , respectively, and are assumed to be finite. ¯ the distribution of S 2 is harder to obtain. Assuming Compared to X, 2 that P is N (µ, σ ), one can show that (n − 1)S 2 /σ 2 has the chi-square

102

2. Fundamentals of Statistics

distribution χ2n−1 (see Example 2.18). An approximate distribution for ¯ and S 2 S 2 can be obtained from the approximate joint distribution of X discussed next. ¯ Under the assumption that P is N (µ, σ 2 ), it can be shown that X and S 2 are independent (Example 2.18). Hence, the joint distribution of ¯ S 2 ) is the product of the marginal distributions of X ¯ and S 2 given in the (X, previous discussion. Without the normality assumption, an approximate joint distribution can be obtained as follows. Assume again that µ = EX1 , σ 2 = Var(X1 ), and E|X1 |4 are finite. Let Yi = (Xi − µ, (Xi − µ)2 ), i = 1, ..., n. Then Y1 , ..., Yn are i.i.d. random 2-vectors with EY1 = (0, σ 2 ) and variance-covariance matrix E(X1 − µ)3 σ2 . Σ= E(X1 − µ)3 E(X1 − µ)4 − σ 4 Pn ¯ − µ, S˜2 ), where S˜2 = n−1 Pn (Xi − µ)2 . Note that Y¯ = n−1 i=1 Yi = (X i=1 Applying the CLT (Corollary 1.2) to Yi ’s, we obtain that √ ¯ − µ, S˜2 − σ 2 ) →d N2 (0, Σ). n(X Since

i n h ˜2 ¯ − µ)2 S − (X n−1 ¯ and X →a.s. µ (the SLLN, Theorem 1.13), an application of Slutsky’s theorem (Theorem 1.11) leads to √ ¯ − µ, S 2 − σ 2 ) →d N2 (0, Σ). n(X S2 =

Example 2.9 (Order statistics). Let X = (X1 , ..., Xn ) with i.i.d. random components and let X(i) be the ith smallest value of X1 , ..., Xn . The statistics X(1) , ..., X(n) are called the order statistics, which is a set of very useful statistics in addition to the sample mean and variance in the previous example. Suppose that Xi has a c.d.f. F having a Lebesgue p.d.f. f . Then the joint Lebesgue p.d.f. of X(1) , ..., X(n) is x1 < x2 < · · · < xn n!f (x1 )f (x2 ) · · · f (xn ) g(x1 , x2 , ..., xn ) = 0 otherwise. The joint Lebesgue p.d.f. of X(i) and X(j) , 1 ≤ i < j ≤ n, is ( i−1 j−i−1 n−j n![F (x)]

gi,j (x, y) =

[F (y)−F (x)] [1−F (y)] (i−1)!(j−i−1)!(n−j)!

f (x)f (y)

0

and the Lebesgue p.d.f. of X(i) is gi (x) =

n! [F (x)]i−1 [1 − F (x)]n−i f (x). (i − 1)!(n − i)!

x 0 and dP/dν > 0 a.e. ν on C. Then there exists a sequence {Ci } ⊂ C P such that ν(Ci ) → supC∈C ν(C). Let C0 be the union of all Ci ’s ∞ and Q = i=1 ci Pi , where Pi is the probability measure corresponding to Ci . Then C0 ∈ C (exercise). Suppose now that Q(A) = 0. Let P ∈ P0 and B = {x : dP/dν > 0}. Since Q(A ∩ C0 ) = 0, ν(A ∩ C0 ) = 0 and P (A ∩ C0 ) = 0. Then P (A) = P (A ∩ C0c ∩ B). If P (A ∩ C0c ∩ B) > 0, then ν(C0 ∪(A∩C0c ∩B)) > ν(C0 ), which contradicts ν(C0 ) = supC∈C ν(C) since A ∩ C0c ∩ B and therefore C0 ∪ (A ∩ C0c ∩ B) is in C. Thus, P (A) = 0 for all P ∈ P0 . Theorem 2.2 (The factorization theorem). Suppose that X is a sample from P ∈ P and P is a family of probability measures on (Rn , B n ) dominated by a σ-finite measure ν. Then T (X) is sufficient for P ∈ P if and only if there are nonnegative Borel functions h (which does not depend on P ) on (Rn , B n ) and gP (which depends on P ) on the range of T such that dP (x) = gP T (x) h(x). dν

(2.11)

2.2. Statistics, Sufficiency, and Completeness

105

Proof. (i) Suppose that T is sufficient for P ∈ P. Then, for any A ∈ B n , P (A|T ) does not depend on P . Let Q be the probability measure in Lemma 2.1. By Fubini’s theorem and the result in Exercise 35 of §1.6, Q(A ∩ B) = =

∞ X j=1 ∞ X

cj Pj (A ∩ B) cj

j=1

=

Z

Z X ∞

P (A|T )dPj B

cj P (A|T )dPj

B j=1

=

Z

P (A|T )dQ

B

for any B ∈ σ(T ). Hence, P (A|T ) = EQ (IA |T ) a.s. Q, where EQ (IA |T ) denotes the conditional expectation of IA given T w.r.t. Q. Let gP (T ) be the Radon-Nikodym derivative dP/dQ on the space (Rn , σ(T ), Q). From Propositions 1.7 and 1.10, Z P (A) = P (A|T )dP Z = EQ (IA |T )gP (T )dQ Z = EQ [IA gP (T )|T ]dQ Z dQ gP (T ) = dν dν A for any A ∈ B n . Hence, (2.11) holds with h = dQ/dν. (ii) Suppose that (2.11) holds. Then ∞ X ∞ dP dP X dPi = = gP (T ) ci gPi (T ) a.s. Q, dQ dν dν i=1 i=1

(2.12)

where the second equality follows from the result in Exercise 35 of §1.6. Let A ∈ σ(X) and P ∈ P. The sufficiency of T follows from P (A|T ) = EQ (IA |T ) a.s. P ,

(2.13)

where EQ (IA |T ) is given in part (i) of the proof. This is because EQ (IA |T ) does not vary with P ∈ P, and result (2.13) and Theorem 1.7 imply that the conditional distribution of X given T is determined by EQ (IA |T ), A ∈ σ(X). By the definition of conditional probability, (2.13) follows from Z Z IA dP = EQ (IA |T )dP (2.14) B

B

106

2. Fundamentals of Statistics

for any B ∈ σ(T ). Let B ∈ σ(T ). By (2.12), dP/dQ is a Borel function of T . Then, by Proposition 1.7(i), Proposition 1.10(vi), and the definition of the conditional expectation, the right-hand side of (2.14) is equal to Z Z Z dP dP dP dQ = T dQ = dQ, EQ (IA |T ) EQ IA IA dQ dQ dQ B B B which equals the left-hand side of (2.14). This proves (2.14) for any B ∈ σ(T ) and completes the proof.

If P is an exponential family with p.d.f.’s given by (2.4) and X(ω) = ω, then we can apply Theorem 2.2 with gθ (t) = exp{[η(θ)]τ t − ξ(θ)} and conclude that T is a sufficient statistic for θ ∈ Θ. In Example P 2.10 the joint distribution of X is in an exponential family with T (X) = ni=1 Xi . Hence, we can conclude that T is sufficient for θ ∈ (0, 1) without computing the conditional distribution of X given T . Example 2.11 (Truncation families). Let φ(x) be a positive Borel function Rb on (R, B) such that a φ(x)dx < ∞ for any a and b, −∞ < a < b < ∞. Let θ = (a, b), Θ = {(a, b) ∈ R2 : a < b}, and fθ (x) = c(θ)φ(x)I(a,b) (x), i−1 hR b . Then {fθ : θ ∈ Θ}, called a truncation where c(θ) = a φ(x)dx family, is a parametric family dominated by the Lebesgue measure on R. Let X1 , ..., Xn be i.i.d. random variables having the p.d.f. fθ . Then the joint p.d.f. of X = (X1 , ..., Xn ) is n Y

i=1

fθ (xi ) = [c(θ)]n I(a,∞) (x(1) )I(−∞,b) (x(n) )

n Y

φ(xi ),

(2.15)

i=1

where x(i) is the ith smallest value of x1 , ..., xn . LetQT (X) = (X(1) , X(n) ), n gθ (t1 , t2 ) = [c(θ)]n I(a,∞) (t1 )I(−∞,b) (t2 ), and h(x) = i=1 φ(xi ). By (2.15) and Theorem 2.2, T (X) is sufficient for θ ∈ Θ. Example 2.12 (Order statistics). Let X = (X1 , ..., Xn ) and X1 , ..., Xn be i.i.d. random variables having a distribution P ∈ P, where P is the family of distributions on R having Lebesgue p.d.f.’s. Let X(1) , ..., X(n) be the order statistics given in Example 2.9. Note that the joint p.d.f. of X is f (x1 ) · · · f (xn ) = f (x(1) ) · · · f (x(n) ). Hence, T (X) = (X(1) , ..., X(n) ) is sufficient for P ∈ P. The order statistics can be shown to be sufficient even when P is not dominated by any σ-finite measure, but Theorem 2.2 is not applicable (see Exercise 31 in §2.6).

107

2.2. Statistics, Sufficiency, and Completeness

There are many sufficient statistics for a given family P. In fact, if T is a sufficient statistic and T = ψ(S), where ψ is measurable and S is another statistic, then S is sufficient. This is obvious from Theorem 2.2 if the population has a p.d.f., but it can be proved directly Pm from PnDefinition 2.4 (Exercise 25). For instance, in Example 2.10, ( i=1 Xi , i=m+1 Xi ) is sufficient for θ, where m is any fixed integer between 1 and n. If T is sufficient and T = ψ(S) with a measurable ψ that is not one-to-one, then σ(T ) ⊂ σ(S) and T is more useful than S, since T provides a further reduction of the data (or σ-field) without loss of information. Is there a sufficient statistic that provides “maximal” reduction of the data? Before introducing the next concept, we need the following notation. If a statement holds except for outcomes in an event A satisfying P (A) = 0 for all P ∈ P, then we say that the statement holds a.s. P. Definition 2.5 (Minimal sufficiency). Let T be a sufficient statistic for P ∈ P. T is called a minimal sufficient statistic if and only if, for any other statistic S sufficient for P ∈ P, there is a measurable function ψ such that T = ψ(S) a.s. P. If both T and S are minimal sufficient statistics, then by definition there is a one-to-one measurable function ψ such that T = ψ(S) a.s. P. Hence, the minimal sufficient statistic is unique in the sense that two statistics that are one-to-one measurable functions of each other can be treated as one statistic. Example 2.13. Let X1 , ..., Xn be i.i.d. random variables from Pθ , the uniform distribution U (θ, θ + 1), θ ∈ R. Suppose that n > 1. The joint Lebesgue p.d.f. of (X1 , ..., Xn ) is fθ (x) =

n Y

i=1

I(θ,θ+1) (xi ) = I(x(n) −1,x(1) ) (θ),

x = (x1 , ..., xn ) ∈ Rn ,

where x(i) denotes the ith smallest value of x1 , ..., xn . By Theorem 2.2, T = (X(1) , X(n) ) is sufficient for θ. Note that x(1) = sup{θ : fθ (x) > 0} and x(n) = 1 + inf{θ : fθ (x) > 0}. If S(X) is a statistic sufficient for θ, then by Theorem 2.2, there are Borel functions h and gθ such that fθ (x) = gθ (S(x))h(x). For x with h(x) > 0, x(1) = sup{θ : gθ (S(x)) > 0} and x(n) = 1 + inf{θ : gθ (S(x)) > 0}. Hence, there is a measurable function ψ such that T (x) = ψ(S(x)) when h(x) > 0. Since h > 0 a.s. P, we conclude that T is minimal sufficient.

108

2. Fundamentals of Statistics

Minimal sufficient statistics exist under weak assumptions, e.g., P contains distributions on Rk dominated by a σ-finite measure (Bahadur, 1957). The next theorem provides some useful tools for finding minimal sufficient statistics. Theorem 2.3. Let P be a family of distributions on Rk . (i) Suppose that P0 ⊂ P and a.s. P0 implies a.s. P. If T is sufficient for P ∈ P and minimal sufficient for P ∈ P0 , then T is minimal sufficient for P ∈ P. (ii) Suppose that PPcontains p.d.f.’s f0 , f1 , f2 , ..., w.r.t. a σ-finite meaP∞ ∞ sure. Let f∞ (x) = i=0 ci fi (x), where ci > 0 for all i and i=0 ci = 1, and let Ti (X) = fi (x)/f∞ (x) when f∞ (x) > 0, i = 0, 1, 2, .... Then T (X) = (T0 , T1 , T2 , ...) is minimal sufficient for P ∈ P. Furthermore, if {x : fi (x) > 0} ⊂ {x : f0 (x) > 0} for all i, then we may replace f∞ by f0 , in which case T (X) = (T1 , T2 , ...) is minimal sufficient for P ∈ P. (iii) Suppose that P contains p.d.f.’s fP w.r.t. a σ-finite measure and that there exists a sufficient statistic T (X) such that, for any possible values x and y of X, fP (x) = fP (y)φ(x, y) for all P implies T (x) = T (y), where φ is a measurable function. Then T (X) is minimal sufficient for P ∈ P. Proof. (i) If S is sufficient for P ∈ P, then it is also sufficient for P ∈ P0 and, therefore, T = ψ(S) a.s. P0 holds for a measurable function ψ. The result follows from the assumption that a.s. P0 implies a.s. P. (ii) Note that f∞ > 0 a.s. P. Let gi (T ) = Ti , i = 0, 1, 2, .... Then fi (x) = gi (T (x))f∞ (x) a.s. P. By Theorem 2.2, T is sufficient for P ∈ P. Suppose that S(X) is another sufficient statistic. By Theorem 2.2, there are Borel functions h and Pg˜∞i such that fi (x) = g˜i (S(x))h(x), i = 0, 1, 2, .... Then Ti (x) = g˜i (S(x))/ j=0 cj g˜j (S(x)) for x’s satisfying f∞ (x) > 0. By Definition 2.5, T is minimal sufficient for P ∈ P. The proof for the case where f∞ is replaced by f0 is the same. (iii) From Bahadur (1957), there exists a minimal sufficient statistic S(X). The result follows if we can show that T (X) = ψ(S(X)) a.s. P for a measurable function ψ. By Theorem 2.2, there are Borel functions gP and h such that fP (x) = gP (S(x))h(x) for all P . Let A = {x : h(x) = 0}. Then P (A) = 0 for all P . For x and y such that S(x) = S(y), x 6∈ A and y 6∈ A, fP (x) = gP (S(x))h(x) = gP (S(y))h(x)h(y)/h(y) = fP (y)h(x)/h(y) for all P . Hence T (x) = T (y). This shows that there is a function ψ such that T (x) = ψ(S(x)) except for x ∈ A. It remains to show that ψ is measurable. Since S is minimal sufficient, g(T (X)) = S(X) a.s. P for a measurable function g. Hence g is one-to-one and ψ = g −1 . The measurability of ψ follows from Theorem 3.9 in Parthasarathy (1967).

109

2.2. Statistics, Sufficiency, and Completeness

Example 2.14. Let P = {fθ : θ ∈ Θ} be an exponential family with p.d.f.’s fθ given by (2.4) and X(ω) = ω. Suppose that there exists Θ0 = {θ0 , θ1 , ..., θp } ⊂ Θ such that the vectors ηi = η(θi ) − η(θ0 ), i = 1, ..., p, are linearly independent in Rp . (This is true if the family is of full rank.) We have shown that T (X) is sufficient for θ ∈ Θ. We now show that T is in fact minimal sufficient for θ ∈ Θ. Let P0 = {fθ : θ ∈ Θ0 }. Note that the set {x : fθ (x) > 0} does not depend on θ. It follows from Theorem 2.3(ii) with f∞ = fθ0 that S(X) = exp{η1τ T (x) − ξ1 }, ..., exp{ηpτ T (x) − ξp }

is minimal sufficient for θ ∈ Θ0 , where ξi = ξ(θi ) − ξ(θ0 ). Since ηi ’s are linearly independent, there is a one-to-one measurable function ψ such that T (X) = ψ(S(X)) a.s. P0 . Hence, T is minimal sufficient for θ ∈ Θ0 . It is easy to see that a.s. P0 implies a.s. P. Thus, by Theorem 2.3(i), T is minimal sufficient for θ ∈ Θ. The results in Examples 2.13 and 2.14 can also be proved by using Theorem 2.3(iii) (Exercise 32). The sufficiency (and minimal sufficiency) depends on the postulated family P of populations (statistical models). Hence, it may not be a useful concept if the proposed statistical model is wrong or at least one has some doubts about the correctness of the proposed model. From the examples in this section and some exercises in §2.6, one can find that for a wide ¯ in (2.1), S 2 in (2.2), (X(1) , X(n) ) in variety of models, statistics such as X Example 2.11, and the order statistics in Example 2.9 are sufficient. Thus, using these statistics for data reduction and summarization does not lose any information when the true model is one of those models but we do not know exactly which model is correct.

2.2.3 Complete statistics A statistic V (X) is said to be ancillary if its distribution does not depend on the population P and first-order ancillary if E[V (X)] is independent of P . A trivial ancillary statistic is the constant statistic V (X) ≡ c ∈ R. If V (X) is a nontrivial ancillary statistic, then σ(V (X)) ⊂ σ(X) is a nontrivial σ-field that does not contain any information about P . Hence, if S(X) is a statistic and V (S(X)) is a nontrivial ancillary statistic, it indicates that σ(S(X)) contains a nontrivial σ-field that does not contain any information about P and, hence, the “data” S(X) may be further reduced. A sufficient statistic T appears to be most successful in reducing the data if no nonconstant function of T is ancillary or even first-order ancillary. This leads to the following concept of completeness.

110

2. Fundamentals of Statistics

Definition 2.6 (Completeness). A statistic T (X) is said to be complete for P ∈ P if and only if, for any Borel f , E[f (T )] = 0 for all P ∈ P implies f (T ) = 0 a.s. P. T is said to be boundedly complete if and only if the previous statement holds for any bounded Borel f . A complete statistic is boundedly complete. If T is complete (or boundedly complete) and S = ψ(T ) for a measurable ψ, then S is complete (or boundedly complete). Intuitively, a complete and sufficient statistic should be minimal sufficient, which was shown by Lehmann and Scheff´e (1950) and Bahadur (1957) (see Exercise 48). However, a minimal sufficient statistic is not necessarily complete; for example, the minimal sufficient statistic (X(1) , X(n) ) in Example 2.13 is not complete (Exercise 47). Proposition 2.1. If P is in an exponential family of full rank with p.d.f.’s given by (2.6), then T (X) is complete and sufficient for η ∈ Ξ. Proof. We have shown that T is sufficient. Suppose that there is a function f such that E[f (T )] = 0 for all η ∈ Ξ. By Theorem 2.1(i), Z f (t) exp{η τ t − ζ(η)}dλ = 0 for all η ∈ Ξ, where λ is a measure on (Rp , B p ). Let η0 be an interior point of Ξ. Then Z Z τ τ (2.16) f+ (t)eη t dλ = f− (t)eη t dλ for all η ∈ N (η0 ), where N (η0 ) = {η ∈ Rp : kη − η0 k < ǫ} for some ǫ > 0. In particular, Z Z τ τ f+ (t)eη0 t dλ = f− (t)eη0 t dλ = c. τ

τ

If c = 0, then f = 0 a.e. λ. If c > 0, then c−1 f+ (t)eη0 t and c−1 f− (t)eη0 t are p.d.f.’s w.r.t. λ and (2.16) implies that their m.g.f.’s are the same in a τ τ neighborhood of 0. By Theorem 1.6(ii), c−1 f+ (t)eη0 t = c−1 f− (t)eη0 t , i.e., f = f+ − f− = 0 a.e. λ. Hence T is complete. Proposition 2.1 is useful for finding a complete and sufficient statistic when the family of distributions is an exponential family of full rank. Example 2.15. Suppose that X1 , ..., Xn are i.i.d. random variables having the N (µ, σ 2 ) distribution, µ ∈ R, σ > 0. From Example 2.6, the joint Pn p.d.f. of X1 , ...,P Xn is (2π)−n/2 exp {η1 T1 + η2 T2 − nζ(η)}, where T = 1 i=1 Xi , n T2 = − i=1 Xi2 , and η = (η1 , η2 ) = σµ2 , 2σ1 2 . Hence, the family of distributions for X = (X1 , ..., Xn ) is a natural exponential family of full rank (Ξ = R × (0, ∞)). By Proposition 2.1, T (X) = (T1 , T2 ) is complete and sufficient for η. Since there is a one-to-one correspondence between η

2.2. Statistics, Sufficiency, and Completeness

111

and θ = (µ, σ 2 ), T is also complete and sufficient for θ. It can be shown that any one-to-one measurable function of a complete and sufficient statistic ¯ S 2 ) is complete and is also complete and sufficient (exercise). Thus, (X, 2 ¯ sufficient for θ, where X and S are the sample mean and variance given by (2.1) and (2.2), respectively. The following examples show how to find a complete statistic for a nonexponential family. Example 2.16. Let X1 , ..., Xn be i.i.d. random variables from Pθ , the uniform distribution U (0, θ), θ > 0. The largest order statistic, X(n) , is complete and sufficient for θ ∈ (0, ∞). The sufficiency of X(n) follows from the fact that the joint Lebesgue p.d.f. of X1 , ..., Xn is θ−n I(0,θ) (x(n) ). From Example 2.9, X(n) has the Lebesgue p.d.f. (nxn−1 /θn )I(0,θ) (x) on R. Let f be a Borel function on [0, ∞) such that E[f (X(n) )] = 0 for all θ > 0. Then Z

θ

f (x)xn−1 dx = 0 for all θ > 0.

0

Let G(θ) be the left-hand side of the previous equation. Applying the result of differentiation of an integral (see, e.g., Royden (1968, §5.3)), we obtain that G′ (θ) = f (θ)θn−1 a.e. m+ , where m+ is the Lebesgue measure on ([0, ∞), B[0,∞) ). Since G(θ) = 0 for all θ > 0, f (θ)θn−1 = 0 a.e. m+ and, hence, f (x) = 0 a.e. m+ . Therefore, X(n) is complete and sufficient for θ ∈ (0, ∞). Example 2.17. In Example 2.12, we showed that the order statistics T (X) = (X(1) , ..., X(n) ) of i.i.d. random variables X1 , ..., Xn is sufficient for P ∈ P, where P is the family of distributions on R having Lebesgue p.d.f.’s. We now show that T (X) is also complete for P ∈ P. Let P0 be the family of Lebesgue p.d.f.’s of the form f (x) = C(θ1 , ..., θn ) exp{−x2n + θ1 x + θ2 x2 + · · · + θn xn }, R where θj ∈ R and C(θ1 , ..., θn ) is a normalizing constant such that f (x)dx = 1. Then P0 ⊂ P and P0 is an exponential family of full rank. Note that the joint distribution of X = (X1 , ..., Xn ) is also in an exponential family of full rank. Thus, by Proposition Pn 2.1, U = (U1 , ..., Un ) is a complete statistic for P ∈ P0 , where Uj = i=1 Xij . Since a.s. P0 implies a.s. P, U (X) is also complete for P ∈ P. The result follows if we can show that there Pnis a one-to-onePcorrespondenceP between T (X) and U (X). Let V1 = i<j Xi Xj , i=1 Xi , V2 = V3 = i<j 0. It can be easily shown that the family {N (µ, σ 2 ) : µ ∈ R} is an exponential family of full rank with natural parameter η = µ/σ 2 . By Proposition 2.1, the sample ¯ in (2.1) is complete and sufficient for η (and µ). Let S 2 be the mean X Pn ¯ 2 , where sample variance given by (2.2). SincePS 2 = (n − 1)−1 i=1 (Zi − Z) n 2 −1 2 Zi = Xi −µ is N (0, σ ) and Z¯ = n Z , S is an ancillary statistic (σ 2 i=1 i 2 ¯ and S are independent w.r.t. N (µ, σ 2 ) is known). By Basu’s theorem, X 2 ¯ and S 2 are independent w.r.t. N (µ, σ 2 ) with µ ∈ R. Since σ is arbitrary, X 2 for any µ ∈ R and σ > 0. ¯ and S 2 , we now show that (n − 1)S 2 /σ 2 Using the independence of X has the chi-square distribution χ2n−1 . Note that 2 2 ¯ n X −µ (n − 1)S 2 X Xi − µ n + = . σ σ2 σ i=1 ¯ − µ)2 /σ 2 has the chiFrom the properties of the normal distributions, n(X P 2 −1/2 square distribution χ1 with the m.g.f. (1 − 2t) and ni=1 (Xi − µ)2 /σ 2

113

2.3. Statistical Decision Theory

has the chi-square distribution χ2n with the m.g.f. (1 − 2t)−n/2 , t < 1/2. By ¯ and S 2 , the m.g.f. of (n − 1)S 2 /σ 2 is the independence of X (1 − 2t)−n/2 /(1 − 2t)−1/2 = (1 − 2t)−(n−1)/2

for t < 1/2. This is the m.g.f. of the chi-square distribution χ2n−1 and, therefore, the result follows.

2.3 Statistical Decision Theory In this section, we describe some basic elements in statistical decision theory. More developments are given in later chapters.

2.3.1 Decision rules, loss functions, and risks Let X be a sample from a population P ∈ P. A statistical decision is an action that we take after we observe X, for example, a conclusion about P or a characteristic of P . Throughout this section, we use A to denote the set of allowable actions. Let FA be a σ-field on A. Then the measurable space (A, FA ) is called the action space. Let X be the range of X and FX be a σ-field on X. A decision rule is a measurable function (a statistic) T from (X, FX ) to (A, FA ). If a decision rule T is chosen, then we take the action T (X) ∈ A whence X is observed. The construction or selection of decision rules cannot be done without any criterion about the performance of decision rules. In statistical decision theory, we set a criterion using a loss function L, which is a function from P × A to [0, ∞) and is Borel on (A, FA ) for each fixed P ∈ P. If X = x is observed and our decision rule is T , then our “loss” (in making a decision) is L(P, T (x)). The average loss for the decision rule T , which is called the risk of T , is defined to be Z RT (P ) = E[L(P, T (X))] = L(P, T (x))dPX (x). (2.17) X

The loss and risk functions are denoted by L(θ, a) and RT (θ) if P is a parametric family indexed by θ. A decision rule with small loss is preferred. But it is difficult to compare L(P, T1 (X)) and L(P, T2 (X)) for two decision rules, T1 and T2 , since both of them are random. For this reason, the risk function (2.17) is introduced and we compare two decision rules by comparing their risks. A rule T1 is as good as another rule T2 if and only if RT1 (P ) ≤ RT2 (P ) for any P ∈ P,

(2.18)

and is better than T2 if and only if (2.18) holds and RT1 (P ) < RT2 (P ) for at least one P ∈ P. Two decision rules T1 and T2 are equivalent if and only

114

2. Fundamentals of Statistics

if RT1 (P ) = RT2 (P ) for all P ∈ P. If there is a decision rule T∗ that is as good as any other rule in ℑ, a class of allowable decision rules, then T∗ is said to be ℑ-optimal (or optimal if ℑ contains all possible rules). Example 2.19. Consider the measurement problem in Example 2.1. Suppose that we need a decision on the value of θ ∈ R, based on the sample X = (X1 , ..., Xn ). If Θ is all possible values of θ, then it is reasonable to consider the action space (A, FA ) = (Θ, BΘ ). An example of a decision rule ¯ the sample mean defined by (2.1). A common loss function is T (X) = X, in this problem is the squared error loss L(P, a) = (θ − a)2 , a ∈ A. Then ¯ is the squared deviation between X ¯ and θ. the loss for the decision rule X Assuming that the population has mean µ and variance σ 2 < ∞, we obtain ¯ the following risk function for X: ¯ 2 RX¯ (P ) = E(θ − X) ¯ 2 + E(E X ¯ − X) ¯ 2 = (θ − E X) 2 ¯ + Var(X) ¯ = (θ − E X) 2

= (µ − θ) +

2

σ n

,

(2.19) (2.20)

¯ in Examwhere result (2.20) follows from the results for the moments of X ple 2.8. If θ is in fact the mean of the population, then the first term on the right-hand side of (2.20) is 0 and the risk is an increasing function of the population variance σ 2 and a decreasing function of the sample size n. Consider another decision rule T1 (X) = (X(1) + X(n) )/2. However, RT1 (P ) does not have an explicit form if there is no further assumption on ¯ (or T1 ) is the population P . Suppose that P ∈ P. Then, for some P, X ¯ ¯ nor T1 is better than T1 (or X) (exercise), whereas for some P, neither X better than the other. A different loss function may also be considered. For example, L(P, a) = |θ − a|, which is called the absolute error loss. However, RX¯ (P ) and RT1 (P ) do not have explicit forms unless P is of some specific form. The problem in Example 2.19 is a special case of a general problem called estimation, in which the action space is the set of all possible values of a population characteristic ϑ to be estimated. In an estimation problem, a decision rule T is called an estimator and result (2.19) holds with θ = ϑ and ¯ replaced by any estimator with a finite variance. The following example X describes another type of important problem called hypothesis testing. Example 2.20. Let P be a family of distributions, P0 ⊂ P, and P1 = {P ∈ P : P 6∈ P0 }. A hypothesis testing problem can be formulated as that of deciding which of the following two statements is true: H0 : P ∈ P0

versus

H1 : P ∈ P 1 .

(2.21)

115

2.3. Statistical Decision Theory

Here, H0 is called the null hypothesis and H1 is called the alternative hypothesis. The action space for this problem contains only two elements, i.e., A = {0, 1}, where 0 is the action of accepting H0 and 1 is the action of rejecting H0 . A decision rule is called a test. Since a test T (X) is a function from X to {0, 1}, T (X) must have the form IC (X), where C ∈ FX is called the rejection region or critical region for testing H0 versus H1 . A simple loss function for this problem is the 0-1 loss: L(P, a) = 0 if a correct decision is made and 1 if an incorrect decision is made, i.e., L(P, j) = 0 for P ∈ Pj and L(P, j) = 1 otherwise, j = 0, 1. Under this loss, the risk is P (T (X) = 1) = P (X ∈ C) P ∈ P0 RT (P ) = P (T (X) = 0) = P (X 6∈ C) P ∈ P1 . See Figure 2.2 on page 127 for an example of a graph of RT (θ) for some T and P in a parametric family. The 0-1 loss implies that the loss for two types of incorrect decisions (accepting H0 when P ∈ P1 and rejecting H0 when P ∈ P0 ) are the same. In some cases, one might assume unequal losses: L(P, j) = 0 for P ∈ Pj , L(P, 0) = c0 when P ∈ P1 , and L(P, 1) = c1 when P ∈ P0 . In the following example the decision problem is neither an estimation nor a testing problem. Another example is given in Exercise 93 in §2.6. Example 2.21. A hazardous toxic waste site requires clean-up when the true chemical concentration θ in the contaminated soil is higher than a given level θ0 ≥ 0. Because of the limitation in resources, we would like to spend our money and efforts more in those areas that pose high risk to public health. In a particular area where soil samples are obtained, we would like to take one of these three actions: a complete clean-up (a1 ), a partial clean-up (a2 ), and no clean-up (a3 ). Then A = {a1 , a2 , a3 }. Suppose that the cost for a complete clean-up is c1 and for a partial clean-up is c2 < c1 ; the risk to public health is c3 (θ − θ0 ) if θ > θ0 and 0 if θ ≤ θ0 ; a complete clean-up can reduce the toxic concentration to an amount ≤ θ0 , whereas a partial clean-up can only reduce a fixed amount of the toxic concentration, i.e., the chemical concentration becomes θ−t after a partial clean-up, where t is a known constant. Then the loss function is given by L(θ, a) θ ≤ θ0 θ 0 < θ ≤ θ0 + t θ > θ0 + t

a1 c1 c1 c1

a2 c2 c2 c2 + c3 (θ − θ0 − t)

a3 0 c3 (θ − θ0 ) c3 (θ − θ0 )

The risk function can be calculated once the decision rule is specified. We discuss this example again in Chapter 4.

116

2. Fundamentals of Statistics

Sometimes it is useful to consider randomized decision rules. Examples are given in §2.3.2, Chapters 4 and 6. A randomized decision rule is a function δ on X × FA such that, for every A ∈ FA , δ(·, A) is a Borel function and, for every x ∈ X, δ(x, ·) is a probability measure on (A, FA ). To choose an action in A when a randomized rule δ is used, we need to simulate a pseudorandom element of A according to δ(x, ·). Thus, an alternative way to describe a randomized rule is to specify the method of simulating the action from A for each x ∈ X. If A is a subset of a Euclidean space, for example, then the result in Theorem 1.7(ii) can be applied. Also, see §7.2.3. A nonrandomized decision rule T previously discussed can be viewed as a special randomized decision rule with δ(x, {a}) = I{a} (T (x)), a ∈ A, x ∈ X. Another example of a randomized rule is a discrete distribution δ(x, ·) assigning probability pj (x) to a nonrandomized decision rule Tj (x), j = 1, 2, ..., in which case the rule δ can be equivalently defined as a rule taking value Tj (x) with probability pj (x). See Exercise 64 for an example. The loss function for a randomized rule δ is defined as Z L(P, δ, x) = L(P, a)dδ(x, a), A

which reduces to the same loss function we discussed when δ is a nonrandomized rule. The risk of a randomized rule δ is then Z Z Rδ (P ) = E[L(P, δ, X)] = L(P, a)dδ(x, a)dPX (x). (2.22) X

A

2.3.2 Admissibility and optimality Consider a given decision problem with a given loss L(P, a). Definition 2.7 (Admissibility). Let ℑ be a class of decision rules (randomized or nonrandomized). A decision rule T ∈ ℑ is called ℑ-admissible (or admissible when ℑ contains all possible rules) if and only if there does not exist any S ∈ ℑ that is better than T (in terms of the risk). If a decision rule T is inadmissible, then there exists a rule better than T . Thus, T should not be used in principle. However, an admissible decision rule is not necessarily good. For example, in an estimation problem a silly estimator T (X) ≡ a constant may be admissible (Exercise 71). The relationship between the admissibility and the optimality defined in §2.3.1 can be described as follows. If T∗ is ℑ-optimal, then it is ℑ-admissible; if T∗ is ℑ-optimal and T0 is ℑ-admissible, then T0 is also ℑ-optimal and is equivalent to T∗ ; if there are two ℑ-admissible rules that are not equivalent, then there does not exist any ℑ-optimal rule.

2.3. Statistical Decision Theory

117

Suppose that we have a sufficient statistic T (X) for P ∈ P. Intuitively, our decision rule should be a function of T , based on the discussion in §2.2.2. This is not true in general, but the following result indicates that this is true if randomized decision rules are allowed. Proposition 2.2. Suppose that A is a subset of Rk . Let T (X) be a sufficient statistic for P ∈ P and let δ0 be a decision rule. Then δ1 (t, A) = E[δ0 (X, A)|T = t],

(2.23)

which is a randomized decision rule depending only on T , is equivalent to δ0 if Rδ0 (P ) < ∞ for any P ∈ P. Proof. Note that δ1 defined by (2.23) is a decision rule since δ1 does not depend on the unknown P by the sufficiency of T . From (2.22), Z Rδ1 (P ) = E L(P, a)dδ1 (X, a) A Z L(P, a)dδ0 (X, a) T =E E Z A L(P, a)dδ0 (X, a) =E A

= Rδ0 (P ),

where the proof of the second equality is left to the reader. Note that Proposition 2.2 does not imply that δ0 is inadmissible. Also, if δ0 is a nonrandomized rule, δ1 (t, A) = E[IA (δ0 (X))|T = t] = P (δ0 (X) ∈ A|T = t) is still a randomized rule, unless δ0 (X) = h(T (X)) a.s. P for some Borel function h (Exercise 75). Hence, Proposition 2.2 does not apply to situations where randomized rules are not allowed. The following result tells us when nonrandomized rules are all we need and when decision rules that are not functions of sufficient statistics are inadmissible. Theorem 2.5. Suppose that A is a convex subset of Rk and that for any P ∈ P, L(P, a) is a convex function of a. R (i) Let δ be a randomized R rule satisfying A kakdδ(x, a) < ∞ for any x ∈ X and let T1 (x) = A adδ(x, a). Then L(P, T1 (x)) ≤ L(P, δ, x) (or L(P, T1 (x)) < L(P, δ, x) if L is strictly convex in a) for any x ∈ X and P ∈ P. (ii) (Rao-Blackwell theorem). Let T be a sufficient statistic for P ∈ P, T0 ∈ Rk be a nonrandomized rule satisfying EkT0 k < ∞, and T1 = E[T0 (X)|T ]. Then RT1 (P ) ≤ RT0 (P ) for any P ∈ P. If L is strictly convex in a and T0 is not a function of T , then T0 is inadmissible.

118

2. Fundamentals of Statistics

The proof of Theorem 2.5 is an application of Jensen’s inequality (1.47) and is left to the reader. The concept of admissibility helps us to eliminate some decision rules. However, usually there are still too many rules left after the elimination of some rules according to admissibility and sufficiency. Although one is typically interested in a ℑ-optimal rule, frequently it does not exist, if ℑ is either too large or too small. The following examples are illustrations. Example 2.22. Let X1 , ..., Xn be i.i.d. random variables from a population P ∈ P that is the family of populations having finite mean µ and variance σ 2 . Consider the estimation of µ (A = R) under the squared error loss. It can be shown that if we let ℑ be the class of all possible estimators, then there is no ℑ-optimal rule (exercise). Next, P let ℑ1 be the class of all linear n functions in X = (X1 , ..., Xn ), i.e., T (X) = i=1 ci Xi with known ci ∈ R, i = 1, ..., n. It follows from (2.19) and the discussion after Example 2.19 that !2 n n X X RT (P ) = µ2 ci − 1 + σ 2 c2i . (2.24) i=1

i=1

Pn

We now show that there does not exist T∗ = i=1 c∗i Xi such that RT∗ (P ) ≤ RT (P ) for any P ∈ P and T ∈ ℑ1 . If there is such a T∗ , then (c∗1 , ..., c∗n ) is a minimum of the function of (c1 , ..., cn ) on the right-hand side of (2.24). Then c∗1 , ..., c∗n must be the same and equal to µ2 /(σ 2 +nµ2 ), which depends on P . Hence T∗ is not a statistic. This shows that there is no ℑ1 -optimal rule. Pn Consider now a subclass ℑ2 ⊂ ℑ1 with ci ’s satisfying P i=1 ci = 1. From P n (2.24), RT (P ) = σ 2 i=1 c2i if T ∈ ℑ2 . Minimizing σ 2 ni=1 c2i subject to Pn −1 for all i. Thus, the i=1 ci = 1 leads to an optimal solution of ci = n ¯ sample mean X is ℑ2 -optimal. There may not be any optimal rule if we consider a small class of decision ¯ then one rules. For example, if ℑ3 contains all the rules in ℑ2 except X, can show that there is no ℑ3 -optimal rule. Example 2.23. Assume that the sample X has the binomial distribution Bi(θ, n) with an unknown θ ∈ (0, 1) and a fixed integer n > 1. Consider the hypothesis testing problem described in Example 2.20 with H0 : θ ∈ (0, θ0 ] versus H1 : θ ∈ (θ0 , 1), where θ0 ∈ (0, 1) is a fixed value. Suppose that we are only interested in the following class of nonrandomized decision rules: ℑ = {Tj : j = 0, 1, ..., n − 1}, where Tj (X) = I{j+1,...,n} (X). From Example 2.20, the risk function for Tj under the 0-1 loss is RTj (θ) = P (X > j)I(0,θ0 ] (θ) + P (X ≤ j)I(θ0 ,1) (θ).

119

2.3. Statistical Decision Theory

For any integers k and j, 0 ≤ k < j ≤ n − 1, −P (k < X ≤ j) < 0 RTj (θ) − RTk (θ) = P (k < X ≤ j) > 0

0 < θ ≤ θ0 θ0 < θ < 1.

Hence, neither Tj nor Tk is better than the other. This shows that every Tj is ℑ-admissible and, thus, there is no ℑ-optimal rule. In view of the fact that an optimal rule often does not exist, statisticians adopt the following two approaches to choose a decision rule. The first approach is to define a class ℑ of decision rules that have some desirable properties (statistical and/or nonstatistical) and then try to find the best rule in ℑ. In Example 2.22, for instance, any estimator T in ℑ2 has the property that T is linear in X and E[T (X)] = µ. In a general estimation problem, we can use the following concept. Definition 2.8 (Unbiasedness). In an estimation problem, the bias of an estimator T (X) of a real-valued parameter ϑ of the unknown population is defined to be bT (P ) = E[T (X)] − ϑ (which is denoted by bT (θ) when P is in a parametric family indexed by θ). An estimator T (X) is said to be unbiased for ϑ if and only if bT (P ) = 0 for any P ∈ P. Thus, ℑ2 in Example 2.22 is the class of unbiased estimators linear in X. In Chapter 3, we discuss how to find a ℑ-optimal estimator when ℑ is the class of unbiased estimators or unbiased estimators linear in X. Another class of decision rules can be defined after we introduce the concept of invariance. Definition 2.9 Let X be a sample from P ∈ P. (i) A class G of one-to-one transformations of X is called a group if and only if gi ∈ G implies g1◦g2 ∈ G and gi−1 ∈ G. (ii) We say that P is invariant under G if and only if g¯(PX ) = Pg(X) is a one-to-one transformation from P onto P for each g ∈ G. (iii) A decision problem is said to be invariant if and only if P is invariant under G and the loss L(P, a) is invariant in the sense that, for every g ∈ G and every a ∈ A, there exists a unique g(a) ∈ A such that L(PX , a) = L Pg(X) , g(a) . (Note that g(X) and g(a) are different functions in general.) (iv) A decision rule T (x) is said to be invariant if and only if, for every g ∈ G and every x ∈ X, T (g(x)) = g(T (x)). Invariance means that our decision is not affected by one-to-one transformations of data. In a problem where the distribution of X is in a location-scale family

120

2. Fundamentals of Statistics

P on Rk , we often consider location-scale transformations of data X of the form g(X) = AX + c, where c ∈ C ⊂ Rk and A ∈ T , a class of invertible k × k matrices. Assume that if Ai ∈ T , i = 1, 2, then A−1 ∈ T and i A1 A2 ∈ T , and that if ci ∈ C, i = 1, 2, then −ci ∈ C and Ac1 + c2 ∈ C for any A ∈ T . Then the collection of all transformations is a group. A special case is given in the following example. Example 2.24. Let X have i.i.d. components from a population in a location family P = {Pµ : µ ∈ R}. Consider the location transformation gc (X) = X +cJk , where c ∈ R and Jk is the k-vector whose components are all equal to 1. The group of transformation is G = {gc : c ∈ R}, which is a location-scale transformation group with T = {Ik } and C = {cJk : c ∈ R}. P is invariant under G with g¯c (Pµ ) = Pµ+c . For estimating µ under the loss L(µ, a) = L(µ − a), where L(·) is a nonnegative Borel function, the decision problem is invariant with gc (a) = a + c. A decision rule T is invariant if and only if T (x + cJk ) = T (x) + c for every x ∈ Rk and c ∈ R. An example of an invariant decision rule is T (x) = lτ x for some l ∈ Rk with lτ Jk = 1. Note that T (x) = lτ x with lτ Jk = 1 is in the class ℑ2 in Example 2.22. In §4.2 and §6.3, we discuss the problem of finding a ℑ-optimal rule when ℑ is a class of invariant decision rules. The second approach to finding a good decision rule is to consider some characteristic RT of RT (P ), for a given decision rule T , and then minimize RT over T ∈ ℑ. The following are two popular ways to carry out this idea. The first one is to consider an average of RT (P ) over P ∈ P: rT (Π) =

Z

RT (P )dΠ(P ),

P

where Π is a known probability measure on (P, FP ) with an appropriate σ-field FP . rT (Π) is called the Bayes risk of T w.r.t. Π. If T∗ ∈ ℑ and rT∗ (Π) ≤ rT (Π) for any T ∈ ℑ, then T∗ is called a ℑ-Bayes rule (or Bayes rule when ℑ contains all possible rules) w.r.t. Π. The second method is to consider the worst situation, i.e., supP ∈P RT (P ). If T∗ ∈ ℑ and supP ∈P RT∗ (P ) ≤ supP ∈P RT (P ) for any T ∈ ℑ, then T∗ is called a ℑ-minimax rule (or minimax rule when ℑ contains all possible rules). Bayes and minimax rules are discussed in Chapter 4. Example 2.25. We usually try to find a Bayes rule or a minimax rule in a parametric problem where P = Pθ for a θ ∈ Rk . Consider the special case of k = 1 and L(θ, a) = (θ − a)2 , the squared error loss. Note that rT (Π) =

Z

R

E[θ − T (X)]2 dΠ(θ),

121

2.3. Statistical Decision Theory

which is equivalent to E[θ − T (X)]2 , where θ is a random variable having the distribution Π and, given θ = θ, the conditional distribution of X is Pθ . Then, the problem can be viewed as a prediction problem for θ using functions of X. Using the result in Example 1.22, the best predictor is E(θ|X), which is the ℑ-Bayes rule w.r.t. Π with ℑ being the class of rules T (X) satisfying E[T (X)]2 < ∞ for any θ. As a more specific example, let X = (X1 , ..., Xn ) with i.i.d. components having the N (µ, σ 2 ) distribution with an unknown µ = θ ∈ R and a known σ 2 , and let Π be the N (µ0 , σ02 ) distribution with known µ0 and σ02 . Then the conditional distribution of θ given X = x is N (µ∗ (x), c2 ) with µ∗ (x) =

σ2 nσ02 µ + x¯ 0 nσ02 + σ 2 nσ02 + σ 2

and

c2 =

σ02 σ 2 nσ02 + σ 2

(2.25)

(exercise). The Bayes rule w.r.t. Π is E(θ|X) = µ∗ (X). ¯ is ℑ-minimax In this special case we can show that the sample mean X with ℑ being the collection of all decision rules. For any decision rule T , Z sup RT (θ) ≥ RT (θ)dΠ(θ) θ∈R ZR ≥ Rµ∗ (θ)dΠ(θ) R = E [θ − µ∗ (X)]2 = E E{[θ − µ∗ (X)]2 |X} = E(c2 )

= c2 , where µ∗ (X) is the Bayes rule given in (2.25) and c2 is also given in (2.25). Since this result is true for any σ02 > 0 and c2 → σ 2 /n as σ02 → ∞, sup RT (θ) ≥

θ∈R

σ2 = sup RX¯ (θ), n θ∈R

¯ under the squared error loss where the equality holds because the risk of X ¯ is minimax. is, by (2.20), σ 2 /n and independent of θ = µ. Thus, X A minimax rule in a general case may be difficult to obtain. It can be seen that if both µ and σ 2 are unknown in the previous discussion, then sup θ∈R×(0,∞)

RX¯ (θ) = ∞,

(2.26)

¯ cannot be minimax unless (2.26) holds with where θ = (µ, σ 2 ). Hence X ¯ X replaced by any decision rule T , in which case minimaxity becomes meaningless.

122

2. Fundamentals of Statistics

2.4 Statistical Inference The loss function plays a crucial role in statistical decision theory. Loss functions can be obtained from a utility analysis (Berger, 1985), but in many problems they have to be determined subjectively. In statistical inference, we make an inference about the unknown population based on the sample X and inference procedures without using any loss function, although any inference procedure can be cast in decision-theoretic terms as a decision rule. There are three main types of inference procedures: point estimators, hypothesis tests, and confidence sets.

2.4.1 Point estimators The problem of estimating an unknown parameter related to the unknown population is introduced in Example 2.19 and the discussion after Example 2.19 as a special statistical decision problem. In statistical inference, however, estimators of parameters are derived based on some principle (such as the unbiasedness, invariance, sufficiency, substitution principle, likelihood principle, Bayesian principle, etc.), not based on a loss or risk function. Since confidence sets are sometimes also called interval estimators or set estimators, estimators of parameters are called point estimators. In Chapters 3 through 5, we consider how to derive a “good” point estimator based on some principle. Here we focus on how to assess performance of point estimators. ˜ ⊂ R be a parameter to be estimated, which is a function of Let ϑ ∈ Θ the unknown population P or θ if P is in a parametric family. An estimator ˜ First, one has to realize that any estimator T (X) is a statistic with range Θ. of ϑ is subject to an estimation error T (x) − ϑ when we observe X = x. This is not just because T (X) is random. In some problems T (x) never equals ϑ. A trivial example is when T (X) has a continuous c.d.f. so that P (T (X) = ϑ) = 0. As a nontrivial example, let X1 , ..., Xn be i.i.d. binary random variables (also called Bernoulli variables) with P (Xi = 1) = p and ¯ is shown to be a good estimator P (Xi = 0) = 1 − p. The sample mean X of ϑ = p in later chapters, but x ¯ never equals ϑ if ϑ is not one of j/n, j = 0, 1, ..., n. Thus, we cannot assess the performance of T (X) by the values of T (x) with particular x’s and it is also not worthwhile to do so. The bias bT (P ) and unbiasedness of a point estimator T (X) is defined in Definition 2.8. Unbiasedness of T (X) means that the mean of T (X) is equal to ϑ. An unbiased estimator T (X) can be viewed as an estimator without “systematic” error, since, on the average, it does not overestimate (i.e., bT (P ) > 0) or underestimate (i.e., bT (P ) < 0). However, an unbiased

123

2.4. Statistical Inference

estimator T (X) may have large positive and negative errors T (x)−ϑ, x ∈ X, although theseRerrors cancel each other in the calculation of the bias, which is the average [T (x) − ϑ]dPX (x). Hence, for an unbiased estimator T (X), it is desired that the values of T (x) be highly concentrated around ϑ. The variance of T (X) is commonly used as a measure of the dispersion of T (X). The mean squared error (mse) of T (X) as an estimator of ϑ is defined to be mseT (P ) = E[T (X) − ϑ]2 = [bT (P )]2 + Var(T (X)),

(2.27)

which is denoted by mseT (θ) if P is in a parametric family. mseT (P ) is equal to the variance Var(T (X)) if and only if T (X) is unbiased. Note that the mse is simply the risk of T in statistical decision theory under the squared error loss. In addition to the variance and the mse, the following are other measures of dispersion that are often used in point estimation problems. The first one is the mean absolute error of an estimator T (X) defined to be E|T (X) − ϑ|. The second one is the probability of falling outside a stated distance of ϑ, i.e., P (|T (X) − ϑ| ≥ ǫ) with a fixed ǫ > 0. Again, these two measures of dispersion are risk functions in statistical decision theory with loss functions |ϑ − a| and I(ǫ,∞) (|ϑ − a|), respectively. For the bias, variance, mse, and mean absolute error, we have implicitly assumed that certain moments of T (X) exist. On the other hand, the dispersion measure P (|T (X)−ϑ| ≥ ǫ) depends on the choice of ǫ. It is possible that some estimators are good in terms of one measure of dispersion, but not in terms of other measures of dispersion. The mse, which is a function of bias and variance according to (2.27), is mathematically easy to handle and, hence, is used the most often in the literature. In this book, we use the mse to assess and compare point estimators unless otherwise stated. Examples 2.19 and 2.22 provide some examples of estimators and their biases, variances, and mse’s. The following are two more examples. Example 2.26. Consider the life-time testing problem in Example 2.2. Let X1 , ..., Xn be i.i.d. from an unknown c.d.f. F . Suppose that the parameter of interest is ϑ = 1 − F (t) for a fixed t > 0. If F is not in a parametric family, then a nonparametric estimator of F (t) is the empirical c.d.f. n

Fn (t) =

1X I(−∞,t] (Xi ), n i=1

t ∈ R.

(2.28)

Since I(−∞,t] (X1 ), ..., I(−∞,t] (Xn ) are i.i.d. binary random variables with P (I(−∞,t] (Xi ) = 1) = F (t), the random variable nFn (t) has the binomial distribution Bi(F (t), n). Consequently, Fn (t) is an unbiased estimator of

124

2. Fundamentals of Statistics

F (t) and Var(Fn (t)) = mseFn (t) (P ) = F (t)[1 − F (t)]/n. Since any linear combination of unbiased estimators is unbiased for the same linear combination of the parameters (by the linearity of expectations), an unbiased estimator of ϑ is U (X) = 1 − Fn (t), which has the same variance and mse as Fn (t). The estimator U (X) = 1 − Fn (t) can be improved in terms of the mse if there is further information about F . Suppose that F is the c.d.f. of the exponential distribution E(0, θ) with an unknown θ > 0. Then ¯ is sufficient for θ > 0. Since ϑ = e−t/θ . From §2.2.2, the sample mean X the squared error loss is strictly convex, an application of Theorem 2.5(ii) ¯ (Rao-Blackwell theorem) shows that the estimator T (X) = E[1 − Fn (t)|X], which is also unbiased, is better than U (X) in terms of the mse. Figure 2.1 shows graphs of the mse’s of U (X) and T (X), as functions of θ, in the special case of n = 10, t = 2, and F (x) = (1 − e−x/θ )I(0,∞) (x). Example 2.27. Consider the sample survey problem in Example 2.3 a Pwith N constant selection probability p(s) and univariate yi . Let ϑ = Y = i=1 yi , P the population total. We now show that the estimator Yˆ = N i∈s yi is n an unbiased estimator of Y . Let a = 1 if i ∈ s and a = 0 otherwise. Thus, i i PN a y . Since p(s) is constant, E(a ) = P (a = 1) = n/N and Yˆ = N i i i=1 i i n ! N N N X N X N X E(Yˆ ) = E ai y i = yi E(ai ) = yi = Y. n i=1 n i=1 i=1 Note that

Var(ai ) = E(ai ) − [E(ai )]2 = and for i 6= j,

n n 1− N N

Cov(ai , aj ) = P (ai = 1, aj = 1) − E(ai )E(aj ) =

n2 n(n − 1) − 2. N (N − 1) N

Hence, the variance or the mse of Yˆ is ! N X N2 ˆ Var(Y ) = 2 Var ai y i n i=1 N 2 X X N = 2 yi2 Var(ai ) + 2 yi yj Cov(ai , aj ) n i=1 1≤i<j≤N N X X n 2 N 1− yi2 − yi yj = n N N −1 i=1 1≤i<j≤N

2 N n X Y N2 1− yi − . = n(N − 1) N i=1 N

125

0.02

mse of T mse of U

0.0

0.01

mse

0.03

2.4. Statistical Inference

0

2

4

6

8

10

q

Figure 2.1: mse’s of U (X) and T (X) in Example 2.26

2.4.2 Hypothesis tests The basic elements of a hypothesis testing problem are described in Example 2.20. In statistical inference, tests for a hypothesis are derived based on some principles similar to those given in an estimation problem. Chapter 6 is devoted to deriving tests for various types of hypotheses. Several key ideas are discussed here. To test the hypotheses H0 versus H1 given in (2.21), there are only two types of statistical errors we may commit: rejecting H0 when H0 is true (called the type I error) and accepting H0 when H0 is wrong (called the type II error). In statistical inference, a test T , which is a statistic from X to {0, 1}, is assessed by the probabilities of making two types of errors: αT (P ) = P (T (X) = 1)

P ∈ P0

(2.29)

and 1 − αT (P ) = P (T (X) = 0)

P ∈ P1 ,

(2.30)

which are denoted by αT (θ) and 1 − αT (θ) if P is in a parametric family indexed by θ. Note that these are risks of T under the 0-1 loss in statistical decision theory. However, an optimal decision rule (test) does not exist even for a very simple problem with a very simple class of tests (Example 2.23).

126

2. Fundamentals of Statistics

That is, error probabilities in (2.29) and (2.30) cannot be minimized simultaneously. Furthermore, these two error probabilities cannot be bounded simultaneously by a fixed α ∈ (0, 1) when we have a sample of a fixed size. Therefore, a common approach to finding an “optimal” test is to assign a small bound α to one of the error probabilities, say αT (P ), P ∈ P0 , and then to attempt to minimize the other error probability 1 − αT (P ), P ∈ P1 , subject to sup αT (P ) ≤ α. (2.31) P ∈P0

The bound α is called the level of significance. The left-hand side of (2.31) is called the size of the test T . Note that the level of significance should be positive, otherwise no test satisfies (2.31) except the silly test T (X) ≡ 0 a.s. P. Example 2.28. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 . Consider the hypotheses H0 : µ ≤ µ0

versus

H1 : µ > µ0 ,

¯ is sufficient for where µ0 is a fixed constant. Since the sample mean X µ ∈ R, it is reasonable to consider the following class of tests: Tc (X) = ¯ i.e., H0 is rejected (accepted) if X ¯ > c (X ¯ ≤ c), where c ∈ R is I(c,∞) (X), a fixed constant. Let Φ be the c.d.f. of N (0, 1). Then, by the property of the normal distributions, √ n(c − µ) αTc (µ) = P (Tc (X) = 1) = 1 − Φ . (2.32) σ Figure 2.2 provides an example of a graph of two types of error probabilities, with µ0 = 0. Since Φ(t) is an increasing function of t, √ n(c − µ0 ) . sup αTc (µ) = 1 − Φ σ P ∈P0 In fact, it is also true that sup [1 − αTc (µ)] = Φ

P ∈P1

√ n(c − µ0 ) . σ

If we would like to use an α as the level of significance, then the most effective way is to choose a cα (a test Tcα (X)) such that α = sup αTcα (µ), P ∈P0

in which case cα must satisfy 1−Φ

√ n(cα − µ0 ) = α, σ

127

0.6 0.4 0.0

0.2

error probability

0.8

1.0

2.4. Statistical Inference

-2

-1

0

1

2

m

Figure 2.2: Error probabilities in Example 2.28 √ i.e., cα = σz1−α / n + µ0 , where za = Φ−1 (a). In Chapter 6, it is shown that for any test T (X) satisfying (2.31), 1 − αT (µ) ≥ 1 − αTcα (µ),

µ > µ0 .

The choice of a level of significance α is usually somewhat subjective. In most applications there is no precise limit to the size of T that can be tolerated. Standard values, such as 0.10, 0.05, or 0.01, are often used for convenience. For most tests satisfying (2.31), a small α leads to a “small” rejection region. It is good practice to determine not only whether H0 is rejected or accepted for a given α and a chosen test Tα , but also the smallest possible level of significance at which H0 would be rejected for the computed Tα (x), i.e., α ˆ = inf{α ∈ (0, 1) : Tα (x) = 1}. Such an α ˆ , which depends on x and the chosen test and is a statistic, is called the p-value for the test Tα . Example 2.29. Consider the problem in Example 2.28. Let us calculate the p-value for Tcα . Note that α=1−Φ

√ √ n(cα − µ0 ) n(¯ x − µ0 ) >1−Φ σ σ

128

2. Fundamentals of Statistics

if and only if x ¯ > cα (or Tcα (x) = 1). Hence √ n(¯ x − µ0 ) = inf{α ∈ (0, 1) : Tcα (x) = 1} = α 1−Φ ˆ (x) σ is the p-value for Tcα . It turns out that Tcα (x) = I(0,α) (ˆ α(x)). With the additional information provided by p-values, using p-values is typically more appropriate than using fixed-level tests in a scientific problem. However, a fixed level of significance is unavoidable when acceptance or rejection of H0 implies an imminent concrete decision. For more discussions about p-values, see Lehmann (1986) and Weerahandi (1995). In Example 2.28, the equality in (2.31) can always be achieved by a suitable choice of c. This is, however, not true in general. In Example 2.23, for instance, it is possible to find an α such that sup P (Tj (X) = 1) 6= α

0j 1 Tj,q (X) = q X =j 0 X < j,

where j = 0, 1, ..., n − 1 and q ∈ [0, 1]. Then

αTj,q (θ) = P (X > j) + qP (X = j)

0 < θ ≤ θ0

and 1 − αTj,q (θ) = P (X < j) + (1 − q)P (X = j)

θ0 < θ < 1.

It can be shown that for any α ∈ (0, 1), there exist an integer j and q ∈ (0, 1) such that the size of Tj,q is α (exercise).

2.4. Statistical Inference

129

2.4.3 Confidence sets Let ϑ be a k-vector of unknown parameters related to the unknown popk ulation P ∈ P and C(X) ∈ BΘ ˜ depending only on the sample X, where k ˜ Θ ∈ B is the range of ϑ. If inf P (ϑ ∈ C(X)) ≥ 1 − α,

P ∈P

(2.33)

where α is a fixed constant in (0, 1), then C(X) is called a confidence set for ϑ with level of significance 1 − α. The left-hand side of (2.33) is called the confidence coefficient of C(X), which is the highest possible level of significance for C(X). A confidence set is a random element that covers the unknown ϑ with certain probability. If (2.33) holds, then the coverage probability of C(X) is at least 1−α, although C(x) either covers or does not cover ϑ whence we observe X = x. The concepts of level of significance and confidence coefficient are very similar to the level of significance and size in hypothesis testing. In fact, it is shown in Chapter 7 that some confidence sets are closely related to hypothesis tests. Consider a real-valued ϑ. If C(X) = [ϑ(X), ϑ(X)] for a pair of realvalued statistics ϑ and ϑ, then C(X) is called a confidence interval for ϑ. If C(X) = (−∞, ϑ(X)] (or [ϑ(X), ∞)), then ϑ (or ϑ) is called an upper (or a lower) confidence bound for ϑ. A confidence set (or interval) is also called a set (or an interval) estimator of ϑ, although it is very different from a point estimator (discussed in §2.4.1). Example 2.31. Consider Example 2.28. Suppose that a confidence inter¯ and ϑ(X), ¯ val for ϑ = µ is needed. Again, we only need to consider ϑ(X) ¯ since the sample mean X is sufficient. Consider confidence intervals of the ¯ − c, X ¯ + c], where c ∈ (0, ∞) is fixed. Note that form [X √ ¯ − c, X ¯ + c] = P |X ¯ − µ| ≤ c = 1 − 2Φ − nc/σ , P µ ∈ [X ¯ −c, X ¯ +c] which is independent of µ. Hence, the confidence coefficient of [X √ is 1 − 2Φ (− nc/σ), which is an increasing function of c and converges to 1 as c → ∞ or 0 as c → 0. Thus, confidence coefficients are positive but less ¯ X] ¯ and (−∞, ∞). We can than 1 except for silly confidence intervals [X, choose a confidence interval with an arbitrarily large confidence coefficient, but the chosen confidence interval may be so wide that it is practically useless. ¯ − c, X ¯ + c] has confidence coefficient 0 If σ2 is also unknown, then [X and, therefore, is not a good inference procedure. In such a case a different confidence interval for µ with positive confidence coefficient can be derived (Exercise 97 in §2.6).

130

2. Fundamentals of Statistics

This example tells us that a reasonable approach is to choose a level of significance 1 − α ∈ (0, 1) (just like the level of significance in hypothesis testing) and a confidence interval or set satisfying √ (2.33). In Example 2.31, when σ2 is known and c is chosen to be σz1−α/2 / n, where za = Φ−1 (a), ¯ − c, X ¯ + c] is exactly the confidence coefficient of the confidence interval [X 1 − α for any fixed α ∈ (0, 1). This is desirable since, for all confidence intervals satisfying (2.33), the one with the shortest interval length is preferred. For a general confidence interval [ϑ(X), ϑ(X)], its length is ϑ(X)−ϑ(X), which may be random. We may consider the expected (or average) length E[ϑ(X)−ϑ(X)]. The confidence coefficient and expected length are a pair of good measures of performance of confidence intervals. Like the two types of error probabilities of a test in hypothesis testing, however, we cannot maximize the confidence coefficient and minimize the length (or expected length) simultaneously. A common approach is to minimize the length (or expected length) subject to (2.33). For an unbounded confidence interval, its length is ∞. Hence we have to define some other measures of performance. For an upper (or a lower) confidence bound, we may consider the distance ϑ(X) − ϑ (or ϑ − ϑ(X)) or its expectation. To conclude this section, we discuss an example of a confidence set for a two-dimensional parameter. General discussions about how to construct and assess confidence sets are given in Chapter 7. Example 2.32. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with both µ ∈ R and σ 2 > 0 unknown. Let θ = (µ, σ2 ) and α ∈ (0, 1) be ¯ be the sample mean and S 2 be the sample variance. Since given. Let X 2 ¯ S ) is sufficient (Example 2.15), we focus on C(X) that is a function of (X, ¯ S 2 ). From Example 2.18, X ¯ and S 2 are independent and (n − 1)S 2 /σ 2 (X, √ ¯ has the chi-square distribution χ2n−1 . Since n(X − µ)/σ has the N (0, 1) distribution (Exercise 43 in §1.6), ¯ −µ √ X √ P −˜ cα ≤ ≤ c˜α = 1 − α, σ/ n √ where c˜α = Φ−1 1+ 21−α (verify). Since the chi-square distribution χ2n−1 is a known distribution, we can always find two constants c1α and c2α such that √ (n − 1)S 2 P c1α ≤ = 1 − α. ≤ c 2α 2 σ Then ¯ −µ X (n − 1)S 2 √ ≤ c˜α , c1α ≤ = 1 − α, ≤ c P −˜ cα ≤ 2α σ/ n σ2

131

4 0

2

variance

6

8

2.5. Asymptotic Criteria and Inference

-4

-2

0

2

4

mean

Figure 2.3: A confidence set for θ in Example 2.32 or P

¯ − µ)2 (n − 1)S 2 (n − 1)S 2 n(X ≤ σ2 , ≤ σ2 ≤ 2 c˜α c2α c1α

= 1 − α.

(2.34)

The left-hand side of (2.34) defines a set in the range of θ = (µ, σ2 ) bounded by two straight lines, σ 2 = (n − 1)S 2 /ciα , i = 1, 2, and a curve σ 2 = ¯ −µ)2 /˜ n(X c2α (see the shadowed part of Figure 2.3). This set is a confidence set for θ with confidence coefficient 1 − α, since (2.34) holds for any θ.

2.5 Asymptotic Criteria and Inference We have seen that in statistical decision theory and inference, a key to the success of finding a good decision rule or inference procedure is being able to find some moments and/or distributions of various statistics. Although many examples are presented (including those in the exercises in §2.6), there are more cases in which we are not able to find exactly the moments or distributions of given statistics, especially when the problem is not parametric (see, e.g., the discussions in Example 2.8). In practice, the sample size n is often large, which allows us to approximate the moments and distributions of statistics that are impossible

132

2. Fundamentals of Statistics

to derive, using the asymptotic tools discussed in §1.5. In an asymptotic analysis, we consider a sample X = (X1 , ..., Xn ) not for fixed n, but as a member of a sequence corresponding to n = n0 , n0 + 1, ..., and obtain the limit of the distribution of an appropriately normalized statistic or variable Tn (X) as n → ∞. The limiting distribution and its moments are used as approximations to the distribution and moments of Tn (X) in the situation with a large but actually finite n. This leads to some asymptotic statistical procedures and asymptotic criteria for assessing their performances, which are introduced in this section. The asymptotic approach is not only applied to the situation where no exact method is available, but also used to provide an inference procedure simpler (e.g., in terms of computation) than that produced by the exact approach (the approach considering a fixed n). Some examples are given in later chapters. In addition to providing more theoretical results and/or simpler inference procedures, the asymptotic approach requires less stringent mathematical assumptions than does the exact approach. The mathematical precision of the optimality results obtained in statistical decision theory, for example, tends to obscure the fact that these results are approximations in view of the approximate nature of the assumed models and loss functions. As the sample size increases, the statistical properties become less dependent on the loss functions and models. However, a major weakness of the asymptotic approach is that typically no good estimates for the precision of the approximations are available and, therefore, we cannot determine whether a particular n in a problem is large enough to safely apply the asymptotic results. To overcome this difficulty, asymptotic results are frequently used in combination with some numerical/empirical studies for selected values of n to examine the finite sample performance of asymptotic procedures.

2.5.1 Consistency A reasonable point estimator is expected to perform better, at least on the average, if more information about the unknown population is available. With a fixed model assumption and sampling plan, more data (larger sample size n) provide more information about the unknown population. Thus, it is distasteful to use a point estimator Tn which, if sampling were to continue indefinitely, could possibly have a nonzero estimation error, although the estimation error of Tn for a fixed n may never equal 0 (see the discussion in §2.4.1). Definition 2.10 (Consistency of point estimators). Let X = (X1 , ..., Xn ) be a sample from P ∈ P and Tn (X) be a point estimator of ϑ for every n. (i) Tn (X) is called consistent for ϑ if and only if Tn (X) →p ϑ w.r.t. any

2.5. Asymptotic Criteria and Inference

133

P ∈ P. (ii) Let {an } be a sequence of positive constants diverging to ∞. Tn (X) is called an -consistent for ϑ if and only if an [Tn (X) − ϑ] = Op (1) w.r.t. any P ∈ P. (iii) Tn (X) is called strongly consistent for ϑ if and only if Tn (X) →a.s. ϑ w.r.t. any P ∈ P. (iv) Tn (X) is called Lr -consistent for ϑ if and only if Tn (X) →Lr ϑ w.r.t. any P ∈ P for some fixed r > 0. Consistency is actually a concept relating to a sequence of estimators, {Tn , n = n0 , n0 + 1, ...}, but we usually just say “consistency of Tn ” for simplicity. Each of the four types of consistency in Definition 2.10 describes the convergence of Tn (X) to ϑ in some sense, as n → ∞. In statistics, consistency according to Definition 2.10(i), which is sometimes called weak consistency since it is implied by any of the other three types of consistency, is the most useful concept of convergence of Tn to ϑ. L2 -consistency is also called consistency in mse, which is the most useful type of Lr -consistency. Example 2.33. Let X1 , ..., Xn be i.i.d. from P ∈ P. If ϑ = µ, which is the mean of P and is assumed to be finite, then by the SLLN (Theorem ¯ is strongly consistent for µ and, therefore, is 1.13), the sample mean X also consistent for µ. If we further assume that the variance of P is finite, ¯ is consistent in mse and is √n-consistent. With the finite then by (2.20), X variance assumption, the sample variance S 2 is strongly consistent for the variance of P , according to the SLLN. Pn Consider estimators of the form Tn = i=1 cni Xi , where {cni } is a double array of constants. If P has a finite variance, by (2.24), Tn Pn Pthen n is consistent in mse if and only if i=1 cni → 1 and i=1 c2ni → 0. If we only assumePthe existence of the mean of P , then Tn with cni = ci /n satn isfying n−1 i=1 ci → 1 and supi |ci | < ∞ is strongly consistent (Theorem 1.13(ii)). One or a combination of the law of large numbers, the CLT, Slutsky’s theorem (Theorem 1.11), and the continuous mapping theorem (Theorems 1.10 and 1.12) are typically applied to establish consistency of point estimators. In particular, Theorem 1.10 implies that if Tn is (strongly) consistent for ϑ and g is a continuous function of ϑ, then g(Tn ) is (strongly) consistent ¯ 2 is strongly for g(ϑ). For example, in Example 2.33 the point estimator X ¯ 2 is √n-consistent under the assumption consistent for µ2 . To show that X that P has a finite variance σ 2 , we can use the identity √ √ ¯ 2 − µ2 ) = n(X ¯ − µ)(X ¯ + µ) n(X

¯ is √n-consistent for µ and X ¯ + µ = Op (1). (Note that and the fact that X

134

2. Fundamentals of Statistics

¯ 2 may not be consistent in mse since we do not assume that P has a finite X √ ¯2 2 fourth moment.) Alternatively, we can use the fact that n( √ X − µ ) →d 2 2 N (0, 4µ σ ) (by the CLT and Theorem 1.12) to show the n-consistency ¯ 2. of X The following example shows another way to establish consistency of some point estimators. Example 2.34. Let X1 , ..., Xn be i.i.d. from an unknown P with a continuous c.d.f. F satisfying F (θ) = 1 for some θ ∈ R and F (x) < 1 for any x < θ. Consider the largest order statistic X(n) . For any ǫ > 0, F (θ −ǫ) < 1 and n P (|X(n) − θ| ≥ ǫ) = P (X(n) ≤ θ − ǫ) = [F (θ − ǫ)] , which imply (according to Theorem 1.8(v)) X(n) →a.s. θ, i.e., X(n) is strongly consistent for θ. If we assume that F (i) (θ−), the ith-order lefthand derivative of F at θ, exists and vanishes for any i ≤ m and that F (m+1) (θ−) exists and is nonzero, where m is a nonnegative integer, then

(−1)m F (m+1) (θ−) (θ − X(n) )m+1 + o |θ − X(n) |m+1 a.s. (m + 1)! This result and the fact that P n[1 − F (X(n) )] ≥ s = (1 − s/n)n imply −1 that (θ − X(n) )m+1 = Op (n−1 ), i.e., X(n) is n(m+1) -consistent. If m = 0, then X(n) is n-consistent, which is the most common situation. If m = 1, √ −1 then X(n) is n-consistent. The limiting distribution of n(m+1) (X(n) − θ) can be derived as follows. Let (m+1)−1 (−1)m (m + 1)! hn (θ) = . nF (m+1) (θ−)

1 − F (X(n) ) =

For t ≤ 0, by Slutsky’s theorem, lim P

n→∞

X(n) − θ ≤ t = lim P n→∞ hn (θ)

θ − X(n) hn (θ)

m+1

m+1

≥ (−t)

= lim P n[1 − F (X(n) )] ≥ (−t)m+1 n→∞ n = lim 1 − (−t)m+1 /n n→∞

= e−(−t)

m+1

!

.

It can be seen from the previous examples that there are many consistent estimators. Like the admissibility in statistical decision theory, consistency is a very essential requirement in the sense that any inconsistent estimators should not be used, but a consistent estimator is not necessarily good. Thus, consistency should be used together with one or a few more criteria.

2.5. Asymptotic Criteria and Inference

135

We now discuss a situation in which finding a consistent estimator is crucial. Suppose that an estimator Tn of ϑ satisfies cn [Tn (X) − ϑ] →d σY,

(2.35)

where Y is a random variable with a known distribution, σ > 0 is an unknown parameter, of constants; for example, in √ ¯and {cn } is a sequence Example 2.33, n(X − µ) →d N (0, σ 2 ); in Example 2.34, (2.35) holds −1 −1 with cn = n(m+1) and σ = [(−1)m (m + 1)!/F (m+1) (θ−)](m+1) . If a consistent estimator σ ˆn of σ can be found, then, by Slutsky’s theorem, σn →d Y cn [Tn (X) − ϑ]/ˆ σn by and, thus, we may approximate the distribution of cn [Tn (X) − ϑ]/ˆ the known distribution of Y .

2.5.2 Asymptotic bias, variance, and mse Unbiasedness as a criterion for point estimators is discussed in §2.3.2 and §2.4.1. In some cases, however, there is no unbiased estimator (Exercise 84 in §2.6). Furthermore, having a “slight” bias in some cases may not be a bad idea (see Exercise 63 in §2.6). Let Tn (X) be a point estimator of ϑ for every n. If ETn exists for every n and limn→∞ E(Tn − ϑ) = 0 for any P ∈ P, then Tn is said to be approximately unbiased. There are many reasonable point estimators whose expectations are not well defined. For example, consider i.i.d. (X1 , Y1 ), ..., (Xn , Yn ) from a bivariate normal distribution with µx = EX1 and µy = EY1 6= 0. Let ¯ Y¯ , the ratio of two sample means. Then ETn is ϑ = µx /µy and Tn = X/ not defined for any n. It is then desirable to define a concept of asymptotic bias for point estimators whose expectations are not well defined. Definition 2.11. (i) Let ξ, ξ1 , ξ2 , ... be random variables and {an } be a sequence of positive numbers satisfying an → ∞ or an → a > 0. If an ξn →d ξ and E|ξ| < ∞, then Eξ/an is called an asymptotic expectation of ξn . (ii) Let Tn be a point estimator of ϑ for every n. An asymptotic expectation of Tn − ϑ, if it exists, is called an asymptotic bias of Tn and denoted by ˜bT (P ) (or ˜bT (θ) if P is in a parametric family). If limn→∞ ˜bT (P ) = 0 for n n n any P ∈ P, then Tn is said to be asymptotically unbiased. Like the consistency, the asymptotic expectation (or bias) is a concept relating to sequences {ξn } and {Eξ/an } (or {Tn } and {˜bTn (P )}). Note that the exact bias bTn (P ) is not necessarily the same as ˜bTn (P ) when both of them exist (Exercise 115 in §2.6). The following result shows that the asymptotic expectation defined in Definition 2.11 is essentially unique.

136

2. Fundamentals of Statistics

Proposition 2.3. Let {ξn } be a sequence of random variables. Suppose that both Eξ/an and Eη/bn are asymptotic expectations of ξn defined according to Definition 2.11(i). Then, one of the following three must hold: (a) Eξ = Eη = 0; (b) Eξ 6= 0, Eη = 0, and bn /an → 0; or Eξ = 0, Eη 6= 0, and an /bn → 0; (c) Eξ 6= 0, Eη 6= 0, and (Eξ/an )/(Eη/bn ) → 1. Proof. According to Definition 2.11(i), an ξn →d ξ and bn ξn →d η. (i) If both ξ and η have nondegenerate c.d.f.’s, then the result follows from Exercise 129 of §1.6. (ii) Suppose that ξ has a nondegenerate c.d.f. but η is a constant. If η 6= 0, then by Theorem 1.11(iii), an /bn → ξ/η, which is impossible since ξ has a nondegenerate c.d.f. If η = 0, then by Theorem 1.11(ii), bn /an → 0. (iii) Suppose that both ξ and η are constants. If ξ = η = 0, the result follows. If ξ 6= 0 and η = 0, then bn /an → 0. If ξ 6= 0 and η 6= 0, then bn /an → η/ξ. If Tn is a consistent estimator of ϑ, then Tn = ϑ + op (1) and, by Definition 2.11(ii), Tn is asymptotically unbiased, although Tn may not be approximately unbiased; in fact, g(Tn ) is asymptotically unbiased for g(ϑ) for ¯ Y¯ , Tn →a.s. µx /µy any continuous function g. For the example of Tn = X/ by the SLLN and Theorem 1.10. Hence Tn is asymptotically unbiased, although ETn may not be defined. In Example 2.34, X(n) has the asymptotic −1 bias ˜bX(n) (P ) = hn (θ)EY , which is of order n−(m+1) . ¯ 2 and ϑ = µ2 in When an (Tn − ϑ) →d Y with EY = 0 (e.g., Tn = X Example 2.33), a more precise order of the asymptotic bias of Tn may be obtained (for comparing different estimators in terms of their asymptotic biases). Suppose that there is a sequence of random variables {ηn } such that an ηn →d Y and a2n (Tn − ϑ − ηn ) →d W, (2.36) where Y and W are random variables with finite means, EY = 0 and ˜ EW 6= 0. Then we may define a−2 n to be the order of bTn (P ) or define EW/a2n to be the a−2 order asymptotic bias of T . However, ηn in (2.36) n n may not be unique. Some regularity conditions have to be imposed so that the order of asymptotic bias of Tn can be uniquely defined. In the following we focus on the case where X1 , ..., Xn are i.i.d. random k-vectors. Suppose that Tn has the following expansion: n n n 1 1X 1 XX , Tn − ϑ = φ(Xi ) + 2 ψ(Xi , Xj ) + op n i=1 n i=1 j=1 n

(2.37)

where φ and ψ are functions that may depend on P , Eφ(X1 ) = 0, E[φ(X1 )]2 < ∞, ψ(x, y) = ψ(y, x), Eψ(x, X1 ) = 0 for all x, E[ψ(Xi , Xj )]2 < ∞, i ≤ j, and Eψ(X1 , X1 ) 6= 0. From the result for V-statistics in §3.5.3 (Theorem

2.5. Asymptotic Criteria and Inference

3.16 and Exercise 113 in §3.6), n

137

n

1 XX ψ(Xi , Xj ) →d W, n i=1 j=1 where W is a random variable with P EW = Eψ(X1 , X1 ). Hence (2.36) √ n holds with an = n and ηn = n−1 i=1 φ(Xi ). Consequently, we can define Eψ(X1 , X1 )/n to be the n−1 order asymptotic bias of Tn . Examples of estimators that have expansion (2.37) are provided in §3.5.3 and §5.2.1. In the following we consider the special case of functions of sample means. ¯ Let PnX1 , ..., Xn be i.i.d. random k-vectors with finite Σ k= Var(X1 ), X = ¯ where g is a function on R that is secondn−1 i=1 Xi , and Tn = g(X), order differentiable at µ = EX1 ∈ Rk . Consider Tn as an estimator of ϑ = g(µ). Using Taylor’s expansion, we obtain expansion (2.37) with φ(x) = [∇g(µ)]τ (x − µ) and ψ(x, y) = (x − µ)τ ∇2 g(µ)(y − µ)/2, where ∇g is the kvector of partial derivatives of g and ∇2 g is the k ×k matrix of second-order partial derivatives of g. By the CLT and Theorem 1.10(iii), n

n

τ 2 1 XX n ¯ ¯ − µ) →d ZΣ ∇ g(µ)ZΣ , − µ)τ ∇2 g(µ)(X ψ(Xi , Xj ) = (X n i=1 j=1 2 2

where ZΣ = Nk (0, Σ). Thus,

tr ∇2 g(µ)Σ E[ZΣτ ∇2 g(µ)ZΣ ] = (2.38) 2n 2n ¯ where tr(A) denotes the is the n−1 order asymptotic bias of Tn = g(X), trace of the matrix A. Note that the quantity in (2.38) is the same as the ¯ obtained under a much more leading term in the exact bias of Tn = g(X) stringent condition on the derivatives of g (Lehmann, 1983, Theorem 2.5.1). Example 2.35. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = p, where p ∈ (0, 1) is unknown. Consider first the estimation ¯ = p(1−p)/n, the n−1 order asymptotic bias of of ϑ = p(1−p). Since Var(X) ¯ − X) ¯ according to (2.38) with g(x) = x(1 − x) is −p(1 − p)/n. On Tn = X(1 ¯ − X)] ¯ = EX ¯ − EX ¯2 = the other hand, a direct computation shows E[X(1 2 ¯ ¯ = p(1 − p) − p(1 − p)/n. Hence, the exact bias of Tn p − (E X) − Var(X) is the same as the n−1 order asymptotic bias. Consider next the estimation of ϑ = p−1 . In this case, there is no ¯ −1 . Then, an unbiased estimator of p−1 (Exercise 84 in §2.6). Let Tn = X −1 n order asymptotic bias of Tn according to (2.38) with g(x) = x−1 is (1 − p)/(p2 n). On the other hand, ETn = ∞ for every n. Like the bias, the mse of an estimator Tn of ϑ, mseTn (P ) = E(Tn − ϑ)2 , is not well defined if the second moment of Tn does not exist. We now

138

2. Fundamentals of Statistics

define a version of asymptotic mean squared error (amse) and a measure of assessing different point estimators of a common parameter. Definition 2.12. Let Tn be an estimator of ϑ for every n and {an } be a sequence of positive numbers satisfying an → ∞ or an → a > 0. Assume that an (Tn − ϑ) →d Y with 0 < EY 2 < ∞. (i) The asymptotic mean squared error of Tn , denoted by amseTn (P ) or amseTn (θ) if P is in a parametric family indexed by θ, is defined to be the asymptotic expectation of (Tn − ϑ)2 , i.e., amseTn (P ) = EY 2 /a2n . The asymptotic variance of Tn is defined to be σT2 n (P ) = Var(Y )/a2n . (ii) Let Tn′ be another estimator of ϑ. The asymptotic relative efficiency of Tn′ w.t.r. Tn is defined to be eTn′ ,Tn (P ) = amseTn (P )/amseTn′ (P ). (iii) Tn is said to be asymptotically more efficient than Tn′ if and only if lim supn eTn′ ,Tn (P ) ≤ 1 for any P and < 1 for some P . The amse and asymptotic variance are the same if and only if EY = 0. By Proposition 2.3, the amse or the asymptotic variance of Tn is essentially unique and, therefore, the concept of asymptotic relative efficiency in Definition 2.12(ii)-(iii) is well defined. 2 2 2 In Example 2.33, amseX¯ 2 (P ) = σX ¯ 2 (P ) = 4µ σ /n. In Example 2.34, 2 2 σX(n) (P ) = [hn (θ)] Var(Y ) and amseX(n) (P ) = [hn (θ)]2 EY 2 . When both mseTn (P ) and mseTn′ (P ) exist, one may compare Tn and Tn′ by evaluating the relative efficiency mseTn (P )/mseTn′ (P ). However, this comparison may be different from the one using the asymptotic relative efficiency in Definition 2.12(ii), since the mse and amse of an estimator may be different (Exercise 115 in §2.6). The following result shows that when the exact mse of Tn exists, it is no smaller than the amse of Tn . It also provides a condition under which the exact mse and the amse are the same. Proposition 2.4. Let Tn be an estimator of ϑ for every n and {an } be a sequence of positive numbers satisfying an → ∞ or an → a > 0. Suppose that an (Tn − ϑ) →d Y with 0 < EY 2 < ∞. Then (i) EY 2 ≤ lim inf n E[a2n (Tn − ϑ)2 ] and (ii) EY 2 = limn→∞ E[a2n (Tn − ϑ)2 ] if and only if {a2n (Tn − ϑ)2 } is uniformly integrable. Proof. (i) By Theorem 1.10(iii), min{a2n (Tn − ϑ)2 , t} →d min{Y 2 , t} for any t > 0. Since min{a2n (Tn − ϑ)2 , t} is bounded by t, lim E(min{a2n (Tn − ϑ)2 , t}) = E(min{Y 2 , t})

n→∞

139

2.5. Asymptotic Criteria and Inference

(Theorem 1.8(viii)). Then EY 2 = lim E(min{Y 2 , t}) t→∞

= lim lim E(min{a2n (Tn − ϑ)2 , t}) t→∞ n→∞

= lim inf E(min{a2n (Tn − ϑ)2 , t}) t,n

≤ lim inf E[a2n (Tn − ϑ)2 ], n

where the third equality follows from the fact that E(min{a2n (Tn − ϑ)2 , t}) is nondecreasing in t for any fixed n. (ii) The result follows from Theorem 1.8(viii). Example 2.36. Let X1 , ..., Xn be i.i.d. from the Poisson distribution P (θ) with an unknown θ > 0. Consider the estimation of ϑ = P (Xi = 0) = e−θ . Let T1n = Fn (0), where Fn is the empirical c.d.f. defined √ in (2.28). Then T1n is unbiased and has mseT1n (θ) = e−θ (1−e−θ )/n. Also, n(T1n −ϑ) →d N (0, e−θ (1 − e−θ )) by the CLT. Thus, in this case amseT1n (θ) = mseT1n (θ). ¯

−1/n

−1) Next, consider T2n = e−X . Note that ET2n = enθ(e . Hence −θ nb (θ) → θe /2. Using Theorem 1.12 and the CLT, we can show that T 2n √ n(T2n − ϑ) →d N (0, e−2θ θ). By Definition 2.12(i), amseT2n (θ) = e−2θ θ/n. Thus, the asymptotic relative efficiency of T1n w.r.t. T2n is

eT1n ,T2n (θ) = θ/(eθ − 1), which is always less than 1. This shows that T2n is asymptotically more efficient than T1n . ¯ of The result for T2n in Example 2.36 is a special case (with Un = X) the following general result. Theorem 2.6. Let g be a function on Rk that is differentiable at θ ∈ Rk and let Un be a k-vector of statistics satisfying an (Un − θ) →d Y for a random k-vector Y with 0 < EkY k2 < ∞ and a sequence of positive numbers {an } satisfying an → ∞. Let Tn = g(Un ) be an estimator of ϑ = g(θ). Then, the amse and asymptotic variance of Tn are, respectively, E{[∇g(θ)]τ Y }2 /a2n and [∇g(θ)]τ Var(Y )∇g(θ)/a2n .

2.5.3 Asymptotic inference Statistical inference based on asymptotic criteria and approximations is called asymptotic statistical inference or simply asymptotic inference. We have previously considered asymptotic estimation. We now focus on asymptotic hypothesis tests and confidence sets.

140

2. Fundamentals of Statistics

Definition 2.13. Let X = (X1 , ..., Xn ) be a sample from P ∈ P and Tn (X) be a test for H0 : P ∈ P0 versus H1 : P ∈ P1 . (i) If lim supn αTn (P ) ≤ α for any P ∈ P0 , then α is an asymptotic significance level of Tn . (ii) If limn→∞ supP ∈P0 αTn (P ) exists, then it is called the limiting size of Tn . (iii) Tn is called consistent if and only if the type II error probability converges to 0, i.e., limn→∞ [1 − αTn (P )] = 0, for any P ∈ P1 . (iv) Tn is called Chernoff-consistent if and only if Tn is consistent and the type I error probability converges to 0, i.e., limn→∞ αTn (P ) = 0, for any P ∈ P0 . Tn is called strongly Chernoff-consistent if and only if Tn is consistent and the limiting size of Tn is 0. Obviously if Tn has size (or significance level) α for all n, then its limiting size (or asymptotic significance level) is α. If the limiting size of Tn is α ∈ (0, 1), then for any ǫ > 0, Tn has size α + ǫ for all n ≥ n0 , where n0 is independent of P . Hence Tn has level of significance α + ǫ for any n ≥ n0 . However, if P0 is not a parametric family, it is likely that the limiting size of Tn is 1 (see, e.g., Example 2.37). This is the reason why we consider the weaker requirement in Definition 2.13(i). If Tn has asymptotic significance level α, then for any ǫ > 0, αTn (P ) < α + ǫ for all n ≥ n0 (P ) but n0 (P ) depends on P ∈ P0 ; and there is no guarantee that Tn has significance level α + ǫ for any n. The consistency in Definition 2.13(iii) only requires that the type II error probability converge to 0. We may define uniform consistency to be limn→∞ supP ∈P1 [1 − αTn (P )] = 0, but it is not satisfied in most problems. If α ∈ (0, 1) is a pre-assigned level of significance for the problem, then a consistent test Tn having asymptotic significance level α is called asymptotically correct, and a consistent test having limiting size α is called strongly asymptotically correct. The Chernoff-consistency (or strong Chernoff-consistency) in Definition 2.13(iv) requires that both types of error probabilities converge to 0. Mathematically, Chernoff-consistency (or strong Chernoff-consistency) is better than asymptotic correctness (or strongly asymptotic correctness). After all, both types of error probabilities should decrease to 0 if sampling can be continued indefinitely. However, if α is chosen to be small enough so that error probabilities smaller than α can be practically treated as 0, then the asymptotic correctness (or strongly asymptotic correctness) is enough, and is probably preferred, since requiring an unnecessarily small type I error probability usually results in an unnecessary increase in the type II error probability, as the following example illustrates. Example 2.37. Consider the testing problem H0 : µ ≤ µ0 versus H1 :

2.5. Asymptotic Criteria and Inference

141

µ > µ0 based on i.i.d. X1 , ..., Xn with EX1 = µ ∈ R. If each Xi has the 2 N (µ, σ 2 ) distribution with √ a known σ , then the test Tcα given in Example 2.28 with cα = σz1−α / n + µ0 and α ∈ (0, 1) has size α (and, therefore, limiting size α). It also follows from (2.32) that for any µ > µ0 , √ n(µ0 − µ) →0 (2.39) 1 − αTcα (µ) = Φ z1−α + σ as n → ∞. This shows that Tcα is consistent and, hence, is strongly asymptotically correct. Note that the convergence in (2.39) is not uniform in µ > µ0 , but is uniform in µ > µ1 for any fixed µ1 > µ0 . Since the size of Tcα is α for all n, Tcα is not Chernoff-consistent. A strongly Chernoff-consistent test can be obtained as follows. Let √ αn = 1 − Φ( nan ), (2.40) √ where an ’s are positive numbers satisfying an → 0 and nan → ∞. Let Tn be Tcα with α = αn for each n. Then, Tn has size αn . Since αn → 0, The limiting size of Tn is 0. On the other hand, (2.39) still holds with α replaced by αn . This follows from the fact that √ n(µ0 − µ) √ µ0 − µ = n an + → −∞ z1−αn + σ σ for any µ > µ0 . Hence Tn is strongly Chernoff-consistent. However, if αn < α, then, from the left-hand side of (2.39), 1 − αTcα (µ) < 1 − αTn (µ) for any µ > µ0 . We now consider the case where the population P is not in a parametric family. We still assume that σ 2 = Var(Xi ) is known. Using the CLT, we can show that for µ > µ0 , √ n(µ0 − µ) = 0, lim [1 − αTcα (µ)] = lim Φ z1−α + n→∞ n→∞ σ i.e., Tcα is still consistent. For µ ≤ µ0 ,

√ n(µ0 − µ) , lim αTcα (µ) = 1 − lim Φ z1−α + n→∞ n→∞ σ

which equals α if µ = µ0 and 0 if µ < µ0 . Thus, the asymptotic significance level of Tcα is α. Combining these two results, we know that Tcα is asymptotically correct. However, if P contains all possible populations on R with finite second moments, then one can show that the limiting size of Tcα is 1 (exercise). For αn defined by (2.40), we can show that Tn = Tcα with α = αn is Chernoff-consistent (exercise). But Tn is not strongly Chernoffconsistent if P contains all possible populations on R with finite second moments.

142

2. Fundamentals of Statistics

Definition 2.14. Let X = (X1 , ..., Xn ) be a sample from P ∈ P, ϑ be a k-vector of parameters related to P , and C(X) be a confidence set for ϑ. (i) If lim inf n P (ϑ ∈ C(X)) ≥ 1 − α for any P ∈ P, then 1 − α is an asymptotic significance level of C(X). (ii) If limn→∞ inf P ∈P P (ϑ ∈ C(X)) exists, then it is called the limiting confidence coefficient of C(X). Note that the asymptotic significance level and limiting confidence coefficient of a confidence set are very similar to the asymptotic significance level and limiting size of a test, respectively. Some conclusions are also similar. For example, in a parametric problem one can often find a confidence set having limiting confidence coefficient 1 − α ∈ (0, 1), which implies that for any ǫ > 0, the confidence coefficient of C(X) is 1 − α − ǫ for all n ≥ n0 , where n0 is independent of P ; in a nonparametric problem the limiting confidence coefficient of C(X) might be 0, whereas C(X) may have asymptotic significance level 1 − α ∈ (0, 1), but for any fixed n, the confidence coefficient of C(X) might be 0. √ The confidence interval in Example 2.31 with c = σz1−α/2 / n and the confidence set in Example 2.32 have confidence coefficient 1 − α for any n and, therefore, have limiting confidence coefficient 1 − α. If we drop the normality assumption and assume EXi4 < ∞, then these confidence sets have asymptotic significance level 1−α; their limiting confidence coefficients may be 0 (exercise).

2.6 Exercises 1. Consider Example 2.3. Suppose that p(s) is constant. Show that Xi and Xj , i 6= j, are not uncorrelated and, hence, X1 , ..., Xn are not independent. Furthermore, when yi ’s are either 0 or 1, show that Pn Z = X i has a hypergeometric distribution and compute the i=1 mean of Z. 2. Consider Example 2.3. Suppose that we do not require that the elements in s be distinct, i.e., we consider sampling with replacement. Define a probability measure p and a sample (X1 , ..., Xn ) such that (2.3) holds. If p(s) is constant, are X1 , ..., Xn independent? If p(s) is constant and Pnyi ’s are either 0 or 1, what are the distribution and mean of Z = i=1 Xi ?

3. Show that {Pθ : θ ∈ Θ} is an exponential family and find its canonical form and natural parameter space, when (a) Pθ is the Poisson distribution P (θ), θ ∈ Θ = (0, ∞); (b) Pθ is the negative binomial distribution N B(θ, r) with a fixed r,

2.6. Exercises

143

θ ∈ Θ = (0, 1); (c) Pθ is the exponential distribution E(a, θ) with a fixed a, θ ∈ Θ = (0, ∞); (d) Pθ is the gamma distribution Γ(α, γ), θ = (α, γ) ∈ Θ = (0, ∞) × (0, ∞); (e) Pθ is the beta distribution B(α, β), θ = (α, β) ∈ Θ = (0, 1)×(0, 1); (f) Pθ is the Weibull distribution W (α, θ) with a fixed α > 0, θ ∈ Θ = (0, ∞). 4. Show that the family of exponential distributions E(a, θ) with two unknown parameters a and θ is not an exponential family. 5. Show that the family of negative binomial distributions N B(p, r) with two unknown parameters p and r is not an exponential family. 6. Show that the family of Cauchy distributions C(µ, σ) with two unknown parameters µ and σ is not an exponential family. 7. Show that the family of Weibull distributions W (α, θ) with two unknown parameters α and θ is not an exponential family. 8. Is the family of log-normal distributions LN (µ, σ 2 ) with two unknown parameters µ and σ 2 an exponential family? 9. Show that the family of double exponential distributions DE(µ, θ) with two unknown parameters µ and θ is not an exponential family, but the family of double exponential distributions DE(µ, θ) with a fixed µ and an unknown parameter θ is an exponential family. 10. Show that the k-dimensional normal family discussed in Example 2.4 is an exponential family. Identify the functions T , η, ξ, and h. 11. Obtain the variance-covariance matrix for (X1 , ..., Xk ) in Example 2.7, using (a) Theorem 2.1(ii) and (b) direct computation. 12. Show that the m.g.f. of the gamma distribution Γ(α, γ) is (1 − γt)−α , t < γ −1 , using Theorem 2.1(ii). 13. A discrete random variable X with P (X = x) = γ(x)θx /c(θ), x = 0, 1, 2, ..., P∞ where γ(x) ≥ 0, θ > 0, and c(θ) = x=0 γ(x)θx , is called a random variable with a power series distribution. (a) Show that {γ(x)θx /c(θ) : θ > 0} is an exponential family. (b) Suppose that X1 , ..., XP n are i.i.d. with a power series distribution n γ(x)θx /c(θ). Show that i=1 Xi has the power series distribution γn (x)θx /[c(θ)]n , where γn (x) is the coefficient of θx in the power series expansion of [c(θ)]n .

144

2. Fundamentals of Statistics

14. Let X be a random variable with a p.d.f. fθ in an exponential family {Pθ : θ ∈ Θ} and let A be a Borel set. Show that the distribution of X truncated on A (i.e., the conditional distribution of X given X ∈ A) has a p.d.f. fθ IA /Pθ (A) that is in an exponential family. 15. Let {P(µ,Σ) : µ ∈ Rk , Σ ∈ Mk } be a location-scale family on Rk . Suppose that P(0,Ik ) has a Lebesgue p.d.f. that is always positive and that the mean and variance-covariance matrix of P(0,Ik ) are 0 and Ik , respectively. Show that the mean and variance-covariance matrix of P(µ,Σ) are µ and Σ, respectively. 16. Show that if the distribution of a positive random variable X is in a scale family, then the distribution of log X is in a location family. 17. Let X be a random variable having the gamma distribution Γ(α, γ) with a known α and an unknown γ > 0 and let Y = σ log X. (a) Show that if σ > 0 is unknown, then the distribution of Y is in a location-scale family. (b) Show that if σ > 0 is known, then the distribution of Y is in an exponential family. 18. Let X1 , ..., Xn be i.i.d. random variables having a finite E|X1 |4 and ¯ and S 2 be the sample mean and variance defined by (2.1) and let X ¯ 3 ), Cov(X, ¯ S 2 ), and Var(S 2 ) in terms of µk = (2.2). Express E(X k ¯ and S 2 are EX1 , k = 1, 2, 3, 4. Find a condition under which X uncorrelated. 19. Let X1 , ..., Xn be i.i.d. random variables having the gamma distribution Γ(α, γx ) and Y1 , ..., Yn be i.i.d. random variables having the gamma distribution Γ(α, γy ), where α > 0, γx > 0, and γy > 0. Assume that Xi ’s and Yi ’s are independent. Derive the distribution of ¯ Y¯ , where X ¯ and Y¯ are the sample means based on the statistic X/ Xi ’s and Yi ’s, respectively. 20. Let X1 , ..., Xn be i.i.d. random variables having the exponential distribution E(a, θ), a ∈ R, and θ > 0. Show that the smallest order statistic, X(1) , has the exponential distribution E(a, θ/n) and that Pn 2 i=1 (Xi − X(1) )/θ has the chi-square distribution χ22n−2 .

21. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors. Suppose that X1 has the Cauchy distribution C(0, 1) and given X1 = x, Y1 has ¯ and Y¯ be the Cauchy distribution C(βx, 1), where β ∈ R. Let X the sample means based on Xi ’s and Yi ’s, respectively. Obtain the ¯ and Y¯ /X. ¯ marginal distributions of Y¯ , Y¯ − β X,

145

2.6. Exercises

22. Let Xi = (Yi , Zi ), i = 1, ..., n, be i.i.d. random 2-vectors. The sample correlation coefficient is defined to be T (X) =

n X 1 ¯ p (Yi − Y¯ )(Zi − Z), (n − 1) SY2 SZ2 i=1

P Pn Pn −1 2 −1 ¯ ¯ 2 where Y¯ = n−1 ni=1 i=1 (Yi−Y ) , PYni , Z = n ¯ 2 i=1 Zi , SY = (n−1) 2 −1 and SZ = (n−1) i=1 (Zi − Z) . (a) Assume that E|Yi |4 < ∞ and E|Zi |4 < ∞. Show that √ n[T (X) − ρ] →d N (0, c2 ), where ρ is the correlation coefficient between Y1 and Z1 and c is a constant depending on some unknown parameters. (b) Assume that Yi and Zi are independently distributed as N (µ1 , σ12 ) and N (µ2 , σ22 ), respectively. Show that T has the Lebesgue p.d.f. Γ n−1 2 (1 − t2 )(n−4)/2 I(−1,1) (t). f (t) = √ πΓ n−2 2 (c) Assume the conditions in (b). Obtain the result in (a) using Scheff´e’s theorem (Proposition 1.18).

23. Let X1 , ..., X√ with EX14 < ∞, P T = (Y, Z), n be i.i.d. random variables Pn n −1 −1 2 and T1 = Y / Z, where Y = n |X | and Z = n i i=1 i=1 Xi . √ √ (a) Show that n(T − θ) →d N2 (0, Σ) and n(T1 − ϑ) →d N (0, c2 ). Identify θ, Σ, ϑ, and c2 in terms of moments of X1 . (b) Repeat (a) when X1 has the normal distribution N (0, σ 2 ). (c) Repeat (a) when X1 has the double exponential distribution D(0, σ). 24. Prove the claims in Example 2.9 for the distributions related to order statistics. 25. Show that if T is a sufficient statistic and T = ψ(S), where ψ is measurable and S is another statistic, then S is sufficient. 26. In the proof of Lemma 2.1, show that C0 ∈ C. Also, prove Lemma 2.1 when P is dominated by a σ-finite measure. 27. Let X1 , ..., Xn be i.i.d. random variables from Pθ ∈ {Pθ : θ ∈ Θ}. In the following cases, find a sufficient statistic for θ ∈ Θ that has the same dimension as θ. (a) Pθ is the Poisson distribution P (θ), θ ∈ (0, ∞). (b) Pθ is the negative binomial distribution N B(θ, r) with a known r, θ ∈ (0, 1).

146

2. Fundamentals of Statistics

(c) Pθ is the exponential distribution E(0, θ), θ ∈ (0, ∞). (d) Pθ is the gamma distribution Γ(α, γ), θ = (α, γ) ∈ (0, ∞)×(0, ∞). (e) Pθ is the beta distribution B(α, β), θ = (α, β) ∈ (0, 1) × (0, 1). (f) Pθ is the log-normal distribution LN (µ, σ 2 ), θ = (µ, σ 2 ) ∈ R × (0, ∞). (g) Pθ is the Weibull distribution W (α, θ) with a known α > 0, θ ∈ (0, ∞). 28. Let X1 , ..., Xn be i.i.d. random variables from P(a,θ) , where (a, θ) ∈ R2 is a parameter. Find a two-dimensional sufficient statistic for (a, θ) in the following cases. (a) P(a,θ) is the exponential distribution E(a, θ), a ∈ R, θ ∈ (0, ∞). (b) P(a,θ) is the Pareto distribution P a(a, θ), a ∈ (0, ∞), θ ∈ (0, ∞). 29. In Example 2.11, show that X(1) (or X(n) ) is sufficient for a (or b) if we consider a subfamily {f(a,b) : a < b} with a fixed b (or a). 30. Let X and Y be two random variables such that Y has the binomial distribution Bi(π, N ) and, given Y = y, X has the binomial distribution Bi(p, y). (a) Suppose that p ∈ (0, 1) and π ∈ (0, 1) are unknown and N is known. Show that (X, Y ) is minimal sufficient for (p, π). (b) Suppose that π and N are known and p ∈ (0, 1) is unknown. Show whether X is sufficient for p and whether Y is sufficient for p. 31. Let X1 , ..., Xn be i.i.d. random variables having a distribution P ∈ P, where P is the family of distributions on R having continuous c.d.f.’s. Let T = (X(1) , ..., X(n) ) be the vector of order statistics. Show that, given T , the conditional distribution of X = (X1 , ..., Xn ) is a discrete distribution putting probability 1/n! on each of the n! points (Xi1 , ..., Xin ) ∈ Rn , where {i1 , ..., in } is a permutation of {1, ..., n}; hence, T is sufficient for P ∈ P. 32. In Example 2.13 and Example 2.14, show that T is minimal sufficient for θ by using Theorem 2.3(iii). 33. A coin has probability p of coming up heads and 1 − p of coming up tails, where p ∈ (0, 1). The first stage of an experiment consists of tossing this coin a known total of M times and recording X, the number of heads. In the second stage, the coin is tossed until a total of X + 1 tails have come up. The number Y of heads observed in the second stage along the way to getting the X + 1 tails is then recorded. This experiment is repeated independently a total of n times and the two counts (Xi , Yi ) for the ith experiment are recorded, i = 1, ..., n. Obtain a statistic that is minimal sufficient for p and derive its distribution.

2.6. Exercises

147

34. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. n o 4 − ξ(θ) , fθ (x) = exp − x−µ σ

where θ = (µ, σ) ∈ Θ = R × (0, ∞). Show that P = {Pθ : θ ∈ Θ} is an exponential family, where P Pθ nis the joint Pn distribution Pn of3 X P1 ,n..., X4n, 2 and that the statistic T = i=1 Xi , i=1 Xi , i=1 Xi , i=1 Xi is minimal sufficient for θ ∈ Θ.

35. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. fθ (x) = (2θ)−1 I(0,θ) (x) + I(2θ,3θ) (x) . Find a minimal sufficient statistic for θ ∈ (0, ∞).

36. Let X1 , ..., Xn be i.i.d. random variables having the Cauchy distribution C(µ, σ) with unknown µ ∈ R and σ > 0. Show that the vector of order statistics is minimal sufficient for (µ, σ). 37. Let X1 , ..., Xn be i.i.d. random variables having the double exponential distribution DE(µ, θ) with unknown µ ∈ R and θ > 0. Show that the vector of order statistics is minimal sufficient for (µ, θ). 38. Let X1 , ..., Xn be i.i.d. random variables having the Weibull distribution W (α, θ) with unknown α > 0 and θ > 0. Show that the vector of order statistics is minimal sufficient for (α, θ). 39. Let X1 , ..., Xn be i.i.d. random variables having the beta distribution B(β, β) with an unknown β > 0. Find a minimal sufficient statistic for β. 40. Let X1 , ..., Xn be i.i.d. random variables having a population P in a parametric family indexed by (θ, j), where θ > 0, j = 1, 2, and n ≥ 2. When j = 1, P is the N (0, θ2 ) distribution. When j = 2, PPis the double Pn exponential distribution DE(0, θ). Show that T = n ( i=1 Xi2 , i=1 |Xi |) is minimal sufficient for (θ, j).

41. Let X1 , ..., Xn be i.i.d. random variables having a population P in a parametric family indexed by (θ, j), where θ ∈ (0, 1), j = 1, 2, and n ≥ 2. When j = 1, P is the Poisson distribution P (θ). When j = 2, P is the binomial distribution Bi(θ, 1). Pn (a) Show that T = i=1 Xi is not sufficient for (θ, j). (b) Find a two-dimensional minimal sufficient statistic for (θ, j).

42. Let X be a sample from P ∈ P = {fθ,j : θ ∈ Θ, j = 1, ..., k}, where fθ,j ’s are p.d.f.’s w.r.t. a common σ-finite measure and Θ is a set of parameters. Assume that {x : fθ,j (x) > 0} ⊂ {x : fθ,k (x) > 0} for all

148

2. Fundamentals of Statistics

θ and j = 1, ..., k − 1. Suppose that for each fixed j, T = T (X) is a statistic sufficient for θ. (a) Obtain a k-dimensional statistic that is sufficient for (θ, j). (b) Derive a sufficient condition under which T is minimal sufficient for (θ, j). 43. A box has an unknown odd number of balls labeled consecutively as −θ, −(θ − 1), ..., −2, −1, 0, 1, 2, ..., (θ − 1), θ, where θ is an unknown nonnegative integer. A simple random sample X1 , ..., Xn is taken without replacement, where Xi is the label on the ith ball selected and n < 2θ + 1. (a) Find a statistic that is minimal sufficient for θ and derive its distribution. (b) Show that the minimal sufficient statistic in (a) is also complete. 44. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. θ−1 e−(x−θ)/θ I(θ,∞) (x), where θ > 0 is an unknown parameter. (a) Find a statistic that is minimal sufficient for θ. (b) Show whether the minimal sufficient statistic in (a) is complete. 45. Let X1 , ..., Xn (n ≥ 2) be i.i.d. random variables having the normal distribution N (θ, 2) when θ = 0 and the normal distribution N (θ, 1) ¯ is a complete when θ ∈ R and θ 6= 0. Show that the sample mean X statistic for θ but it is not a sufficient statistic for θ. 46. Let X be a random variable with a distribution Pθ in {Pθ : θ ∈ Θ}, fθ be the p.d.f. of Pθ w.r.t. a measure ν, A be an event, and PA = {fθ IA /Pθ (A) : θ ∈ Θ}. (a) Show that if T (X) is sufficient for Pθ ∈ P, then it is sufficient for Pθ ∈ PA . (b) Show that if T is sufficient and complete for Pθ ∈ P, then it is complete for Pθ ∈ PA . 47. Show that (X(1) , X(n) ) in Example 2.13 is not complete. 48. Let T be a complete (or boundedly complete) and sufficient statistic. Suppose that there is a minimal sufficient statistic S. Show that T is minimal sufficient and S is complete (or boundedly complete). 49. Let T and S be two statistics such that S = ψ(T ) for a measurable ψ. Show that (a) if T is complete, then S is complete; (b) if T is complete and sufficient and ψ is one-to-one, then S is complete and sufficient; (c) the results in (a) and (b) still hold if the completeness is replaced by the bounded completeness.

149

2.6. Exercises

50. Find complete and sufficient statistics for the families in Exercises 27 and 28. 51. Show that (X(1) , X(n) ) in Example 2.11 is complete. 52. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors having the following Lebesgue p.d.f. p fθ (x, y) = (2πγ 2 )−1 I(0,γ) (x − a)2 + (y − b)2 , (x, y) ∈ R2 , where θ = (a, b, γ) ∈ R2 × (0, ∞). (a) If a = 0 and b = 0, find a complete and sufficient statistic for γ. (b) If all parameters are unknown, show that the convex hull of the sample points is a sufficient statistic for θ.

53. Let X be a discrete random variable with θ fθ (x) = (1 − θ)2 θx−1 0

p.d.f. x=0 x = 1, 2, ... otherwise,

where θ ∈ (0, 1). Show that X is boundedly complete, but not complete.

54. Show that the sufficient statistic T in Example 2.10 is also complete without using Proposition 2.1. 55. Let Y1 , ..., Yn be i.i.d. random variables having the Lebesgue p.d.f. λxλ−1 I(0,1) (x) with an unknown λ > 0 and let Z1 , ..., Zn be i.i.d. discrete random variables having the power series distribution given in Exercise 13 with an unknown θ > 0. Assume that Yi ’s and Zj ’s are independent. Let Xi = Yi + Zi , i = 1, ..., n. Find a complete and sufficient statistic for the unknown parameter (θ, λ) based on the sample X = (X1 , ..., Xn ). 56. Suppose that (X1 , Y1 ), ..., (Xn , Yn ) are i.i.d. random 2-vectors and 2 Xi and Yi are independently distributed as N (µ, σX ) and N (µ, σY2 ), 2 2 ¯ and respectively, with θ = (µ, σX , σY ) ∈ R × (0, ∞) × (0, ∞). Let X 2 SX be the sample mean and variance given by (2.1) and (2.2) for Xi ’s and Y¯ and SY2 be the sample mean and variance for Yi ’s. Show that ¯ Y¯ , S 2 , S 2 ) is minimal sufficient for θ but T is not boundedly T = (X, X Y complete. 57. Let X1 , ..., Xn be i.i.d. from the N (θ, θ2 ) distribution, where θ > 0 is a parameter. Find a minimal sufficient statistic for θ and show whether it is complete.

150

2. Fundamentals of Statistics

58. Suppose that (X1 , Y1 ), ..., (Xn , Yn ) are i.i.d. random 2-vectors having the normal distribution with EX1 = EY1 = 0, Var(X1 ) = Var(Y1 ) = 1, and Cov(X1 , Y1 ) = θ ∈ (−1, 1). (a) Find a minimal sufficient statistic for θ. (b) Show whether the minimal sufficient statistic in (a) is complete or not. Pn Pn (c) Prove that T1 = i=1 Xi2 and T2 = i=1 Yi2 are both ancillary but (T1 , T2 ) is not ancillary. 59. Let X1 , ..., Xn be i.i.d. random variables having the exponential distribution E(a, θ). Pn (a) Show that i=1 (Xi − X(1) ) and X(1) are independent for any (a, θ). (b) Show that Zi = (X(n)P− X(i) )/(X(n) − X(n−1) ), i = 1, ..., n − 2, n are independent of (X(1) , i=1 (Xi − X(1) )).

60. Let X1 , ..., Xn be i.i.d. random having the gamma distriP variablesP bution Γ(α, γ). Show that ni=1 Xi and ni=1 [log Xi − log X(1) ] are independent for any (α, γ).

61. Let X1 , ..., Xn be i.i.d. random variables having the uniform distribution on the interval (a, b), where −∞ < a < b < ∞. Show that (X(i) − X(1) )/(X(n) − X(1) ), i = 2, ..., n − 1, are independent of (X(1) , X(n) ) for any a and b. 62. Consider Example 2.19. Assume that n > 2. ¯ is better than T1 if P = N (θ, σ 2 ), θ ∈ R, σ > 0. (a) Show that X ¯ if P is the uniform distribution on (b) Show that T1 is better than X 1 1 the interval (θ − 2 , θ + 2 ), θ ∈ R. ¯ nor T1 is better than the (c) Find a family P for which neither X other. 63. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution, where µ ∈ R and σ > 0. Consider the estimation of σ 2 with the squared error loss. 2 2 Show that n−1 n S is better than S , the sample variance. Can you 2 find an estimator of the form cS with a nonrandom c such that it is 2 better than n−1 n S ? 64. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = θ ∈ (0, 1). Consider estimating θ with the squared error loss. Calculate the risks of the following estimators: ¯ (the sample mean) and (a) the nonrandomized estimators X if more than half of Xi ’s are 0 0 T0 (X) = 1 if more than half of Xi ’s are 1 1 if exactly half of Xi ’s are 0; 2

151

2.6. Exercises

(b) the randomized estimators ¯ X T1 (X) = T0 and T2 (X) =

¯ X 1 2

with probability with probability

1 2 1 2

¯ with probability X ¯ with probability 1 − X.

65. Let X1 , ..., Xn be i.i.d. random variables having the exponential distribution E(0, θ), θ ∈ (0, ∞). Consider estimating θ with the squared ¯ and cX(1) , where error loss. Calculate the risks of the sample mean X ¯ better than cX(1) for some c? c is a positive constant. Is X 66. Consider the estimation of an unknown parameter θ ≥ 0 under the squared error loss. Show that if T and U are two estimators such that T ≤ U and RT (P ) < RU (P ), then RT+ (P ) < RU+ (P ), where RT (P ) is the risk of an estimator T and T+ denotes the positive part of T . 67. Let X1 , ..., Xn be i.i.d. random variables having the exponential distribution E(0, θ), θ ∈ (0, ∞). Consider the hypotheses H0 : θ ≤ θ 0

versus H1 : θ > θ0 ,

where θ0 > 0 is a fixed constant. Obtain the risk function (in terms ¯ under the 0-1 loss. of θ) of the test rule Tc (X) = I(c,∞) (X), 68. Let X1 , ..., Xn be i.i.d. random variables having the Cauchy distribution C(µ, σ) with unknown µ ∈ R and σ > 0. Consider the hypotheses H0 : µ ≤ µ0

versus H1 : µ > µ0 ,

where µ0 is a fixed constant. Obtain the risk function of the test rule ¯ under the 0-1 loss. Tc (X) = I(c,∞) (X), 69. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = θ, where θ ∈ (0, 1) is unknown and n is an even integer. Consider the problem of testing H0 : θ ≤ 0.5 versus H1 : θ > 0.5 with action space {0, 1} (0 means H0 is accepted and 1 means H1 is accepted). Let the loss function be L(θ, a) = 0 if Hj is true and a = j, j = 0, 1; L(θ, 0) = C0 when θ > 0.5; and L(θ, 1) = C1 when θ ≤ 0.5, where C0 > C1 > 0 are some constants. Calculate the risk function of the following randomized test (decision rule): if more than half of Xi ’s are 0 0 T = 1 if more than half of Xi ’s are 1 1 if exactly half of Xi ’s are 0. 2

152

2. Fundamentals of Statistics

70. Consider Example 2.21. Suppose that our decision rule, based on a sample X = (X1 , ..., Xn ) with i.i.d. components from the N (θ, 1) distribution with an unknown θ > 0, is ¯ b1 < X a1 ¯ ≤ b1 T (X) = b0 < X a 2 ¯ a3 X ≤ b0 . Express the risk of T in terms of θ.

71. Consider an estimation problem with P = {Pθ : θ ∈ Θ} (a parametric family), A = Θ, and the squared error loss. If θ0 ∈ Θ satisfies that Pθ ≪ Pθ0 for any θ ∈ Θ, show that the estimator T ≡ θ0 is admissible. 72. Let ℑ be a class of decision rules. A subclass ℑ0 ⊂ ℑ is called ℑcomplete if and only if, for any T ∈ ℑ and T 6∈ ℑ0 , there is a T0 ∈ ℑ0 that is better than T , and ℑ0 is called ℑ-minimal complete if and only if ℑ0 is ℑ-complete and no proper subclass of ℑ0 is ℑ-complete. Show that if a ℑ-minimal complete class exists, then it is exactly the class of ℑ-admissible rules. 73. Let X1 , ..., Xn be i.i.d. random variables having a distribution P ∈ P. Assume that EX12 < ∞. Consider estimating µ = EX1 under the squared error loss. ¯ + b is inadmissible, where (a) Show that any estimator of the form aX ¯ X is the sample mean, a and b are constants, and a > 1. ¯ + b is inadmissible, where (b) Show that any estimator of the form X b 6= 0 is a constant. 74. Consider an estimation problem with ϑ ∈ [c, d] ⊂ R, where c and d are known. Suppose that the action space is A ⊃ [c, d] and the loss function is L(|ϑ − a|), where L(·) is an increasing function on [0, ∞). Show that any decision rule T with P (T (X) 6∈ [c, d]) > 0 for some P ∈ P is inadmissible. k 75. Suppose that the action space is (Ω, BΩ ), where Ω ∈ B k . Let X be a sample from P ∈ P, δ0 (X) be a nonrandomized rule, and T be a sufficient statistic for P ∈ P. Show that if E[IA (δ0 (X))|T ] is a k nonrandomized rule, i.e., E[IA (δ0 (X))|T ] = IA (h(T )) for any A ∈ BΩ , where h is a Borel function, then δ0 (X) = h(T (X)) a.s. P .

76. Let T , δ0 , and δ1 be as given in the statement of Proposition 2.2. Show that Z Z L(P, a)dδ1 (X, a) = E L(P, a)dδ0 (X, a) T a.s. P . A

A

153

2.6. Exercises

77. Prove Theorem 2.5. 78. In Exercise 64, use Theorem 2.5 to find decision rules that are better than Tj , j = 0, 1, 2. 79. In Exercise 65, use Theorem 2.5 to find a decision rule better than cX(1) . 80. Consider Example 2.22. (a) Show that there is no optimal rule if ℑ contains all possible estimators. (Hint: consider constant estimators.) (b) Find a ℑ2 -optimal rule if X1 , ..., Xn are independent random variables having a common mean µ and Var(Xi ) = σ 2 /ai with known ai , i = 1, ..., n. (c) Find a ℑ2 -optimal rule if X1 , ..., Xn are identically distributed but are correlated with a common correlation coefficient ρ. 81. Let Xij = µ + ai + ǫij , i = 1, ..., m, j = 1, ..., n, where ai ’s and ǫij ’s 2 are independent random variables, ai is N (0, σa2 ), ǫij is N (0, Pσne ), and 2 2 −1 ¯ µ, σa , and σe are unknown parameters. Define Xi = n j=1 Xij , ¯ i − X) ¯ 2 , and MSE ¯ = m−1 Pm X ¯ i , MSA = n(m − 1)−1 Pm (X X i=1 P i=1 m Pn ¯ i )2 . Assume that m(n − 1) > 4. = m−1 (n − 1)−1 i=1 j=1 (Xij − X Consider the following class of estimators of θ = σa2 /σe2 : 1 MSA ˆ θ(δ) = (1 − δ) −1 :δ ∈R . n MSE (a) Show that MSA and MSE are independent. ˆ (b) Obtain a δ ∈ R such that θ(δ) is unbiased for θ. ˆ (c) Show that the risk of θ(δ) under the squared error loss is a function of (δ, θ). (d) Show that there is a constant δ ∗ such that for any fixed θ, the risk ˆ is strictly decreasing in δ for δ < δ ∗ and strictly increasing for of θ(δ) δ > δ∗. (e) Show that the unbiased estimator of θ derived in (b) is inadmissible. 82. Let T0 (X) be an unbiased estimator of ϑ in an estimation problem. Show that any unbiased estimator of ϑ is of the form T (X) = T0 (X)− U (X), where U (X) is an “unbiased estimator” of 0. 83. Let X be a discrete random variable with P (X = −1) = p,

P (X = k) = (1 − p)2 pk ,

k = 0, 1, 2, ...,

where p ∈ (0, 1) is unknown. (a) Show that U (X) is an unbiased estimator of 0 if and only if U (k) =

154

2. Fundamentals of Statistics

ak for all k = −1, 0, 1, 2, ... and some a. (b) Show that T0 (X) = I{0} (X) is unbiased for ϑ = (1 − p)2 and that, under the squared error loss, T0 is a ℑ-optimal rule, where ℑ is the class of all unbiased estimators of ϑ. (c) Show that T0 (X) = I{−1} (X) is unbiased for ϑ = p and that, under the squared error loss, there is no ℑ-optimal rule, where ℑ is the class of all unbiased estimators of ϑ. 84. (Nonexistence of an unbiased estimator). Let X be a random variable having the binomial distribution Bi(p, n) with an unknown p ∈ (0, 1) and a known n. Consider the problem of estimating ϑ = p−1 . Show that there is no unbiased estimator of ϑ. 85. Let X1 , ..., Xn be i.i.d. random variables having the normal distribution N (θ, 1), where θ = 0 or 1. Consider the estimation of θ. (a) Let ℑ be the class of nonrandomized rules (estimators), i.e., estimators that take values 0 and 1 only. Show that there does not exist any unbiased estimator of θ in ℑ. (b) Find an estimator in ℑ that is approximately unbiased. 86. Let X1 , ..., Xn be i.i.d. from the Poisson distribution P (θ) with an ¯ unknown θ > 0. Find the bias and mse of Tn = (1 − a/n)nX as an −aθ estimator of ϑ = e , where a 6= 0 is a known constant.

87. Let X1 , ..., Xn be i.i.d. (n ≥ 3) from N (µ, σ 2 ), where µ > 0 and σ > 0 ¯ are unknown parameters. Let T1 = X/S be an estimator of µ/σ and 2 2 ¯ ¯ and S 2 are the sample mean T2 = X be an estimator of µ , where X and variance, respectively. Calculate the mse’s of T1 and T2 . 88. Consider a location family {Pµ : µ ∈ Rk } on Rk , where Pµ = P(µ,Ik ) is given in (2.10). Let l0 ∈ Rk be a fixed vector and L(P, a) = L(kµ − ak), where a ∈ A = Rk and L(·) is a nonnegative Borel function on [0, ∞). Show that the family is invariant and the decision problem is invariant under the transformation g(X) = X + cl0 , c ∈ R. Find an invariant decision rule. 89. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with unknown µ ∈ R and σ 2 > 0. Consider the scale transformation aX, a ∈ (0, ∞). (a) For estimating σ 2 under the loss function L(P, a) = (1 − a/σ 2 )2 , show that the problem is invariant and that the sample variance S 2 is invariant. (b) For testing H0 : µ ≤ 0 versus H1 : µ > 0 under the loss |µ| µ I(0,∞) (µ) and L(P, 1) = I(−∞,0] (µ), σ σ show p that the problem is invariant and any test that is a function of ¯ S 2 /n is invariant. X/ L(P, 0) =

2.6. Exercises

155

90. Let X1 , ..., Xn be i.i.d. random variables having the c.d.f. F (x − θ), where F is symmetric aboutP 0 and θ ∈ R is unknown. n (a) Show that the c.d.f. of i=1 wi X(i) − θ is symmetric about 0, where X(i) is the ith Pnorder statistic and wi ’s are constants satisfying wi = wn−i+1 and Pn i=1 wi = 1. (b) Show that i=1 wi X(i) in (a) is unbiased for θ if the mean of F exists. Pn Pn (c) Show that i=1 wi X(i) is location invariant when i=1 wi = 1.

91. In Example 2.25, show that the conditional distribution of θ given X = x is N (µ∗ (x), c2 ) with µ∗ (x) and c2 given by (2.25). 92. A median of a random variable Y (or its distribution) is any value m such that P (Y ≤ m) ≥ 12 and P (Y ≥ m) ≥ 12 . (a) Show that the set of medians is a closed interval [m0 , m1 ]. (b) Suppose that E|Y | < ∞. If c is not a median of Y , show that E|Y − c| ≥ E|Y − m| for any median m of Y . (c) Let X be a sample from Pθ , where θ ∈ Θ ⊂ R. Consider the estimation of θ under the absolute error loss function |a − θ|. Let Π be a given distribution on Θ with finite mean. Find the ℑ-Bayes rule w.r.t. Π, where ℑ is the class of all rules. 93. (Classification). Let X be a sample having a p.d.f. fj (x) w.r.t. a σfinite measure ν, where j is unknown and j ∈ {1, ..., J} with a known integer J ≥ 2. Consider a decision problem in which the action space A = {1, ..., J} and the loss function is 0 if a = j L(j, a) = 1 if a 6= j. (a) Let ℑ be the class of all nonrandomized decision rules. Obtain the risk of a δ ∈ ℑ. (b) Let Π be a probability measure on {1, ..., J} with Π({j}) = πj , j = 1, ..., J. Obtain the Bayes risk of δ ∈ ℑ w.r.t. Π. (c) Obtain a ℑ-Bayes rule w.r.t. Π in (b). (d) Assume that J = 2, π1 = π2 = 0.5, and fj (x) = φ(x − µj ), where φ(x) is the p.d.f. of the standard normal distribution and µj , j = 1, 2, are known constants. Obtain the Bayes rule in (c) and compute the Bayes risk. (e) Obtain the risk and the Bayes risk (w.r.t. Π in (b)) of a randomized decision rule. (f) Obtain a Bayes rule w.r.t. Π. (g) Obtain a minimax rule. 94. Let θˆ be an unbiased estimator of an unknown θ ∈ R. (a) Under the squared error loss, show that the estimator θˆ + c is not

156

2. Fundamentals of Statistics

minimax unless supθ RT (θ) = ∞ for any estimator T , where c 6= 0 is a known constant. (b) Under the squared error loss, show that the estimator cθˆ is not minimax unless supθ RT (θ) = ∞ for any estimator T , where c ∈ (0, 1) is a known constant. (c) Consider the loss function L(θ, a) = (a − θ)2 /θ2 (assuming θ 6= 0). Show that θˆ is not minimax unless supθ RT (θ) = ∞ for any T . 95. Let X be a binary observation with P (X = 1) = θ1 or θ2 , where 0 < θ1 < θ2 < 1 are known values. Consider the estimation of θ with action space {a1 , a2 } and loss function L(θi , aj ) = lij , where l21 ≥ l12 > l11 = l22 = 0. For a decision rule δ(X), the vector (Rδ (θ1 ), Rδ (θ2 )) is defined to be its risk point. (a) Show that the set of risk points of all decision rules is the convex hull of the set of risk points of all nonrandomized rules. (b) Find a minimax rule. (c) Let Π be a distribution on {θ1 , θ2 }. Obtain the class of all Bayes rules w.r.t. Π. Discuss when there is a unique Bayes rule. 96. Consider the decision problem in Example 2.23. (a) Let Π be the uniform distribution on (0, 1). Show that a ℑ-Bayes rule w.r.t. Π is Tj ∗ (X), where j ∗ is the largest integer in {0, 1, ..., n−1} such that Bj+1,n−j+1 (θ0 ) ≥ 21 and Ba,b (·) denotes the c.d.f. of the beta distribution B(a, b). (b) Derive a ℑ-minimax rule. 97. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with unknown µ ∈ R and σ 2 > 0. To test the hypotheses H0 : µ ≤ µ0

versus

H1 : µ > µ0 ,

where µ0 is a fixed constant, consider p a test of the form Tc (X) = ¯ − µ0 )/ S 2 /n and c is a fixed constant. I(c,∞) (Tµ0 ), where Tµ0 = (X (a) Find the size of Tc . (Hint: Tµ0 has the t-distribution tn−1 .) (b) If α is a given level of significance, find a cα such that Tcα has size α. (c) Compute the p-value for Tcp α derived in (b). p ¯ − cα S 2 /n, X ¯ + cα S 2 /n] is a confidence (d) Find a cα such that [X interval for µ with confidence coefficient 1 − α. What is the expected interval length? 98. In Exercise 67, calculate the size of Tc (X); find a cα such that Tcα has size α, a given level of significance; and find the p-value for Tcα . 99. In Exercise 68, assume that σ is known. Calculate the size of Tc (X); find a cα such that Tcα has size α, a given level of significance; and find the p-value for Tcα .

2.6. Exercises

157

100. Let α ∈ (0, 1) be given and Tj,q (X) be the test given in Example 2.30. Show that there exist integer j and q ∈ (0, 1) such that the size of Tj,q is α. 101. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(a, θ) with unknown a ∈ R andPθ > 0. Let α ∈ (0, 1) be given. (a) Using T1 (X) = ni=1 (Xi − X(1) ), construct a confidence interval for θ with confidence coefficient 1 − α and find the expected interval length. (b) Using T1 (X) and T2 (X) = X(1) , construct a confidence interval for a with confidence coefficient 1 − α and find the expected interval length. (c) Using the method in Example 2.32, construct a confidence set for the two-dimensional parameter (a, θ) with confidence coefficient 1−α. 102. Suppose that X is a sample and a statistic T (X) has a distribution in a location family {Pµ : µ ∈ R}. Using T (X), derive a confidence interval for µ with level of significance 1 − α and obtain the expected interval length. Show that if the c.d.f. of T (X) is continuous, then we can always find a confidence interval for µ with confidence coefficient 1 − α for any α ∈ (0, 1). 103. Let X = (X1 , ..., Xn ) be a sample from Pθ , where θ ∈ {θ1 , ..., θk } with a fixed integer k. Let Tn (X) be an estimator of θ with range {θ1 , ..., θk }. (a) Show that Tn (X) is consistent if and only if Pθ (Tn (X) = θ) → 1. (b) Show that if Tn (X) is consistent, then it is an -consistent for any {an }. 104. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (θ − 21 , θ + 12 ), where θ ∈ R is unknown. Show that (X(1) + X(n) )/2 is strongly consistent for θ and also consistent in mse. 105. Let X1 , ..., Xn be i.i.d. from a population with the Lebesgue p.d.f. fθ (x) = 2−1 (1 + θx)I(−1,1) (x), where θ ∈ (−1, 1) is an unknown√parameter. Find a consistent estimator of θ. Is your estimator nconsistent? 106. Let X1 , ..., Xn be i.i.d. observations. Suppose that Tn is an unbiased estimator of ϑ based on X1 , ..., Xn such that for any n, Var(Tn ) < ∞ and Var(Tn ) ≤ Var(Un ) for any other unbiased estimator Un of ϑ based on X1 , ..., Xn . Show that Tn is consistent in mse. 107. Consider the Bayes rule √ µ∗ (X) in Example 2.25. Show that µ∗ (X) is a strongly consistent, n-consistent, and L2 -consistent estimator of µ. What is the order of the bias of µ∗ (X) as an estimator of µ?

158

2. Fundamentals of Statistics

108. In Exercise 21, show that ¯ is an inconsistent estimator of β; (a) Y¯ /X ˆ (b) β = Z(m) is a consistent estimator of β, where m = n/2 when n is even, m = (n + 1)/2 when n is odd, and Z(i) is the ith smallest value of Yi /Xi , i = 1, ..., n. 109. Show that the estimator T0 of θ in Exercise 64 is inconsistent. 110. Let g1 , g2 ,... be continuous functions on (a, b) ⊂ R such that gn (x) → g(x) uniformly for x in any closed subinterval of (a, b). Let Tn be a consistent estimator of θ ∈ (a, b). Show that gn (Tn ) is consistent for ϑ = g(θ). 111. Let X1 , ..., Xn be i.i.d. from P with unknown mean µ ∈ R and variance σ 2 > 0, and let g(µ) = 0 if µ 6= 0 and g(0) = 1. Find a consistent estimator of ϑ = g(µ). 112. Establish results for the smallest order statistic X(1) (based on i.i.d. random variables X1 , ..., Xn ) similar to those in Example 2.34. 113. (Consistency for finite population). In Example 2.27, show that Yˆ →p Y as n → N for any fixed N and population. Is Yˆ still consistent if sampling is with replacement? 114. Assume that Xi = θti + ei , i = 1, ..., n, where θ ∈ Θ is an unknown parameter, Θ is a closed subset of R, ei ’s are i.i.d. on the interval [−τ, τ ] with some unknown τ > 0 and Eei = 0, and ti ’s are fixed constants. Let Tn = Sn (θ˜n ) = min Sn (γ), γ∈Θ

where Sn (γ) = 2 max |Xi − γti |/ i≤n

p 1 + γ 2.

(a) Assume that supi |ti | < ∞ and supi ti − inf i ti > 2τ . Show that the sequence {θ˜n , n = 1, 2, ...} is bounded a.s. (b) Let θn ∈ Θ, n = 1, 2, .... If θn → θ, show that Sn (θn ) − Sn (θ) = O(|θn − θ|) a.s. (c) Under the conditions in (a), show that Tn is a strongly consistent estimator of ϑ = minγ∈Θ S(γ), where S(γ) = limn→∞ Sn (γ) a.s. ¯ be 115. Let X1 , ..., Xn be i.i.d. random variables with EX12 < ∞ and X the sample mean. Consider the estimation of µ = EX1 . ¯ + ξn /√n, where ξn is a random variable satisfying (a) Let Tn = X ξn = 0 with probability 1 − n−1 and ξn = n3/2 with probability n−1 .

2.6. Exercises

159

(P ) for any P . Show that bTn (P ) 6= ˜bTn√ ¯ + ηn / n, where ηn is a random variable that is (b) Let Tn = X independent of X1 , ..., Xn and equals 0 with probability 1 − 2n−1 and √ ± n with probability n−1 . Show that amseTn (P ) = amseX¯ (P ) = mseX¯ (P ) and mseTn (P ) > amseTn (P ) for any P . 116. Let X1 , ..., Xn be i.i.d. random variables with finite θ = EX1 and Var(X θ > 0 is unknown. Consider the estimation of √1 ) = θ, where √ ¯ ¯ and S 2 are the ¯ and T2n = X/S, where X ϑ = θ. Let T1n = X sample mean and sample variance. (a) Obtain the n−1 order asymptotic biases of T1n and T2n according to (2.38). (b) Obtain the asymptotic relative efficiency of T1n w.r.t. T2n . 117. Let X1 , ..., Xn be i.i.d. according to N (µ, 1) with an unknown µ ∈ R. Let ϑ = P (X1 ≤ c) for a fixed constant c. Consider the following estimators of ϑ: T1n = Fn (c), where Fn is the empirical c.d.f. defined ¯ where Φ is the c.d.f. of N (0, 1). in (2.28), and T2n = Φ(c − X), −1 (a) Find the n order asymptotic bias of T2n according to (2.38). (b) Find the asymptotic relative efficiency of T1n w.r.t. T2n . 118. Let X1 , ..., Xn be i.i.d. from the N (0, σ 2 ) distribution with an unknown σ > 0. Consider estimation of ϑ = σ.PFind the asymptotic p theP relative efficiency of π/2 ni=1 |Xi |/n w.r.t. ( ni=1 Xi2 /n)1/2 .

119. Let X1 , ..., Xn be i.i.d. from P with EX14 < ∞ and unknown mean µ ∈ R and variance σ 2 > 0. Consider the estimation of ϑ = µ2 and ¯ 2 , T2n = X ¯ 2 − S 2 /n, T3n = the following three estimators: T1n = X 2 ¯ max{0, T2n }, where X and S are the sample mean and variance. Show that the amse’s of Tjn , j = 1, 2, 3, are the same when µ 6= 0 but may be different when µ = 0. Which estimator is the best in terms of the asymptotic relative efficiency when µ = 0? 120. Prove Theorem 2.6. 121. Let X1 , ..., Xn be EXi = µ, Var(Xi ) = 1, and EXi4 < ∞. Pni.i.d. with −1 2 −1 ¯2 Let T1n = n be estimators of i=1 Xi − 1 and T2n = X − n 2 ϑ=µ . (a) Find the asymptotic relative efficiency of T1n w.r.t. T2n . (b) Show that eT1n ,T2n (P ) ≤ 1 if the c.d.f. of Xi − µ is symmetric about 0 and µ 6= 0. (c) Find a distribution P for which eT1n ,T2n (P ) > 1.

122. Let X1 , ..., Xn be i.i.d. binary random variables with unknown p = P (Xi = 1) ∈ (0, 1). Consider the estimation of p. Let a and b be two positive constants. Find the asymptotic relative efficiency of the ¯ ¯ estimator (a + nX)/(a + b + n) w.r.t. X.

160

2. Fundamentals of Statistics

123. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with an unknown µ ∈ R and a ¯ be the sample mean and T2 = µ∗ (X) be the known σ 2 . Let T1 = X Bayes estimator given in (2.25). Assume that EX14 < ∞. (a) Calculate the exact mse of both estimators. Can you conclude that one estimator is better than the other in terms of the mse? (b) Find the asymptotic relative efficiency of T1 w.r.t. T2 . 124. In Example 2.37, show that (a) the limiting size of Tcα is 1 if P contains all possible populations on R with finite second moments; (b) Tn = Tcα with α = αn (given by (2.40)) is Chernoff-consistent; (c) Tn in (b) is not strongly Chernoff-consistent if P contains all possible populations on R with finite second moments. 125. Let X1 , ..., Xn be i.i.d. with unknown mean µ ∈ R and variance σ 2 > 0. For testing H0 : µ ≤ µ0 versus H1 : µ > µ0 , consider the test Tcα obtained in Exercise 97(b). (a) Show that Tcα has asymptotic significance level α and is consistent. (b) Find a test that is Chernoff-consistent. 126. Consider the test Tj in Example 2.23. For each n, find a j = jn such that Tjn has asymptotic significance level α ∈ (0, 1). 127. Show that the test Tcα in Exercise 98 is consistent, but Tcα in Exercise 99 is not consistent. 128. In Example 2.31, suppose that we drop the normality assumption but assume that µ = EXi and σ 2 = Var(Xi ) are finite. (a) Show that when σ 2 is known, the asymptotic significance level ¯ − cα , X ¯ + cα ] is 1 − α, where cα = of the confidence interval [X √ −1 σz1−α/2 / n and za = Φ (a). (b) Show that when σ 2 is known, the limiting confidence coefficient of the interval in (a) might be 0 if P contains all possible populations on R. (c) Show that the confidence interval in Exercise 97(d) has asymptotic significance level 1 − α. 129. Let X1 , ..., Xn be i.i.d. with unknown mean µ ∈ R and variance σ 2 > 0. Assume that EX14 < ∞. Using the sample variance S 2 , construct a confidence interval for σ 2 that has asymptotic significance level 1 − α. 130. Consider the sample correlation coefficient T defined in Exercise 22. Construct a confidence interval for ρ that has asymptotic significance level 1 − α, assuming that (Yi , Zi ) is normally distributed. (Hint: show that the asymptotic variance of T is (1 − ρ2 )2 .)

Chapter 3

Unbiased Estimation Unbiased or asymptotically unbiased estimation plays an important role in point estimation theory. Unbiasedness of point estimators is defined in §2.3.2. In this chapter, we discuss in detail how to derive unbiased estimators and, more importantly, how to find the best unbiased estimators in various situations. Although an unbiased estimator (even the best unbiased estimator if it exists) is not necessarily better than a slightly biased estimator in terms of their mse’s (see Exercise 63 in §2.6), unbiased estimators can be used as “building blocks” for the construction of better estimators. Furthermore, one may give up the exact unbiasedness, but cannot give up asymptotic unbiasedness since it is necessary for consistency (see §2.5.2). Properties and the construction of asymptotically unbiased estimators are studied in the last part of this chapter.

3.1 The UMVUE Let X be a sample from an unknown population P ∈ P and ϑ be a realvalued parameter related to P . Recall that an estimator T (X) of ϑ is unbiased if and only if E[T (X)] = ϑ for any P ∈ P. If there exists an unbiased estimator of ϑ, then ϑ is called an estimable parameter. Definition 3.1. An unbiased estimator T (X) of ϑ is called the uniformly minimum variance unbiased estimator (UMVUE) if and only if Var(T (X)) ≤ Var(U (X)) for any P ∈ P and any other unbiased estimator U (X) of ϑ. Since the mse of any unbiased estimator is its variance, a UMVUE is ℑ-optimal in mse with ℑ being the class of all unbiased estimators. One 161

162

3. Unbiased Estimation

can similarly define the uniformly minimum risk unbiased estimator in statistical decision theory when we use an arbitrary loss instead of the squared error loss that corresponds to the mse.

3.1.1 Sufficient and complete statistics The derivation of a UMVUE is relatively simple if there exists a sufficient and complete statistic for P ∈ P. Theorem 3.1 (Lehmann-Scheff´e theorem). Suppose that there exists a sufficient and complete statistic T (X) for P ∈ P. If ϑ is estimable, then there is a unique unbiased estimator of ϑ that is of the form h(T ) with a Borel function h. (Two estimators that are equal a.s. P are treated as one estimator.) Furthermore, h(T ) is the unique UMVUE of ϑ. This theorem is a consequence of Theorem 2.5(ii) (Rao-Blackwell theorem). One can easily extend this theorem to the case of the uniformly minimum risk unbiased estimator under any loss function L(P, a) that is strictly convex in a. The uniqueness of the UMVUE follows from the completeness of T (X). There are two typical ways to derive a UMVUE when a sufficient and complete statistic T is available. The first one is solving for h when the distribution of T is available. The following are two typical examples. Example 3.1. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (0, θ), θ > 0. Let ϑ = g(θ), where g is a differentiable function on (0, ∞). Since the sufficient and complete statistic X(n) has the Lebesgue p.d.f. nθ−n xn−1 I(0,θ) (x), an unbiased estimator h(X(n) ) of ϑ must satisfy θn g(θ) = n

Z

θ

h(x)xn−1 dx

for all θ > 0.

0

Differentiating both sizes of the previous equation and applying the result of differentiation of an integral (Royden (1968, §5.3)) lead to nθn−1 g(θ) + θn g ′ (θ) = nh(θ)θn−1 . Hence, the UMVUE of ϑ is h(X(n) ) = g(X(n) ) + n−1 X(n) g ′ (X(n) ). In particular, if ϑ = θ, then the UMVUE of θ is (1 + n−1 )X(n) . Example 3.2. Let X1 , ..., Xn be i.i.d. from Pn the Poisson distribution P (θ) with an unknown θ > 0. Then T (X) = i=1 Xi is sufficient and complete for θ > 0 and has the Poisson distribution P (nθ). P Suppose that ϑ = g(θ), ∞ j where g is a smooth function such that g(x) = j=0 aj x , x > 0. An

163

3.1. The UMVUE

unbiased estimator h(T ) of ϑ must satisfy ∞ X h(t)nt t=0

t!

θt = enθ g(θ) ∞ X nk

θk

∞ X

aj θ j k! j=0 k=0 ∞ X X n k aj θt = k! t=0 =

j,k:j+k=t

for any θ > 0. Thus, a comparison of coefficients in front of θt leads to h(t) =

t! nt

X

j,k:j+k=t

n k aj , k!

i.e., h(T ) is the UMVUE of ϑ. In particular, if ϑ = θr for some fixed integer r ≥ 1, then ar = 1 and ak = 0 if k 6= r and ( 0 t 0 and I(t,∞) (X1 ) is unbiased for ϑ, ¯ = P (X1 > t|X) ¯ T (X) = E[I(t,∞) (X1 )|X] ¯ is availis the UMVUE of ϑ. If the conditional distribution of X1 given X ¯ directly. But the following techable, then we can calculate P (X1 > t|X) nique can be applied to avoid the derivation of conditional distributions. ¯ and X ¯ are independent. By By Basu’s theorem (Theorem 2.4), X1 /X Proposition 1.10(vii), ¯ =x ¯ > t/X| ¯X ¯ =x ¯ > t/¯ P (X1 > t|X ¯) = P (X1 /X ¯) = P (X1 /X x).

164

3. Unbiased Estimation

To compute this unconditional probability, we need the distribution of ! X n n X X1 + X1 Xi = X1 Xi . i=1

i=2

Using Pn the transformation technique discussed in §1.3.1 and the fact that is independent of X1 and has a gamma distribution, we obtain i=2 Xi P n that X1 / i=1 Xi has the Lebesgue p.d.f. (n − 1)(1 − x)n−2 I(0,1) (x). Hence n−1 Z 1 t ¯ =x P (X1 > t|X ¯) = (n − 1) (1 − x)n−2 dx = 1 − n¯ x t/(n¯ x) and the UMVUE of ϑ is n−1 t T (X) = 1 − ¯ . nX We now show more examples of applying these two methods to find UMVUE’s. Example 3.4. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ ∈ R ¯ S 2 ) is sufficient and comand σ 2 > 0. From Example 2.18, T = (X, 2 2 ¯ plete for θ = (µ, σ ) and X and (n − 1)S /σ 2 are independent and have the N (µ, σ 2 /n) and chi-square distribution χ2n−1 , respectively. Using the ¯ the method of solving for h directly, we find that the UMVUE for µ is X; 2 2 2 r ¯ UMVUE of µ is X − S /n; the UMVUE for σ with r > 1 − n is kn−1,r S r , where nr/2 Γ(n/2) kn,r = r/2 n+r 2 Γ 2 ¯ (exercise); and the UMVUE of µ/σ is kn−1,−1 X/S, if n > 2. Suppose that ϑ satisfies P (X1 ≤ ϑ) = p with a fixed p ∈ (0, 1). Let Φ be the c.d.f. of the standard normal distribution. Then ϑ = µ + σΦ−1 (p) ¯ + kn−1,1 SΦ−1 (p). and its UMVUE is X Let c be a fixed constant and ϑ = P (X1 ≤ c) = Φ c−µ . We can σ find the UMVUE of ϑ using the method of conditioning and the technique used in Example 3.3. Since I(−∞,c) (X1 ) is an unbiased estimator of ϑ, the UMVUE of ϑ is E[I(−∞,c) (X1 )|T ] = P (X1 ≤ c|T ). By Basu’s theorem, ¯ ¯ S 2 ). the ancillary statistic Z(X) = (X1 − X)/S is independent of T = (X, Then, by Proposition 1.10(vii), ¯ c−X 2 2 P X1 ≤ c|T = (¯ T = (¯ x, s ) x, s ) = P Z ≤ S c−x ¯ . =P Z≤ s

165

3.1. The UMVUE

It can be shown that Z has the Lebesgue p.d.f. √ (n/2)−2 nΓ n−1 nz 2 2 f (z) = √ 1− I(0,(n−1)/√n) (|z|) (3.1) (n − 1)2 π(n − 1)Γ n−2 2

(exercise). Hence the UMVUE of ϑ is Z (c−X)/S ¯ P (X1 ≤ c|T ) = f (z)dz √

(3.2)

−(n−1)/ n

with f given by (3.1). Suppose that we would like to estimate ϑ = σ1 Φ′ c−µ , the Lebesgue σ p.d.f. of X1 evaluated at a fixed c, where Φ′ is the first-order derivative ¯ = x¯ and S 2 = s2 is of Φ. By (3.2), the conditional p.d.f. of X1 given X x−¯ x −1 2 ¯ s f s . Let fT be the joint p.d.f. of T = (X, S ). Then Z Z ¯ c − x¯ c−X 1 1 f fT (t)dt = E f . ϑ= s s S S Hence the UMVUE of ϑ is 1 f S

¯ c−X . S

Example 3.5. Let X1 , ..., Xn be i.i.d. from a power series distribution (see Exercise 13 in §2.6), i.e., P (Xi = x) = γ(x)θx /c(θ),

x = 0, 1, 2, ...,

with a known function γ(x) ≥ 0 and an unknown parameter θ > 0. It turns out that the joint distribution of X = (X1 , ..., Xn ) isP in an exponential famn ily with a sufficient and complete statistic T (X) = i=1 Xi . Furthermore, the distribution of T is also in a power series family, i.e., P (T = t) = γn (t)θt /[c(θ)]n ,

t = 0, 1, 2, ...,

where γn (t) is the coefficient of θt in the power series expansion of [c(θ)]n (Exercise 13 in §2.6). This result can help us to find the UMVUE of ϑ = g(θ). For example, by comparing both sides of ∞ X

h(t)γn (t)θt = [c(θ)]n−p θr ,

t=0

we conclude that the UMVUE of θr /[c(θ)]p is ( 0 T 1,

168

3. Unbiased Estimation

where c and b are some constants. From the previous discussion, E[h(X(n) )U (X(n) )] = 0,

θ ≥ 1.

Since E[h(X(n) )] = θ, we obtain that θ = cP (X(n) ≤ 1) + bE[X(n) I(1,∞) (X(n) )] = cθ−n + [bn/(n + 1)](θ − θ−n ).

Thus, c = 1 and b = (n + 1)/n. The UMVUE of θ is then 1 0 ≤ X(n) ≤ 1 T = X(n) > 1. (1 + n−1 )X(n) This estimator is better than (1 + n−1 )X(n) , which is the UMVUE when Θ = (0, ∞) and does not make use of the information about θ ≥ 1. Example 3.8. Let X be a sample (of size 1) from the uniform distribution U (θ − 12 , θ + 12 ), θ ∈ R. We now apply Theorem 3.2 to show that there is no UMVUE of ϑ = g(θ) for any nonconstant function g. Note that an unbiased estimator U (X) of 0 must satisfy Z θ+ 12 U (x)dx = 0 for all θ ∈ R. θ− 12

Differentiating both sizes of the previous equation and applying the result of differentiation of an integral lead to U (x) = U (x + 1) a.e. m, where m is the Lebesgue measure on R. If T is a UMVUE of g(θ), then T (X)U (X) is unbiased for 0 and, hence, T (x)U (x) = T (x+1)U (x+1) a.e. m, where U (X) is any unbiased estimator of 0. Since this is true for all U , T (x) = T (x + 1) a.e. m. Since T is unbiased for g(θ), Z θ+ 21 g(θ) = T (x)dx for all θ ∈ R. θ− 12

Differentiating both sizes of the previous equation and applying the result of differentiation of an integral, we obtain that g ′ (θ) = T θ + 12 − T θ − 12 = 0 a.e. m. As a consequence of Theorem 3.2, we have the following useful result. Corollary 3.1. (i) Let Tj be a UMVUE of ϑj , j = 1, ..., k, where k is a Pk Pk fixed positive integer. Then j=1 cj Tj is a UMVUE of ϑ = j=1 cj ϑj for any constants c1 , ..., ck . (ii) Let T1 and T2 be two UMVUE’s of ϑ. Then T1 = T2 a.s. P for any P ∈ P.

169

3.1. The UMVUE

3.1.3 Information inequality Suppose that we have a lower bound for the variances of all unbiased estimators of ϑ and that there is an unbiased estimator T of ϑ whose variance is always the same as the lower bound. Then T is a UMVUE of ϑ. Although this is not an effective way to find UMVUE’s (compared with the methods introduced in §3.1.1 and §3.1.2), it provides a way of assessing the performance of UMVUE’s. The following result provides such a lower bound in some cases. Theorem 3.3 (Cram´er-Rao lower bound). Let X = (X1 , ..., Xn ) be a sample from P ∈ P = {Pθ : θ ∈ Θ}, where Θ is an open set in Rk . Suppose that T (X) is an estimator with E[T (X)] = g(θ) being a differentiable function of θ; Pθ has a p.d.f. fθ w.r.t. a measure ν for all θ ∈ Θ; and fθ is differentiable as a function of θ and satisfies Z Z ∂ ∂ h(x)fθ (x)dν = h(x) fθ (x)dν, θ ∈ Θ, (3.3) ∂θ ∂θ for h(x) ≡ 1 and h(x) = T (x). Then ∂ τ ∂ Var(T (X)) ≥ ∂θ g(θ) [I(θ)]−1 ∂θ g(θ), where

I(θ) = E

τ ∂ ∂ log fθ (X) log fθ (X) ∂θ ∂θ

(3.4)

(3.5)

is assumed to be positive definite for any θ ∈ Θ. Proof. We prove the univariate case (k = 1) only. The proof for the multivariate case (k > 1) is left to the reader. When k = 1, (3.4) reduces to [g ′ (θ)]2 Var(T (X)) ≥ (3.6) 2 . ∂ E ∂θ log fθ (X)

From inequality (1.37), we only need to show that

2 ∂ ∂ E log fθ (X) = Var log fθ (X) ∂θ ∂θ and

∂ log fθ (X) . g (θ) = Cov T (X), ∂θ ′

These two results are consequences of condition (3.3). The k × k matrix I(θ) in (3.5) is called the Fisher information matrix. The greater I(θ) is, the easier it is to distinguish θ from neighboring values

170

3. Unbiased Estimation

and, therefore, the more accurately θ can be estimated. In fact, if the equality in (3.6) holds for an unbiased estimator T (X) of g(θ) (which is then a UMVUE), then the greater I(θ) is, the smaller Var(T (X)) is. Thus, I(θ) is a measure of the information that X contains about the unknown θ. The inequalities in (3.4) and (3.6) are called information inequalities. The following result is helpful in finding the Fisher information matrix. Proposition 3.1. (i) Let X and Y be independent with the Fisher information matrices IX (θ) and IY (θ), respectively. Then, the Fisher information about θ contained in (X, Y ) is IX (θ) + IY (θ). In particular, if X1 , ..., Xn are i.i.d. and I1 (θ) is the Fisher information about θ contained in a single Xi , then the Fisher information about θ contained in X1 , ..., Xn is nI1 (θ). (ii) Suppose that X has the p.d.f. fθ that is twice differentiable in θ and that (3.3) holds with h(x) ≡ 1 and fθ replaced by ∂fθ /∂θ. Then ∂2 log fθ (X) . (3.7) I(θ) = −E ∂θ∂θτ Proof. Result (i) follows from the independence of X and Y and the definition of the Fisher information. Result (ii) follows from the equality τ ∂2 ∂2 ∂ ∂ ∂θ∂θ τ fθ (X) − log f log f log f (X) = (X) (X) . θ θ θ ∂θ∂θτ fθ (X) ∂θ ∂θ The following example provides a formula for the Fisher information matrix for many parametric families with a two-dimensional parameter θ. Example 3.9. Let X1 , ..., Xn be i.i.d. with the Lebesgue p.d.f. σ1 f x−µ , σ where f (x) > 0 and f ′ (x) exists for all x ∈ R, µ ∈ R, and σ > 0 (a location-scale family). Let θ = (µ, σ). Then, the Fisher information about θ contained in X1 , ..., Xn is (exercise) R [f ′ (x)]2 R f ′ (x)[xf ′ (x)+f (x)] dx f (x) dx f (x) n I(θ) = 2 . σ R [xf ′ (x)+f (x)]2 R f ′ (x)[xf ′ (x)+f (x)] dx dx f (x) f (x) Note that I(θ) depends on the particular parameterization. If θ = ψ(η) and ψ is differentiable, then the Fisher information that X contains about η is h i ∂ ∂η ψ(η)I(ψ(η))

∂ ∂η ψ(η)

τ

.

However, it is easy to see that the Cram´er-Rao lower bound in (3.4) or (3.6) is not affected by any one-to-one reparameterization.

171

3.1. The UMVUE

If we use inequality (3.4) or (3.6) to find a UMVUE T (X), then we obtain a formula for Var(T (X)) at the same time. On the other hand, the Cram´er-Rao lower bound in (3.4) or (3.6) is typically not sharp. Under some regularity conditions, the Cram´er-Rao lower bound is attained if and only if fθ is in an exponential family; see Propositions 3.2 and 3.3 and the discussion in Lehmann (1983, p. 123). Some improved information inequalities are available (see, e.g., Lehmann (1983, Sections 2.6 and 2.7)). Proposition 3.2. Suppose that the distribution of X is from an exponential family {fθ : θ ∈ Θ}, i.e., the p.d.f. of X w.r.t. a σ-finite measure is fθ (x) = exp [η(θ)]τ T (x) − ξ(θ) c(x) (3.8)

(see §2.1.3), where Θ is an open subset of Rk . (i) The regularity condition (3.3) is satisfied for any h with E|h(X)| < ∞ and (3.7) holds. (ii) If I(η) is the Fisher information matrix for the natural parameter η, then the variance-covariance matrix Var(T ) = I(η). (iii) If I(ϑ) is the Fisher information matrix for the parameter ϑ = E[T (X)], then Var(T ) = [I(ϑ)]−1 . Proof. (i) This is a direct consequence of Theorem 2.1. (ii) From (2.6), the p.d.f. under the natural parameter η is fη (x) = exp {η τ T (x) − ζ(η)} c(x). From Theorem 2.1 and result (1.54) in §1.3.3, E[T (X)] = result follows from ∂ ∂η

log fη (x) = T (x) −

(iii) Since ϑ = E[T (X)] = I(η) =

∂ϑ ∂η I(ϑ)

∂ ∂η ζ(η).

The

∂ ∂η ζ(η).

∂ ∂η ζ(η), ∂ϑ ∂η

τ

=

∂2 ∂η∂η τ

ζ(η)I(ϑ)

h

∂2 ∂η∂η τ

By Theorem 2.1, result (1.54), and the result in (ii), I(η). Hence

ζ(η)

∂2 ∂η∂η τ

iτ

.

ζ(η) = Var(T ) =

I(ϑ) = [I(η)]−1 I(η)[I(η)]−1 = [I(η)]−1 = [Var(T )]−1 . A direct consequence of Proposition 3.2(ii) is that the variance of any linear function of T in (3.8) attains the Cram´er-Rao lower bound. The following result gives a necessary condition for Var(U (X)) of an estimator U (X) to attain the Cram´er-Rao lower bound.

172

3. Unbiased Estimation

Proposition 3.3. Assume that the conditions in Theorem 3.3 hold with T (X) replaced by U (X) and that Θ ⊂ R. (i) If Var(U (X)) attains the Cram´er-Rao lower bound in (3.6), then a(θ)[U (X) − g(θ)] = g ′ (θ)

∂ log fθ (X) a.s. Pθ ∂θ

for some function a(θ), θ ∈ Θ. (ii) Let fθ and T be given by (3.8). If Var(U (X)) attains the Cram´er-Rao lower bound, then U (X) is a linear function of T (X) a.s. Pθ , θ ∈ Θ. Example 3.10. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 . Let fµ be the joint distribution of X = (X1 , ..., Xn ). Then ∂ ∂µ

log fµ (X) =

n X i=1

(Xi − µ)/σ 2 .

¯ attains the Cram´er-Rao lower Thus, I(µ) = n/σ 2 . It is obvious that Var(X) ¯2 = bound in (3.6). Consider now the estimation of ϑ = µ2 . Since E X 2 2 2 2 ¯ ¯ µ + σ /n, the UMVUE of ϑ is h(X) = X − σ /n. A straightforward calculation shows that ¯ = Var(h(X))

4µ2 σ 2 2σ 4 + 2 . n n

On the other hand, the Cram´er-Rao lower bound in this case is 4µ2 σ 2 /n. ¯ does not attain the Cram´er-Rao lower bound. The difHence Var(h(X)) ference is 2σ 4 /n2 . Condition (3.3) is a key regularity condition for the results in Theorem 3.3 and Proposition 3.3. If fθ is not in an exponential family, then (3.3) has to be checked. Typically, it does not hold if the set {x : fθ (x) > 0} depends on θ (Exercise 37). More discussions can be found in Pitman (1979).

3.1.4 Asymptotic properties of UMVUE’s UMVUE’s are typically consistent (see Exercise 106 in §2.6). If there is an unbiased estimator of ϑ whose mse is of the order a−2 n , where {an } is a sequence of positive numbers diverging to ∞, then the UMVUE of ϑ (if it exists) has an mse of order a−2 n and is an -consistent. For instance, in Example 3.3, the mse of U (X) = 1 − Fn (t) is Fθ (t)[1 − Fθ (t)]/n; hence the √ UMVUE T (X) is n-consistent and its mse is of the order n−1 . UMVUE’s are exactly unbiased so that there is no need to discuss their asymptotic biases. Their variances (or mse’s) are finite, but amse’s can be

173

3.1. The UMVUE

used to assess their performance if the exact forms of mse’s are difficult to obtain. In many cases, although the variance of a UMVUE Tn does not attain the Cram´er-Rao lower bound, the limit of the ratio of the amse (or mse) of Tn over the Cram´er-Rao lower bound (if it is not 0) is 1. For instance, in Example 3.10, ¯ 2 − σ 2 /n) σ2 Var(X =1+ 2 →1 the Cram´er-Rao lower bound 2µ n if µ 6= 0. In general, under the conditions in Theorem 3.3, if Tn (X) is unbiased for g(θ) and if, for any θ ∈ Θ, ∂ τ ∂ Tn (X) − g(θ) = ∂θ g(θ) [I(θ)]−1 ∂θ log fθ (X) [1 + op (1)] a.s. Pθ , (3.9) then

amseTn (θ) = the Cram´er-Rao lower bound

(3.10)

whenever the Cram´er-Rao lower bound is not 0. Note that the case of zero Cram´er-Rao lower bound is not of interest since a zero lower bound does not provide any information on the performance of estimators. n−1 Consider the UMVUE Tn = 1 − ntX¯ of e−t/θ in Example 3.3. Using the fact that log(1 − x) = − we obtain that

∞ X xj j=1

j

|x| ≤ 1,

,

¯ Tn − e−t/X = Op n−1 .

Using Taylor’s expansion, we obtain that ¯ ¯ − θ)[1 + op (1)], e−t/X − e−t/θ = g ′ (θ)(X where g(θ) = e−t/θ . On the other hand,

∂ ¯ − θ. [I(θ)]−1 ∂θ log fθ (X) = X

Hence (3.9) and (3.10) hold. Note that the exact variance of Tn is not easy to obtain. In this example, it can be shown that {n[Tn − g(θ)]2 } is uniformly integrable and, therefore, lim nVar(Tn ) = lim n[amseTn (θ)]

n→∞

n→∞

= lim n[g ′ (θ)]2 [I(θ)]−1 =

n→∞ 2 −2t/θ

t e θ2

.

It is shown in Chapter 4 that if (3.10) holds, then Tn is asymptotically optimal in some sense. Hence UMVUE’s satisfying (3.9), which is often true, are asymptotically optimal, although they may be improved in terms of the exact mse’s.

174

3. Unbiased Estimation

3.2 U-Statistics Let X1 , ..., Xn be i.i.d. from an unknown population P in a nonparametric family P. In Example 3.6 we argued that if the vector of order statistic is sufficient and complete for P ∈ P, then a symmetric unbiased estimator of any estimable ϑ is the UMVUE of ϑ. In a large class of problems, parameters to be estimated are of the form ϑ = E[h(X1 , ..., Xm )] with a positive integer m and a Borel function h that is symmetric and satisfies E|h(X1 , ..., Xm )| < ∞ for any P ∈ P. It is easy to see that a symmetric unbiased estimator of ϑ is Un =

−1 X n h(Xi1 , ..., Xim ), m c

P where c denotes the summation over the elements {i1 , ..., im } from {1, ..., n}.

n m

(3.11)

combinations of m distinct

Definition 3.2. The statistic Un in (3.11) is called a U -statistic with kernel h of order m.

3.2.1 Some examples The use of U-statistics is an effective way of obtaining unbiased estimators. In nonparametric problems, U-statistics are often UMVUE’s, whereas in parametric problems, U-statistics can be used as initial estimators to derive more efficient estimators. If m = 1, Un in (3.11) is simply a type of sample mean. Examples include the empirical c.d.f. (2.28) evaluated at a particular t and the sample Pn moments n−1 i=1 Xik for a positive integer k. We now consider some examples with m > 1. Consider the estimation of ϑ = µm , where µ = EX1 and m is a positive integer. Using h(x1 , ..., xm ) = x1 · · · xm , we obtain the following U-statistic unbiased for ϑ = µm : −1 X n Xi1 · · · Xim . Un = m c Consider next the estimation of ϑ = σ 2 = Var(X1 ). Since σ 2 = [Var(X1 ) + Var(X2 )]/2 = E[(X1 − X2 )2 /2],

(3.12)

175

3.2. U-Statistics

we obtain the following U-statistic with kernel h(x1 , x2 ) = (x1 − x2 )2 /2: ! n X X (Xi − Xj )2 1 2 2 2 ¯ = = S2, Un = X − nX n(n − 1) 2 n − 1 i=1 i 1≤i<j≤n

which is the sample variance in (2.2). In some cases, we would like to estimate ϑ = E|X1 − X2 |, a measure of concentration. Using kernel h(x1 , x2 ) = |x1 − x2 |, we obtain the following U-statistic unbiased for ϑ = E|X1 − X2 |: Un =

2 n(n − 1)

X

1≤i<j≤n

|Xi − Xj |,

which is known as Gini’s mean difference. Let ϑ = P (X1 + X2 ≤ 0). Using kernel h(x1 , x2 ) = I(−∞,0] (x1 + x2 ), we obtain the following U-statistic unbiased for ϑ: Un =

2 n(n − 1)

X

I(−∞,0] (Xi + Xj ),

1≤i<j≤n

which is known as the one-sample Wilcoxon statistic. Let Tn = Tn (X1 , ..., Xn ) be a given statistic and let r and d be two positive integers such that r + d = n. For any s = {i1 , ..., ir } ⊂ {1, ..., n}, define Tr,s = Tr (Xi1 , ..., Xir ), which is the statistic Tn computed after Xi , i 6∈ s, are deleted from the original sample. Let −1 X n 2 r Un = (3.13) d (Tr,s − Tn ) . r c Then Un is a U-statistic with kernel hn (x1 , ..., xr ) = dr [Tr (x1 , ..., xr ) − Tn (x1 , ..., xn )]2 . Unlike the kernels in the previous examples, the kernel in this example depends on n. The order of the kernel, r, may also depend on n. The statistic Un in (3.13) is known as the delete-d jackknife variance estimator for Tn (see, e.g., Shao and Tu (1995)), since it is often true that E[hn (X1 , ..., Xr )] ≈ Var(Tn ). ¯ then nUn in (3.13) is exactly the same as It can be shown that if Tn = X, 2 the sample variance S (exercise).

176

3. Unbiased Estimation

3.2.2 Variances of U-statistics If E[h(X1 , ..., Xm )]2 < ∞, then the variance of Un in (3.11) with kernel h has an explicit form. To derive Var(Un ), we need some notation. For k = 1, ..., m, let hk (x1 , ..., xk ) = E[h(X1 , ..., Xm )|X1 = x1 , ..., Xk = xk ] = E[h(x1 , ..., xk , Xk+1 , ..., Xm )]. Note that hm = h. It can be shown that hk (x1 , ..., xk ) = E[hk+1 (x1 , ..., xk , Xk+1 )].

(3.14)

˜ k = hk − E[h(X1 , ..., Xm )], h

(3.15)

Define

˜=˜ k = 1, ..., m, and h hm . Then, for any Un defined by (3.11), −1 X n ˜ i1 , ..., Xim ). h(X Un − E(Un ) = m c

(3.16)

Theorem 3.4 (Hoeffding’s theorem). For a U-statistic Un given by (3.11) with E[h(X1 , ..., Xm )]2 < ∞, −1 X m m n−m n ζk , Var(Un ) = k m−k m k=1

where ζk = Var(hk (X1 , ..., Xk )). Proof. Consider two sets {i1 , ..., im } and {j1 , ..., jm } of m distinct integers from {1, ..., n} with exactly k integers in common. The number of distinct n m n−m ˜ choices of two such sets is m k m−k . By the symmetry of hm and independence of X1 , ..., Xn , ˜ j , ..., Xj )] = ζk ˜ i , ..., Xi )h(X E[h(X 1 m 1 m for k = 1, ..., m (exercise). Then, by (3.16), −2 X X n ˜ i , ..., Xi )h(X ˜ j , ..., Xj )] E[h(X Var(Un ) = 1 m 1 m m c c −2 X m n m n−m n ζk . = m k m−k m k=1

This proves the result.

(3.17)

177

3.2. U-Statistics

Corollary 3.2. Under the condition of Theorem 3.4, 2 (i) mn ζ1 ≤ Var(Un ) ≤ m n ζm ; (ii) (n + 1)Var(Un+1 ) ≤ nVar(Un ) for any n > m; (iii) For any fixed m and k = 1, ..., m, if ζj = 0 for j < k and ζk > 0, then 2 k! m ζk 1 k Var(Un ) = . + O nk nk+1 It follows from Corollary 3.2 that a U-statistic Un as an estimator of its mean is consistent in mse (under the finite second moment assumption on h). In fact, for any fixed m, if ζj = 0 for j < k and ζk > 0, then the mse of Un is of the order n−k and, therefore, Un is nk/2 -consistent. Example 3.11. Consider first h(x1 , x2 ) = x1 x2 , which leads to a U˜ 1 (x1 ) = statistic unbiased for µ2 , µ = EX1 . Note that h1 (x1 ) = µx1 , h ˜ 1 (X1 )]2 = µ2 Var(X1 ) = µ2 σ 2 , h(x ˜ 1 , x2 ) = x1 x2 − µ2 , µ(x1 − µ), ζ1 = E[h and ζ2 = Var(X1 X2 ) = E(X1 X2 )2 − µ4 = (µ2 + σ 2 )2 − µ4 . By Theorem −1 P 3.4, for Un = n2 1≤i<j≤n Xi Xj , −1 2 n−2 n 2 n−2 Var(Un ) = ζ1 + ζ2 1 1 2 2 0 2 2(n − 2)µ2 σ 2 + (µ2 + σ 2 )2 − µ4 = n(n − 1) 2σ 4 4µ2 σ 2 + . = n n(n − 1)

¯ 2 − σ 2 /n in Example 3.10, which is the UMVUE Comparing Un with X under the normality and known σ 2 assumption, we find that ¯ 2 − σ 2 /n) = Var(Un ) − Var(X

2σ 4 . n2 (n − 1)

Next, consider h(x1 , x2 ) = I(−∞,0] (x1 + x2 ), which leads to the onesample Wilcoxon statistic. Note that h1 (x1 ) = P (x1 + X2 ≤ 0) = F (−x1 ), where F is the c.d.f. of P . Then ζ1 = Var(F (−X1 )). Let ϑ = E[h(X1 , X2 )]. Then ζ2 = Var(h(X1 , X2 )) = ϑ(1 − ϑ). Hence, for Un being the one-sample Wilcoxon statistic, Var(Un ) =

2 [2(n − 2)ζ1 + ϑ(1 − ϑ)] . n(n − 1)

If F is continuous and symmetric about 0, then ζ1 can be simplified as ζ1 = Var(F (−X1 )) = Var(1 − F (X1 )) = Var(F (X1 )) =

1 12 ,

178

3. Unbiased Estimation

since F (X1 ) has the uniform distribution on [0, 1]. Finally, consider h(x1 , x2 ) = |x1 − x2 |, which leads to Gini’s mean difference. Note that Z h1 (x1 ) = E|x1 − X2 | = |x1 − y|dP (y),

and

ζ1 = Var(h1 (X1 )) = where ϑ = E|X1 − X2 |.

Z Z

2 |x − y|dP (y) dP (x) − ϑ2 ,

3.2.3 The projection method Since P is nonparametric, the exact distribution of any U-statistic is hard to derive. In this section, we study asymptotic distributions of U-statistics by using the method of projection. Definition 3.3. Let Tn be a given statistic based on X1 , ..., Xn . The projection of Tn on kn random elements Y1 , ..., Ykn is defined to be Tˇn = E(Tn ) +

kn X i=1

[E(Tn |Yi ) − E(Tn )].

Let ψn (Xi ) = E(Tn |Xi ). If Tn is symmetric (as a function of X1 , ..., Xn ), then ψn (X1 ), ..., ψn (Xn ) are i.i.d. with mean E[ψn (Xi )] = E[E(Tn |Xi )] = E(Tn ). If E(Tn2 ) < ∞ and Var(ψn (Xi )) > 0, then n

X 1 p [ψn (Xi ) − E(Tn )] →d N (0, 1) nVar(ψn (X1 )) i=1

(3.18)

by the CLT. Let Tˇn be the projection of Tn on X1 , ..., Xn . Then Tn − Tˇn = Tn − E(Tn ) −

n X i=1

[ψn (Xi ) − E(Tn )].

(3.19)

If we can show that Tn − Tˇn has a negligible order of magnitude, then we can derive the asymptotic distribution of Tn by using (3.18)-(3.19) and Slutsky’s theorem. The order of magnitude of Tn − Tˇn can be obtained with the help of the following lemma. Lemma 3.1. Let Tn be a symmetric statistic with Var(Tn ) < ∞ for every n and Tˇn be the projection of Tn on X1 , ..., Xn . Then E(Tn ) = E(Tˇn ) and E(Tn − Tˇn )2 = Var(Tn ) − Var(Tˇn ).

179

3.2. U-Statistics Proof. Since E(Tn ) = E(Tˇn ), E(Tn − Tˇn )2 = Var(Tn ) + Var(Tˇn ) − 2Cov(Tn , Tˇn ). From Definition 3.3 with Yi = Xi and kn = n, Var(Tˇn ) = nVar(E(Tn |Xi )). The result follows from Cov(Tn , Tˇn ) = E(Tn Tˇn ) − [E(Tn )]2 = nE[Tn E(Tn |Xi )] − n[E(Tn )]2

= nE{E[Tn E(Tn |Xi )|Xi ]} − n[E(Tn )]2 = nE{[E(Tn |Xi )]2 } − n[E(Tn )]2 = nVar(E(Tn |Xi )) = Var(Tˇn ).

This method of deriving the asymptotic distribution of Tn is known as the method of projection and is particularly effective for U-statistics. For a U-statistic Un given by (3.11), one can show (exercise) that n

X ˜ 1 (Xi ), ˇn = E(Un ) + m U h n i=1

(3.20)

ˇn is the projection of Un on X1 , ..., Xn and ˜h1 is defined by (3.15). where U Hence ˇn ) = m2 ζ1 /n Var(U and, by Corollary 3.2 and Lemma 3.1, ˇn )2 = O(n−2 ). E(Un − U If ζ1 > 0, then (3.18) holds with ψn (Xi ) = mh1 (Xi ), which leads to the result in Theorem 3.5(i) stated later. ˜ 1 ≡ 0 and we have to use another projection of Un . If ζ1 = 0, then h Suppose that ζ1 = · · · = ζk−1 = 0 and ζk > 0 for an integer k > 1. ˇkn of Un on n random vectors {Xi1 , ..., Xi }, Consider the projection U k k 1 ≤ i1 < · · · < ik ≤ n. We can establish a result similar to that in Lemma 3.1 (exercise) and show that ˇn )2 = O(n−(k+1) ). E(Un − U Also, see Serfling (1980, §5.3.4). With these results, we obtain the following theorem.

180

3. Unbiased Estimation

Theorem 3.5. Let Un be given by (3.11) with E[h(X1 , ..., Xm )]2 < ∞. (i) If ζ1 > 0, then √ n[Un − E(Un )] →d N (0, m2 ζ1 ). (ii) If ζ1 = 0 but ζ2 > 0, then ∞

m(m − 1) X λj (χ21j − 1), n[Un − E(Un )] →d 2 j=1

(3.21)

where χ21j ’s are i.i.d. random variables having the chi-square distribution χ21 P∞ and λj ’s are some constants (which may depend on P ) satisfying j=1 λ2j = ζ2 . We have actually proved Theorem 3.5(i). A proof for Theorem 3.5(ii) is given in Serfling (1980, §5.5.2). One may derive results for the cases where ζ2 = 0, but the case of either ζ1 > 0 or ζ2 > 0 is the most interesting case in applications. If ζ1 > 0, it follows from Theorem 3.5(i) and Corollary 3.2(iii) that amseUn (P ) = m2 ζ1 /n = Var(Un ) + O(n−2 ). By Proposition 2.4(ii), {n[Un − E(Un )]2 } is uniformly integrable. If ζ1 = 0 but ζ2 > 0, it follows from Theorem 3.5(ii) that amseUn (P ) = EY 2 /n2 , where Y denotes the random variable on the right-hand side of (3.21). The following result provides the value of EY 2 . Lemma 3.2. Let Y be the random variable on the right-hand side of 2 2 ζ2 . (3.21). Then EY 2 = m (m−1) 2 Proof. Define k

Yk =

m(m − 1) X λj (χ21j − 1), 2 j=1

k = 1, 2, ....

It can be shown (exercise) that {Yk2 } is uniformly integrable. Since Yk →d Y as k → ∞, limk→∞ EYk2 = EY 2 (Theorem 1.8(viii)). Since χ21j ’s are independent chi-square random variables with Eχ21j = 1 and Var(χ21j ) = 2, EYk = 0 for any k and k m2 (m − 1)2 X 2 λj Var(χ21j ) 4 j=1 k m2 (m − 1)2 X 2 = λj 2 4 j=1

EYk2 =

→

m2 (m − 1)2 ζ2 . 2

181

3.2. U-Statistics

It follows from Corollary 3.2(iii) and Lemma 3.2 that amseUn (P ) = = Var(Un ) + O(n−3 ) if ζ1 = 0. Again, by Proposition 2.4(ii), the sequence {n2 [Un − E(Un )]2 } is uniformly integrable. We now apply in Example 3.11. For P Theorem 3.5 to the2 U-statistics 2 2 Un = n(n−1) X X , ζ = µ σ . Thus, if µ 6= 0, the result in i j 1 1≤i<j≤n Theorem 3.5(i) holds with ζ1 = µ2 σ 2 . If µ = 0, then ζ1 = 0, ζ2 = σ 4 > 0, and Theorem 3.5(ii) applies. However, it is not convenient to use Theorem 3.5(ii) to find the limiting distribution of Un . We may derive this limiting distribution using the following technique, which is further discussed in §3.5. By the CLT and Theorem 1.10, m2 (m−1)2 ζ2 /n2 2

¯ 2 /σ 2 →d χ2 nX 1 when µ = 0, where χ21 is a random variable having the chi-square distribution χ21 . Note that n

¯2 nX 1 X 2 (n − 1)Un = 2 X + . 2 σ σ n i=1 i σ2

By the SLLN, leads to

1 σ2 n

Pn

i=1

Xi2 →a.s. 1. An application of Slutsky’s theorem nUn /σ 2 →d χ21 − 1.

Since µ = 0, this implies that the right-hand side of (3.21) is σ 2 (χ21 − 1), i.e., λ1 = σ 2 and λj = 0 when j > 1. For the one-sample Wilcoxon statistic, ζ1 = Var(F (−X1 )) > 0 unless F is degenerate. Similarly, for Gini’s mean difference, ζ1 > 0 unless F is degenerate. Hence Theorem 3.5(i) applies to these two cases. Theorem 3.5 does not apply to Un defined by (3.13) if r, the order of the kernel, depends on n and diverges to ∞ as n → ∞. We consider the simple case where n 1X Tn = ψ(Xi ) + Rn (3.22) n i=1 for some Rn satisfying E(Rn2 ) = o(n−1 ). Note that (3.22) is satisfied for Tn being a U-statistic (exercise). Assume that r/d is bounded. Let Sψ2 = P P (n − 1)−1 ni=1 [ψ(Xi ) − n−1 ni=1 ψ(Xi )]2 . Then nUn = Sψ2 + op (1)

(3.23)

(exercise). Under (3.22), if 0 < E[ψ(Xi )]2 < ∞, then amseTn (P ) = E[ψ(Xi )]2 /n. Hence, the jackknife estimator Un in (3.13) provides a consistent estimator of amseTn (P ), i.e., Un /amseTn (P ) →p 1.

182

3. Unbiased Estimation

3.3 The LSE in Linear Models One of the most useful statistical models for non-i.i.d. data in applications is the general linear model Xi = β τ Z i + ε i ,

i = 1, ..., n,

(3.24)

where Xi is the ith observation and is often called the ith response; β is a p-vector of unknown parameters, p < n; Zi is the ith value of a pvector of explanatory variables (or covariates); and ε1 , ..., εn are random errors. Our data in this case are (X1 , Z1 ), ..., (Xn , Zn ) (εi ’s are not observed). Throughout this book Zi ’s are considered to be nonrandom or given values of a random p-vector, in which case our analysis is conditioned on Z1 , ..., Zn . Each εi can be viewed as a random measurement error in measuring the unknown mean of Xi when the covariate vector is equal to Zi . The main parameter of interest is β. More specific examples of model (3.24) are provided in this section. Other examples and examples of data from model (3.24) can be found in many standard books for linear models, for example, Draper and Smith (1981) and Searle (1971).

3.3.1 The LSE and estimability Let X = (X1 , ..., Xn ), ε = (ε1 , ..., εn ), and Z be the n × p matrix whose ith row is the vector Zi , i = 1, ..., n. Then, a matrix form of model (3.24) is X = Zβ + ε.

(3.25)

Definition 3.4. Suppose that the range of β in model (3.25) is B ⊂ Rp . A least squares estimator (LSE) of β is defined to be any βˆ ∈ B such that ˆ 2 = min kX − Zbk2 . kX − Z βk b∈B

(3.26)

For any l ∈ Rp , lτ βˆ is called an LSE of lτ β. Throughout this book, we consider B = Rp unless otherwise stated. Differentiating kX − Zbk2 w.r.t. b, we obtain that any solution of Z τ Zb = Z τ X

(3.27)

is an LSE of β. If the rank of the matrix Z is p, in which case (Z τ Z)−1 exists and Z is said to be of full rank, then there is a unique LSE, which is βˆ = (Z τ Z)−1 Z τ X.

(3.28)

183

3.3. The LSE in Linear Models

If Z is not of full rank, then there are infinitely many LSE’s of β. It can be shown (exercise) that any LSE of β is of the form βˆ = (Z τ Z)− Z τ X, τ

−

where (Z Z)

(3.29) τ

is called a generalized inverse of Z Z and satisfies Z τ Z(Z τ Z)− Z τ Z = Z τ Z.

Generalized inverse matrices are not unique unless Z is of full rank, in which case (Z τ Z)− = (Z τ Z)−1 and (3.29) reduces to (3.28). To study properties of LSE’s of β, we need some assumptions on the distribution of X. Since Zi ’s are nonrandom, assumptions on the distribution of X can be expressed in terms of assumptions on the distribution of ε. Several commonly adopted assumptions are stated as follows. Assumption A1: ε is distributed as Nn (0, σ 2 In ) with an unknown σ 2 > 0. Assumption A2: E(ε) = 0 and Var(ε) = σ 2 In with an unknown σ 2 > 0. Assumption A3: E(ε) = 0 and Var(ε) is an unknown matrix. Assumption A1 is the strongest and implies a parametric model. We may assume a slightly more general assumption that ε has the Nn (0, σ 2 D) distribution with unknown σ 2 but a known positive definite matrix D. Let D−1/2 be the inverse of the square root matrix of D. Then model (3.25) with assumption A1 holds if we replace X, Z, and ε by the transformed ˜ = D−1/2 X, Z˜ = D−1/2 Z, and ε˜ = D−1/2 ε, respectively. A variables X similar conclusion can be made for assumption A2. Under assumption A1, the distribution of X is Nn (Zβ, σ 2 In ), which is in an exponential family P with parameter θ = (β, σ 2 ) ∈ Rp × (0, ∞). However, if the matrix Z is not of full rank, then P is not identifiable (see §2.1.2), since Zβ1 = Zβ2 does not imply β1 = β2 . Suppose that the rank of Z is r ≤ p. Then there is an n × r submatrix Z∗ of Z such that Z = Z∗ Q (3.30) and Z∗ is of rank r, where Q is a fixed r × p matrix. Then Zβ = Z∗ Qβ and P is identifiable if we consider the reparameterization β˜ = Qβ. Note that the new parameter β˜ is in a subspace of Rp with dimension r. In many applications, we are interested in estimating some linear functions of β, i.e., ϑ = lτ β for some l ∈ Rp . From the previous discussion, however, estimation of lτ β is meaningless unless l = Qτ c for some c ∈ Rr so that ˜ lτ β = cτ Qβ = cτ β.

184

3. Unbiased Estimation

The following result shows that lτ β is estimable if l = Qτ c, which is also necessary for lτ β to be estimable under assumption A1. Theorem 3.6. Assume model (3.25) with assumption A3. (i) A necessary and sufficient condition for l ∈ Rp being Qτ c for some c ∈ Rr is l ∈ R(Z) = R(Z τ Z), where Q is given by (3.30) and R(A) is the smallest linear subspace containing all rows of A. (ii) If l ∈ R(Z), then the LSE lτ βˆ is unique and unbiased for lτ β. (iii) If l 6∈ R(Z) and assumption A1 holds, then lτ β is not estimable. Proof. (i) Note that a ∈ R(A) if and only if a = Aτ b for some vector b. If l = Qτ c, then l = Qτ c = Qτ Z∗τ Z∗ (Z∗τ Z∗ )−1 c = Z τ [Z∗ (Z∗τ Z∗ )−1 c]. Hence l ∈ R(Z). If l ∈ R(Z), then l = Z τ ζ for some ζ and l = (Z∗ Q)τ ζ = Qτ c with c = Z∗τ ζ. (ii) If l ∈ R(Z) = R(Z τ Z), then l = Z τ Zζ for some ζ and by (3.29), ˆ = E[lτ (Z τ Z)− Z τ X] E(lτ β) = ζ τ Z τ Z(Z τ Z)− Z τ Zβ = ζ τ Z τ Zβ = lτ β. If β¯ is any other LSE of β, then, by (3.27), ¯ = ζ τ (Z τ X − Z τ X) = 0. lτ βˆ − lτ β¯ = ζ τ (Z τ Z)(βˆ − β) (iii) Under assumption A1, if there is an estimator h(X, Z) unbiased for lτ β, then Z h(x, Z)(2π)−n/2 σ −n exp − 2σ1 2 kx − Zβk2 dx. lτ β = Rn

Differentiating w.r.t. β and applying Theorem 2.1 lead to Z h(x, Z)(2π)−n/2 σ −n−2 (x − Zβ) exp − 2σ1 2 kx − Zβk2 dx, l = Zτ Rn

which implies l ∈ R(Z).

Theorem 3.6 shows that LSE’s are unbiased for estimable parameters lτ β. If Z is of full rank, then R(Z) = Rp and, therefore, lτ β is estimable for any l ∈ Rp .

185

3.3. The LSE in Linear Models

Example 3.12 (Simple linear regression). Let β = (β0 , β1 ) ∈ R2 and Zi = (1, ti ), ti ∈ R, i = 1, ..., n. Then model (3.24) or (3.25) is called a simple linear regression model. It turns out that Pn t n Pn Pni=1 2i Zτ Z = . i=1 ti i=1 ti

This matrix is invertible if and only if some ti ’s are different. Thus, if some ti ’s are different, then the unique unbiased LSE of lτ β for any l ∈ R2 is lτ (Z τ Z)−1 Z τ X, which has the normal distribution if assumption A1 holds. The result can be easily extended to the case of polynomial regression of order p in which β = (β0 , β1 , ..., βp−1 ) and Zi = (1, ti , ..., tp−1 ). i Example 3.13 (One-way ANOVA). Suppose that n = positive integers n1 , ..., nm and that

Pm

j=1

nj with m

Xi = µj + εi , i = kj−1 + 1, ..., kj , j = 1, ..., m, P where k0 = 0, kj = jl=1 nl , j = 1, ..., m, and (µ1 , ..., µm ) = β. Let Jm be the m-vector of ones. Then the matrix Z in this case is a block diagonal matrix with Jnj as the jth diagonal column. Consequently, Z τ Z is an m × m diagonal matrix whose jth diagonal element is nj . Thus, Z τ Z is invertible and the unique LSE of β is the m-vector whose jth component Pkj is n−1 j i=kj−1 +1 Xi , j = 1, ..., m. Sometimes it is more convenient to use the following notation: Xij = Xki−1 +j , εij = εki−1 +j ,

j = 1, ..., ni , i = 1, ..., m,

and µi = µ + αi ,

i = 1, ..., m.

Then our model becomes Xij = µ + αi + εij ,

j = 1, ..., ni , i = 1, ..., m,

(3.31)

which is called a one-way analysis of variance (ANOVA) model. Under model (3.31), β = (µ, α1 , ..., αm ) ∈ Rm+1 . The matrix Z under model (3.31) is not of full rank (exercise). An LSE of β under model (3.31) is ¯ X ¯ 1· − X, ¯ ..., X ¯ m· − X ¯ , βˆ = X, ¯ is still the sample mean of Xij ’s and X ¯ i· is the sample mean of the where X ith group {Xij , j = 1, ..., ni }. The problem of finding the form of l ∈ R(Z) under model (3.31) is left as an exercise.

The notation used in model (3.31) allows us to generalize the one-way ANOVA model to any s-way ANOVA model with a positive integer s under

186

3. Unbiased Estimation

the so-called factorial experiments. The following example is for the twoway ANOVA model. Example 3.14 (Two-way balanced ANOVA). Suppose that Xijk = µ + αi + βj + γij + εijk ,

i = 1, ..., a, j = 1, ..., b, k = 1, ..., c, (3.32)

where a, b, and c are some positive integers. Model (3.32) is called a twoway balanced ANOVA model. If we view model (3.32) as a special case of model (3.25), then the parameter vector β is β = (µ, α1 , ..., αa , β1 , ..., βb , γ11 , ..., γ1b , ..., γa1 , ..., γab ).

(3.33)

One can obtain the matrix Z and show that it is n × p, where n = abc and p = 1 + a + b + ab, and is of rank ab < p (exercise). It can also be shown (exercise) that an LSE of β is given by the right-hand side of (3.33) with µ, ¯ ··· , αi , βj , and γij replaced by µ ˆ, α ˆ i , βˆj , and γˆij , respectively, where µ ˆ=X ¯ i·· − X ¯ ··· , βˆj = X ¯ ·j· − X ¯ ··· , γˆij = X ¯ ij· − X ¯ i·· − X ¯ ·j· + X ¯ ··· , and a dot α ˆi = X is used to denote averaging over the indicated subscript, e.g., a

c

XX ¯ ·j· = 1 Xijk X ac i=1 k=1

with a fixed j.

3.3.2 The UMVUE and BLUE We now study UMVUE’s in model (3.25) with assumption A1. Theorem 3.7. Consider model (3.25) with assumption A1. (i) The LSE lτ βˆ is the UMVUE of lτ β for any estimable lτ β. ˆ 2 , where r is the rank ˆ 2 = (n − r)−1 kX − Z βk (ii) The UMVUE of σ 2 is σ of Z. Proof. (i) Let βˆ be an LSE of β. By (3.27), ˆ τ Z(βˆ − β) = (X τ Z − X τ Z)(βˆ − β) = 0 (X − Z β) and, hence, kX − Zβk2 = kX − Z βˆ + Z βˆ − Zβk2 ˆ 2 + kZ βˆ − Zβk2 = kX − Z βk

ˆ 2 − 2β τ Z τ X + kZβk2 + kZ βk ˆ 2. = kX − Z βk

Using this result and assumption A1, we obtain the following joint Lebesgue p.d.f. of X: o n τ τ ˆ 2 ˆ 2 βk kZβk2 . − (2πσ 2 )−n/2 exp β σZ2 x − kx−Z βk2σ+kZ 2 2σ2

187

3.3. The LSE in Linear Models

By Proposition 2.1 and the fact that Z βˆ = Z(Z τ Z)− Z τ X is a function of ˆ 2 ) is complete and sufficient for θ = (β, σ 2 ). Note Z τ X, (Z τ X, kX − Z βk ˆ that β is a function of Z τ X and, hence, a function of the complete sufficient statistic. If lτ β is estimable, then lτ βˆ is unbiased for lτ β (Theorem 3.6) and, hence, lτ βˆ is the UMVUE of lτ β. ˆ 2 + kZ βˆ − Zβk2 and E(Z β) ˆ = Zβ (ii) From kX − Zβk2 = kX − Z βk (Theorem 3.6), ˆ 2 = E(X − Zβ)τ (X − Zβ) − E(β − β) ˆ τ Z τ Z(β − β) ˆ EkX − Z βk ˆ = tr Var(X) − Var(Z β) = σ 2 [n − tr Z(Z τ Z)− Z τ Z(Z τ Z)− Z τ ] = σ 2 [n − tr (Z τ Z)− Z τ Z ].

Since each row of Z ∈ R(Z), Z βˆ does not depend on the choice of (Z τ Z)− in βˆ = (Z τ Z)− Z τ X (Theorem 3.6). Hence, we can evaluate tr((Z τ Z)− Z τ Z) using a particular (Z τ Z)− . From the theory of linear algebra, there exists a p × p matrix C such that CC τ = Ip and Λ 0 , C τ (Z τ Z)C = 0 0 where Λ is an r × r diagonal matrix whose diagonal elements are positive. Then, a particular choice of (Z τ Z)− is −1 0 Λ τ − Cτ (Z Z) = C (3.34) 0 0 and τ

−

τ

(Z Z) Z Z = C

Ir 0

0 0

Cτ

whose trace is r. Hence σ ˆ 2 is the UMVUE of σ 2 , since it is a function of the complete sufficient statistic and ˆ 2 = σ2 . Eσ ˆ 2 = (n − r)−1 EkX − Z βk In general, ˆ = lτ (Z τ Z)− Z τ Var(ε)Z(Z τ Z)− l. Var(lτ β)

(3.35)

If l ∈ R(Z) and Var(ε) = σ 2 In (assumption A2), then the use of the genˆ = σ 2 lτ (Z τ Z)− l, which eralized inverse matrix in (3.34) leads to Var(lτ β) attains the Cram´er-Rao lower bound under assumption A1 (Proposition 3.2).

188

3. Unbiased Estimation

ˆ 2 is called The vector X − Z βˆ is called the residual vector and kX − Z βk the sum of squared residuals and is denoted by SSR. The estimator σ ˆ 2 is then equal to SSR/(n − r). Since X − Z βˆ = [In − Z(Z τ Z)− Z τ ]X and lτ βˆ = lτ (Z τ Z)− Z τ X are linear in X, they are normally distributed under assumption A1. Also, using the generalized inverse matrix in (3.34), we obtain that [In − Z(Z τ Z)− Z τ ]Z(Z τ Z)− = Z(Z τ Z)− − Z(Z τ Z)− Z τ Z(Z τ Z)− = 0, which implies that σ ˆ 2 and lτ βˆ are independent (Exercise 58 in §1.6) for any τ estimable l β. Furthermore, [Z(Z τ Z)− Z τ ]2 = Z(Z τ Z)− Z τ (i.e., Z(Z τ Z)− Z τ is a projection matrix) and SSR = X τ [In − Z(Z τ Z)− Z τ ]X. The rank of Z(Z τ Z)− Z τ is tr(Z(Z τ Z)− Z τ ) = r. Similarly, the rank of the projection matrix In − Z(Z τ Z)− Z τ is n − r. From X τ X = X τ [Z(Z τ Z)− Z τ ]X + X τ [In − Z(Z τ Z)− Z τ ]X and Theorem 1.5 (Cochran’s theorem), SSR/σ 2 has the chi-square distribution χ2n−r (δ) with δ = σ −2 β τ Z τ [In − Z(Z τ Z)− Z τ ]Zβ = 0. Thus, we have proved the following result. Theorem 3.8. Consider model (3.25) with assumption A1. For any esˆ 2 are independent; the timable parameter lτ β, the UMVUE’s lτ βˆ and σ τˆ τ 2 τ τ − distribution of l β is N (l β, σ l (Z Z) l); and (n − r)ˆ σ 2 /σ 2 has the chi2 square distribution χn−r . Example 3.15. In Examples 3.12-3.14, UMVUE’s of estimable lτ β are the ˆ under assumption A1. In Example 3.13, LSE’s lτ β, SSR =

ni m X X i=1 j=1

¯ i· )2 ; (Xij − X

in Example 3.14, if c > 1, SSR =

b X c a X X i=1 j=1 k=1

¯ ij· )2 . (Xijk − X

189

3.3. The LSE in Linear Models

ˆ 2 under assumption A2, i.e., withWe now study properties of lτ βˆ and σ out the normality assumption on ε. From Theorem 3.6 and the proof of Theorem 3.7(ii), lτ βˆ (with an l ∈ R(Z)) and σ ˆ 2 are still unbiased without ˆ 2 optimal beyond the normality assumption. In what sense are lτ βˆ and σ ˆ Some disbeing unbiased? We have the following result for the LSE lτ β. 2 cussion about σ ˆ can be found, for example, in Rao (1973, p. 228). Theorem 3.9. Consider model (3.25) with assumption A2. (i) A necessary and sufficient condition for the existence of a linear unbiased estimator of lτβ (i.e., an unbiased estimator that is linear in X) is l ∈ R(Z). (ii) (Gauss-Markov theorem). If l ∈ R(Z), then the LSE lτ βˆ is the best linear unbiased estimator (BLUE) of lτ β in the sense that it has the minimum variance in the class of linear unbiased estimators of lτ β. Proof. (i) The sufficiency has been established in Theorem 3.6. Suppose now a linear function of X, cτ X with c ∈ Rn , is unbiased for lτ β. Then lτ β = E(cτ X) = cτ EX = cτ Zβ.

Since this equality holds for all β, l = Z τ c, i.e., l ∈ R(Z). (ii) Let l ∈ R(Z) = R(Z τ Z). Then l = (Z τ Z)ζ for some ζ and lτ βˆ = ζ τ (Z τ Z)βˆ = ζ τ Z τ X by (3.27). Let cτ X be any linear unbiased estimator of lτ β. From the proof of (i), Z τ c = l. Then Cov(ζ τ Z τ X, cτ X − ζ τ Z τ X) = E(X τ Zζcτ X) − E(X τ Zζζ τ Z τ X) = σ 2 tr(Zζcτ ) + β τ Z τ Zζcτ Zβ − σ 2 tr(Zζζ τ Z τ ) − β τ Z τ Zζζ τ Z τ Zβ = σ 2 ζ τ l + (lτ β)2 − σ 2 ζ τ l − (lτ β)2

= 0. Hence

Var(cτ X) = Var(cτ X − ζ τ Z τ X + ζ τ Z τ X) = Var(cτ X − ζ τ Z τ X) + Var(ζ τ Z τ X) + 2Cov(ζ τ Z τ X, cτ X − ζ τ Z τ X) ˆ = Var(cτ X − ζ τ Z τ X) + Var(lτ β) τˆ ≥ Var(l β).

3.3.3 Robustness of LSE’s Consider now model (3.25) under assumption A3. An interesting question is under what conditions on Var(ε) is the LSE of lτ β with l ∈ R(Z) ˆ considered still the BLUE. If lτ βˆ is still the BLUE, then we say that lτ β, as a BLUE, is robust against violation of assumption A2. In general, a

190

3. Unbiased Estimation

statistical procedure having certain properties under an assumption is said to be robust against violation of the assumption if and only if the statistical procedure still has the same properties when the assumption is (slightly) violated. For example, the LSE of lτ β with l ∈ R(Z), as an unbiased estimator, is robust against violation of assumption A1 or A2, since the LSE is unbiased as long as E(ε) = 0, which can be always assumed without loss of generality. On the other hand, the LSE as a UMVUE may not be robust against violation of assumption A1 (see §3.5). Theorem 3.10. Consider model (3.25) with assumption A3. The following are equivalent. (a) lτ βˆ is the BLUE of lτ β for any l ∈ R(Z). ˆ τ X) = 0 for any l ∈ R(Z) and any η such that E(η τ X) = 0. (b) E(lτ βη τ (c) Z Var(ε)U = 0, where U is a matrix such that Z τ U = 0 and R(U τ ) + R(Z τ ) = Rn . (d) Var(ε) = ZΛ1 Z τ + U Λ2 U τ for some Λ1 and Λ2 . (e) The matrix Z(Z τ Z)− Z τ Var(ε) is symmetric. Proof. We first show that (a) and (b) are equivalent, which is an analogue of Theorem 3.2(i). Suppose that (b) holds. Let l ∈ R(Z). If cτ X is unbiased for lτ β, then E(η τ X) = 0 with η = c − Z(Z τ Z)− l. Hence ˆ Var(cτ X) = Var(cτ X − lτ βˆ + lτ β)

ˆ = Var(cτ X − lτ (Z τ Z)− Z τ X + lτ β) ˆ = Var(η τ X + lτ β) ˆ + 2Cov(η τ X, lτ β) ˆ = Var(η τ X) + Var(lτ β) ˆ + 2E(lτ βη ˆ τ X) = Var(η τ X) + Var(lτ β) ˆ = Var(η τ X) + Var(lτ β) ˆ ≥ Var(lτ β).

Suppose now that there are l ∈ R(Z) and η such that E(η τ X) = 0 but ˆ τ X) 6= 0. Let ct = tη + Z(Z τ Z)− l. From the previous proof, δ = E(lτ βη ˆ + 2δt. Var(cτt X) = t2 Var(η τ X) + Var(lτ β)

ˆ This As long as δ 6= 0, there exists a t such that Var(cτt X) < Var(lτ β). τˆ shows that l β cannot be a BLUE and, therefore, (a) implies (b). We next show that (b) implies (c). Suppose that (b) holds. Since l ∈ R(Z), l = Z τ γ for some γ. Let η ∈ R(U τ ). Then E(η τ X) = η τ Zβ = 0 and, hence, ˆ τ X) = E[γ τ Z(Z τ Z)− Z τ XX τ η] = γ τ Z(Z τ Z)− Z τ Var(ε)η. 0 = E(lτ βη Since this equality holds for all l ∈ R(Z), it holds for all γ. Thus, Z(Z τ Z)− Z τ Var(ε)U = 0,

3.3. The LSE in Linear Models

191

which implies Z τ Z(Z τ Z)− Z τ Var(ε)U = Z τ Var(ε)U = 0, since Z τ Z(Z τ Z)− Z τ = Z τ . Thus, (c) holds. To show that (c) implies (d), we need to use the following facts from the theory of linear algebra: there exists a nonsingular matrix C such that Var(ε) = CC τ and C = ZC1 + U C2 for some matrices Cj (since R(U τ ) + R(Z τ ) = Rn ). Let Λ1 = C1 C1τ , Λ2 = C2 C2τ , and Λ3 = C1 C2τ . Then Var(ε) = ZΛ1 Z τ + U Λ2 U τ + ZΛ3 U τ + U Λτ3 Z τ (3.36) and Z τ Var(ε)U = Z τ ZΛ3 U τ U , which is 0 if (c) holds. Hence, (c) implies 0 = Z(Z τ Z)− Z τ ZΛ3 U τ U (U τ U )− U τ = ZΛ3 U τ , which with (3.36) implies (d). If (d) holds, then Z(Z τ Z)− Z τ Var(ε) = ZΛ1 Z τ , which is symmetric. Hence (d) implies (e). To complete the proof, we need to show that (e) implies (b), which is left as an exercise. As a corollary of this theorem, the following result shows when the UMVUE’s in model (3.25) with assumption A1 are robust against the violation of Var(ε) = σ 2 In . Corollary 3.3. Consider model (3.25) with a full rank Z, ε = Nn (0, Σ), and an unknown positive definite matrix Σ. Then lτ βˆ is a UMVUE of lτ β for any l ∈ Rp if and only if one of (b)-(e) in Theorem 3.10 holds. Example 3.16. Consider model (3.25) with β replaced by a random vector β that is independent of ε. Such a model is called a linear model with random coefficients. Suppose that Var(ε) = σ 2 In and E(β) = β. Then X = Zβ + Z(β − β) + ε = Zβ + e,

(3.37)

where e = Z(β − β) + ε satisfies E(e) = 0 and Var(e) = ZVar(β)Z τ + σ 2 In . Since Z(Z τ Z)− Z τ Var(e) = ZVar(β)Z τ + σ 2 Z(Z τ Z)− Z τ is symmetric, by Theorem 3.10, the LSE lτ βˆ under model (3.37) is the BLUE for any lτ β, l ∈ R(Z). If Z is of full rank and ε is normal, then, by Corollary 3.3, lτ βˆ is the UMVUE of lτ β for any l ∈ Rp .

192

3. Unbiased Estimation

Example 3.17 (Random effects models). Suppose that Xij = µ + Ai + eij ,

j = 1, ..., ni , i = 1, ..., m,

(3.38)

where µ ∈ R is an unknown parameter, Ai ’s are i.i.d. random variables having mean 0 and variance σa2 , eij ’s are i.i.d. random errors with mean 0 and variance σ 2 , and Ai ’s and eij ’s are independent. Model (3.38) is called a one-way random effects model and Ai ’s are unobserved random effects. Let εij = Ai + eij . Then (3.38) is a special case of the general model (3.25) with Var(ε) = σa2 Σ + σ 2 In , τ where Σ is a block diagonal matrix whose ith block Pm is Jni Jni and Jτ k is−theτ kvector of ones. Under this model, Z = Jn , n = i=1 ni , and Z(Z Z) Z = n−1 Jn Jnτ . Note that

n1 Jn1 Jnτ1 n1 Jn2 Jnτ1 Jn Jnτ Σ = ······ n1 Jnm Jnτ1

n2 Jn1 Jnτ2 n2 Jn2 Jnτ2 ······ n2 Jnm Jnτ2

· · · nm Jn1 Jnτm · · · nm Jn2 Jnτm , ··· ······ · · · nm Jnm Jnτm

which is symmetric if and only if n1 = n2 = · · · = nm . Since Jn Jnτ Var(ε) is symmetric if and only if Jn Jnτ Σ is symmetric, a necessary and sufficient condition for the LSE of µ to be the BLUE is that all ni ’s are the same. This condition is also necessary and sufficient for the LSE of µ to be the UMVUE when εij ’s are normal. In some cases, we are interested in some (not all) linear functions of β. For example, consider lτ β with l ∈ R(H), where H is an n × p matrix such that R(H) ⊂ R(Z). We have the following result. Proposition 3.4. Consider model (3.25) with assumption A3. Suppose that H is a matrix such that R(H) ⊂ R(Z). A necessary and sufficient condition for the LSE l τ βˆ to be the BLUE of lτ β for any l ∈ R(H) is H(Z τ Z)− Z τ Var(ε)U = 0, where U is the same as that in (c) of Theorem 3.10. Example 3.18. Consider model (3.25) with assumption A3 and Z = (H1 H2 ), where H1τ H2 = 0. Suppose that under the reduced model X = H1 β1 + ε, lτ βˆ1 is the BLUE for any lτ β1 , l ∈ R(H1 ), and that under the reduced model X = H2 β2 + ε,

193

3.3. The LSE in Linear Models

lτ βˆ2 is not a BLUE for some lτ β2 , l ∈ R(H2 ), where β = (β1 , β2 ) and βˆj ’s are LSE’s under the reduced models. Let H = (H1 0) be n × p. Note that H(Z τ Z)− Z τ Var(ε)U = H1 (H1τ H1 )− H1τ Var(ε)U, which is 0 by Theorem 3.10 for the U given in (c) of Theorem 3.10, and Z(Z τ Z)− Z τ Var(ε)U = H2 (H2τ H2 )− H2τ Var(ε)U, which is not 0 by Theorem 3.10. This implies that some LSE lτ βˆ is not a BLUE of lτ β but lτ βˆ is the BLUE of lτ β if l ∈ R(H). Finally, we consider model (3.25) with Var(ε) being a diagonal matrix whose ith diagonal element is σi2 , i.e., εi ’s are uncorrelated but have unequal variances. A straightforward calculation shows that condition (e) in Theorem 3.10 holds if and only if, for all i 6= j, σi2 6= σj2 only when hij = 0, where hij is the (i, j)th element of the projection matrix Z(Z τ Z)− Z τ . Thus, an LSE is not a BLUE in general, although it is still unbiased for estimable lτ β. Suppose that the unequal variances of εi ’s are caused by some small perturbations, i.e., εi = ei + ui , where Var(ei ) = σ 2 , Var(ui ) = δi , and ei and ui are independent so that σi2 = σ 2 + δi . From (3.35), ˆ = lτ (Z τ Z)− Var(lτ β)

n X

σi2 Zi Ziτ (Z τ Z)− l.

i=1

If δi = 0 for all i (no perturbations), then assumption A2 holds and lτ βˆ ˆ = σ 2 lτ (Z τ Z)− l. Suppose is the BLUE of any estimable lτ β with Var(lτ β) 2 that 0 < δi ≤ σ δ. Then ˆ ≤ (1 + δ)σ 2 lτ (Z τ Z)− l. Var(lτ β) This indicates that the LSE is robust in the sense that its variance increases slightly when there is a slight violation of the equal variance assumption (small δ).

3.3.4 Asymptotic properties of LSE’s We consider first the consistency of the LSE lτ βˆ with l ∈ R(Z) for every n. Theorem 3.11. Consider model (3.25) with assumption A3. Suppose that supn λ+ [Var(ε)] < ∞, where λ+ [A] is the largest eigenvalue of the matrix A, and that limn→∞ λ+ [(Z τ Z)− ] = 0. Then lτ βˆ is consistent in mse for

194

3. Unbiased Estimation

any l ∈ R(Z). Proof. The result follows from the fact that lτ βˆ is unbiased and ˆ = lτ (Z τ Z)− Z τ Var(ε)Z(Z τ Z)− l Var(lτ β) ≤ λ+ [Var(ε)]lτ (Z τ Z)− l. Without the normality assumption on ε, the exact distribution of lτ βˆ is very hard to obtain. The asymptotic distribution of lτ βˆ is derived in the following result. Theorem 3.12. Consider model (3.25) with assumption A3. Suppose that 0 < inf n λ− [Var(ε)], where λ− [A] is the smallest eigenvalue of the matrix A, and that lim max Ziτ (Z τ Z)− Zi = 0. (3.39) n→∞ 1≤i≤n

Pk Suppose further that n = j=1 mj for some integers k, mj , j = 1, ..., k, with mj ’s bounded by a fixed integer m, ε = (ξ1 , ..., ξk ), ξj ∈ Rmj , and ξj ’s are independent. (i) If supi E|εi |2+δ < ∞, then for any l ∈ R(Z), q ˆ →d N (0, 1). lτ (βˆ − β) Var(lτ β) (3.40) (ii) Suppose that when mi = mj , 1 ≤ i < j ≤ k, ξi and ξj have the same distribution. Then result (3.40) holds for any l ∈ R(Z). Proof. Let l ∈ R(Z). Then lτ (Z τ Z)− Z τ Zβ − lτ β = 0 and l (βˆ − β) = lτ (Z τ Z)− Z τ ε = τ

k X

cτnj ξj ,

j=1

where cnj is the mj -vector whose components are lτ (Z τ Z)− Zi , i = kj−1 + P 1, ..., kj , k0 = 0, and kj = jt=1 mt , j = 1, ..., k. Note that k X j=1

kcnj k2 = lτ (Z τ Z)− Z τ Z(Z τ Z)− l = lτ (Z τ Z)− l.

Also, max kcnj k2 ≤ m max [lτ (Z τ Z)− Zi ]2

1≤j≤k

1≤i≤n

≤ mlτ (Z τ Z)− l max Ziτ (Z τ Z)− Zi , 1≤i≤n

(3.41)

195

3.4. Unbiased Estimators in Survey Problems

which, together with (3.41) and condition (3.39), implies that X k kcnj k2 = 0. lim max kcnj k2 n→∞

1≤j≤k

j=1

The results then follow from Corollary 1.3.

Under the conditions of Theorem 3.12, Var(ε) is a diagonal block matrix with Var(ξj ) as the jth diagonal block, which includes the case of independent εi ’s as a special case. The following lemma tells us how to check condition (3.39). Lemma 3.3. The following are sufficient conditions for (3.39). (a) λ+ [(Z τ Z)− ] → 0 and Znτ (Z τ Z)− Zn → 0, as n → ∞. (b) There is an increasing sequence {an } such that an → ∞, an /an+1 → 1, and Z τ Z/an converges to a positive definite matrix. P P If n−1 ni=1 t2i → c and n−1 ni=1 ti → d in the simple linear regression model (Example 3.12), where c is positive and c > d2 , then condition (b) in Lemma 3.3 is satisfied with an = n and, therefore, Theorem 3.12 applies. In the one-way ANOVA model (Example 3.13), max Ziτ (Z τ Z)− Zi = λ+ [(Z τ Z)− ] = max n−1 j .

1≤i≤n

1≤j≤m

Hence conditions related to Z in Theorem 3.12 are satisfied if and only if minj nj → ∞. Some similar conclusions can be drawn in the two-way ANOVA model (Example 3.14).

3.4 Unbiased Estimators in Survey Problems In this section, we consider unbiased estimation for another type of noni.i.d. data often encountered in applications: survey data from finite populations. A description of the problem is given in Example 2.3 of §2.1.1. Examples and a fuller account of theoretical aspects of survey sampling can be found, for example, in Cochran (1977) and S¨arndal, Swensson, and Wretman (1992).

3.4.1 UMVUE’s of population totals We use the same notation as in Example 2.3. Let X = (X1 , ..., Xn ) be a sample from a finite population P = {y1 , ..., yN } with P (X1 = yi1 , ..., Xn = yin ) = p(s)/n!,

196

3. Unbiased Estimation

where s = {i1 , ..., in } is a subset of distinct elements of {1, ..., N } and p is a selection probability measure. We consider univariate yi , although most of our conclusions are valid for the case of multivariate P yi . In many survey problems the parameter to be estimated is Y = N i=1 yi , the population total. ¯ =NP In Example 2.27, it is shown that Yˆ = N X i∈s yi is unbiased for n Y if p(s) is constant (simple random sampling); a formula of Var(Yˆ ) is also given. We now show that Yˆ is in fact the UMVUE of Y under simple QN random sampling. Let Y be the range of yi , θ = (y1 , ..., yN ) and Θ = i=1 Y. Under simple random sampling, the population under consideration is a parametric family indexed by θ ∈ Θ. Theorem 3.13 (Watson-Royall theorem). (i) If p(s) > 0 for all s, then the vector of order statistics X(1) ≤ · · · ≤ X(n) is complete for θ ∈ Θ. (ii) Under simple random sampling, the vector of order statistics is sufficient for θ ∈ Θ. (iii) Under simple random sampling, for any estimable function of θ, its unique UMVUE is the unbiased estimator g(X1 , ..., Xn ), where g is symmetric in its n arguments. Proof. (i) Let h(X) be a function of the order statistics. Then h is symmetric in its n arguments. We need to show that if X E[h(X)] = p(s)h (yi1 , ..., yin ) /n! = 0 (3.42) s={i1 ,...,in }⊂{1,...,N } for all θ ∈ Θ, then h(yi1 , ..., yin ) = 0 for all yi1 , ..., yin . First, suppose that all N elements of θ are equal to a ∈ Y. Then (3.42) implies h(a, ..., a) = 0. Next, suppose that N − 1 elements in θ are equal to a and one is b > a. Then (3.42) reduces to q1 h(a, ..., a) + q2 h(a, ..., a, b), where q1 and q2 are some known numbers in (0, 1). Since h(a, ..., a) = 0 and q2 6= 0, h(a, ..., a, b) = 0. Using the same argument, we can show that h(a, ..., a, b, ..., b) = 0 for any k a’s and n − k b’s. Suppose next that elements of θ are equal to a, b, or c, a < b < c. Then we can show that h(a, ..., a, b, ..., b, c, ..., c) = 0 for any k a’s, l b’s, and n−k −l c’s. Continuing inductively, we see that h(y1 , ..., yn ) = 0 for all possible y1 , ..., yn . This completes the proof of (i). (ii) The result follows from the factorization theorem (Theorem 2.2), the fact that p(s) is constant under simple random sampling, and P (X1 = yi1 , ..., Xn = yin ) = P (X(1) = y(i1 ) , ..., X(n) = y(in ) )/n!, where y(i1 ) ≤ · · · ≤ y(in ) are the ordered values of yi1 , ..., yin . (iii) The result follows directly from (i) and (ii).

3.4. Unbiased Estimators in Survey Problems

197

It is interesting to note the following two issues. (1) Although we have a parametric problem under simple random sampling, the sufficient and complete statistic is the same as that in a nonparametric problem (Example 2.17). (2) For the completeness of the order statistics, we do not need the assumption of simple random sampling. ¯ is unbiased for Y . Since Yˆ Example 3.19. From Example 2.27, Yˆ = N X is symmetric in its arguments, it is the UMVUE of Y . We now derive the UMVUE for Var(Yˆ ). From Example 2.27, Var(Yˆ ) = where σ2 =

N2 n 2 1− σ , n N

(3.43)

2 N 1 X Y yi − . N − 1 i=1 N

It can be shown (exercise) that E(S 2 ) = σ 2 , where S 2 is the usual sample variance !2 n X ˆ 1 X 1 Y 2 2 ¯ = yi − S = (Xi − X) . n − 1 i=1 n − 1 i∈s N 2 2 n Since S 2 is symmetric in its arguments, Nn 1 − N S is the UMVUE of Var(Yˆ ). Simple random sampling is simple and easy to use, but it is inefficient unless the population is fairly homogeneous w.r.t. the yi ’s. A sampling plan often used in practice is the stratified sampling plan, which can be described as follows. The population P is divided into nonoverlapping subpopulations P1 , ..., PH called strata; a sample is drawn from each stratum Ph , independently across the strata. There are many reasons for stratification: (1) it may produce a gain in precision in parameter estimation when a heterogeneous population is divided into strata, each of which is internally homogeneous; (2) sampling problems may differ markedly in different parts of the population; and (3) administrative considerations may also lead to stratification. More discussions can be found, for example, in Cochran (1977). In stratified sampling, if a simple random sample (without replacement), Xh = (Xh1 , ..., Xhnh ), is drawn from each stratum, where nh is the sample size in stratum h, then the joint distribution of X = (X1 , ..., XH ) is in a parametric family indexed by θ = (θ1 , ..., θH ), where θh = (yi , i ∈ Ph ), h = QNh Yh , where 1, ..., H. Let Yh be the range of yi ’s in stratum h and Θh = i=1 QH Nh is the size of Ph . We assume that the parameter space is Θ = i=1 Θh . The following result is similar to Theorem 3.13.

198

3. Unbiased Estimation

Theorem 3.14. Let X be a sample obtained using the stratified simple random sampling plan described previously. (i) For each h, let Zh be the vector of the ordered values of the sample in stratum h. Then (Z1 , ..., ZH ) is sufficient and complete for θ ∈ Θ. (ii) For any estimable function of θ, its unique UMVUE is the unbiased estimator g(X) that is symmetric in its first n1 arguments, symmetric in its second n2 arguments,..., and symmetric in its last nH arguments. Example 3.20. Consider the estimation of the population total Y based on a sample X = (Xhi , i = 1, ..., nh , h = 1, ..., H) obtained by stratified simple random sampling. Let Yh be the population total of the hth stratum and ¯ h· is the sample mean of the sample from stratum ¯ h· , where X let Yˆh = Nh X h, h = 1, ..., H. From Example 2.27, each Yˆh is an unbiased estimator of Yh . Let nh H H X X X Nh Xhi . Yˆh = Yˆst = nh i=1 h=1

h=1

Then, by Theorem 3.14, Yˆst is the UMVUE of Y . Since Yˆ1 , ..., YˆH are independent, it follows from (3.43) that Var(Yˆst ) =

H X Nh2 nh 1− σh2 , nh Nh

(3.44)

h=1

P where σh2 = (Nh − 1)−1 i∈Ph (yi − Yh /Nh )2 . An argument similar to that in Example 3.19 shows that the UMVUE of Var(Yˆst ) is 2 Sst

H X Nh2 nh 1− Sh2 , = nh Nh

(3.45)

h=1

where Sh2 is the usual sample variance based on Xh1 , ..., Xhnh . It is interesting to compare the mse of the UMVUE Yˆst with the mse of the UMVUE Yˆ under simple random sampling (Example 3.19). Let σ 2 be given in (3.43). Then (N − 1)σ 2 =

H X

(Nh − 1)σh2 +

h=1

H X

h=1

Nh (µh − µ)2 ,

where µh = Yh /Nh is the population mean of the hth stratum and µ = Y /N is the overall population mean. By (3.43), (3.44), and (3.45), Var(Yˆ ) ≥ Var(Yˆst ) if and only if H X

h=1

N 2 Nh n(N −1)

1−

n N

H h X Nh2 2 (µh − µ) ≥ nh 1 − h=1

nh Nh

−

N 2 (Nh −1) n(N −1)

1−

n N

i 2 σh .

199

3.4. Unbiased Estimators in Survey Problems

This means that stratified simple random sampling is better than simple nh n random sampling if the deviations µh − µ are sufficiently large. If N ≡N h (proportional allocation), then this condition simplifies to H X

h=1

H X Nh σh2 , 1− Nh (µh − µ) ≥ N 2

(3.46)

h=1

which is usually true when µh ’s are different and some Nh ’s are large. Note that the variances Var(Yˆ ) and Var(Yˆst ) are w.r.t. different sampling plans under which Yˆ and Yˆst are obtained.

3.4.2 Horvitz-Thompson estimators If some elements of the finite population P are groups (called clusters) of subunits, then sampling from P is cluster sampling. Cluster sampling is used often because of administrative convenience or economic considerations. Although sometimes the first intention may be to use the subunits as sampling units, it is found that no reliable list of the subunits in the population is available. For example, in many countries there are no complete lists of the people or houses in a region. From the maps of the region, however, it can be divided into units such as cities or blocks in the cities. In cluster sampling, one may greatly increase the precision of estimation by using sampling with probability proportional to cluster size. Thus, unequal probability sampling is often used. Suppose that a sample of clusters is obtained. If subunits within a selected cluster give similar results, then it may be uneconomical to measure them all. A sample of the subunits in any chosen cluster may be selected. This is called two-stage sampling. One can continue this process to have a multistage sampling (e.g., cities → blocks → houses → people). Of course, at each stage one may use stratified sampling and/or unequal probability sampling. When the sampling plan is complex, so is the structure of the observations. We now introduce a general method of deriving unbiased estimators of population totals, which are called Horvitz-Thompson estimators. Theorem 3.15. Let X = {yi , i ∈ s} denote a sample from P = {y1 , ..., yN } that is selected, without replacement, by some method. Define πi = probability that i ∈ s,

i = 1, ..., N.

(i) (Horvitz-Thompson). If πi > 0 for i = 1, ..., N and πi is known when P i ∈ s, then Yˆht = i∈s yi /πi is an unbiased estimator of the population

200

3. Unbiased Estimation

total Y . (ii) Define πij = probability that i ∈ s and j ∈ s,

i = 1, ..., N, j = 1, ..., N.

Then Var(Yˆht ) =

N X 1 − πi i=1

=

πi

N N X X

i=1 j=i+1

yi2 + 2

N N X X πij − πi πj yi yj πi πj i=1 j=i+1

(πi πj − πij )

yj yi − πi πj

2

.

(3.47)

(3.48)

Proof. (i) Let ai = 1 if i ∈ s and ai = 0 if i 6∈ s, i = 1, ..., N . Then E(ai ) = πi and ! N N X X ai y i ˆ = yi = Y. E(Yht ) = E πi i=1 i=1 (ii) Since a2i = ai , Var(ai ) = E(ai ) − [E(ai )]2 = πi (1 − πi ). For i 6= j, Cov(ai , aj ) = E(ai aj ) − E(ai )E(aj ) = πij − πi πj . Then Var(Yˆht ) = Var

N X ai y i i=1

=

N X i=1

=

N N X X yi2 yi yj Var(ai ) + 2 Cov(ai , aj ) πi2 π π i=1 j=i+1 i j

N X 1 − πi i=1

πi

!

πi

yi2

N N X X πij − πi πj +2 yi yj . πi πj i=1 j=i+1

Hence (3.47) follows. To show (3.48), note that N X

πi = n

i=1

which implies X

and

X

j=1,...,N,j6=i

πij = (n − 1)πi ,

(πij − πi πj ) = (n − 1)πi − πi (n − πi ) = −πi (1 − πi ).

j=1,...,N,j6=i

201

3.4. Unbiased Estimators in Survey Problems

Hence N X 1 − πi

πi

i=1

yi2 =

N X

X

(πi πj − πij )

i=1 j=1,...,N,j6=i

=

N N X X

i=1 j=i+1

(πi πj − πij )

yi2 πi2

yj2 yi2 + πi2 πj2

!

and, by (3.47), Var(Yˆht ) =

N N X X

i=1 j=i+1

=

N N X X

i=1 j=i+1

yj2 2yi yj yi2 + 2− 2 πi πj πi πj

(πij − πi πj ) (πi πj − πij )

yi yj − πi πj

2

!

.

Using the same idea, we can obtain unbiased estimators of Var(Yˆht ). Suppose that πij > 0 for all i and j and πij is known when i ∈ s and j ∈ s. By (3.47), an unbiased estimator of Var(Yˆht ) is v1 =

X 1 − πi i∈s

πi2

yi2 + 2

X X

πij − πi πj yi yj . πi πj πij i∈s j∈s,j>i

(3.49)

By (3.48), an unbiased estimator of Var(Yˆht ) is v2 =

X X

πi πj − πij πij i∈s j∈s,j>i

yi yj − πi πj

2

.

(3.50)

Variance estimators v1 and v2 may not be the same in general, but they are the same in some special cases (Exercise 92). A more serious problem is that they may take negative values. Some discussions about deriving better estimators of Var(Yˆht ) are provided in Cochran (1977, Chapter 9A). Some special cases of Theorem 3.15 are considered as follows. Under simple random sampling, πi = n/N . Thus, Yˆ in Example 3.19 is the Horvitz-Thompson estimator. Under stratified simple random sampling, πi = nh /Nh if unit i is in stratum h. Hence, the estimator Yˆst in Example 3.20 is the Horvitz-Thompson estimator. Suppose now each yi ∈ P is a cluster, i.e., yi = (yi1 , ..., yiMi ), where Mi is the size ofP the ith cluster, i = 1, ..., N . The total number of units in N P is then M = i=1 Mi . Consider a single-stage sampling plan, i.e., if yi is selected, then every yij is observed. If simple random sampling is used,

202

3. Unbiased Estimation

then πi = k/N , where k is the first-stage sample size (the total sample size Pk is n = i=1 Mi ), and the Horvitz-Thompson estimator is Mi N XX N X Yˆs = yij = Yi , k i∈s j=1 k i∈s 1 1

where s1 is the index set of first-stage sampled clusters and Yi is the total of the ith cluster. In this case, Var(Yˆs ) =

N2 k(N − 1)

N 2 k X Y 1− Yi − . N i=1 N

If the selection probability is proportional to the cluster size, then πi = kMi /M and the Horvitz-Thompson estimator is Mi M X 1 X M X Yi yij = Yˆpps = k i∈s Mi j=1 k i∈s Mi 1 1

whose variance is given by (3.47) or (3.48). Usually Var(Yˆpps ) is smaller than Var(Yˆs ); see the discussions in Cochran (1977, Chapter 9A). Consider next a two-stage sampling in which k first-stage clusters are selected and a simple random sample of size mi is selected from each sampled cluster yi , where sampling is independent across clusters. If the first-stage sampling plan is simple random sampling, then πi = kmi /(N Mi ) and the Horvitz-Thompson estimator is N X Mi X Yˆs = yij , k i∈s mi j∈s 1 2i where s2i denotes the second-stage sample from cluster i. If the first-stage selection probability is proportional to the cluster size, then πi = kmi /M and the Horvitz-Thompson estimator is M X 1 X Yˆpps = yij . k i∈s mi j∈s 1 2i Finally, let us consider another popular sampling method called systematic sampling. Suppose that P = {y1 , ..., yN } and the population size N = nk for two integers n and k. To select a sample of size n, we first draw a j randomly from {1, ..., k}. Our sample is then {yj , yj+k , yj+2k , ..., yj+(n−1)k }.

203

3.4. Unbiased Estimators in Survey Problems

Systematic sampling is used mainly because it is easier to draw a systematic sample and often easier to execute without mistakes. It is also likely that systematic sampling provides more efficient point estimators than simple random sampling or even stratified sampling, since the sample units are spread more evenly over the population. Under systematic sampling, πi = k −1 for every i and the Horvitz-Thompson estimator of the population total is n X yj+(t−1)k . Yˆsy = k t=1

The unbiasedness of this estimator is a direct consequence of Theorem 3.15, but it can be easily shown as follows. Since j takes value i ∈ {1, ..., k} with probability k −1 , ! n k X N X X 1 E(Yˆsy ) = k yi+(t−1)k = yi = Y. k i=1 t=1 i=1

The variance of Yˆsy is simply Var(Yˆsy ) =

k N2 X (µi − µ)2 , k i=1

Pn Pk where µi = n−1 t=1 yi+(t−1)k and µ = k −1 i=1 µi = Y /N . Let σ 2 be given in (3.43) and k

2 σsy =

Then (N − 1)σ 2 = n Thus, and

n

XX 1 (yi+(t−1)k − µi )2 . k(n − 1) i=1 t=1 k X i=1

(µi − µ)2 +

n k X X i=1 t=1

(yi+(t−1)k − µi )2 .

2 (N − 1)σ 2 = N −1 Var(Yˆsy ) + k(n − 1)σsy 2 Var(Yˆsy ) = N (N − 1)σ 2 − N (N − k)σsy .

Since the variance of the Horvitz-Thompson estimator of the population total under simple random sampling is, by (3.43), N2 n 2 1− σ = N (k − 1)σ 2 , n N

the Horvitz-Thompson estimator under systematic sampling has a smaller 2 variance if and only if σsy > σ2 .

204

3. Unbiased Estimation

3.5 Asymptotically Unbiased Estimators As we discussed in §2.5, we often need to consider biased but asymptotically unbiased estimators. A large and useful class of such estimators are smooth functions of some exactly unbiased estimators such as UMVUE’s, U-statistics, LSE’s, and Horvitz-Thompson estimators. Some other methods of constructing asymptotically unbiased estimators are also introduced in this section.

3.5.1 Functions of unbiased estimators If the parameter to be estimated is ϑ = g(θ) with a vector-valued parameter θ and Un is a vector of unbiased estimators of components of θ (i.e., EUn = θ), then Tn = g(Un ) is often asymptotically unbiased for ϑ. Assume that g is differentiable and cn (Un − θ) →d Y . Then amseTn (P ) = E{[∇g(θ)]τ Y }2 /c2n

(Theorem 2.6). Hence, Tn has a good performance in terms of amse if Un is optimal in terms of mse (such as the UMVUE). The following are some examples. Example 3.21 (Ratio estimators). Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors with EX1 = µx and EY1 = µy . Consider the estimation of ¯ the ratio of two population means: ϑ = µy /µx (µx 6= 0). Note that (Y¯ , X), the vector of sample means, is unbiased for (µy , µx ). The sample means are UMVUE’s under some statistical models (§3.1 and §3.2) and are BLUE’s ¯ Assume in general (Example 2.22). The ratio estimator is Tn = Y¯ /X. that σx2 = Var(X1 ), σy2 = Var(Y1 ), and σxy = Cov(X1 , Y1 ) exist. A direct calculation shows that the n−1 order asymptotic bias of Tn according to (2.38) is 2 ˜bT (P ) = ϑσx − σxy n µ2x n (verify). Using the CLT and the delta-method (Corollary 1.1), we obtain that ! σy2 − 2ϑσxy + ϑ2 σx2 √ n(Tn − ϑ) →d N 0, µ2x (verify), which implies amseTn (P ) =

σy2 − 2ϑσxy + ϑ2 σx2 . µ2x n

In some problems, we are not interested in the ratio, but the use of a ratio estimator to improve an estimator of a marginal mean. For example,

205

3.5. Asymptotically Unbiased Estimators

suppose that µx is known and we are interested in estimating µy . Consider the following estimator: ¯ x. µ ˆy = (Y¯ /X)µ Note that µ ˆy is not unbiased; its n−1 order asymptotic bias according to (2.38) is 2 ˜bµˆ (P ) = ϑσx − σxy ; y µx n and

σy2 − 2ϑσxy + ϑ2 σx2 . n ˆy is asympComparing µ ˆy with the unbiased estimator Y¯ , we find that µ totically more efficient if and only if amseµˆy (P ) =

2ϑσxy > ϑ2 σx2 , which means that µ ˆy is a better estimator if and only if the correlation between X1 and Y1 is large enough to pay off the extra variability caused ¯ by using µx /X. Another example related to a bivariate sample is the sample correlation coefficient defined in Exercise 22 in §2.6. Example 3.22. Consider a polynomial regression of order p: Xi = β τ Zi + εi ,

i = 1, ..., n,

), and εi ’s are i.i.d. with where β = (β0 , β1 , ..., βp−1 ), Zi = (1, ti , ..., tp−1 i mean 0 and variance σ 2 > 0. Suppose that the parameter to be estimated is tβ ∈ T ⊂ R such that p−1 X

βj tjβ

= max

j=0

t∈T

p−1 X

βj t j .

j=0

Note that tβ = g(β) for some function g. Let βˆ be the LSE of β. Then the ˆ is asymptotically unbiased and its amse can be derived estimator tˆβ = g(β) under some conditions (Exercise 98). Example 3.23. In the study of the reliability of a system component, we assume that Xij = θ τi z(tj ) + εij ,

i = 1, ..., k, j = 1, ..., m.

Here Xij is the measurement of the ith sample component at time tj ; z(t) is a q-vector whose components are known functions of the time t; θ i ’s

206

3. Unbiased Estimation

are unobservable random q-vectors that are i.i.d. from Nq (θ, Σ), where θ and Σ are unknown; εij ’s are i.i.d. measurement errors with mean zero and variance σ 2 ; and θi ’s and εij ’s are independent. As a function of t, θτ z(t) is the degradation curve for a particular component and θτ z(t) is the mean degradation curve. Suppose that a component will fail to work if θτ z(t) < η, a given critical value. Assume that θτ z(t) is always a decreasing function of t. Then the reliability function of a component is τ θ z(t) − η τ R(t) = P (θ z(t) > η) = Φ , s(t) p where s(t) = [z(t)]τ Σz(t) and Φ is the standard normal distribution function. For a fixed t, estimators of R(t) can be obtained by estimating θ and Σ, since Φ is a known function. It can be shown (exercise) that the BLUE of θ is the LSE ¯ θˆ = (Z τ Z)−1 Z τ X, where Z is the m × q matrix whose jth row is the vector z(tj ), Xi = ¯ is the sample mean of Xi ’s. The estimation of Σ is (Xi1 , ..., Xim ), and X more difficult. It can be shown (exercise) that a consistent (as k → ∞) estimator of Σ is k X τ −1 ˆ= 1 ¯ ¯ τ Σ (Z τ Z)−1 Z τ (Xi − X)(X −σ ˆ 2 (Z τ Z)−1 , i − X) Z(Z Z) k i=1

where

k

σ ˆ2 =

X 1 [X τ Xi − Xiτ Z(Z τ Z)−1 Z τ Xi ]. k(m − q) i=1 i

Hence an estimator of R(t) is

ˆ =Φ R(t) where sˆ(t) =

θˆτ z(t) − η sˆ(t)

!

,

q ˆ [z(t)]τ Σz(t).

If we define Yi1 = Xiτ Z(Z τ Z)−1 z(t), Yi2 = [Xiτ Z(Z τ Z)−1 z(t)]2 , Yi3 = [Xiτ Xi − Xiτ Z(Z τ Z)−1 Z τ Xi ]/(m − q), and Yi = (Yi1 , Yi2 , Yi3 ), then it is ˆ apparent that R(t) can be written as g(Y¯ ) for a function ! y1 − η . g(y1 , y2 , y3 ) = Φ p y2 − y12 − y3 [z(t)]τ (Z τ Z)−1 z(t)

Suppose that εij has a finite fourth moment, which implies the existence of ˆ can be derived (exercise). Var(Yi ). The amse of R(t)

3.5. Asymptotically Unbiased Estimators

207

3.5.2 The method of moments The method of moments is the oldest method of deriving point estimators. It almost always produces some asymptotically unbiased estimators, although they may not be the best estimators. Consider a parametric problem where X1 , ..., Xn are i.i.d. random variables from Pθ , θ ∈ Θ ⊂ Rk , and E|X1 |k < ∞. Let µj = EX1j be the jth moment of P and let n 1X j µ ˆj = X n i=1 i

be the jth sample moment, which is an unbiased estimator of µj , j = 1, ..., k. Typically, µj = hj (θ), j = 1, ..., k, (3.51)

for some functions hj on Rk . By substituting µj ’s on the left-hand side of ˆ i.e., θˆ (3.51) by the sample moments µ ˆj , we obtain a moment estimator θ, satisfies ˆ j = 1, ..., k, µ ˆ j = hj (θ), which is a sample analogue of (3.51). This method of deriving estimators is called the method of moments. Note that an important statistical principle, the substitution principle, is applied in this method. ˆ If the inverse Let µ ˆ = (ˆ µ1 , ..., µ ˆk ) and h = (h1 , ..., hk ). Then µ ˆ = h(θ). −1 µ). function h exists, then the unique moment estimator of θ is θˆ = h−1 (ˆ ˆ ˆ = h(θ) When h−1 does not exist (i.e., h is not one-to-one), any solution of µ ˆ is a moment estimator of θ; if possible, we always choose a solution θ in the parameter space Θ. In some cases, however, a moment estimator does not exist (see Exercise 111). Assume that θˆ = g(ˆ µ) for a function g. If h−1 exists, then g = h−1 . If g is continuous at µ = (µ1 , ..., µk ), then θˆ is strongly consistent for θ, since µ ˆj →a.s. µj by the SLLN. If g is differentiable at µ and E|X1 |2k < ∞, then θˆ is asymptotically normal, by the CLT and Theorem 1.12, and amseθˆ(θ) = n−1 [∇g(µ)]τ Vµ ∇g(µ), where Vµ is a k × k matrix whose (i, j)th element is µi+j − µi µj . Furthermore, it follows from (2.38) that the n−1 order asymptotic bias of θˆ is (2n)−1 tr ∇2 g(µ)Vµ .

Example 3.24. Let X1 , ..., Xn be i.i.d. from a population Pθ indexed by the parameter θ = (µ, σ 2 ), where µ = EX1 ∈ R and σ 2 = Var(X1 ) ∈ (0, ∞). This includes cases such as the family of normal distributions,

208

3. Unbiased Estimation

double exponential distributions, or logistic distributions (Table 1.2, page 20). Since EX1 = µ and EX12 = Var(X1 ) + (EX1 )2 = σ 2 + µ2 , setting µ ˆ1 = µ and µ ˆ2 = σ 2 + µ2 we obtain the moment estimator ! n n−1 2 1X 2 ˆ ¯ ¯ ¯ S . = X, θ = X, (Xi − X) n i=1 n ¯ is unbiased, but n−1 S 2 is not. If Xi is normal, then θˆ is suffiNote that X n cient and is nearly the same as an optimal estimator such as the UMVUE. On the other hand, if Xi is from a double exponential or logistic distribution, then θˆ is not sufficient and can often be improved. Consider now the estimation of σ 2 when we know that µ = 0. Obviously we cannot use the equation µ ˆ1 = µ to solve the problem.PUsing µ ˆ2 = µ2 = n σ 2 , we obtain the moment estimator σ ˆ2 = µ ˆ2 = n−1 i=1 Xi2 . This is still a good estimator when Xi is normal, but is not a function of sufficient statistic when Xi is from a double exponential distribution. For the double exponential case one can argue that we should first make a transformation Yi = |Xi | and then obtain the moment estimator based on the transformed data.P The moment estimator of σ 2 based on the transformed data is Y¯ 2 = n −1 2 2 (n i=1 |Xi |) , which is sufficient for σ . Note that this estimator can also be obtained based on absolute moment equations. Example 3.25. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (θ1 , θ2 ), −∞ < θ1 < θ2 < ∞. Note that EX1 = (θ1 + θ2 )/2 and EX12 = (θ12 + θ22 + θ1 θ2 )/3. ˆ2 = EX12 and substituting θ1 in the second equaSetting µ ˆ1 = EX1 and µ tion by 2ˆ µ1 − θ2 (the first equation), we obtain that µ1 − θ2 )θ2 = 3ˆ µ2 , (2ˆ µ1 − θ2 )2 + θ22 + (2ˆ which is the same as ˆ1 )2 = 3(ˆ µ2 − µ ˆ21 ). (θ2 − µ Since θ2 > EX1 , we obtain that q q ¯ + 3(n−1) S 2 ˆ1 + 3(ˆ µ2 − µ ˆ21 ) = X θˆ2 = µ n and

θˆ1 = µ ˆ1 −

q q ¯ − 3(n−1) S 2 . 3(ˆ µ2 − µ ˆ21 ) = X n

209

3.5. Asymptotically Unbiased Estimators

These estimators are not functions of the sufficient and complete statistic (X(1) , X(n) ). Example 3.26. Let X1 , ..., Xn be i.i.d. from the binomial distribution Bi(p, k) with unknown parameters k ∈ {1, 2, ...} and p ∈ (0, 1). Since EX1 = kp and EX12 = kp(1 − p) + k 2 p2 , we obtain the moment estimators ˆ21 − µ ˆ2 )/ˆ µ1 = 1 − pˆ = (ˆ µ1 + µ and

n−1 2 ¯ n S /X

¯ kˆ = µ ˆ21 /(ˆ µ1 + µ ˆ21 − µ ˆ2 ) = X/(1 −

n−1 2 ¯ n S /X).

The estimator pˆ is in the range of (0, 1). But kˆ may not be an integer. It can be improved by an estimator that is kˆ rounded to the nearest positive integer. Example 3.27. Suppose that X1 , ..., Xn are i.i.d. from the Pareto distribution P a(a, θ) with unknown a > 0 and θ > 2 (Table 1.2, page 20). Note that EX1 = θa/(θ − 1) and EX12 = θa2 /(θ − 2). From the moment equation, (θ−1)2 θ(θ−2)

Note that

(θ−1)2 θ(θ−2)

−1=

1 θ(θ−2) .

=µ ˆ 2 /ˆ µ21 .

Hence

µ2 − µ ˆ21 ). θ(θ − 2) = µ ˆ21 /(ˆ Since θ > 2, there is a unique solution in the parameter space: q q n ¯2 ˆ2 /(ˆ µ2 − µ ˆ21 ) = 1 + 1 + n−1 X /S 2 θˆ = 1 + µ and

µ ˆ 1 (θˆ − 1) ˆ q qθ . ¯ 1+ n X ¯ 2 /S 2 1 + 1+ =X n−1

a ˆ=

n ¯2 2 n−1 X /S

.

210

3. Unbiased Estimation

The method of moments can also be applied to nonparametric problems. Consider, for example, the estimation of the central moments cj = E(X1 − µ1 )j , Since cj =

j = 2, ..., k.

j X j (−µ1 )t µj−t , t t=0

the moment estimator of cj is cˆj =

j X j ¯ tµ (−X) ˆj−t , t t=0

where µ ˆ0 = 1. It can be shown (exercise) that n

cˆj =

1X ¯ j, (Xi − X) n i=1

j = 2, ..., k,

(3.52)

which are sample central moments. From the SLLN, cˆj ’s are strongly consistent. If E|X1 |2k < ∞, then √ n (ˆ c2 − c2 , ..., cˆk − ck ) →d Nk−1 (0, D) (3.53) (exercise), where the (i, j)th element of the (k − 1) × (k − 1) matrix D is ci+j+2 − ci+1 cj+1 − (i + 1)ci cj+2 − (j + 1)ci+2 cj + (i + 1)(j + 1)ci cj c2 .

3.5.3 V-statistics Let X1 , ..., Xn be i.i.d. from P . For every U-statistic Un defined in (3.11) as an estimator of ϑ = E[h(X1 , ..., Xm )], there is a closely related V-statistic defined by n n X 1 X Vn = m ··· h(Xi1 , ..., Xim ). (3.54) n i =1 i =1 1

m

As an estimator of ϑ, Vn is biased; but the bias is small asymptotically as the following results show. For a fixed sample size n, Vn may be better than Un in terms of their mse’s. Consider, for example, the kernel h(x1 , x2 ) = (x1 − x2 )2 /2 in §3.2.1, which leads to ϑ = σ 2 = Var(X1 ) and Un = S 2 , the sample variance. The corresponding V-statistic is n n 1 X X (Xi − Xj )2 1 = 2 n2 i=1 j=1 2 n

X

1≤i<j≤n

(Xi − Xj )2 =

n−1 2 S , n

211

3.5. Asymptotically Unbiased Estimators

which is the moment estimator of σ 2 discussed in Example 3.24. In Exercise 2 2 63 in §2.6, n−1 n S is shown to have a smaller mse than S when Xi is normally distributed. Of course, there are situations where U-statistics are better than their corresponding V-statistics. The following result provides orders of magnitude of the bias and variance of a V-statistic as an estimator of ϑ. Proposition 3.5. Let Vn be defined by (3.54). (i) Assume that E|h(Xi1 , ..., Xim )| < ∞ for all 1 ≤ i1 ≤ · · · ≤ im ≤ m. Then the bias of Vn satisfies bVn (P ) = O(n−1 ). (ii) Assume that E[h(Xi1 , ..., Xim )]2 < ∞ for all 1 ≤ i1 ≤ · · · ≤ im ≤ m. Then the variance of Vn satisfies Var(Vn ) = Var(Un ) + O(n−2 ), where Un is given by (3.11). Proof. (i) Note that h Un − Vn = 1 −

n!

nm (n−m)!

i

(Un − Wn ),

(3.55)

where Wn is the average of all terms h(Xi1 , ..., Xim ) with at least one equality im = il , m 6= l. The result follows from E(Un − Wn ) = O(1). (ii) The result follows from E(Un − Wn )2 = O(1), E[Wn (Un − ϑ)] = O(n−1 ) (exercise), and (3.55). To study the asymptotic behavior of a V-statistic, we consider the following representation of Vn in (3.54): m X m Vnj , Vn = j j=1 where Vnj = ϑ +

n n X 1 X · · · gj (Xi1 , ..., Xij ) nj i =1 i =1 1

is a “V-statistic” with

j

gj (x1 , ..., xj ) = hj (x1 , ..., xj ) − +

X

1≤i1 0, then √ n(Vn − ϑ) →d N (0, m2 ζ1 ). (ii) If ζ1 = 0 but ζ2 = Var(h2 (X1 , X2 )) > 0, then n(Vn − ϑ) →d

∞

m(m − 1) X λj χ21j , 2 j=1

where χ21j ’s and λj ’s are the same as those in (3.21). Result (3.57) and Theorem 3.16 imply that Vn has expansion (2.37) −1 and, therefore, the nP order asymptotic bias of Vn is E[g2 (X1 , X1 )]/n = ∞ nEVn2 = m(m − 1) j=1 λj /(2n) (exercise). Theorem 3.16 shows that if ζ1 > 0, then the amse’s of Un and Vn are the same. If ζ1 = 0 but ζ2 > 0, then an argument similar to that in the proof of Lemma 3.2 leads to 2 ∞ m2 (m − 1)2 ζ2 m2 (m − 1)2 X amseVn (P ) = + λj 2n2 4n2 j=1 2 ∞ X m (m − 1) = amseUn (P ) + λj 4n2 j=1 2

2

(see Lemma 3.2). Hence Un is asymptotically more efficient than Vn , unless P ∞ j=1 λj = 0. Technically, the proof of the asymptotic results for Vn also requires moment conditions stronger than those for Un . Example 3.28. Consider the estimation ofP µ2 , where µ = EX1 . From the 1 results in §3.2, the U-statistic Un = n(n−1) 1≤i<j≤n Xi Xj is unbiased for

3.5. Asymptotically Unbiased Estimators

213

¯ 2 . If µ 6= 0, then ζ1 6= 0 µ2 . The corresponding V-statistic is simply Vn = X and the asymptotic relative efficiency of Vn w.r.t. Un is 1. If µ = 0, then nVn →d σ 2 χ21

and

nUn →d σ 2 (χ21 − 1),

where χ21 is a random variable having the chi-square distribution χ21 . Hence the asymptotic relative efficiency of Vn w.r.t. Un is E(χ21 − 1)2 /E(χ21 )2 = 2/3.

3.5.4 The weighted LSE In linear model (3.25), the unbiased LSE of lτ β may be improved by a slightly biased estimator when Var(ε) is not σ 2 In and the LSE is not BLUE. Assume that Z in (3.25) is of full rank so that every lτ β is estimable. For simplicity, let us denote Var(ε) by V . If V is known, then the BLUE ˘ where of lτ β is lτ β, β˘ = (Z τ V −1 Z)−1 Z τ V −1 X

(3.58)

(see the discussion after the statement of assumption A3 in §3.3.1). If V is unknown and Vˆ is an estimator of V , then an application of the substitution principle leads to a weighted least squares estimator βˆw = (Z τ Vˆ −1 Z)−1 Z τ Vˆ −1 X.

(3.59)

The weighted LSE is not linear in X and not necessarily unbiased for β. If the distribution of ε is symmetric about 0 and Vˆ remains unchanged when ε changes to −ε (Examples 3.29 and 3.30), then the distribution of βˆw − β is symmetric about 0 and, if E βˆw is well defined, βˆw is unbiased for β. In such a case the LSE lτ βˆ may not be a UMVUE (when ε is normal), since ˆ Var(lτ βˆw ) may be smaller than Var(lτ β). Asymptotic properties of the weighted LSE depend on the asymptotic behavior of Vˆ . We say that Vˆ is consistent for V if and only if kVˆ −1 V − In kmax →p 0,

(3.60)

where kAkmax = maxi,j |aij | for a matrix A whose (i, j)th element is aij . Theorem 3.17. Consider model (3.25) with a full rank Z. Let β˘ and βˆw be defined by (3.58) and (3.59), respectively, with a Vˆ consistent in the sense of (3.60). Assume the conditions in Theorem 3.12. Then lτ (βˆw − β)/an →d N (0, 1),

214

3. Unbiased Estimation

where l ∈ Rp , l 6= 0, and ˘ = lτ (Z τ V −1 Z)−1 l. a2n = Var(lτ β) Proof. Using the same argument as in the proof of Theorem 3.12, we obtain that lτ (β˘ − β)/an →d N (0, 1). By Slutsky’s theorem, the result follows from lτ βˆw − lτ β˘ = op (an ). Define and

(3.61)

ξn = lτ (Z τ Vˆ −1 Z)−1 Z τ (Vˆ −1 − V −1 )ε ζn = lτ [(Z τ Vˆ −1 Z)−1 − (Z τ V −1 Z)−1 ]Z τ V −1 ε.

Then

lτ βˆw − lτ β˘ = ξn + ζn .

Let Bn = (Z τ Vˆ −1 Z)−1 Z τ V −1 Z − Ip and Cn = Vˆ 1/2p V −1 Vˆ 1/2 − In . By (3.60), kCn kmax = op (1). For any matrix A, denote tr(Aτ A) by kAk. Then kBn k2 = k(Z τ Vˆ −1 Z)−1 Z τ Vˆ −1/2 Cn Vˆ −1/2 Zk2 = tr (Z τ Vˆ −1 Z)−1 (Z τ Vˆ −1/2 Cn Vˆ −1/2 Z)2 (Z τ Vˆ −1 Z)−1 ≤ kCn k2max tr (Z τ Vˆ −1 Z)−1 (Z τ Vˆ −1 Z)2 (Z τ Vˆ −1 Z)−1 = op (1)tr(Ip ).

This proves that kBn kmax = op (1). Let An = V 1/2 Vˆ −1 V 1/2 − In . Using inequality (1.37) and the previous results, we obtain that ξn2 = [lτ (Z τ Vˆ −1 Z)−1 Z τ V −1/2 An V −1/2 ε]2 ≤ εV −1 ετ lτ (Z τ Vˆ −1 Z)−1 Z τ V −1/2 A2n V −1/2 Z(Z τ Vˆ −1 Z)−1 l ≤ Op (1)kAn k2 lτ (Z τ Vˆ −1 Z)−1 Z τ V −1 Z(Z τ Vˆ −1 Z)−1 l max

= op (1)lτ (Bn + Ip )2 (Z τ V −1 Z)−1 l

= op (an ). Since Ek(Z τ V −1 Z)−1/2 Z τ V −1 εk2 = p, k(Z τ V −1 Z)−1/2 Z τ V −1 εk = Op (1). Define B1n = (Z τ Vˆ −1 Z)1/2 Bn (Z τ Vˆ −1 Z)−1/2 . Then B1n = (Z τ Vˆ −1 Z)−1/2 Z τ Vˆ −1/2 Cn Vˆ −1/2 Z(Z τ Vˆ −1 Z)−1/2 ≤ kCn kmax (Z τ Vˆ −1 Z)−1/2 Z τ Vˆ −1 Z(Z τ Vˆ −1 Z)−1/2 = op (1)Ip .

215

3.5. Asymptotically Unbiased Estimators Let B2n = (Z τ V −1 Z)1/2 (Z τ Vˆ −1 Z)−1/2 . Since kB2n k2 = tr (Z τ V −1 Z)1/2 (Z τ Vˆ −1 Z)−1 (Z τ V −1 Z)1/2 = tr (Z τ Vˆ −1 Z)−1 Z τ V −1 Z = tr(Bn + Ip ) = p + op (1),

we obtain that τ k = op (1). kB2n B1n B2n

Then ζn2 = [lτ Bn (Z τ V −1 Z)−1 Z τ V −1 ε]2 τ = [lτ (Z τ V −1 Z)−1/2 B2n B1n B2n (Z τ V −1 Z)−1/2 Z τ V −1 ε]2 τ 2 ≤ lτ (Z τ V −1 Z)−1 lkB2n B1n B2n k k(Z τ V −1 Z)−1/2 Z τ V −1 εk2 = op (a2n ).

This proves (3.61) and thus completes the proof. Theorem 3.17 shows that as long as Vˆ is consistent in the sense of (3.60), ˘ which is the BLUE the weighted LSE βˆw is asymptotically as efficient as β, ˘ attains the if V is known. If V is known and ε is normal, then Var(lτ β) Cram´er-Rao lower bound (Proposition 3.2) and, thus, (3.10) holds with Tn = lτ βˆw . By Theorems 3.12 and 3.17, the asymptotic relative efficiency of the LSE lτ βˆ w.r.t. the weighted LSE lτ βˆw is lτ (Z τ V −1 Z)−1 l , Z(Z τ Z)−1 l

lτ (Z τ Z)−1 Z τ V

which is always less than 1 and equals 1 if lτ βˆ is a BLUE (in which case ˘ βˆ = β). Finding a consistent Vˆ is possible when V has a certain type of structure. We consider three examples. Example 3.29. Consider model (3.25). Suppose that V = Var(ε) is a block diagonal matrix with the ith diagonal block σ 2 Imi + Ui ΣUiτ ,

i = 1, ..., k,

(3.62)

where mi ’s are integers bounded by a fixed integer m, σ 2 > 0 is an unknown parameter, Σ is a q×q unknown nonnegative definite matrix, Ui is an mi ×q

216

3. Unbiased Estimation

full rank matrix whose columns are in R(Wi ), q < inf i mi , and Wi is the p × mi matrix such that Z τ = ( W1 W2 ... Wk ). Under (3.62), a consistent Vˆ can be obtained if we can obtain consistent estimators of σ 2 and Σ. Let X = (Y1 , ..., Yk ), where Yi is an mi -vector, and let Ri be the matrix whose columns are linearly independent rows of Wi . Then σ ˆ2 =

k 1 X τ Y [Imi − Ri (Riτ Ri )−1 Riτ ]Yi n − kq i=1 i

(3.63)

is an unbiased estimator of σ 2 . Assume that Yi ’s are independent and that supi E|εi |2+δ < ∞ for some δ > 0. Then σ ˆ 2 is consistent for σ 2 (exercise). τˆ Let ri = Yi − Wi β and k X τ ˆ= 1 ˆ 2 (Uiτ Ui )−1 . (Ui Ui )−1 Uiτ ri riτ Ui (Uiτ Ui )−1 − σ Σ k i=1

(3.64)

ˆ is consistent for Σ in the sense that It can be shown (exercise) that Σ ˆ ˆ − Σk →p 0 (see Exercise 116). kΣ − Σkmax →p 0 or, equivalently, kΣ Example 3.30. Suppose that V is a block diagonal matrix with the ith diagonal block matrix Vmi , i = 1, ..., k, where Vt is an unknown t × t matrix and mi ∈ {1, ..., m} with a fixed positive integer m. Thus, we need to obtain consistent estimators of at most m different matrices V1 , ..., Vm . It can be shown (exercise) that the following estimator is consistent for Vt when kt → ∞ as k → ∞: 1 X Vˆt = ri riτ , kt

t = 1, ..., m,

i∈Bt

where ri is the same as that in Example 3.29, Bt is the set of i’s such that mi = t, and kt is the number of i’s in Bt . Example 3.31. Suppose that V is diagonal with the ith diagonal element σi2 = ψ(Zi ), where ψ is an unknown function. The simplest case is ψ(t) = θ0 + θ1 v(Zi ) for a known function v and some unknown θ0 and θ1 . One can then obtain a consistent estimator Vˆ by using the LSE of θ0 and θ1 under the “model” Eri2 = θ0 + θ1 v(Zi ), i = 1, ..., n, (3.65) where ri = Xi − Ziτ βˆ (exercise). If ψ is nonlinear or nonparametric, some results are given in Carroll (1982) and M¨ uller and Stadrm¨ uller (1987). Finally, if Vˆ is not consistent (i.e., (3.60) does not hold), then the weighted LSE lτ βˆw can still be consistent and asymptotically normal, but

217

3.6. Exercises

its asymptotic variance is not lτ (Z τ V −1 Z)−1 l; in fact, lτ βˆw may not be asymptotically as efficient as the LSE lτ βˆ (Carroll and Cline, 1988; Chen and Shao, 1993). For example, if kVˆ −1 U − In kmax →p 0, where U is positive definite, 0 < inf n λ− [U ] ≤ supn λ+ [U ] < ∞, and U 6= V (i.e., Vˆ is inconsistent for V ), then, using the same argument as that in the proof of Theorem 3.17, we can show (exercise) that lτ (βˆw − β)/bn →d N (0, 1)

(3.66)

for any l 6= 0, where b2n = lτ (Z τ U −1 Z)−1 Z τ U −1 V U −1 Z(Z τ U −1 Z)−1 l. Hence, the asymptotic relative efficiency of the LSE lτ βˆ w.r.t. lτ βˆw can be less than 1 or larger than 1.

3.6 Exercises 1. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = p ∈ (0, 1). (a) Find the UMVUE of pm , m ≤ n. (b) Find the UMVUE of P (X1 + · · · + Xm = k), where m and k are positive integers ≤ n. (c) Find the UMVUE of P (X1 + · · · + Xn−1 > Xn ). 2. Let X1 , ..., Xn be i.i.d. having the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 > 0. (a) Find the UMVUE’s of µ3 and µ4 . d (b) Find the UMVUE’s of P (X1 ≤ t) and dt P (X1 ≤ t) with a fixed t ∈ R. 3. In Example 3.4, (a) show that the UMVUE of σ r is kn−1,r S r , where r > 1 − n; ¯ (b) prove that (X1 − X)/S has the p.d.f. given by (3.1); ¯ (c) show that (X1 − X)/S →d N (0, 1) by using (i) the SLLN and (ii) Scheff´e’s theorem (Proposition 1.18). 4. Let X1 , ..., Xm be i.i.d. having the N (µx , σx2 ) distribution and let Y1 , ..., Yn be i.i.d. having the N (µy , σy2 ) distribution. Assume that Xi ’s and Yj ’s are independent. (a) Assume that µx ∈ R, µy ∈ R, σx2 > 0, and σy2 > 0. Find the UMVUE’s of µx − µy and (σx /σy )r , where r > 0 and r < n. (b) Assume that µx ∈ R, µy ∈ R, and σx2 = σy2 > 0. Find the UMVUE’s of σx2 and (µx − µy )/σx .

218

3. Unbiased Estimation (c) Assume that µx = µy ∈ R, σx2 > 0, σy2 > 0, and σx2 /σy2 = γ is known. Find the UMVUE of µx . (d) Assume that µx = µy ∈ R, σx2 > 0, and σy2 > 0. Show that a UMVUE of µx does not exist. (e) Assume that µx ∈ R, µy ∈ R, σx2 > 0, and σy2 > 0. Find the UMVUE of P (X1 ≤ Y1 ). (f) Repeat (e) under the assumption that σx = σy .

5. Let X1 , ..., Xn be i.i.d. having the uniform distribution on the interval (θ1 −θ2 , θ1 +θ2 ), where θ1 ∈ R, θ2 > 0, and n > 2. Find the UMVUE’s of θj , j = 1, 2, and θ1 /θ2 . 6. Let X1 , ..., Xn be i.i.d. having the exponential distribution E(a, θ) with parameters θ > 0 and a ∈ R. (a) Find the UMVUE of a when θ is known. (b) Find the UMVUE of θ when a is known. (c) Find the UMVUE’s of θ and a. (d) Assume that θ is known. Find the UMVUE of P (X1 ≥ t) and d dt P (X1 ≥ t) for a fixed t > a. (e) Find the UMVUE of P (X1 ≥ t) for a fixed t > a. 7. Let X1 , ..., Xn be i.i.d. having the Pareto distribution P a(a, θ) with θ > 0 and a > 0. (a) Find the UMVUE of θ when a is known. (b) Find the UMVUE of a when θ is known. (c) Find the UMVUE’s of a and θ. 8. Consider Exercise 52(a) of §2.6. Find the UMVUE of γ. 9. Let X1 , ..., Xm be i.i.d. having the exponential distribution E(ax , θx ) with θx > 0 and ax ∈ R and Y1 , ..., Yn be i.i.d. having the exponential distribution E(ay , θy ) with θy > 0 and ay ∈ R. Assume that Xi ’s and Yj ’s are independent. (a) Find the UMVUE’s of ax − ay and θx /θy . (b) Suppose that θx = θy but it is unknown. Find the UMVUE’s of θx and (ax − ay )/θx . (c) Suppose that ax = ay but it is unknown. Show that a UMVUE of ax does not exist. (d) Suppose that n = m and ax = ay = 0 and that our sample is (Z1 , ∆1 ), ..., (Zn , ∆n ), where Zi = min{Xi , Yi } and ∆i = 1 if Xi ≥ Yi and 0 otherwise, i = 1, ..., n. Find the UMVUE of θx − θy . 10. Let X1 , ..., Xm be i.i.d. having the uniform distribution U (0, θx ) and Y1 , ..., Yn be i.i.d. having the uniform distribution U (0, θy ). Suppose that Xi ’s and Yj ’s are independent and that θx > 0 and θy > 0. Find the UMVUE of θx /θy when n > 1.

219

3.6. Exercises

11. Let X be a random variable having the negative binomial distribution N B(p, r) with an unknown p ∈ (0, 1) and a known r. (a) Find the UMVUE of pt , t < r. (b) Find the UMVUE of Var(X). (c) Find the UMVUE of log p. 12. Let X1 , ..., Xn be i.i.d. random variables having the Poisson distribution P (θ) truncated at 0, i.e., P (Xi = x) = (eθ − 1)−1 θx /x!, x = 1, 2, ..., θ > 0. Find the UMVUE of θ when n = 1, 2. 13. Let X be a random variable having the negative binomial distribution N B(p, r) truncated at r, where r is known and p ∈ (0, 1) is unknown. Let k be a fixed positive integer > r. For r = 1, 2, 3, find the UMVUE of pk . 14. Let X1 , ..., Xn be i.i.d. having the log-distribution L(p) with an unknown p ∈ (0, 1). Let k be a fixed positive integer. (a) For n = 1, 2, 3, find the UMVUE of pk . (b) For n = 1, 2, 3, find the UMVUE of P (X = k). 15. Consider Exercise 43 of §2.6. (a) Show that the estimator U = 2(|X1 | − 41 )I{X1 6=0} is unbiased for θ. (b) Derive the UMVUE of θ. 16. Derive the UMVUE of p in Exercise 33 of §2.6. 17. Derive the UMVUE’s of θ and λ in Exercise 55 of §2.6, based on data X1 , ..., Xn . 18. Suppose that (X0 , X1 , ..., Xk ) has the multinomial distribution in ExPk ample 2.7 with pi ∈ (0, 1), j=0 pj = 1. Find the UMVUE of pr00 · · · prkk , where rj ’s are nonnegative integers with r0 + · · · + rk ≤ n. 19. Let Y1 , ..., Yn be i.i.d. from the uniform distribution U (0, θ) with an unknown θ ∈ (1, ∞). (a) Suppose that we only observe if Yi ≥ 1 Yi Xi = i = 1, ..., n. 1 if Yi < 1, Derive a UMVUE of θ. (b) Suppose that we only observe if Yi ≤ 1 Yi Xi = 1 if Yi > 1,

i = 1, ..., n.

Derive a UMVUE of the probability P (Y1 > 1).

220

3. Unbiased Estimation

20. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors distributed as bivariate normal with EXi = EYi = βzi , Var(Xi ) = Var(Yi ) = σ 2 , and Cov(Xi , Yi ) = ρσ 2 , i = 1, ..., n, where β ∈ R, σ > 0, and ρ ∈ (−1, 1) are unknown parameters, and zi ’s are known constants. (a) Obtain a UMVUE of β and calculate its variance. (b) Obtain a UMVUE of σ 2 and calculate its variance. 21. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors from a population P ∈ P that is the family of all bivariate populations with Lebesgue p.d.f.’s. (a) Show that the set of n pairs (Xi , Yi ) ordered according to the value of their first coordinate constitutes a sufficient and complete statistic for P ∈ P. (b) A statistic T is a function of the complete and sufficient statistic if and only if T is invariant of the n pairs. Pnunder permutation ¯ i − Y¯ ) is the UMVUE of (c) Show that (n − 1)−1 i=1 (Xi − X)(Y Cov(X1 , Y1 ). (d) Find the UMVUE’s of P (Xi ≤ Yi ) and P (Xi ≤ Xj and Yi ≤ Yj ), i 6= j. 22. Let X1 , ..., Xn be i.i.d. from P ∈ P containing all symmetric c.d.f.’s with finite means and with Lebesgue p.d.f.’s on R. Show that there is no UMVUE of µ = EX1 when n > 1. 23. Prove Corollary 3.1. 24. Suppose that T is a UMVUE of an unknown parameter ϑ. Show that T k is a UMVUE of E(T k ), where k is any positive integer for which E(T 2k ) < ∞. 25. Consider the problem in Exercise 83 of §2.6. Use Theorem 3.2 to show that I{0} (X) is a UMVUE of (1 − p)2 and that there is no UMVUE of p. 26. Let X1 , ..., Xn be i.i.d. from a discrete distribution with P (Xi = θ − 1) = P (Xi = θ) = P (Xi = θ + 1) = 31 , where θ is an unknown integer. Show that no nonconstant function of θ has a UMVUE. 27. Let X be a random variable having the Lebesgue p.d.f. √ [(1 − θ) + θ/(2 x)]I(0,1) (x), where θ ∈ [0, 1]. Show that there is no UMVUE of θ.

3.6. Exercises

221

28. Let X be a discrete random variable with P (X = −1) = 2p(1 − p) and P (X = k) = pk (1 − p)3−k , k = 0, 1, 2, 3, where p ∈ (0, 1). (a) Determine whether there is a UMVUE of p. (b) Determine whether there is a UMVUE of p(1 − p). 29. Let X1 , ..., Xn be i.i.d. observations. Obtain a UMVUE of a in the following cases. (a) Xi has the exponential distribution E(a, θ) with a known θ and an unknown a ≤ 0. (b) Xi has the Pareto distribution P a(a, θ) with a known θ > 1 and an unknown a ∈ (0, 1]. 30. In Exercise 41 of §2.6, find a UMVUE of θ and show that it is unique a.s. 31. Prove Theorem 3.3 for the multivariate case (k > 1). 32. Let X be a single sample from Pθ . Find the Fisher information I(θ) in the following cases. (a) Pθ is the N (µ, σ 2 ) distribution with θ = µ ∈ R. (b) Pθ is the N (µ, σ 2 ) distribution with θ = σ 2 > 0. (c) Pθ is the N (µ, σ 2 ) distribution with θ = σ > 0. (d) Pθ is the N (σ, σ 2 ) distribution with θ = σ > 0. (e) Pθ is the N (µ, σ 2 ) distribution with θ = (µ, σ 2 ) ∈ R × (0, ∞). (f) Pθ is the negative binomial distribution N B(θ, r) with θ ∈ (0, 1). (g) Pθ is the gamma distribution Γ(α, γ) with θ = (α, γ) ∈ (0, ∞) × (0, ∞). (h) Pθ is the beta distribution B(α, β) with θ = (α, β) ∈ (0, 1)×(0, 1). 33. Find a function of θ for which the amount of information is independent of θ, when Pθ is (a) the Poisson distribution P (θ) with θ > 0; (b) the binomial distribution Bi(θ, r) with θ ∈ (0, 1); (c) the gamma distribution Γ(α, θ) with θ > 0. 34. Prove the result in Example 3.9. 35. Obtain the Fisher information matrix for a random variable with (a) the Cauchy distribution C(µ, σ), µ ∈ R, σ > 0; (b) the double exponential distribution DE(µ, θ), µ ∈ R, θ > 0; (c) the logistic distribution LG(µ, σ), µ ∈ R, σ > 0; (d) the c.d.f. Fr x−µ , where Fr is the c.d.f. of the t-distribution tr σ with a known r, µ ∈ R, σ > 0; (e) the Lebesgue p.d.f. fθ (x) = (1 − ǫ)φ(x − µ) + σǫ φ x−µ , θ = σ (µ, σ, ǫ) ∈ R × (0, ∞) × (0, 1), where φ is the standard normal p.d.f.

222

3. Unbiased Estimation

36. Let X be a sample having a p.d.f. satisfying the conditions in Theorem 3.3, where θ is a k-vector of unknown parameters, and let T (X) be a statistic. If T has a p.d.f. gθ satisfying the conditions in Theorem ∂ ∂ 3.3, then we define IT (θ) = E{ ∂θ log gθ (T )[ ∂θ log gθ (T )]τ } to be the Fisher information about θ contained in T . (a) Show that IX (θ) − IT (θ) is nonnegative definite, where IX (θ) is the Fisher information about θ contained in X. (b) Show that IX (θ) = IT (θ) if T is sufficient for θ. 37. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (0, θ) with θ > 0. (a) Show that condition (3.3) does not hold for h(X) = X(n) . (b) Show that the inequality in (3.6) does not hold for the UMVUE of θ. 38. Prove Proposition 3.3. 39. Let X be a single sample from the double exponential distribution DE(µ, θ) with µ = 0 and θ > 0. Find the UMVUE’s of the following parameters and, in each case, determine whether the variance of the UMVUE attains the Cram´er-Rao lower bound. (a) ϑ = θ; (b) ϑ = θr , where r > 1; (c) ϑ = (1 + θ)−1 . 40. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = p ∈ (0, 1). ¯ − X)/(n ¯ (a) Show that the UMVUE of p(1 − p) is Tn = nX(1 − 1). (b) Show that Var(Tn ) does not attain the Cram´er-Rao lower bound. (c) Show that (3.10) holds. 41. Let X1 , ..., Xn be i.i.d. having the Poisson distribution P (θ) with θ > 0. Find the amse of the UMVUE of e−tθ with a fixed t > 0 and show that (3.10) holds. 42. Let X1 , ..., Xn be i.i.d. having the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 > 0. (a) Find the UMVUE of ϑ = etµ with a fixed t 6= 0. (b) Determine whether the variance of the UMVUE in (a) attains the Cram´er-Rao lower bound. (c) Show that (3.10) holds. 43. Show that if X1 , ..., Xn are i.i.d. binary random variables, Un in (3.12) equals P T (T − 1) · · · (T − m + 1)/[n(n − 1) · · · (n − m + 1)], where T = ni=1 Xi .

223

3.6. Exercises

¯ then Un in (3.13) is the same as the sample 44. Show that if Tn = X, 2 variance S in (2.2). Show that (3.23) holds for Tn given by (3.22) with E(Rn2 ) = o(n−1 ). 45. Prove (3.14), (3.16), and (3.17). 46. Let ζk be given in Theorem 3.4. Show that ζ1 ≤ ζ2 ≤ · · · ≤ ζm . 47. Prove Corollary 3.2. ˇn is also a U-statistic. 48. Prove (3.20) and show that Un − U

ˇ 49. Let Tn be a symmetric statistic with Var(Tn ) < ∞ for every n and Tn be the projection of Tn on nk random vectors {Xi1 , ..., Xik }, 1 ≤ i1 < · · · < ik ≤ n. Show that E(Tn ) = E(Tˇn ) and calculate E(Tn − Tˇn )2 .

50. Let Yk be defined in Lemma 3.2. Show that {Yk2 } is uniformly integrable. 51. Show that (3.22) with E(Rn2 ) = o(n−1 ) is satisfied for Tn being a U-statistic with E[h(X1 , ..., Xm )]2 < ∞. 52. Let S 2 be the sample variance given by (2.2), which is also a Ustatistic (§3.2.1). Find the corresponding h1 , h2 , ζ1 , and ζ2 . Discuss how to apply Theorem 3.5 to this case. 53. Let h(x1 , x2 , x3 ) = I(−∞,0) (x1 + x2 + x3 ). Define the U-statistic with this kernel and find hk and ζk , k = 1, 2, 3. 54. Let X1 , ..., Xn be i.i.d. random variables having finite µ = EX1 and µ ¯ = EX1−1 . Find a U-statistic that is an unbiased estimator of µ¯ µ and derive its variance and asymptotic distribution. 55. Show that βˆ is an LSE of β if and only if it is given by (3.29). 56. Obtain explicit forms for the LSE’s of βj , j = 0, 1, and SSR, under the simple linear regression model in Example 3.11, assuming that some ti ’s are different. 57. Consider the polynomial model Xi = β0 + β1 ti + β2 t2i + εi ,

i = 1, ..., n.

Find explicit forms for the LSE’s of βj , j = 0, 1, 2, and SSR, assuming that some ti ’s are different. 58. Suppose that Xij = αi + βtij + εij ,

i = 1, ..., a, j = 1, ..., b.

Find explicit forms for the LSE’s of β, αi , i = 1, ..., a, and SSR.

224

3. Unbiased Estimation

59. Consider the polynomial model Xi = β0 + β1 ti + β2 t2i + β3 t3i + εi ,

i = 1, ..., n,

where εi ’s are i.i.d. from N (0, σ 2 ). Suppose that n = 12, ti = −1, i = 1, ..., 4, ti = 0, i = 5, ..., 8, and ti = 1, i = 9, ..., 12. (a) Obtain the matrix Z τ Z when this polynomial model is considered as a special case of model (3.24). (b) Show whether the following parameters are estimable: β0 + β2 , β1 , β0 − β1 , β1 + β3 , and β0 + β1 + β2 + β3 . 60. Find the matrix Z, Z τ Z, and the form of l ∈ R(Z) under the one-way ANOVA model (3.31). 61. Obtain the matrix Z under the two-way balanced ANOVA model (3.32). Show that the rank of Z is ab. Verify the form of the LSE of β given in Example 3.14. Find the form of l ∈ R(Z). 62. Consider the following model as a special case of model (3.25): Xijk = µ + αi + βj + εijk ,

i = 1, ..., a, j = 1, ..., b, k = 1, ..., c.

Obtain the matrix Z, the parameter vector β, and the form of LSE’s of β. Discuss conditions under which l ∈ R(Z). 63. Under model (3.25) and assumption A1, find the UMVUE’s of (lτ β)2 , lτ β/σ, and (lτ β/σ)2 for an estimable lτ β. 64. Verify the formulas for SSR’s in Example 3.15. 65. Consider the one-way random effects model in Example 3.17. Assume that ni = n for all i and that Ai ’s and eij ’s are normally distributed. Show that the family of populations is an exponential family with ¯ ·· , SA = n Pm (X ¯ i· − X ¯ ·· )2 , and sufficient and complete statistics X i=1 Pm Pn 2 2 ¯ SE = i=1 j=1 (Xij − Xi· ) . Find the UMVUE’s of µ, σa , and σ 2 .

66. Consider model (3.25). Suppose that εi ’s are i.i.d. with Eεi = 0 and a Lebesgue p.d.f. σ −1 f (x/σ), where f is a known Lebesgue p.d.f. and σ > 0 is unknown. (a) Show that X is from a location-scale family given by (2.10). (b) Find the Fisher information about (β, σ) contained in Xi . (c) Find the Fisher information about (β, σ) contained in X.

67. Consider model (3.25) with assumption A2. Let c ∈ Rp . Show that if the equation c = Z τ y has a solution, then there is a unique solution y0 ∈ R(Z τ ) such that Var(y0τ X) ≤ Var(y τ X) for any other solution of c = Z τ y.

225

3.6. Exercises

68. Consider model (3.25). Show that the number of independent linear functions of X with mean 0 is n − r, where r is the rank of Z. ˆ which ˆ i = Z τ β, 69. Consider model (3.25) with assumption A2. Let X i is called the least squares prediction of Xi . Let hij be the (i, j)th element of Z(Z τ Z)− Z τ and hi = hii . Show that ˆ i ) = σ 2 hi ; (a) Var(X ˆ i ) = σ 2 (1 − hi ); (b) Var(Xi − X ˆi , X ˆ j ) = σ 2 hij ; (c) Cov(X ˆ i , Xj − X ˆ j ) = −σ 2 hij , i 6= j; (d) Cov(Xi − X ˆ ˆ (e) Cov(Xi , Xj − Xj ) = 0.

70. Consider model (3.25) with assumption A2. Let Z = (Z1 , Z2 ) and β = (β1 , β2 ), where Zj is n × pj and βj is a pj -vector, j = 1, 2. Assume that (Z1τ Z1 )−1 and [Z2τ Z2 − Z2τ Z1 (Z1τ Z1 )−1 Z1τ Z2 ]−1 exist. (a) Derive the LSE of β in terms of Z1 , Z2 , and X. (b) Let βˆ = (βˆ1 , βˆ2 ) be the LSE in (a). Calculate the covariance between βˆ1 and βˆ2 . (c) Suppose that it is known that β2 = 0. Let β˜1 be the LSE of β1 under the reduced model X = Z1 β1 + ε. Show that, for any l ∈ Rp1 , lτ β˜1 is better than lτ βˆ1 in terms of their mse’s. 71. Prove that (e) implies (b) in Theorem 3.10. 72. Show that (a) in Theorem 3.10 is equivalent to either (f) Var(ε)Z = ZB for some matrix B, or (g) R(Z τ ) is generated by r eigenvectors of Var(ε), where r is the rank of Z. 73. Prove Corollary 3.3. 74. Suppose that X = µJn + Hξ + e, where µ ∈ R is an unknown parameter, Jn is the n-vector of 1’s, H is an n × p known matrix of full rank, ξ is a random p-vector with E(ξ) = 0 and Var(ξ) = σξ2 Ip , e is a random n-vector with E(e) = 0 and Var(e) = σ 2 In , and ξ and e are independent. Show that the LSE of µ is the BLUE if and only if the row totals of HH τ are the same. 75. Consider a special case of model (3.25): Xij = µ + αi + βj + εij ,

i = 1, ..., a, j = 1, ..., b,

where µ, αi ’s, and βj ’s are unknown parameters, E(εij ) = 0, Var(εij ) = σ 2 , Cov(εij , εi′ j ′ ) = 0 if i 6= i′ , and Cov(εij , εij ′ ) = σ 2 ρ if j 6= j ′ . Show that the LSE of lτ β is the BLUE for any l ∈ R(Z).

226

3. Unbiased Estimation

76. Consider model (3.25) under assumption A3 with Var(ε) = a block diagonal matrix whose ith block diagonal Vi is ni ×ni and has a single eigenvalue λi with eigenvector Jni (the ni -vector of 1’s) and a repeated Pk eigenvalue ρi with multiplicity ni − 1, i = 1, ..., k, i=1 ni = n. Let U be the n × k matrix whose ith column is Ui , where U1 = (Jnτ1 , 0, ..., 0), U2 = (0, Jnτ2 , ..., 0),..., Uk = (0, 0, ..., Jnτk ). (a) If R(Z τ ) ⊂ R(U τ ) and λi ≡ λ, show that lτ βˆ is the BLUE of lτ β for any l ∈ R(Z). (b) If Z τ Ui = 0 for all i and ρi ≡ ρ, show that lτ βˆ is the BLUE of lτ β for any l ∈ R(Z). 77. Prove Proposition 3.4. 78. Show that the condition supn λ+ [Var(ε)] < ∞ is equivalent to the condition supi Var(εi ) < ∞. 79. Find a condition under which the mse of lτ βˆ is of the order n−1 . Apply it to problems in Exercises 56, 58, and 60-62. 80. Consider model (3.25) with i.i.d. ε1 , ..., εn having E(εi ) = 0 and ˆ i = Z τ βˆ and hi = Z τ (Z τ Z)− Zi . Var(εi ) = σ 2 . Let X i i (a) Show that for any ǫ > 0, ˆi − E X ˆ i | ≥ ǫ) ≥ min{P (εi ≥ ǫ/hi ), P (εi ≤ −ǫ/hi )}. P (|X (Hint: for independent random variables X and Y , P (|X + Y | ≥ ǫ) ≥ P (X ≥ ǫ)P (Y ≥ 0) + P (X ≤ −ǫ)P (Y < 0).) ˆi − EX ˆ i →p 0 if and only if hi → 0. (b) Show that X 81. Prove Lemma 3.3 and show that condition (a) is implied by {kZi k} being bounded and λ+ (Z τ Z)− → 0. 82. Consider the problem in Exercise 58. Suppose that {tij } is bounded. Find a condition under which (3.39) holds. 83. Under the two-way ANOVA models in Example 3.14 and Exercise 62, find sufficient conditions for (3.39). 84. Consider the one-way random effects model in Example 3.17. Assume that {ni } is bounded and E|eij |2+δ < ∞ for some δ > 0. Show that the LSE µ ˆ of µ is asymptotically normal and derive an explicit form of Var(ˆ µ). 85. Suppose that Xi = ρti + εi ,

i = 1, ..., n,

where ρ ∈ R is an unknown parameter, ti ’s are known and in (a, b), a and b are known positive constants, and εi ’s are independent random

3.6. Exercises

227

variables satisfying E(εi ) = 0, E|εi |2+δ < ∞ for some δ > 0, and Var(εi ) = σ 2 ti with an unknown σ 2 > 0. (a) Obtain the LSE of ρ. (b) Obtain the BLUE of ρ. (c) Show that both the LSE and BLUE are asymptotically normal and obtain the asymptotic relative efficiency of the BLUE w.r.t. the LSE. 86. In Example 3.19, show that E(S 2 ) = σ 2 given in (3.43). 87. Suppose that X = (X1 , ..., Xn ) is a simple random sample (without replacement) from a finite population P = {y1 , ..., yN } with univariate yi . (a) Show that a necessary condition for h(θ) to be estimable is that h is symmetric in its N arguments. (b) Find the UMVUE of Y m , where m is a fixed positive integer < n and Y is the population total. (c) Find the UMVUE of P (Xi ≤ Xj ), i 6= j. (d) Find the UMVUE of Cov(Xi , Xj ), i 6= j. 88. Prove Theorem 3.14. 89. Under stratified simple random sampling described in §3.4.1, show that the vector of ordered values of all Xhi ’s is neither sufficient nor complete for θ ∈ Θ. 90. Let P = {y1 , ..., yN } be a population with univariate yi . Define the P population c.d.f. by F (t) = N −1 N i=1 I(−∞,t] (yi ). Find the UMVUE of F (t) under (a) simple random sampling and (b) stratified simple random sampling. 91. Consider the estimation of F (t) in the previous exercise. Suppose that a sample of size n is selected with πi > 0. Find the Horvitz-Thompson estimator of F (t). Is it a c.d.f.? 92. Show that v1 in (3.49) and v2 in (3.50) are unbiased estimators of Var(Yˆht ). Prove that v1 = v2 under (a) simple random sampling and (b) stratified simple random sampling. 93. Consider the following two-stage stratified sampling plan. In the first stage, the population is stratified into H strata and kh clusters are selected from stratum h with probability proportional to cluster size, where sampling is independent across strata. In the second stage, a sample of mhi units is selected from sampled cluster i in stratum h, and sampling is independent across clusters. Find πi and the HorvitzThompson estimator Yˆht of the population total.

228

3. Unbiased Estimation

94. In the previous exercise, prove the unbiasedness of Yˆht directly (without using Theorem 3.15). 95. Under systematic sampling, show that Var(Yˆsy ) is equal to k 1 σ2 2 X 1− + N n nN i=1

X

1≤t 0. Consider the estimation of ϑ = EΦ(a + bX1 ), where Φ is the standard normal c.d.f. and a and b are known constants. Obtain an explicit ¯ S 2 ). form of a function g(µ, σ 2 ) = ϑ and the amse of ϑˆ = g(X, 101. Let X1 , ..., Xn be i.i.d. with mean µ, variance σ 2 , and finite µj = EX1j , ¯ j = 2, 3, 4. The sample coefficient of variation is defined to be S/X, 2 where S is the squared root sample variance S . √ of the ¯ − σ/µ) →d N (0, τ ) and obtain an (a) If µ 6= 0, show that n(S/X explicit formula of τ in terms of µ, σ 2 , and µj . ¯ →d [N (0, 1)]−1 . (b) If µ = 0, show that n−1/2 S/X 102. Prove (3.52) and (3.53). 103. Let X1 , ..., Xn be i.i.d. from P in a parametric family. Obtain moment estimators of parameters in the following cases. (a) P is the gamma distribution Γ(α, γ), α > 0, γ > 0. (b) P is the exponential distribution E(a, θ), a ∈ R, θ > 0. (c) P is the beta distribution B(α, β), α > 0, β > 0. (d) P is the log-normal distribution LN (µ, σ 2 ), µ ∈ R, σ > 0. (e) P is the uniform distribution U (θ − 12 , θ + 12 ), θ ∈ R. (f) P is the negative binomial distribution N B(p, r), p ∈ (0, 1), r = 1, 2,....

229

3.6. Exercises

(g) P is the log-distribution L(p), p ∈ (0, 1). (h) P is the log-normal distribution LN (µ, σ2 ), µ ∈ R, σ = 1. (i) P is the chi-square distribution χ2k with an unknown k = 1, 2, .... 104. Obtain moment estimators of λ and p in Exercise 55 of §2.6, based on data X1 , ..., Xn . 105. Obtain the asymptotic distributions of the moment estimators in Exercise 103(a), (c), (e), and (g), and the asymptotic relative efficiencies of moment estimators w.r.t. UMVUE’s in Exercise 103(b) and (h). 106. In Exercise 19(a), find a moment estimator of θ and derive its asymptotic distribution. In Exercise 19(b), obtain a moment estimator of θ−1 and its asymptotic relative efficiency w.r.t. the UMVUE of θ−1 . 107. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. fα,β (x) = αβ −α xα−1 I(0,β) (x), where α > 0 and β > 0 are unknown. (a) Obtain moment estimators of α and β. (b) Obtain the asymptotic distribution of the moment estimators of α and β derived in (a). 108. Let X1 , ..., Xn be i.i.d. from the following discrete distribution: P (X1 = 1) =

2(1 − θ) , 2−θ

P (X1 = 2) =

θ , 2−θ

where θ ∈ (0, 1) is unknown. (a) Obtain an estimator of θ using the method of moments. (b) Obtain the amse of the moment estimator in (a). 109. Let X1 , ..., Xn (n > 1) be i.i.d. from a population having the Lebesgue p.d.f. x−µ ǫ fθ (x) = (1 − ǫ)φ(x − µ) + φ , σ σ where φ is the standard normal p.d.f., θ = (µ, σ) ∈ R × (0, ∞) is unknown, and ǫ ∈ (0, 1) is a known constant. (a) Obtain an estimator of θ using the method of moments. (b) Obtain the asymptotic distribution of the moment estimator in part (a). 110. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. x>0 (θ1 + θ2 )−1 e−x/θ1 fθ1 ,θ2 (x) = x ≤ 0, (θ1 + θ2 )−1 ex/θ2 where θ1 > 0 and θ2 > 0 are unknown. (a) Obtain an estimator of (θ1 , θ2 ) using the method of moments.

230

3. Unbiased Estimation

(b) Obtain the asymptotic distribution of the moment estimator in part (a). 111. (Nonexistence of a moment estimator). Consider X1 , ..., Xn and the parametric family indexed by (θ, j) ∈ (0, 1) × {1, 2} in Exercise 41 of §2.6. Let hi (θ, j) = EX1i , i = 1, 2. Show that P (ˆ µi = hi (θ, j) has a solution) → 0 as n → ∞, when Xi ’s are from the Poisson distribution P (θ). 112. In the proof of Proposition 3.5, show that E[Wn (Un − ϑ)] = O(n−1 ). 113. Assume the conditions of Theorem 3.16. (a) Prove (3.56). P∞ (b) Show that E[g2 (X1 , X1 )]/n = nEVn2 = m(m − 1) j=1 λj /(2n).

114. Let X1 , ..., Xn be i.i.d.Rwith a c.d.f. F and Un and Vn be the U- and Vstatistics with kernel [I(−∞,y] (x1 ) − F0 (y)][I(−∞,y] (x2 ) − F0 (y)]dF0 , where F0 is a known c.d.f. (a) Obtain the asymptotic distributions of Un and Vn when F 6= F0 . (b) Obtain the asymptotic relative efficiency of Un w.r.t. Vn when F = F0 . 115. Let X1 , ..., Xn be i.i.d. with a c.d.f. F having a finite sixth moment. Consider the estimation of µ3 , where µ = EX1 . When µ = 0, find −1 P amseX¯ 3 (P )/amseUn (P ), where Un = n3 1≤i<j 0. Under the ˆ in (3.64) is consistent for Σ in the sense same conditions, show that Σ ˆ that kΣ − Σkmax →p 0. 118. In Example 3.30, show that Vˆt is consistent for Vt when kt → ∞ as k → ∞. 119. Show how to use equation (3.65) to obtain consistent estimators of θ0 and θ1 . 120. Prove (3.66) under the assumed conditions in §3.5.4.

Chapter 4

Estimation in Parametric Models In this chapter, we consider point estimation methods in parametric models. One such method, the moment method, has been introduced in §3.5.2. It is assumed in this chapter that the sample X is from a population in a parametric family P = {Pθ : θ ∈ Θ}, where Θ ⊂ Rk for a fixed integer k ≥ 1.

4.1 Bayes Decisions and Estimators Bayes rules are introduced in §2.3.2 as decision rules minimizing the average risk w.r.t. a given probability measure Π on Θ. Bayes rules, however, are optimal rules in the Bayesian approach, which is fundamentally different from the classical frequentist approach that we have been adopting.

4.1.1 Bayes actions In the Bayesian approach, θ is viewed as a realization of a random vector θ whose prior distribution is Π. The prior distribution is based on past experience, past data, or a statistician’s belief and thus may be very subjective. A sample X is drawn from Pθ = Px|θ , which is viewed as the conditional distribution of X given θ = θ. The sample X = x is then used to obtain an updated prior distribution, which is called the posterior distribution and can be derived as follows. By Theorem 1.7, the joint distribution of X and θ is a probability measure on X × Θ determined by Z P (A × B) = Px|θ (A)dΠ(θ), A ∈ BX , B ∈ BΘ , B

231

232

4. Estimation in Parametric Models

where X is the range of X. The posterior distribution of θ, given X = x, is the conditional distribution Pθ|x whose existence is guaranteed by Theorem 1.7 for almost all x ∈ X. When Px|θ has a p.d.f., the following result provides a formula for the p.d.f. of the posterior distribution Pθ|x . Theorem 4.1 (Bayes formula). Assume that P = {Px|θ : θ ∈ Θ} is dP

dominated by a σ-finite measure ν and fθ (x) = dνx|θ (x) is a Borel function on (X × RΘ, σ(BX × BΘ )). Let Π be a prior distribution on Θ. Suppose that m(x) = Θ fθ (x)dΠ > 0. (i) The posterior distribution Pθ|x ≪ Π and dPθ|x fθ (x) = . dΠ m(x) (ii) If Π ≪ λ and

dΠ dλ

= π(θ) for a σ-finite measure λ, then dPθ|x fθ (x)π(θ) = . dλ m(x)

(4.1)

Proof. Result (ii) follows from result (i) and Proposition 1.7(iii). To show (i), we first show that m(x) < ∞ a.e. ν. Note that Z Z Z Z Z m(x)dν = fθ (x)dΠdν = fθ (x)dνdΠ = 1, (4.2) X

X

Θ

Θ

X

where the second equality follows from Fubini’s theorem. Thus, m(x) is integrable w.r.t. ν and m(x) < ∞ a.e. ν. For x ∈ X with m(x) < ∞, define Z 1 P (B, x) = fθ (x)dΠ, B ∈ BΘ . m(x) B Then P (·, x) is a probability measure on Θ a.e. ν. By Theorem 1.7, it remains to show that P (B, x) = P (θ ∈ B|X = x). By Fubini’s theorem, P (B, ·) is a measurable function of x. Let Px,θ denote the “joint” distribution of (X, θ). For any A ∈ σ(X), Z Z Z IB (θ)dPx,θ = fθ (x)dΠdν A×Θ A B Z Z Z fθ (x) dΠ fθ (x)dΠ dν = A B m(x) Θ Z Z Z fθ (x) dΠ fθ (x)dνdΠ = m(x) ZΘ A B = P (B, x)dPx,θ , A×Θ

233

4.1. Bayes Decisions and Estimators

where the third equality follows from Fubini’s theorem. This completes the proof. Because of (4.2), m(x) is called the marginal p.d.f. of X w.r.t. ν. If m(x) = 0 for an x ∈ X, then fθ (x) = 0 a.s. Π. Thus, either x should be eliminated from X or the prior Π is incorrect and a new prior should be specified. Therefore, without loss of generality we may assume that the assumption of m(x) > 0 in Theorem 4.1 is always satisfied. If both X and θ are discrete and ν and λ are the counting measures, then (4.1) becomes P (θ = θ|X = x) = P

P (X = x|θ = θ)P (θ = θ) , θ∈Θ P (X = x|θ = θ)P (θ = θ)

which is the Bayes formula that appears in elementary probability. In the Bayesian approach, the posterior distribution Pθ|x contains all the information we have about θ and, therefore, statistical decisions and inference should be made based on Pθ|x , conditional on the observed X = x. In the problem of estimating θ, Pθ|x can be viewed as a randomized decision rule under the approach discussed in §2.3. Definition 4.1. Let A be an action space in a decision problem and L(θ, a) ≥ 0 be a loss function. For any x ∈ X, a Bayes action w.r.t. Π is any δ(x) ∈ A such that E[L(θ, δ(x))|X = x] = min E[L(θ, a)|X = x], a∈A

(4.3)

where the expectation is w.r.t. the posterior distribution Pθ|x . The existence and uniqueness of Bayes actions can be discussed under some conditions on the loss function and the action space. Proposition 4.1. Assume that the conditions in Theorem 4.1 hold; L(θ, a) is convex in a for each fixed θ; and for each x ∈ X, E[L(θ, a)|X = x] < ∞ for some a. (i) If A is a compact subset of Rp for some integer p ≥ 1, then a Bayes action δ(x) exists for each x ∈ X. (ii) If A = Rp and L(θ, a) tends to ∞ as kak → ∞ uniformly in θ ∈ Θ0 ⊂ Θ with Π(Θ0 ) > 0, then a Bayes action δ(x) exists for each x ∈ X. (iii) In (i) or (ii), if L(θ, a) is strictly convex in a for each fixed θ, then the Bayes action is unique. Proof. The convexity of the loss function implies the convexity and continuity of E[L(θ, a)|X = x] as a function of a with any fixed x. Then, the result in (i) follows from the fact that any continuous function on a compact

234

4. Estimation in Parametric Models

set attains its minimum. The result in (ii) follows from the fact that Z L(θ, a)dPθ|x = ∞ lim E[L(θ, a)|X = x] ≥ lim kak→∞

kak→∞

Θ0

under the assumed condition in (ii). Finally, the result in (iii) follows from the fact that E[L(θ, a)|X = x] is strictly convex in a for any fixed x under the assumed conditions. Other conditions on L under which a Bayes action exists can be found, for example, in Lehmann (1983, §1.6 and §4.1). Example 4.1. Consider the estimation of ϑ = g(θ) for some real-valued R function g such that Θ [g(θ)]2 dΠ < ∞. Suppose that A = the range of g(θ) and L(θ, a) = [g(θ) − a]2 (squared error loss). Using the same argument as in Example 1.22, we obtain the Bayes action R R g(θ)fθ (x)dΠ Θ g(θ)fθ (x)dΠ δ(x) = = ΘR , (4.4) m(x) f (x)dΠ Θ θ

which is the posterior expectation of g(θ), given X = x. More specifically, let us consider the case where g(θ) = θj for some integer j ≥ 1, fθ (x) = e−θ θx I{0,1,2,...} (x)/x! (the Poisson distribution) with θ > 0, and Π has a Lebesgue p.d.f. π(θ) = θα−1 e−θ/γ I(0,∞) (θ)/[Γ(α)γ α ] (the gamma distribution Γ(α, γ) with known α > 0 and γ > 0). Then, for x = 0, 1, 2, ..., fθ (x)π(θ) = c(x)θx+α−1 e−θ(γ+1)/γ I(0,∞) (θ), m(x)

(4.5)

where c(x) is some function of x. By using Theorem 4.1 and matching the right-hand side of (4.5) with that of the p.d.f. of the gamma distribution, we know that the posterior is the gamma distribution Γ(x + α, γ/(γ + 1)). Hence, without actually working out the integral m(x), we know that c(x) = (1 + γ −1 )x+α /Γ(x + α). Then Z ∞ θj+x+α−1 e−θ(γ+1)/γ dθ. δ(x) = c(x) 0

Note that the integrand is proportional to the p.d.f. of the gamma distribution Γ(j + x + α, γ/(γ + 1)). Hence δ(x) = c(x)Γ(j + x + α)/(1 + γ −1 )j+x+α = (j + x + α − 1) · · · (x + α)/(1 + γ −1 )j . In particular, δ(x) = (x + α)γ/(γ + 1) when j = 1.

235

4.1. Bayes Decisions and Estimators

An interesting phenomenon in Example 4.1 is that the prior and the posterior are in the same parametric family of distributions. Such a prior is called a conjugate prior. Under a conjugate prior, Bayes actions often have explicit forms (in x) when the loss function is simple. Whether a prior is conjugate involves a pair of families; one is the family P = {fθ : θ ∈ Θ} and the other is the family from which Π is chosen. Example 4.1 shows that the Poisson family and the gamma family produce conjugate priors. It can be shown (exercise) that many pairs of families in Table 1.1 (page 18) and Table 1.2 (pages 20-21) produce conjugate priors. In general, numerical methods have to be used in evaluating the integrals in (4.4) or Bayes actions under general loss functions. Even under a conjugate prior, the integral in (4.4) involving a general g may not have an explicit form. More discussions on the computation of Bayes actions are given in §4.1.4. As an example of deriving a Bayes action in a general decision problem, we consider Example 2.21. Example 4.2. Consider the decision problem in Example 2.21. Let Pθ|x be the posterior distribution of θ, given X = x. In this problem, A = {a1 , a2 , a3 }, which is compact in R. By Proposition 4.1, we know that there is a Bayes action if the mean of Pθ|x is finite. Let Eθ|x be the expectation w.r.t. Pθ|x . Since A contains only three elements, a Bayes action can be obtained by comparing j=1 c1 Eθ|x [L(θ, aj )] = j=2 c2 + c3 Eθ|x [ψ(θ, t)] j = 3, c3 Eθ|x [ψ(θ, 0)] where ψ(θ, t) = (θ − θ0 − t)I(θ0 +t,∞) (θ).

The minimization problem (4.3) is the same as the minimization problem Z Z L(θ, δ(x))fθ (x)dΠ = min L(θ, a)fθ (x)dΠ. (4.6) Θ

a∈A

Θ

The minimization problem (4.6) is still defined even if Π is not a probability measure but a σ-finite measure on Θ, in which case m(x) may not be finite. If Π(Θ) 6= 1, Π is called an improper prior. A prior with Π(Θ) = 1 is then called a proper prior. An action δ(x) that satisfies (4.6) with an improper prior is called a generalized Bayes action. The following is a reason why we need to discuss improper priors and generalized Bayes actions. In many cases, one has no past information and has to choose a prior subjectively. In such cases, one would like to select a noninformative prior that tries to treat all parameter values in Θ

236

4. Estimation in Parametric Models

equitably. A noninformative prior is often improper. We only provide one example here. For more detailed discussions of the use of improper priors, see Jeffreys (1939, 1948, 1961), Box and Tiao (1973), and Berger (1985). Example 4.3. Suppose that X = (X1 , ..., Xn ) and Xi ’s are i.i.d. from N (µ, σ 2 ), where µ ∈ Θ ⊂ R is unknown and σ 2 is known. Consider the estimation of ϑ = µ under the squared error loss. If Θ = [a, b] with −∞ < a < b < ∞, then a noninformative prior that treats all parameter values equitably is the uniform distribution on [a, b]. If Θ = R, however, the corresponding “uniform distribution” is the Lebesgue measure on R, which is an improper prior. If Π is the Lebesgue measure on R, then ) ( n Z ∞ X (xi − µ)2 2 −n/2 2 (2πσ ) dµ < ∞. µ exp − 2σ 2 −∞ i=1 By differentiating a in (2πσ 2 )−n/2

Z

∞

−∞

(

(µ − a)2 exp −

n X (xi − µ)2 i=1

2σ 2

)

dµ

P P and using the fact that ni=1 (xi − µ)2 = ni=1 (xi − x ¯)2 + n(¯ x − µ)2 , where x ¯ is the sample mean of the observations x1 , ..., xn , we obtain that R∞ µ exp −n(¯ x − µ)2 /(2σ 2 ) dµ −∞ =x ¯. δ(x) = R ∞ exp {−n(¯ x − µ)2 /(2σ 2 )} dµ −∞

Thus, the sample mean is a generalized Bayes action under the squared error loss. From Example 2.25 and Exercise 91 in §2.6, if Π is N (µ0 , σ02 ), then the Bayes action is µ∗ (x) in (2.25). Note that in this case x ¯ is a limit of µ∗ (x) as σ02 → ∞.

4.1.2 Empirical and hierarchical Bayes methods A Bayes action depends on the chosen prior that may depend on some parameters called hyperparameters. In §4.1.1, hyperparameters are assumed to be known. If hyperparameters are unknown, one way to solve the problem is to estimate them using data x1 , ..., xn ; the resulting Bayes action is called an empirical Bayes action. The simplest empirical Bayes method is to estimate prior parameters by viewing x = (x1 , ..., xn ) as a “sample” from the marginal distribution Z Px|ξ (A) = Px|θ (A)dΠθ|ξ , A ∈ BX , Θ

237

4.1. Bayes Decisions and Estimators

where Πθ|ξ is a prior depending on an unknown vector ξ of hyperparameters, or from the marginal p.d.f. m(x) in (4.2), if Px|θ has a p.d.f. fθ . The method of moments introduced in §3.5.3, for example, can be applied to estimate ξ. We consider an example. Example 4.4. Let X = (X1 , ..., Xn ) and Xi ’s be i.i.d. from N (µ, σ 2 ) with an unknown µ ∈ R and a known σ 2 . Consider the prior Πµ|ξ = N (µ0 , σ02 ) with ξ = (µ0 , σ02 ). To obtain a moment estimate of ξ, we need to calculate Z Z x1 m(x)dx and x21 m(x)dx, Rn

Rn

where x = (x1 , ..., xn ). These two integrals can be obtained without calculating m(x). Note that Z Z Z Z x1 m(x)dx = x1 fµ (x)dxdΠµ|ξ = µdΠµ|ξ = µ0 Rn

and Z

Rn

x21 m(x)dx =

Rn

Θ

Z Z Θ

Rn

x21 fµ (x)dxdΠµ|ξ = σ 2 +

R

Z

R

µ2 dΠµ|ξ = σ 2 +µ20 +σ02 .

Thus, by viewing x1 , ..., xn as a sample from m(x), we obtain the moment estimates n 1X µ ˆ0 = x¯ and σ ˆ02 = (xi − x¯)2 − σ 2 , n i=1

where x ¯ is the sample mean of xi ’s. Replacing µ0 and σ02 in formula (2.25) (Example 2.25) by µ ˆ0 and σ ˆ02 , respectively, we find that the empirical Bayes action under the squared error loss is simply the sample mean x¯ (which is a generalized Bayes action; see Example 4.3). Note that σ ˆ02 in Example 4.4 can be negative. Better empirical Bayes methods can be found, for example, in Berger (1985, §4.5). The following method, called the hierarchical Bayes method, is generally better than empirical Bayes methods. Instead of estimating hyperparameters, in the hierarchical Bayes approach we put a prior on hyperparameters. Let Πθ|ξ be a (first-stage) prior with a hyperparameter vector ξ and let Λ be a prior on Ξ, the range of ξ. Then the “marginal” prior for θ is defined by Z Π(B) = Πθ|ξ (B)dΛ(ξ), B ∈ BΘ . (4.7) Ξ

If the second-stage prior Λ also depends on some unknown hyperparameters, then one can go on to consider a third-stage prior. In most applications,

238

4. Estimation in Parametric Models

however, two-stage priors are sufficient, since misspecifying a second-stage prior is much less serious than misspecifying a first-stage prior (Berger, 1985, §4.6). In addition, the second-stage prior can be chosen to be noninformative (improper). Bayes actions can be obtained in the same way as before using the prior in (4.7). Thus, the hierarchical Bayes method is simply a Bayes method with a hierarchical prior. Empirical Bayes methods, however, deviate from the Bayes method since x1 , ..., xn are used to estimate hyperparameters. Suppose that X has a p.d.f. fθ (x) w.r.t. a σ-finite measure ν and Πθ|ξ has a p.d.f. πθ|ξ (θ) w.r.t. a σ-finite measure κ. Then the prior Π in (4.7) has a p.d.f. Z π(θ) = πθ|ξ (θ)dΛ(ξ) Ξ

w.r.t. κ and

m(x) =

Z Z Θ

fθ (x)πθ|ξ (θ)dΛdκ. Ξ

Let Pθ|x,ξ be the posterior distribution of θ given x and ξ (or ξ is assumed known) and Z mx|ξ (x) = fθ (x)πθ|ξ (θ)dκ, Θ

which is the marginal of X given ξ (or ξ is assumed known). Then the posterior distribution Pθ|x has a p.d.f. dPθ|x fθ (x)π(θ) = dκ m(x) Z fθ (x)πθ|ξ (θ) dΛ(ξ) = m(x) Ξ Z fθ (x)πθ|ξ (θ) mx|ξ (x) = dΛ(ξ) mx|ξ (x) m(x) Ξ Z dPθ|x,ξ dPξ|x , = dκ Ξ

where Pξ|x is the posterior distribution of ξ given x. Thus, under the estimation problem considered in Example 4.1, the (hierarchical) Bayes action is Z δ(x) = δ(x, ξ)dPξ|x , (4.8) Ξ

where δ(x, ξ) is the Bayes action when ξ is known. A result similar to (4.8) is given in Lemma 4.1.

Example 4.5. Consider Example 4.4 again. Suppose that one of the parameters in the first-stage prior N (µ0 , σ02 ), µ0 , is unknown and σ02 is

239

4.1. Bayes Decisions and Estimators

known. Let the second-stage prior for ξ = µ0 be the Lebesgue measure on R (improper prior). From Example 2.25, δ(x, ξ) =

σ2 nσ02 ξ+ x ¯. 2 +σ nσ02 + σ 2

nσ02

To obtain the Bayes action δ(x), it suffices to calculate Eξ|x (ξ), where the expectation is w.r.t. Pξ|x . Note that the p.d.f. of Pξ|x is proportional to Z ∞ o n 2 −µ)2 dµ. ψ(ξ) = exp − n(¯x2σ − (µ−ξ) 2 2 2σ 0

−∞

Using the properties of normal distributions, one can show that −1 2 ξ ξ2 n 1 n¯ x − 2σ2 ψ(ξ) = C1 exp 2σ2 + 2σ02 2σ2 + 2σ02 0 o n 2 xξ = C2 exp − 2(nσnξ2 +σ2 ) + nσn¯ 2 2 0 0 +σ o n 2 n(ξ−¯ x) = C3 exp − 2(nσ , 2 +σ2 ) 0

where C1 , C2 , and C3 are quantities not depending on ξ. Hence Eξ|x (ξ) = x¯. The (hierarchical) generalized Bayes action is then δ(x) =

σ2 nσ02 E (ξ) + x ¯=x ¯. ξ|x nσ02 + σ 2 nσ02 + σ 2

4.1.3 Bayes rules and estimators The discussion in §4.1.1 and §4.1.2 is more general than point estimation and adopts an approach that is different from the frequentist approach used in the rest of this book. In the frequentist approach, if a Bayes action δ(x) is a measurable function of x, then δ(X) is a nonrandomized decision rule. It can be shown (exercise) that δ(X) defined in Definition 4.1 (if it exists R for X = x ∈ A with Θ Pθ (A)dΠ = 1) also minimizes the Bayes risk Z RT (θ)dΠ rT (Π) = Θ

over all decision rules T (randomized or nonrandomized), where RT (θ) is the risk function of T defined in (2.22). Thus, δ(X) is a Bayes rule (§2.3.2). In an estimation problem, a Bayes rule is called a Bayes estimator. Generalized Bayes risks, generalized Bayes rules (or estimators), and empirical Bayes rules (or estimators) can be defined similarly. In view of the discussion in §2.3.2, even if we do not adopt the Bayesian approach, the method described in §4.1.1 can be used as a way of generating

240

4. Estimation in Parametric Models

decision rules. In this section, we study a Bayes rule or estimator in terms of its risk (and bias and consistency for a Bayes estimator). Bayes rules are typically admissible since, if there is a rule better than a Bayes rule, then that rule has the same Bayes risk as the Bayes rule and, therefore, is itself a Bayes rule. This actually proves part (i) of the following result. The proof of the other parts of the following result is left as an exercise. Theorem 4.2. In a decision problem, let δ(X) be a Bayes rule w.r.t. a prior Π. (i) If δ(X) is a unique Bayes rule, then δ(X) is admissible. (ii) If Θ is a countable set, the Bayes risk rδ (Π) < ∞, and Π gives positive probability to each θ ∈ Θ, then δ(X) is admissible. (iii) Let ℑ be the class of decision rules having continuous risk functions. If δ(X) ∈ ℑ, rδ (Π) < ∞, and Π gives positive probability to any open subset of Θ, then δ(X) is ℑ-admissible. Generalized Bayes rules or estimators are not necessarily admissible. Many generalized Bayes rules are limits of Bayes rules (see Examples 4.3 and 4.7). Limits of Bayes rules are often admissible (Farrell, 1968a,b). The following result shows a technique of proving admissibility using limits of generalized Bayes risks. Theorem 4.3. Suppose that Θ is an open set of Rk . In a decision problem, let ℑ be the class of decision rules having continuous risk functions. A decision rule T ∈ ℑ is ℑ-admissible if there exists a sequence {Πj } of (possibly improper) priors such that (a) the generalized Bayes risks rT (Πj ) are finite for all j; (b) for any θ0 ∈ Θ and η > 0, rT (Πj ) − rj∗ (Πj ) = 0, j→∞ Πj (Oθ0 ,η ) lim

where rj∗ (Πj ) = inf T ∈ℑ rT (Πj ) and Oθ0 ,η = {θ ∈ Θ : kθ − θ0 k < η} with Πj (Oθ0 ,η ) < ∞ for all j. Proof. Suppose that T is not ℑ-admissible. Then there exists T0 ∈ ℑ such that RT0 (θ) ≤ RT (θ) for all θ and RT0 (θ0 ) < RT (θ0 ) for a θ0 ∈ Θ. From the continuity of the risk functions, we conclude that RT0 (θ) < RT (θ) − ǫ for all θ ∈ Oθ0 ,η and some constants ǫ > 0 and η > 0. Then, for any j, rT (Πj ) − rj∗ (Πj ) ≥ rT (Πj ) − rT0 (Πj ) Z [RT (θ) − RT0 (θ)]dΠj (θ) ≥ Oθ0 ,η

≥ ǫΠj (Oθ0 ,η ),

which contradicts condition (b). Hence, T is ℑ-admissible.

241

4.1. Bayes Decisions and Estimators

Example 4.6. Consider Example 4.3 and the estimation of µ under the squared error loss. From Theorem 2.1, the risk function of any decision rule is continuous in µ if the risk is finite. We now apply Theorem 4.3 to show ¯ is admissible. Let Πj = N (0, j). Since RX¯ (µ) = that the sample mean X 2 2 σ /n, rX¯ (Πj ) = σ /n for any j. Hence, condition (a) in Theorem 4.3 is satisfied. From Example 2.25, the Bayes estimator w.r.t. Πj is δj (X) = nj ¯ nj+σ2 X (see formula (2.25)). Thus, Rδj (µ) = and rj∗ (Πj )

=

Z

σ 2 nj 2 + σ 4 µ2 (nj + σ 2 )2

Rδj (µ)dΠj =

σ2 j . nj + σ 2

For any Oµ0 ,η = {µ : |µ − µ0 | < η}, µ0 − η 2ηΦ′ (ξj ) µ0 + η √ √ √ −Φ = Πj (Oµ0 ,η ) = Φ j j j √ √ for some ξj satisfying (µ0 − η)/ j ≤ ξj ≤ (µ0 + η)/ j, where Φ is the standard normal c.d.f. and Φ′ is its derivative. Since Φ′ (ξj ) → Φ′ (0) = (2π)−1/2 , √ rX¯ (Πj ) − rj∗ (Πj ) σ4 j = →0 Πj (Oµ0 ,η ) 2ηΦ′ (ξj )n(nj + σ 2 ) as j → ∞. Thus, condition (b) in Theorem 4.3 is satisfied and, hence, the ¯ is admissible. sample mean X More results in admissibility can be found in §4.2 and §4.3. The following result concerns the bias of a Bayes estimator. Proposition 4.2. Let δ(X) be a Bayes estimator of ϑ = g(θ) under the squared error loss. Then δ(X) is not unbiased unless the Bayes risk rδ (Π) = 0. Proof. Suppose that δ(X) is unbiased, i.e., E[δ(X)|θ] = g(θ). Conditioning on θ and using Proposition 1.10, we obtain that E[g(θ)δ(X)] = E{g(θ)E[δ(X)|θ]} = E[g(θ)]2 . Since δ(X) = E[g(θ)|X], conditioning on X and using Proposition 1.10, we obtain that E[g(θ)δ(X)] = E{δ(X)E[g(θ)|X]} = E[δ(X)]2 . Then rδ (Π) = E[δ(X) − g(θ)]2 = E[δ(X)]2 + E[g(θ)]2 − 2E[g(θ)δ(X)] = 0.

242

4. Estimation in Parametric Models

Since rδ (Π) = 0 occurs usually in some trivial cases, a Bayes estimator is typically not unbiased. Hence, Proposition 4.2 can be used to check whether an estimator can be a Bayes estimator w.r.t. some prior under the squared error loss. However, a generalized Bayes estimator may be unbiased; see, for instance, Examples 4.3 and 4.7. Bayes estimators are usually consistent and approximately unbiased. In a particular problem, it is usually easy to check directly whether Bayes estimators are consistent and approximately unbiased (Examples 4.7-4.9), especially when Bayes estimators have explicit forms. Bayes estimators also have some other good asymptotic properties, which are studied in §4.5.3. Let us consider some examples. Example 4.7. Let X = (X1 , ..., Xn ) and Xi ’s be i.i.d. from the exponential distribution E(0, θ) with an unknown θ > 0. Let the prior be such that θ−1 has the gamma distribution Γ(α, γ) with known α > 0 and γ > 0. Then ¯ + γ −1 )−1 ) the posterior of ω = θ−1 is the gamma distribution Γ(n + α, (nX ¯ (verify), where X is the sample mean. Consider first the estimation of θ = ω −1 . The Bayes estimator of θ under the squared error loss is ¯ + γ −1 )n+α Z ∞ ¯ + γ −1 −1 (nX nX ¯ . δ(X) = ω n+α−2 e−(nX+γ )ω dω = Γ(n + α) n+α−1 0 The bias of δ(X) is nθ + γ −1 γ −1 − (α − 1)θ −θ = =O n+α−1 n+α−1

1 . n

¯ It is also easy to see that δ(X) is consistent. The UMVUE of θ is X. 2 ¯ ¯ Since Var(X) = θ /n, rX¯ (Π) > 0 for any Π and, hence, X is not a Bayes ¯ is the generalized Bayes estimator w.r.t. the estimator. In this case, X dΠ improper prior dω = I(0,∞) (ω) and is a limit of Bayes estimators δ(X) as α → 1 and γ → ∞ (exercise). The admissibility of δ(X) is considered in Exercises 32 and 80. Consider next the estimation of e−t/θ = e−tω (see Examples 2.26 and 3.3). The Bayes estimator under the squared error loss is ¯ + γ −1 )n+α Z ∞ −1 (nX ¯ δt (X) = ω n+α−1 e−(nX+γ +t)ω dω Γ(n + α) 0 −(n+α) t = 1+ ¯ . nX + γ −1 Again, this estimator is biased and it is easy to show that δt (X) is consistent as n → ∞. In this case, the UMVUE given in Example 3.3 is neither a Bayes estimator nor a limit of δt (X).

243

4.1. Bayes Decisions and Estimators

Example 4.8. Let X = (X1 , ..., Xn ) and Xi ’s be i.i.d. from N (µ, σ 2 ) with unknown µ ∈ R and σ 2 > 0. Let the prior for ω = (2σ 2 )−1 be the gamma distribution Γ(α, γ) with known α and γ and let the prior for µ be N (µ0 , σ02 /ω) (conditional on ω). Then the posterior p.d.f. of (µ, ω) is proportional to i o n h 2 ¯ − µ)2 + (µ−µ20 ) ω , ω (n+1)/2+α−1 exp − γ −1 + Y + n(X 2σ 0

Pn

¯ 2 and X ¯ is the sample mean. Note that where Y = i=1 (Xi − X) 2 ¯2 + ¯ + µ02 µ + nX ¯ − µ)2 + (µ−µ20 ) = n + 1 2 µ2 − 2 nX n(X 2σ 2σ 2σ 0

0

0

µ20 . 2σ02

Hence, the posterior p.d.f. of (µ, ω) is proportional to i o n h ω (n+1)/2+α−1 exp − γ −1 + W + n + 2σ1 2 (µ − ζ(X))2 ω , 0

where

ζ(X) =

¯+ nX n+

µ0 2σ02 1 2σ02

µ20 1 2 ¯ and W = Y + nX + 2 − n + 2 [ζ(X)]2 . 2σ0 2σ0

Thus, the posterior of ω is the gamma distribution Γ(n/2+α, (γ −1 +W )−1 ) and the posterior of µ (given ω and X) is N ζ(X), [(2n+σ0−2 )ω]−1 . Under the squared error loss, the Bayes estimator of µ is ζ(X) and the Bayes estimator of σ 2 = (2ω)−1 is (γ −1 +W )/(n+2α−2), provided that n+2α > 2. Apparently, these Bayes estimators are biased but the biases are of the order n−1 ; and they are consistent as n → ∞. To consider the last example, we need the following useful lemma whose proof is similar to the proof of result (4.8). Lemma 4.1. Suppose that X has a p.d.f. fθ (x) w.r.t. a σ-finite measure ν. Suppose that θ = (θ1 , θ2 ), θj ∈ Θj , and that the prior has a p.d.f. π(θ) = πθ1 |θ2 (θ1 )πθ2 (θ2 ), where πθ2 (θ2 ) is a p.d.f. w.r.t. a σ-finite measure ν2 on Θ2 and for any given θ2 , πθ1 |θ2 (θ1 ) is a p.d.f. w.r.t. a σ-finite measure ν1 on Θ1 . Suppose further that if θ2 is given, the Bayes estimator of h(θ1 ) = g(θ1 , θ2 ) under the squared error loss is δ(X, θ2 ). Then the Bayes estimator of g(θ1 , θ2 ) under the squared error loss is δ(X) with Z δ(x) = δ(x, θ2 )pθ2 |x (θ2 )dν2 , Θ2

where pθ2 |x (θ2 ) is the posterior p.d.f. of θ 2 given X = x.

244

4. Estimation in Parametric Models

Example 4.9. Consider a linear model Xij = β τ Zi + εij ,

j = 1, ..., ni , i = 1, ..., k,

where β ∈ Rp is unknown, Zi ’s are known vectors, εij ’s are independent, and εij is N (0, σi2 ), j = 1, ..., ni , i = 1, ..., k. Let X be the sample vector containing all Xij ’s. The parameter vector is then θ = (β, ω), where ω = (ω1 , ..., ωk ) and ωi = (2σi2 )−1 . Assume that the prior for θ has the Lebesgue p.d.f. k Y c π(β) ωiα e−ωi /γ , (4.9) i=1

where α > 0, γ > 0, and c > 0 are known constants and π(β) is a known Lebesgue p.d.f. on Rp . The posterior p.d.f. of θ is then proportional to h(X, θ) = π(β)

k Y

n /2+α −[γ −1 +vi (β)]ωi

ωi i

e

,

i=1

Pni (Xij − β τ Zi )2 . If β is known, the Bayes estimator of where vi (β) = j=1 σi2 under the squared error loss is Z 1 h(X, θ) γ −1 + vi (β) R dω = . 2ωi h(X, θ)dω 2α + ni By Lemma 4.1, the Bayes estimator of σi2 is Z −1 γ + vi (β) fβ|X (β)dβ, σ ˆi2 = 2α + ni

(4.10)

where fβ|X (β) ∝

Z

h(X, θ)dω

∝ π(β) ∝ π(β)

k Z Y

α+ni /2 −[γ −1 +vi (β)]ωi

ωi

e

dωi

i=1

k Y −(α+1+ni /2) −1 γ + vi (β)

(4.11)

i=1

is the posterior p.d.f. of β. The Bayes estimator of lτ β for any l ∈ Rp is then the posterior mean of lτ β w.r.t. the p.d.f. fβ|X (β). In this problem, Bayes estimators do not have explicit forms. A numerical method (such as one of those in §4.1.4) has to be used to evaluate Bayes estimators (see Example 4.10).

245

4.1. Bayes Decisions and Estimators

¯ i· and S 2 be the sample mean and variance of Xij , j = 1, ..., ni Let X i 2 (Si is defined to be 0 if ni = 1), and let σ02 = (2αγ)−1 (the prior mean of σi2 ). Then the Bayes estimator σ ˆi2 in (4.10) can be written as ni − 1 2 ni 2α σ2 + S + 2α + ni 0 2α + ni i 2α + ni

Z

¯ i· − β τ Zi )2 fβ|X (β)dβ. (X

(4.12)

The Bayes estimator in (4.12) is a weighted average of prior information, “within group” variation, and averaged squared “residuals”. If ni → ∞, then the first term in (4.12) converges to 0 and the second term in (4.12) is consistent and approximately unbiased for σi2 . Hence, the Bayes estimator σ ˆi2 is consistent and approximately unbiased for σi2 if the mean of the last term in (4.12) tends to 0, which is true under some conditions (see, e.g., Exercise 36). It is easy to see that σ ˆi2 is consistent and 2 approximately unbiased for σi w.r.t. the joint distribution of (X, θ), since the mean of the last term in (4.12) w.r.t. the joint distribution of (X, θ) is bounded by σ02 /ni .

4.1.4 Markov chain Monte Carlo As we discussed previously, Bayes actions or estimators have to be computed numerically in many applications. Typically we need to compute an integral of the form Z Ep (g) =

g(θ)p(θ)dν

Θ

with some function g, where p(θ) is a p.d.f. w.r.t. a σ-finite measure ν on (Θ, BΘ ) and Θ ⊂ Rk . For example, if g is an indicator function of A ∈ BΘ and p(θ) is the posterior p.d.f. of θ given X = x, then Ep (g) is the posterior probability of A; under the squared error loss, Ep (g) is the Bayes action (4.4) if p(θ) is the posterior p.d.f. There are many numerical methods for computing integrals Ep (g); see, for example, §4.5.3 and Berger (1985, §4.9). In this section, we discuss the Markov chain Monte Carlo (MCMC) methods, which are powerful numerical methods not only for Bayesian computations, but also for general statistical computing (see, e.g., §4.4.1). We start with the simple Monte Carlo method, which can be viewed as a special case of the MCMC. Suppose that we can generate i.i.d. θ(1) , ..., θ(m) from a p.d.f. h(θ) > 0 w.r.t. ν. By the SLLN (Theorem 1.13(ii)), as m → ∞, Z m X g(θ(j) )p(θ(j) ) g(θ)p(θ) ˆp (g) = 1 h(θ)dν = Ep (g). → E a.s. (j) m j=1 h(θ) h(θ ) Θ ˆp (g) can be used as a numerical approximation to Ep (g). The Hence E process of generating θ(j) according to h is called importance sampling and

246

4. Estimation in Parametric Models

h(θ) is called the importance function. More discussions on importance sampling can be found, for example, in Berger (1985), Geweke (1989), Shao (1989), and Tanner (1996). When p(θ) is intractable or complex, it is often difficult to choose a function h that is simple enough for importance ˆp (g) as well. sampling and results in a fast convergence of E The simple Monte Carlo method, however, may not work well when k, the dimension of Θ, is large. This is because, when k is large, the converˆp (g) requires a very large m; generating a random vector from gence of E a k-dimensional distribution is usually expensive, if not impossible. More sophisticated MCMC methods are different from the simple Monte Carlo in two aspects: generating random vectors can be done using distributions whose dimensions are much lower than k; and θ(1) , ..., θ(m) are not independent, but form a Markov chain. Let {Y (t) : t = 0, 1, ...} be a Markov chain (§1.4.4) taking values in Y ⊂ Rk . {Y (t) } is homogeneous if and only if P (Y (t+1) ∈ A|Y (t) ) = P (Y (1) ∈ A|Y (0) ) for any t. For a homogeneous Markov chain {Y (t) }, define P (y, A) = P (Y (1) ∈ A|Y (0) = y),

y ∈ Y, A ∈ BY ,

which is called the transition kernel of the Markov chain. Note that P (y, ·) is a probability measure for every y ∈ Y; P (·, A) is a Borel function for every A ∈ BY ; and the distribution of a homogeneous Markov chain is determined by P (y, A) and the distribution ofR Y (0) (initial distribution). Pm MCMC approximates an integral of the form Y g(y)p(y)dν by m−1 t=1 g(Y (t) ) with a Markov chain {Y (t) : t = 0, 1, ...}. The basic justification of the MCMC approximation is given in the following result. Theorem 4.4. Let p(y) be Ra p.d.f. on Y w.r.t. a σ-finite measure ν and g be a Borel function on Y with Y |g(y)|p(y)dν < ∞. Let {Y (t) : t = 0, 1, ...} be a homogeneous Markov chain taking values on Y ⊂ Rk with the transition kernel P (y, A). Then m

1 X g(Y (t) ) →a.s. m t=1

Z

g(y)p(y)dν

(4.13)

Y

and, as t → ∞, P t (y, A) = P (Y (t) ∈ A|Y (0) = y) →a.s.

Z

p(y)dν,

(4.14)

A

provided that (a) the Markov chain is aperiodic in the sense that there does not exist d ≥ 2

247

4.1. Bayes Decisions and Estimators

nonempty disjoint events A0 , ..., Ad−1 in BY such that for all i = 0, ..., d − 1 and all y ∈ Ai , P (y, Aj ) = 1 for j = i + 1 (mod d); R R(b) the Markov chain is p-invariant in the sense that P (y, A)p(y)dν = p(y)dν for all A ∈ BY ; A (c) the Markov chain is p-irreducible in the sense that for any y ∈ Y and any R A with A p(y)dν > 0, there exists a positive integer t such that P t (y, A) in (4.14) is positive; and recurrent in the sense that for any A with R(d) the Markov chain P∞is Harris (t) p(y)dν > 0, P ) = ∞|Y (0) = y = 1 for all y. t=1 IA (Y A

The proof of these results is beyond the scope of this book and, hence, is omitted. It can be found, for example, in Nummelin (1984), Chan (1993), and Tierney (1994). A homogeneous Markov chain satisfying conditions (a)-(d) in Theorem 4.4 is called ergodic with equilibrium distribution p. Result (4.13) means that the MCMC approximation is consistent and result (4.14) indicates that p is the limiting p.d.f. of the Markov chain. One of the key issues in MCMC is the choice of the kernel P (y, A). The first requirement on P (y, A) is that conditions (a)-(d) in Theorem 4.4 be satisfied. Condition (a) is usually easy to check for any given P (y, A). In the following, we consider two popular MCMC methods satisfying conditions (a)-(d). Gibbs sampler One way to construct a p-invariant homogeneous Markov chain is to use conditioning. Suppose that Y has the p.d.f. p(y). Let Yi (or yi ) be the ith component of Y (or y) and let Y−i (or y−i ) be the (k − 1)-vector containing all components of Y (or y) except Yi (or yi ). Then Pi (y−i , A) = P (Y ∈ A|Y−i = y−i ) is a transition kernel for any i. The MCMC method using this kernel is called the single-site Gibbs sampler. Note that Z Z Pi (y−i , A)p(y)dν = E[P (Y ∈ A|Y−i )] = P (Y ∈ A) = p(y)dν A

and, therefore, the chain with kernel Pi (y−i , A) is p-invariant. However, this chain is not p-irreducible since Pi (y−i , ·) puts all its mass on the set ψi−1 (y−i ), where ψi (y) = y−i . Gelfand and Smith (1990) considered a systematic scan Gibbs sampler whose kernel P (y, A) is a composite of k kernels Pi (y−i , A), i = 1, ..., k. More precisely, the chain is defined as follows. Given (t) (t−1) (t−1) (t) Y (t−1) = y (t−1) , we generate y1 from P1 (y2 , ..., yk , · ),..., yj from (t)

(t)

(t−1)

(t−1)

Pj (y1 , ..., yj−1 , yj+1 , ..., yk

(t)

(t)

(t)

, · ),..., yk from Pk (y1 , ..., yk−1 , · ). It can

248

4. Estimation in Parametric Models

be shown that this Markov chain is still p-invariant. We illustrate this with (1) (0) the case of k = 2. Note that Y1 is generated from P2 (y2 , ·), the con(0) (1) (0) ditional distribution of Y given Y2 = y2 . Hence (Y1 , Y2 ) has p.d.f. p. (1) (1) Similarly, we can show that Y (1) = (Y1 , Y2 ) has p.d.f. p. Thus, Z Z P (y, A)p(y)dν = P (Y (1) ∈ A|Y (0) = y)p(y)dν = E[P (Y (1) ∈ A|Y (0) )] = P (Y (1) ∈ A) Z = p(y)dν. A

This Markov chain is also p-irreducible and aperiodic if p(y) > 0 for all y ∈ Y; see, for example, Chan (1993). Finally, if p(y) > 0 for all y ∈ Y, then P (y, A) ≪ the distribution with p.d.f. p for all y and, by Corollary 1 of Tierney (1994), the Markov chain is Harris recurrent. Thus, Theorem 4.4 applies and (4.13) and (4.14) hold. The previous Gibbs sampler can obviously be extended to the case where yi ’s are subvectors (of possibly different dimensions) of y. Let us now return to Bayesian computation and consider the following example. Example 4.10. Consider Example 4.9. Under the given prior for θ = (β, ω), it is difficult to generate random vectors directly from the posterior p.d.f., given X = x (which does not have a familiar form). To apply a Gibbs sampler with y = θ, y1 = β, and y2 = ω, we need to generate random vectors from the posterior of β, given x and ω, and the posterior of ω, given x and β. From (4.9) and (4.11), the posterior of ω = (ω1 , ..., ωk ), given x and β, is a product of marginals of ωi ’s that are the gamma distributions Γ(α + 1 + ni /2, [γ −1 + vi (β)]−1 ), i = 1, ..., k. Assume now that π(β) ≡ 1 (noninformative prior for β). It follows from (4.9) that the posterior p.d.f. of β, given x and ω, is proportional to k Y

i=1

e−ωi vi (β) ∝ e−kW

1/2

Zβ−W 1/2 Xk2

,

where W is the diagonal on the diagonal Pk block matrix whose ith block 1/2 is ωi Ini . Let n = Zβ, given X i=1 ni . Then, the posterior of W and ω, is Nn (W 1/2 X, 2−1 In ) and the posterior of β, given X and ω, is Np ((Z τ W Z)−1 Z τ W X, 2−1 (Z τ W Z)−1 ) (Z τ W Z is assumed of full rank for simplicity), since β = [(Z τ W Z)−1 Z τ W 1/2 ]W 1/2 Zβ. Note that random generation using these two posterior distributions is fairly easy.

249

4.1. Bayes Decisions and Estimators

The Metropolis algorithm A large class of MCMC methods are obtained using the Metropolis algorithm (Metropolis et al., 1953). We introduce Hastings’ version of the algorithm. Let Q(y, A) be a transition kernel of a homogeneous Markov chain satisfying Z Q(y, A) =

q(y, z)dν(z)

A

for a measurable function q(y, z) ≥ 0 on Y × Y and R a σ-finite measure ν on (Y, BY ). Without loss of generality, assume that Y p(y)dν = 1 and that p is not concentrated on a single point. Define o n ( p(z)q(z,y) ,1 p(y)q(y, z) > 0 min p(y)q(y,z) α(y, z) = 1 p(y)q(y, z) = 0

and p(y, z) =

q(y, z)α(y, z) 0

y= 6 z y = z.

The Metropolis kernel P (y, A) is defined by Z p(y, z)dν(z) + r(y)διy (A), P (y, A) =

(4.15)

A

R where r(y) = 1 − p(y, z)dν(z) and διy is the point mass at y defined in (1.22). The corresponding Markov chain can be described as follows. If the chain is currently at a point Y (t) = y, then it generates a candidate value z for the next location Y (t+1) from Q(y, ·). With probability α(y, z), the chain moves to Y (t+1) = z. Otherwise, the chain remains at Y (t+1) = y. Note that this algorithm only depends on p(y) through p(y)/p(z). Thus, it can be used when p(y) is known up to a normalizing constant, which often occurs in Bayesian analysis. We now show that a Markov chain with a Metropolis kernel P (y, A) is p-invariant. First, by the definition of p(y, z) and α(y, z), p(y)p(y, z) = p(z)p(z, y)

for any y and z. Then, for any A ∈ BY , Z Z Z Z p(y, z)dν(z) p(y)dν(y) + r(y)διy (A)p(y)dν(y) P (y, A)p(y)dν = A Z Z Z = p(y, z)p(y)dν(y) dν(z) + r(y)p(y)dν(y) A A Z Z Z = p(z, y)p(z)dν(y) dν(z) + r(y)p(y)dν(y) A

A

250

4. Estimation in Parametric Models

=

Z

A

=

Z

[1 − r(z)]p(z)dν(z) +

Z

r(z)p(z)dν(z)

A

p(z)dν(z).

A

If a MarkovR chain with a Metropolis kernel defined by (4.15) is pirreducible and r(y)>0 p(y)dν > 0, then, by the results of Nummelin (1984, §2.4), the chain is aperiodic; by Corollary 2 of Tierney (1994), the chain is Harris recurrent. Hence, to apply Theorem 4.4 to a Markov chain with a Metropolis kernel, it suffices to show that the chain is p-irreducible. Lemma 4.2. Suppose that Q(y, A) is the transition kernel of a p-irreducible Markov chain and that either q(y, z) > 0 for all y and z or q(y, z) = q(z, y) for all y and z. Then the chain with the Metropolis kernel p(y, A) in (4.15) is p-irreducible. Proof. It can be shown (exercise) that if Q is any transition kernel of a homogeneous Markov chain, then Z Y Z Z t Qt (y, A) = ··· q(zn−j+1 , zn−j )dν(zn−j ), (4.16) A

j=1

R where zn = y, y ∈ Y, and A ∈ RBY . Let y ∈ Y, A ∈ BY with A p(z)dν > 0, and By = {z : α(y, z) = 1}. If A∩B c p(z)dν > 0, then y Z Z q(z, y)p(z) q(y, z)α(y, z)dν(z) = P (y, A) ≥ dν(z) > 0, p(y) A∩Byc A∩Byc

which follows from either Rq(z, y) > 0 or q(z, y) = q(y, z) > 0 on Byc . If R A∩B c p(z)dν = 0, then A∩By p(z)dν > 0. From the irreducibility of y

Q(y, A), there exists a t ≥ 1 such that Qt (y, A ∩ By ) > 0. Then, by (4.15) and (4.16), P t (y, A) ≥ P t (y, A ∩ By ) ≥ Qt (y, A ∩ By ) > 0.

Two examples of q(y, z) given by Tierney (1994) are q(y, z) = f (z − y) with a Lebesgue p.d.f. f on Rk , which corresponds to a random walk chain, and q(y, z) = f (z) with a p.d.f. f , which corresponds to an independence chain and is closely related to the importance sampling discussed earlier. Although the MCMC methods have been used over the last 50 years, the research on the theory of MCMC is still very active. Important topics include the choice of the transition kernel for MCMC; the rate of the convergence in (4.13); the choice of the Monte Carlo size m; and the estimation of the errors due to Monte Carlo. See more results and discussions in Tierney (1994), Basag et al. (1995), Tanner (1996), and the references therein.

251

4.2. Invariance

4.2 Invariance The concept of invariance is introduced in §2.3.2 (Definition 2.9). In this section, we study the best invariant estimators and their properties in one-parameter location families (§4.2.1), in one-parameter scale families (§4.2.2), and in general location-scale families (§4.2.3). Note that invariant estimators are also called equivariant estimators.

4.2.1 One-parameter location families Assume that the sample X = (X1 , ..., Xn ) has a joint distribution Pµ with a Lebesgue p.d.f. f (x1 − µ, ..., xn − µ), (4.17) where f is known and µ ∈ R is an unknown location parameter. The family P = {Pµ : µ ∈ R} is called a one-parameter location family, a special case of the general location-scale family described in Definition 2.3. It is invariant under the location transformations gc (X) = (X1 + c,..., Xn + c), c ∈ R. We consider the estimation of µ as a statistical decision problem with action space A = R and loss function L(µ, a). It is natural to consider the same transformation in the action space, i.e., if Xi is transformed to Xi + c, then our action a is transformed to a+ c. Consequently, the decision problem is invariant under location transformation if and only if L(µ, a) = L(µ + c, a + c)

for all c ∈ R,

which is equivalent to L(µ, a) = L(a − µ)

(4.18)

for a Borel function L(·) on R. According to Definition 2.9 (see also Example 2.24), an estimator T (decision rule) of µ is location invariant if and only if T (X1 + c, ..., Xn + c) = T (X1 , ..., Xn ) + c.

(4.19)

Many estimators of µ, such as the sample mean and weighted average of the order statistics, are location invariant. The following result provides a characterization of location invariant estimators. Proposition 4.3. Let T0 be a location invariant estimator of µ. Let di = xi − xn , i = 1, ..., n − 1, and d = (d1 , ..., dn−1 ). A necessary and sufficient condition for an estimator T to be location invariant is that there exists a Borel function u on Rn−1 (u ≡ a constant if n = 1) such that T (x) = T0 (x) − u(d)

for all x ∈ Rn .

(4.20)

252

4. Estimation in Parametric Models

Proof. It is easy to see that T given by (4.20) satisfies (4.19) and, therefore, is location invariant. Suppose that T is location invariant. Let u ˜(x) = T (x) − T0 (x) for any x ∈ Rn . Then u ˜(x1 + c, ..., xn + c) = T (x1 + c, ..., xn + c) − T0 (x1 + c, ..., xn + c) = T (x1 , ..., xn ) − T0 (x1 , ..., xn ) for all c ∈ R and xi ∈ R. Putting c = −xn leads to u ˜(x1 − xn , ..., xn−1 − xn , 0) = T (x) − T0 (x),

x ∈ Rn .

The result follows with u(d1 , ..., dn−1 ) = u˜(x1 − xn , ..., xn−1 − xn , 0). Therefore, once we have a location invariant estimator T0 of µ, any other location invariant estimator of µ can be constructed by taking the difference between T0 and a Borel function of the ancillary statistic D = (X1 − Xn , ..., Xn−1 − Xn ). The next result states an important property of location invariant estimators. Proposition 4.4. Let X be distributed with the p.d.f. given by (4.17) and let T be a location invariant estimator of µ under the loss function given by (4.18). If the bias, variance, and risk of T are well defined, then they are all constant (do not depend on µ). Proof. The result for the bias follows from Z bT (µ) = T (x)f (x1 − µ, ..., xn − µ)dx − µ Z = T (x1 + µ, ..., xn + µ)f (x)dx − µ Z = [T (x) + µ]f (x)dx − µ Z = T (x)f (x)dx. The proof of the result for variance or risk is left as an exercise. An important consequence of this result is that the problem of finding the best location invariant estimator reduces to comparing constants instead of risk functions. The following definition can be used not only for location invariant estimators, but also for general invariant estimators. Definition 4.2. Consider an invariant estimation problem in which all invariant estimators have constant risks. An invariant estimator T is called the minimum risk invariant estimator (MRIE) if and only if T has the smallest risk among all invariant estimators.

253

4.2. Invariance

Theorem 4.5. Let X be distributed with the p.d.f. given by (4.17) and consider the estimation of µ under the loss function given by (4.18). Suppose that there is a location invariant estimator T0 of µ with finite risk. Let D = (X1 − Xn , ..., Xn−1 − Xn ). (i) Assume that for each d there exists a u∗ (d) that minimizes h(d) = E0 [L(T0 (X) − u(d))|D = d] over all functions u, where the expectation E0 is calculated under the assumption that X has p.d.f. f (x1 , ..., xn ). Then an MRIE exists and is given by T∗ (X) = T0 (X) − u∗ (D). (ii) The function u∗ in (i) exists if L(t) is convex and not monotone; it is unique if L is strictly convex. (iii) If T0 and D are independent, then u∗ is a constant that minimizes E0 [L(T0 (X) − u)]. If, in addition, the distribution of T0 is symmetric about µ and L is convex and even, then u∗ = 0. Proof. By Theorem 1.7 and Propositions 4.3 and 4.4, RT (µ) = E0 [h(D)], where T (X) = T0 (X) − u(D). This proves part (i). If L is (strictly) convex and not monotone, then E0 [L(T0 (x)−a)|D = d] is (strictly) convex and not monotone in a (exercise). Hence lim|a|→∞ E0 [L(T0 (x) − a)|D = d] = ∞. This proves part (ii). The proof of part (iii) is left as an exercise. Theorem 4.6. Assume the conditions of Theorem 4.5 and that the loss is the squared error loss. (i) The unique MRIE of µ is R∞ tf (X1 − t, ..., Xn − t)dt T∗ (X) = R−∞ , ∞ f (X1 − t, ..., Xn − t)dt −∞

which is known as the Pitman estimator of µ. (ii) The MRIE of µ is unbiased. Proof. (i) Under the squared error loss,

u∗ (d) = E0 [T0 (X)|D = d]

(4.21)

(exercise). Let T0 (X) = Xn (the nth observation). Then Xn is location invariant. If there exists a location invariant estimator of µ with finite risk, then E0 (Xn |D = d) is finite a.s. P (exercise). By Proposition 1.8, when µ = 0, the joint Lebesgue p.d.f. of (D, Xn ) is f (d1 + xn , ..., dn−1 + xn , xn ), d = (d1 , ..., dn−1 ). The conditional p.d.f. of Xn given D = d is then f (d + xn , ..., dn−1 + xn , xn ) R∞ 1 f (d1 + t, ..., dn−1 + t, t)dt −∞

254

4. Estimation in Parametric Models

(see (1.61)). By Proposition 1.9, R∞

tf (d1 + t, ..., dn−1 + t, t)dt

E0 (Xn |D = d) = R−∞ ∞

−∞

R∞

= R−∞ ∞

f (d1 + t, ..., dn−1 + t, t)dt tf (x1 − xn + t, ..., xn−1 − xn + t, t)dt

f (x1 − xn + t, ..., xn−1 − xn + t, t)dt R∞ uf (x1 − u, ..., xn − u)du = xn − R−∞ ∞ −∞ f (x1 − u, ..., xn − u)du −∞

by letting u = xn −t. The result in (i) follows from T∗ (X) = Xn −E(Xn |D) (Theorem 4.5). (ii) Let b be the constant bias of T∗ (Proposition 4.4). Then T1 (X) = T∗ (X) − b is a location invariant estimator of µ and RT1 = E[T∗ (X) − b − µ]2 = Var(T∗ ) ≤ Var(T∗ ) + b2 = RT∗ . Since T∗ is the MRIE, b = 0, i.e., T∗ is unbiased. Theorem 4.6(ii) indicates that we only need to consider unbiased location invariant estimators in order to find the MRIE, if the loss is the squared error loss. In particular, a location invariant UMVUE is an MRIE. Example 4.11. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with an unknown ¯ is location invariant. Since X ¯ is the µ ∈ R and a known σ 2 . Note that X UMVUE of µ (§2.1), it is the MRIE under the squared error loss. Since the ¯ is symmetric about µ and X ¯ is independent of D (Basu’s distribution of X ¯ is an MRIE if L is convex theorem), it follows from Theorem 4.5(iii) that X and even. Example 4.12. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(µ, θ), where θ is known and µ ∈ R is unknown. Since X(1) − θ/n is location invariant and is the UMVUE of µ, it is the MRIE under the squared error loss. Note that X(1) is independent of D (Basu’s theorem). By Theorem 4.5(iii), an MRIE is of the form X(1) − u∗ with a constant u∗ . For the absolute error loss, X(1) − θ log 2/n is an MRIE (exercise). Example 4.13. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (µ − 12 , µ + 12 ) with an unknown µ ∈ R. Consider the squared error loss. Note that f (x1 − µ, ..., xn − µ) =

1 0

µ − 12 ≤ x(1) ≤ x(n) ≤ µ + otherwise.

1 2

255

4.2. Invariance

By Theorem 4.6(i), the MRIE of µ is T∗ (X) =

Z

X(1) + 12

tdt X(n) − 12

Z

X(1) + 12

dt =

X(n) − 12

X(1) + X(n) . 2

We end this section with a brief discussion of the admissibility of MRIE’s in a one-parameter location problem. Under the squared error loss, the MRIE (Pitman’s estimator) is admissible if there exists a location invariant estimator T0 with E|T0 (X)|3 < ∞ (Stein, 1959). Under a general loss function, an MRIE is admissible when it is a unique MRIE (under some other minor conditions). See Farrell (1964), Brown (1966), and Brown and Fox (1974) for further discussions.

4.2.2 One-parameter scale families Assume that the sample X = (X1 , ..., Xn ) has a joint distribution Pσ with a Lebesgue p.d.f. x1 xn 1 (4.22) σn f σ , ..., σ ,

where f is known and σ > 0 is an unknown scale parameter. The family P = {Pσ : σ > 0} is called a one-parameter scale family and is a special case of the general location-scale family in Definition 2.3. This family is invariant under the scale transformations gr (X) = rX, r > 0. We consider the estimation of σ h with A = [0, ∞), where h is a nonzero constant. The transformation gr induces the transformation gr (σ h ) = rh σ h . Hence, a loss function L is scale invariant if and only if L(rσ, rh a) = L(σ, a)

for all r > 0,

which is equivalent to L(σ, a) = L

a σh

(4.23)

for a Borel function L(·) on [0, ∞). An example of a loss function satisfying (4.23) is p a |a − σ h |p L(σ, a) = h − 1 = , (4.24) σ σ ph

where p ≥ 1 is a constant. However, the squared error loss does not satisfy (4.23). An estimator T of σ h is scale invariant if and only if T (rX1 , ..., rXn ) = rh T (X1 , ..., Xn ).

2 Examples of scale invariant estimators are √ the sample variance S (for h = 2), the sample standard deviation S = S 2 (for h = 1), the sample range

256

4. Estimation in Parametric Models

P ¯ X(n) − X(1) (for h = 1), and the sample mean deviation n−1 ni=1 |Xi − X| (for h = 1). The following result is an analogue of Proposition 4.3. Its proof is left as an exercise. Proposition 4.5. Let T0 be a scale invariant estimator of σ h . A necessary and sufficient condition for an estimator T to be scale invariant is that there exists a positive Borel function u on Rn such that T (x) = T0 (x)/u(z)

for all x ∈ Rn ,

where z = (z1 , ..., zn ), zi = xi /xn , i = 1, ..., n − 1, and zn = xn /|xn |. The next result is similar to Proposition 4.4. It applies to any invariant problem defined in Definition 2.9. We use the notation in Definition 2.9. Theorem 4.7. Let P be a family invariant under G (a group of transformations). Suppose that the loss function is invariant and T is an invariant decision rule. Then the risk function of T is a constant. The proof is left as an exercise. Note that a special case of Theorem 4.7 is that any scale invariant estimator of σ h has a constant risk and, therefore, an MRIE (Definition 4.2) of σ h usually exists. However, Proposition 4.4 is not a special case of Theorem 4.7, since the bias of a scale invariant estimator may not be a constant in general. For example, the bias of the sample standard deviation is a function of σ. The next result and its proof are analogues of those of Theorem 4.5. Theorem 4.8. Let X be distributed with the p.d.f. given by (4.22) and consider the estimation of σ h under the loss function given by (4.23). Suppose that there is a scale invariant estimator T0 of σ h with finite risk. Let Z = (Z1 , ..., Zn ) with Zi = Xi /Xn , i = 1, ..., n − 1, and Zn = Xn /|Xn |. (i) Assume that for each z there exists a u∗ (z) that minimizes E1 [L(T0 (X)/u(z))|Z = z] over all positive Borel functions u, where the conditional expectation E1 is calculated under the assumption that X has p.d.f. f (x1 , ..., xn ). Then, an MRIE exists and is given by T∗ (X) = T0 (X)/u∗ (Z). (ii) The function u∗ in (i) exists if γ(t) = L(et ) is convex and not monotone; it is unique if γ(t) is strictly convex.

257

4.2. Invariance

The loss function given by (4.24) satisfies the condition in Theorem 4.8(ii). A loss function corresponding to the squared error loss in this problem is the loss function (4.24) with p = 2. We have the following result similar to Theorem 4.6 (its proof is left as an exercise). Corollary 4.1. Under the conditions of Theorem 4.8 and the loss function (4.24) with p = 2, the unique MRIE of σ h is R ∞ n+h−1 t f (tX1 , ..., tXn )dt T0 (X)E1 [T0 (X)|Z] 0 = R∞ , T∗ (X) = 2 n+2h−1 E1 {[T0 (X)] |Z} t f (tX1 , ..., tXn )dt 0

which is known as the Pitman estimator of σ h .

Example 4.14. Let X1 , ...,P Xn be i.i.d. from N (0, σ 2 ) and consider the esn 2 timation of σ . Then T0 = i=1 Xi2 is scale invariant. By Basu’s theorem, T0 is independent of Z. Hence u∗ in Theorem 4.8 is a constant minimizing E1 [L(T0 /u)] over u > 0. When the loss is given by (4.24) with p = 2, by Corollary 4.1, the MRIE (Pitman’s estimator) is n

T∗ (X) =

T0 (X)E1 [T0 (X)] 1 X 2 = X , 2 E1 [T0 (X)] n + 2 i=1 i

since T0 has the chi-square distribution χ2n when σ = 1. Note that the UMVUE of σ 2 is T0 /n, which is different from the MRIE. Example 4.15. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (0, σ) and consider the estimation of σ. By Basu’s theorem, the scale invariant estimator X(n) is independent of Z. Hence u∗ in Theorem 4.8 is a constant minimizing E1 [L(X(n) /u)] over u > 0. When the loss is given by (4.24) with p = 2, by Corollary 4.1, the MRIE (Pitman’s estimator) is T∗ (X) =

X(n) E1 X(n) (n + 2)X(n) . = 2 E1 X(n) n+1

4.2.3 General location-scale families Assume that X = (X1 , ..., Xn ) has a joint distribution Pθ with a Lebesgue p.d.f. x1 −µ xn −µ 1 , (4.25) σn f σ , ..., σ

where f is known, θ = (µ, σ) ∈ Θ, and Θ = R × (0, ∞). The family P = {Pθ : θ ∈ Θ} is a location-scale family defined by Definition 2.3 and is invariant under the location-scale transformations of the form gc,r (X) = (rX1 + c, ..., rXn + c), c ∈ R, r > 0, which induce similar transformations on Θ: gc,r (θ) = (rµ + c, rσ), c ∈ R, r > 0.

258

4. Estimation in Parametric Models

Consider the estimation of σ h with a fixed h 6= 0 under the loss function (4.23), which is invariant under the location-scale transformations gc,r . An estimator T of σ h is location-scale invariant if and only if T (rX1 + c, ..., rXn + c) = rh T (X1 , ..., Xn ).

(4.26)

By Theorem 4.7, any location-scale invariant T has a constant risk. Letting r = 1 in (4.26), we obtain that T (X1 + c, ..., Xn + c) = T (X1 , ..., Xn ) for all c ∈ R. Therefore, T is a function of D = (D1 , ..., Dn−1 ), Di = Xi − Xn , i = 1, ..., n − 1. From (4.25), the joint Lebesgue p.d.f. of D is R ∞ d1 dn−1 1 f + t, ..., + t, t dt, (4.27) n−1 σ σ σ −∞

which is of the form (4.22) with n replaced by n−1 and xi ’s replaced by di ’s. It follows from Theorem 4.8 that if T0 (D) is any finite risk scale invariant estimator of σ h based on D, then an MRIE of σ h is T∗ (D) = T0 (D)/u∗ (W ),

(4.28)

where W = (W1 , ..., Wn−1 ), Wi = Di /Dn−1 , i = 1, ..., n − 2, Wn−1 = ˜1 [L(T0 (D)/u(w))|W = w] Dn−1 /|Dn−1 |, u∗ (w) is any number minimizing E ˜ over all positive Borel functions u, and E1 is the conditional expectation calculated under the assumption that D has p.d.f. (4.27) with σ = 1. Consider next the estimation of µ. Under the location-scale transformation gc,r , it can be shown (exercise) that a loss function is invariant if and only if it is of the form L a−µ . (4.29) σ An estimator T of µ is location-scale invariant if and only if T (rX1 + c, ..., rXn + c) = rT (X1 , ..., Xn ) + c. Again, by Theorem 4.7, the risk of an invariant T is a constant. The following result is an analogue of Proposition 4.3 or 4.5. Proposition 4.6. Let T0 be any estimator of µ invariant under locationscale transformation and let T1 be any estimator of σ satisfying (4.26) with h = 1 and T1 > 0. Then an estimator T of µ is location-scale invariant if and only if there is a Borel function u on Rn−1 such that T (X) = T0 (X) − u(W )T1 (X), where W is given in (4.28).

259

4.2. Invariance

The proofs of Proposition 4.6 and the next result, an analogue of Theorem 4.5 or 4.8, are left as exercises. Theorem 4.9. Let X be distributed with p.d.f. given by (4.25) and consider the estimation of µ under the loss function given by (4.29). Suppose that there is a location-scale invariant estimator T0 of µ with finite risk. Let T1 be given in Proposition 4.6. Then an MRIE of µ is T∗ (X) = T0 (X) − u∗ (W )T1 (X), where W is given in (4.28), u∗ (w) is any number minimizing E0,1 [L(T0 (X) − u(w)T1 (X))|W = w] over all Borel functions u, and E0,1 is computed under the assumption that X has the p.d.f. (4.25) with µ = 0 and σ = 1. Corollary 4.2. Under the conditions of Theorem 4.9 and the loss function (a − µ)2 /σ 2 , u∗ (w) in Theorem 4.9 is equal to u∗ (w) =

E0,1 [T0 (X)T1 (X)|W = w] . E0,1 {[T1 (X)]2 |W = w}

Example 4.16. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ), where µ ∈ R and σ 2 > 0 are unknown. Consider first the estimation of σ 2 under loss function (4.23). The sample variance S 2 is location-scale invariant and is independent of W in (4.28) (Basu’s theorem). Thus, by (4.28), S 2 /u∗ is an MRIE, ˜1 [L(S 2 /u)] over all u > 0. If the loss where u∗ is a constant minimizing E function is given by (4.24) with p = 2, then by Corollary 4.1, the MRIE of σ 2 is n

T∗ (X) =

˜1 (S 2 ) S 2E S2 1 X ¯ 2, = 2 = (Xi − X) 2 2 2 ˜ (n − 1)/(n − 1) n + 1 i=1 E1 (S )

since (n − 1)S 2 has a chi-square distribution χ2n−1 when σ = 1. Next, consider the estimation of µ under the loss function (4.29). Since ¯ is a location-scale invariant estimator of µ and is independent of W in X (4.28) (Basu’s theorem), by Theorem 4.9, an MRIE of µ is ¯ − u∗ S 2 , T∗ (X) = X where u∗ is a constant. If L in (4.29) is convex and even, then u∗ = 0 (see ¯ is an MRIE of µ. Theorem 4.5(iii)) and, hence, X Example 4.17. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (µ − 12 σ, µ + 12 σ), where µ ∈ R and σ > 0 are unknown. Consider first the

260

4. Estimation in Parametric Models

estimation of σ under the loss function (4.24) with p = 2. The sample range X(n) − X(1) is a location-scale invariant estimator of σ and is independent of W in (4.28) (Basu’s theorem). By (4.28) and Corollary 4.1, the MRIE of σ is T∗ (X) =

˜1 (X(n) − X(1) ) (X(n) − X(1) )E (n + 2)(X(n) − X(1) ) = . 2 ˜ n E1 (X(n) − X(1) )

Consider now the estimation of µ under the loss function (4.29). Since (X(1) + X(n) )/2 is a location-scale invariant estimator of µ and is independent of W in (4.28) (Basu’s theorem), by Theorem 4.9, an MRIE of µ is X(1) + X(n) T∗ (X) = − u∗ (X(n) − X(1) ), 2 where u∗ is a constant. If L in (4.29) is convex and even, then u∗ = 0 (see Theorem 4.5(iii)) and, hence, (X(1) + X(n) )/2 is an MRIE of µ. Finding MRIE’s in various location-scale families under transformations AX +c, where A ∈ T and c ∈ C with given T and C, can be done in a similar way. We only provide some brief discussions for two important cases. The first case is the two-sample location-scale problem in which two samples, X = (X1 , ..., Xm ) and Y = (Y1 , ..., Yn ), are taken from a distribution with Lebesgue p.d.f. yn −µy x1 −µx xm −µx y1 −µy 1 , (4.30) σm σn f σx , ..., σx , σy , ..., σy x

y

where f is known, µx ∈ R and µy ∈ R are unknown location parameters, and σx > 0 and σy > 0 are unknown scale parameters. The family of distributions is invariant under the transformations g(X, Y ) = (rX1 + c, ..., rXm + c, r′ Y1 + c′ , ..., r′ Yn + c′ ),

(4.31)

where r > 0, r′ > 0, c ∈ R, and c′ ∈ R. The parameters to be estimated in this problem are usually ∆ = µy − µx and η = (σy /σx )h with a fixed h 6= 0. If X and Y are from two populations, ∆ and η are measures of the difference between the two populations. For estimating η, results similar to those in this section can be established. For estimating ∆, MRIE’s can be obtained under some conditions. See Exercises 63-65. The second case is the general linear model (3.25) under the assumption that εi ’s are i.i.d. with the p.d.f. σ −1 f (x/σ), where f is a known Lebesgue p.d.f. The family of populations is invariant under the transformations g(X) = rX + Zc,

r ∈ (0, ∞), c ∈ Rp

(4.32)

261

4.3. Minimaxity and Admissibility

(exercise). The estimation of lτ β with l ∈ R(Z) is invariant under the a−lτ β and the LSE lτ βˆ is an invariant estimator of lτ β loss function L σ (exercise). When f is normal, the following result can be established using an argument similar to that in Example 4.16. Theorem 4.10. Consider model (3.25) with assumptionA1. τ (i) Under transformations (4.32) and the loss function L a−lσ β , where L is convex and even, the LSE lτ βˆ is an MRIE of lτ β for any l ∈ R(Z). (ii) Under transformations (4.32) and the loss function (a − σ 2 )2 /σ 4 , the MRIE of σ 2 is SSR/(n − r + 2), where SSR is given by (3.35) and r is the rank of Z. MRIE’s in a parametric family with a multi-dimensional θ are often inadmissible. See Lehmann (1983, p. 285) for more discussions.

4.3 Minimaxity and Admissibility Consider the estimation of a real-valued ϑ = g(θ) based on a sample X from Pθ , θ ∈ Θ, under a given loss function. A minimax estimator minimizes the maximum risk supθ∈Θ RT (θ) over all estimators T (see §2.3.2). A unique minimax estimator is admissible, since any estimator better than a minimax estimator is also minimax. This indicates that we should consider minimaxity and admissibility together. The situation is different for a UMVUE (or an MRIE), since if a UMVUE (or an MRIE) is inadmissible, it is dominated by an estimator that is not unbiased (or invariant).

4.3.1 Estimators with constant risks By minimizing the maximum risk, a minimax estimator tries to do as well as possible in the worst case. Such an estimator can be very unsatisfactory. However, if a minimax estimator has some other good properties (e.g., it is a Bayes estimator), then it is often a reasonable estimator. Here we study when estimators having constant risks (e.g., MRIE’s) are minimax. Theorem 4.11. Let Π be a proper prior on Θ and δ be a Bayes estimator of ϑ w.r.t. Π. Let ΘΠ = {θ : Rδ (θ) = supθ∈Θ Rδ (θ)}. If Π(ΘΠ ) = 1, then δ is minimax. If, in addition, δ is the unique Bayes estimator w.r.t. Π, then it is the unique minimax estimator. Proof. Let T be any other estimator of ϑ. Then Z Z sup RT (θ) ≥ RT (θ)dΠ ≥ Rδ (θ)dΠ = sup Rδ (θ). θ∈Θ

ΘΠ

ΘΠ

θ∈Θ

262

4. Estimation in Parametric Models

If δ is the unique Bayes estimator, then the second inequality in the previous expression should be replaced by > and, therefore, δ is the unique minimax estimator. The condition of Theorem 4.11 essentially means that δ has a constant risk. Thus, a Bayes estimator having constant risk is minimax. Example 4.18. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). Consider the estimation of p under the squared er¯ has risk p(1 − p)/n which is not constant. In fact, ror loss. The UMVUE X ¯ is not minimax (Exercise 67). To find a minimax estimator by applying X Theorem 4.11, we consider the Bayes estimator w.r.t. the beta distribution B(α, β) with known α and β (Exercise 1): ¯ (α + β + n). δ(X) = (α + nX)

A straightforward calculation shows that

Rδ (p) = [np(1 − p) + (α − αp − βp)2 ] (α + β + n)2 .

To apply Theorem 4.11, we need to find values of α > 0 and β > 0 such that Rδ (p) √ is constant. It can be shown that Rδ (p) is constant if and only if α = β = n/2, which leads to the unique minimax estimator √ √ ¯ + n/2) (n + n). T (X) = (nX √ The risk of T is RT = 1/[4(1 + n)2 ]. Note that T is a Bayes estimator and has some good properties. Com¯ we find that T has smaller risk if and paring the risk of T with that of X, only if q q 1 1 n 1 1 n p ∈ 2 − 2 1 − (1+√n)2 , 2 + 2 1 − (1+√n)2 . (4.33)

¯ for most Thus, for a small n, T is better (and can be much better) than X of the range of p (Figure 4.1). When n → ∞, the interval in (4.33) shrinks ¯ is better than T toward 12 . Hence, for a large (and even moderate) n, X for most of the range of p (Figure 4.1). The limit of the asymptotic relative ¯ is 4p(1 − p), which is always smaller than 1 when efficiency of T w.r.t. X 1 p 6= 2 and equals 1 when p = 12 . The minimax estimator depends strongly on the loss function. To see this, let us consider the loss function L(p, a) = (a−p)2 /[p(1−p)]. Under this ¯ has constant risk and is the unique Bayes estimator w.r.t. loss function, X ¯ is the unique minimax the uniform prior on (0, 1). By Theorem 4.11, X √ estimator. On the other hand, the risk of T is equal to 1/[4(1+ n)2 p(1−p)], which is unbounded.

263

4.3. Minimaxity and Admissibility n=4

n=9

n=16

0.0

0.10

mse

0.20

0.30

0.0

0.10

mse

0.20

0.30

n=1

0.0

0.25

0.5 p

0.75

1.0

0.0

0.25

0.5

0.75

1.0

p

¯ (curve) and T (X) (straight line) in Example 4.18 Figure 4.1: mse’s of X In many cases a constant risk estimator is not a Bayes estimator (e.g., an unbiased estimator under the squared error loss), but a limit of Bayes estimators w.r.t. a sequence of priors. Then the following result may be used to find a minimax estimator. Theorem 4.12. Let Πj , j = 1, 2, ..., be a sequence of priors and rj be the Bayes risk of a Bayes estimator of ϑ w.r.t. Πj . Let T be a constant risk estimator of ϑ. If lim inf j rj ≥ RT , then T is minimax. The proof of this theorem is similar to that of Theorem 4.11. Although Theorem 4.12 is more general than Theorem 4.11 in finding minimax estimators, it does not provide uniqueness of the minimax estimator even when there is a unique Bayes estimator w.r.t. each Πj . In Example 2.25, we actually applied the result in Theorem 4.12 to show ¯ as an estimator of µ = EX1 when X1 , ..., Xn are i.i.d. the minimaxity of X from a normal distribution with a known σ 2 = Var(X1 ), under the squared ¯ in the case where σ 2 is unknown, error loss. To discuss the minimaxity of X we need the following lemma.

264

4. Estimation in Parametric Models

Lemma 4.3. Let Θ0 be a subset of Θ and T be a minimax estimator of ϑ when Θ0 is the parameter space. Then T is a minimax estimator if sup RT (θ) = sup RT (θ). θ∈Θ

θ∈Θ0

Proof. If there is an estimator T0 with supθ∈Θ RT0 (θ) < supθ∈Θ RT (θ), then sup RT0 (θ) ≤ sup RT0 (θ) < sup RT (θ) = sup RT (θ), θ∈Θ0

θ∈Θ

θ∈Θ

θ∈Θ0

which contradicts the minimaxity of T when Θ0 is the parameter space. Hence, T is minimax when Θ is the parameter space. Example 4.19. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown θ = (µ, σ 2 ). Consider the estimation of µ under the squared error loss. Suppose first that Θ = R × (0, c] with a constant c > 0. Let Θ0 = R × {c}. From ¯ is a minimax estimator of µ when the parameter space Example 2.25, X ¯ is also minimax when is Θ0 . An application of Lemma 4.3 shows that X 2 the parameter space is Θ. Although σ is assumed to be bounded by c, the ¯ does not depend on c. minimax estimator X Consider next the case where Θ = R × (0, ∞), i.e., σ2 is unbounded. Let T be any estimator of µ. For any fixed σ 2 , σ2 ≤ sup RT (θ), n µ∈R ¯ that is minimax when σ 2 is known (Example since σ 2 /n is the risk of X 2 2.25). Letting σ → ∞, we obtain that supθ RT (θ) = ∞ for any estimator T . Thus, minimaxity is meaningless (any estimator is minimax). Theorem 4.13. Suppose that T as an estimator of ϑ has constant risk and is admissible. Then T is minimax. If the loss function is strictly convex, then T is the unique minimax estimator. Proof. By the admissibility of T , if there is another estimator T0 with supθ RT0 (θ) ≤ RT , then RT0 (θ) = RT for all θ. This proves that T is minimax. If the loss function is strictly convex and T0 is another minimax estimator, then R(T +T0 )/2 (θ) < (RT0 + RT )/2 = RT for all θ and, therefore, T is inadmissible. This shows that T is unique if the loss is strictly convex. Combined with Theorem 4.7, Theorem 4.13 tells us that if an MRIE is admissible, then it is minimax. From the discussion at the end of §4.2.1, MRIE’s in one-parameter location families (such as Pitman’s estimators) are usually minimax.

265

4.3. Minimaxity and Admissibility

4.3.2 Results in one-parameter exponential families The following result provides a sufficient condition for the admissibility of a class of estimators when the population Pθ is in a one-parameter exponential family. Using this result and Theorem 4.13, we can obtain a class of minimax estimators. The proof of this result is an application of the information inequality introduced in §3.1.3. Theorem 4.14. Suppose that X has the p.d.f. c(θ)eθT (x) w.r.t. a σ-finite measure ν, where T (x) is real-valued and θ ∈ (θ− , θ+ ) ⊂ R. Consider the estimation of ϑ = E[T (X)] under the squared error loss. Let λ ≥ 0 and γ be known constants and let Tλ,γ (X) = (T + γλ)/(1 + λ). Then a sufficient condition for the admissibility of Tλ,γ is that Z

θ+

θ0

e−γλθ dθ = [c(θ)]λ

Z

θ0

θ−

e−γλθ dθ = ∞, [c(θ)]λ

(4.34)

where θ0 ∈ (θ− , θ+ ). Proof. From Theorem 2.1, ϑ = E[T (X)] = −c′ (θ)/c(θ) and dϑ dθ = Var(T ) = I(θ), the Fisher information defined in (3.5). Suppose that there is an estimator δ of ϑ such that for all θ, Rδ (θ) ≤ RTλ,γ (θ) = [I(θ) + λ2 (ϑ − γ)2 ]/(1 + λ)2 . Let bδ (θ) be the bias of δ. From the information inequality (3.6), Rδ (θ) ≥ [bδ (θ)]2 + [I(θ) + b′δ (θ)]2 /I(θ). Let h(θ) = bδ (θ) − λ(γ − ϑ)/(1 + λ). Then [h(θ)]2 −

2λh(θ)(ϑ − γ) + 2h′ (θ) [h′ (θ)]2 + ≤ 0, 1+λ I(θ)

which implies [h(θ)]2 −

2λh(θ)(ϑ − γ) + 2h′ (θ) ≤ 0. 1+λ

(4.35)

Let a(θ) = h(θ)[c(θ)]λ eγλθ . Differentiation of a(θ) reduces (4.35) to [a(θ)]2 e−γλθ 2a′ (θ) ≤ 0. + [c(θ)]λ 1+λ

(4.36)

Suppose that a(θ0 ) < 0 for some θ0 ∈ (θ− , θ+ ). From (4.36), a′ (θ) ≤ 0 for all θ. Hence a(θ) < 0 for all θ ≥ θ0 and, for θ > θ0 , (4.36) can be written as d 1 (1 + λ)e−γλθ ≥ . dθ a(θ) 2[c(θ)]λ

266

4. Estimation in Parametric Models

Integrating both sides from θ0 to θ gives 1+λ 2

Z

θ

θ0

e−γλθ 1 1 1 − ≤− . dθ ≤ λ [c(θ)] a(θ) a(θ0 ) a(θ0 )

Letting θ → θ+ , the left-hand side of the previous expression diverges to ∞ by condition (4.34), which is impossible. This shows that a(θ) ≥ 0 for all θ. Similarly, we can show that a(θ) ≤ 0 for all θ. Thus, a(θ) = 0 for all θ. This means that h(θ) = 0 for all θ and b′δ (θ) = −λϑ′ /(1 + λ) = −λI(θ)/(1 + λ), which implies Rδ (θ) ≡ RTλ,γ (θ). This proves the admissibility of Tλ,γ . The reason why Tλ,γ is considered is that it is often a Bayes estimator w.r.t. some prior; see, for example, Examples 2.25, 4.1, 4.7, and 4.8. To find minimax estimators, we may use the following result. Corollary 4.3. Assume that X has the p.d.f. as described in Theorem 4.14 with θ− = −∞ and θ+ = ∞. (i) As an estimator of ϑ = E(T ), T (X) is admissible under the squared error loss and the loss (a − ϑ)2 /Var(T ). (ii) T is the unique minimax estimator of ϑ under the loss (a − ϑ)2 /Var(T ). Proof. (i) With λ = 0, condition (4.34) is clearly satisfied. Hence, Theorem 4.14 applies under the squared error loss. The admissibility of T under the loss (a − ϑ)2 /Var(T ) follows from the fact that T is admissible under the squared error loss and Var(T ) 6= 0. (ii) This is a consequence of part (i) and Theorem 4.13. Example 4.20. P Let X1 , ..., Xn be i.i.d. from N (0, σ 2 ) with an unknown 2 σ > 0. Let Y = ni=1 Xi2 . From Example 4.14, Y /(n+2) is the MRIE of σ 2 and has constant risk under the loss (a − σ 2 )2 /σ 4 . We now apply Theorem 4.14 to show that Y /(n + 2) is admissible. Note that the joint p.d.f. of Xi ’s is of the form c(θ)eθT (x) with θ = −n/(4σ 2 ), c(θ) = (−2θ/n)n/2 , T (X) = 2Y /n, θ− = −∞, and θ+ = 0. By Theorem 4.14, Tλ,γ = (T + γλ)/(1 + λ) is admissible under the squared error loss if Z

−c

−∞

e

−γλθ

−2θ n

−nλ/2

dθ =

Z

0

c

eγλθ θ−nλ/2 dθ = ∞

for some c > 0. This means that Tλ,γ is admissible if γ = 0 and λ = 2/n, or if γ > 0 and λ ≥ 2/n. In particular, 2Y /(n + 2) is admissible for estimating E(T ) = 2E(Y )/n = 2σ 2 , under the squared error loss. It is easy to see that Y /(n + 2) is then an admissible estimator of σ 2 under the squared error loss and the loss (a − σ 2 )2 /σ 4 . Hence Y /(n + 2) is minimax under the loss (a − σ 2 )2 /σ 4 . Note that we cannot apply Corollary 4.3 directly since θ+ = 0.

267

4.3. Minimaxity and Admissibility

Example 4.21. Let X1 , ..., Xn be i.i.d. from the Poisson distribution P (θ) with an unknown θ > 0. The joint p.d.f. of Xi ’s w.r.t. the counting measure is (x1 ! · · · xn !)−1 e−nθ en¯x log θ . For η = n log θ, the conditions of Corollary ¯ Since E(T ) = θ and Var(T ) = θ/n, 4.3 are satisfied with T (X) = X. ¯ by Corollary 4.3, X is the unique minimax estimator of θ under the loss function (a − θ)2 /θ.

4.3.3 Simultaneous estimation and shrinkage estimators In this chapter (and most of Chapter 3) we have focused on the estimation of a real-valued ϑ. The problem of estimating a vector-valued ϑ under the decision theory approach is called simultaneous estimation. Many results for the case of a real-valued ϑ can be extended to simultaneous estimation in a straightforward manner. ˜ A Let ϑ be a p-vector of parameters (functions of θ) with range Θ. vector-valued estimator T (X) can be viewed as a decision rule taking values ˜ Let L(θ, a) be a given nonnegative loss function in the action space A = Θ. on Θ × A. A natural generalization of the squared error loss is L(θ, a) = ka − ϑk2 =

p X i=1

(ai − ϑi )2 ,

(4.37)

where ai and ϑi are the ith components of a and ϑ, respectively. A vector-valued estimator T is called unbiased if and only if E(T ) = ϑ for all θ ∈ Θ. If there is an unbiased estimator of ϑ, then ϑ is called estimable. It can be seen that the result in Theorem 3.1 extends to the case of vector-valued ϑ with any L strictly convex in a. If the loss function is given by (4.37) and Ti is a UMVUE of ϑi for each i, then T = (T1 , ..., Tp ) is a UMVUE of ϑ. If there is a sufficient and complete statistic U (X) for θ, then by Theorem 2.5 (Rao-Blackwell theorem), T must be a function of U (X) and is the unique best unbiased estimator of ϑ. Example 4.22. Consider the general linear model (3.25) with assumption A1 and a full rank Z. Let ϑ = β. An unbiased estimator of β is then the ˆ From the proof of Theorem 3.7, βˆ is a function of the sufficient and LSE β. complete statistic for θ = (β, σ 2 ). Hence, βˆ is the unique best unbiased estimator of ϑ under any strictly convex loss function. In particular, βˆ is the UMVUE of β under the loss function (4.37). Next, we consider Bayes estimators of ϑ, which is still defined to be Bayes actions considered as functions of X. Under the loss function (4.37), the Bayes estimator is still given by (4.4) with vector-valued g(θ) = ϑ.

268

4. Estimation in Parametric Models

Example 4.23. Let X = (X0 , X1 , ..., Xk ) have the multinomial distribution given in Example 2.7. Consider the estimation of the vector θ = (p0 , p1 , ..., pk ) under the loss function (4.37), and the Dirichlet prior for θ that has the Lebesgue p.d.f. Γ(α0 + · · · + αk ) α0 −1 k −1 p · · · pα IA (θ), k Γ(α0 ) · · · Γ(αk ) 0

(4.38)

Pk where αj ’s are known positive constants and A = {θ : 0 ≤ pj , j=0 pj = 1}. It turns out that the Dirichlet prior is conjugate so that the posterior of θ given X = x is also a Dirichlet distribution having the p.d.f. given by (4.38) with αj replaced by αj + xj , j = 0, 1, ..., k. Thus, the Bayes estimator of θ is δ = (δ0 , δ1 , ..., δk ) with δj (X) =

αj + Xj , α0 + α1 + · · · + αk + n

j = 0, 1, ..., k.

After a suitable class of transformations is defined, the results in §4.2 for invariant estimators and MRIE’s are still valid. This is illustrated by the following example. Example 4.24. Let X be a sample with the Lebesgue p.d.f. f (x − θ), where f is a known Lebesgue p.d.f. on Rp with a finite second moment and θ ∈ Rp is an unknown parameter. Consider the estimation of θ under the loss function (4.37). This problem is invariant under the location transformations g(X) = X + c, where c ∈ Rp . Invariant estimators of θ are of the form X + l, l ∈ Rp . It is easy to show that any invariant estimator has constant bias and risk (a generalization of Proposition 4.4) and the MRIE of θ is the unbiased invariant estimator. In particular, if f is the p.d.f. of Np (0, Ip ), then the MRIE is X. The definition of minimax estimators applies without changes. Example 4.25. Let X be a sample from Np (θ, Ip ) with an unknown θ ∈ Rp . Consider the estimation of θ under the loss function (4.37). A modification of the proof of Theorem 4.12 with independent priors for θi ’s shows that X is a minimax estimator of θ (exercise). Example 4.26. Consider Example 4.23. If we choose α0 = · · · = αk = √ n/(k + 1), then the Bayes estimator of θ in Example 4.23 has constant risk. Using the same argument in the proof of Theorem 4.11, we can show that this Bayes estimator is minimax. The previous results for simultaneous estimation are fairly straightforward generalizations of those for the case of a real-valued ϑ. Results for

4.3. Minimaxity and Admissibility

269

admissibility in simultaneous estimation, however, are quite different. A surprising result, due to Stein (1956), is that in estimating the vector mean θ = EX of a normally distributed p-vector X (Example 4.25), X is inadmissible under the loss function (4.37) when p ≥ 3, although X is the UMVUE, MRIE (Example 4.24), and minimax estimator (Example 4.25). Since any estimator better than a minimax estimator is also minimax, there exist many (in fact, infinitely many) minimax estimators in Example 4.25 when p ≥ 3, which is different from the case of p = 1 in which X is the unique admissible minimax estimator (Example 4.6 and Theorem 4.13). We start with the simple case where X is from Np (θ, Ip ) with an unknown θ ∈ Rp . James and Stein (1961) proposed the following class of estimators of ϑ = θ having smaller risks than X when the loss is given by (4.37) and p ≥ 3: p−2 δc = X − (X − c), (4.39) kX − ck2 where c ∈ Rp is fixed. The choice of c is discussed next and at the end of this section. Before we prove that δc in (4.39) is better than X, we try to motivate δc from two viewpoints. First, suppose that it were thought a priori likely, though not certain, that θ = c. Then we might first test a hypothesis H0 : θ = c and estimate θ by c if H0 is accepted and by X otherwise. The best rejection region has the form kX − ck2 > t for some constant t > 0 (see Chapter 6) so that we might estimate θ by I(t,∞) (kX − ck2 )X + [1 − I(t,∞) (kX − ck2 )]c. It can be seen that δc in (4.39) is a smoothed version of this estimator, since δc = ψ(kX − ck2 )X + [1 − ψ(kX − ck2 )]c (4.40) for some function ψ. Any estimator having the form of the right-hand side of (4.40) shrinks the observations toward a given point c and, therefore, is called a shrinkage estimator. Next, δc in (4.40) can be viewed as an empirical Bayes estimator (§4.1.2). In view of (2.25) in Example 2.25, a Bayes estimator of θ is of the form δ = (1 − B)X + Bc, where c is the prior mean of θ and B involves prior variances. If 1 − B is “estimated” by ψ(kX − ck2 ), then δc is an empirical Bayes estimator. Theorem 4.15. Suppose that X is from Np (θ, Ip ) with p ≥ 3. Then, under the loss function (4.37), the risks of the following estimators of θ, δc,r = X −

r(p − 2) (X − c), kX − ck2

(4.41)

270

4. Estimation in Parametric Models

are given by Rδc,r (θ) = p − (2r − r2 )(p − 2)2 E(kX − ck−2 ),

(4.42)

where c ∈ Rp and r ∈ R are known. Proof. Let Z = X − c. Then

2

r(p − 2)

. Rδc,r (θ) = Ekδc,r − E(X)k2 = E Z − E(Z) 1 −

2 kZk

Hence, we only need to show the case of c = 0. Let h(θ) = Rδ0,r (θ), g(θ) be 2 the right-hand side of (4.42) with c = 0, and πα (θ) = (2πα)−p/2 e−kθk /(2α) , which is the p.d.f. of Np (0, αIp ). Note that the distribution of X can be viewed as the conditional distribution of X given θ = θ, where θ has the Lebesgue p.d.f. πα (θ). Then Z g(θ)πα (θ)dθ = p − (2r − r2 )(p − 2)2 E[E(kXk−2 |θ)] Rp

= p − (2r − r2 )(p − 2)2 E(kXk−2 ) = p − (2r − r2 )(p − 2)/(α + 1),

where the expectation in the second line of the previous expression is w.r.t. the joint distribution of (X, θ) and the last equality follows from the fact that the marginal distribution of X is Np (0, (α+1)Ip ), kXk2/(α+1) has the chi-square distribution χ2p and, therefore, E(kXk−2) = 1/[(p − 2)(α + 1)]. ˆ = r(p − 2)/kXk2. Then Let B = 1/(α + 1) and B Z ˆ h(θ)πα (θ)dθ = Ek(1 − B)X − θk2 Rp

ˆ = E{E[k(1 − B)X − θk2 |X]}

= E{E[kθ − E(θ|X)k2 |X] 2 ˆ } + kE(θ|X) − (1 − B)Xk ˆ − B)2 kXk2 } = E{p(1 − B) + (B

= E{p(1 − B) + B 2 kXk2 − 2Br(p − 2) + r2 (p − 2)2 kXk−2 }

= p − (2r − r2 )(p − 2)B,

where the fourth equality follows from the fact that the conditional distribution of θ given X is Np (1 − B)X, (1 − B)Ip and the last equality follows from EkXk−2 = B/(p − 2) and EkXk2 = p/B. This proves Z Z g(θ)πα (θ)dθ = h(θ)πα (θ)dθ, α > 0. (4.43) Rp

Rp

4.3. Minimaxity and Admissibility

271

Note that h(θ) and g(θ) are expectations of functions of kXk2 , θτ X, and kθk2 . Make an orthogonal transformation from X to Y such that Y1 = θτ X/kθk, EYj = 0 for j > 1, and P Var(Y ) = Ip . Then h(θ) and g(θ) are expectations of functions of Y1 , pj=2 Yj2 , and kθk2 . Thus, both h and g are functions of kθk2 . For the family of p.d.f.’s {πα (θ) : α > 0}, kθk2 is a complete and sufficient “statistic”. Hence, (4.43) and the fact that h and g are functions of kθk2 imply that h(θ) = g(θ) a.e. w.r.t. the Lebesgue measure. From Theorem 2.1, both h and g are continuous functions of kθk2 and, therefore, h(θ) = g(θ) for all θ ∈ Rp . This completes the proof. It follows from Theorem 4.15 that the risk of δc,r is smaller than that of X (for every value of θ) when p ≥ 3 and 0 < r < 2, since the risk of X is p under the loss function (4.37). From Example 4.6, X is admissible when p = 1. When p = 2, X is still admissible (Stein, 1956). But we have just shown that X is inadmissible when p ≥ 3. The James-Stein estimator δc in (4.39), which is a special case of (4.41) with r = 1, is better than any δc,r in (4.41) with r 6= 1, since the factor 2r − r2 takes on its maximum value 1 if and only if r = 1. To see that δc may have a substantial improvement over X in terms of risks, consider the special case where θ = c. Since kX − ck2 has the chi-square distribution χ2p when θ = c, EkX −ck−2 = (p−2)−1 and the right-hand side of (4.42) equals 2. Thus, the ratio RX (θ)/Rδc (θ) equals p/2 when θ = c and, therefore, can be substantially larger than 1 near θ = c when p is large. Since X is minimax (Example 4.25), any shrinkage estimator of the form (4.41) is minimax provided that p ≥ 3 and 0 < r < 2. Unfortunately, the James-Stein estimator with any c is also inadmissible. It is dominated by p−2 δc+ = X − min 1, (X − c); (4.44) kX − ck2 see, for example, Lehmann (1983, Theorem 4.6.2). This estimator, however, is still inadmissible. An example of an admissible estimator of the form (4.40) is provided by Strawderman (1971); see also Lehmann (1983, p. 304). Although neither the James-Stein estimator δc nor δc+ in (4.44) is admissible, it is found that no substantial improvements over δc+ are possible (Efron and Morris, 1973). To extend Theorem 4.15 to general Var(X), we consider the case where Var(X) = σ 2 D with an unknown σ 2 > 0 and a known positive definite matrix D. If σ 2 is known, then an extended James-Stein estimator is δ˜c,r = X −

r(p − 2)σ 2 D−1 (X − c). kD−1 (X − c)k2

(4.45)

272

4. Estimation in Parametric Models

One can show (exercise) that under the loss (4.37), the risk of δ˜c,r is σ 2 tr(D) − (2r − r2 )(p − 2)2 σ 2 E(kD−1 (X − c)k−2 ) . (4.46)

When σ 2 is unknown, we assume that there exists a statistic S02 such that S02 is independent of X and S02 /σ 2 has the chi-square distribution χ2m (see Example 4.27). Replacing rσ 2 in (4.45) by σ ˆ 2 = tS02 with a constant t > 0 leads to the following extended James-Stein estimator: δ˜c = X −

(p − 2)ˆ σ2 D−1 (X − c). kD−1 (X − c)k2

(4.47)

By (4.46) and the independence of σ ˆ 2 and X, the risk of δ˜c (as an estimator of ϑ = EX) is h i Rδ˜c (θ) = E E(kδ˜c − ϑk2 |ˆ σ2 ) h i = E E(kδ˜c,(ˆσ2 /σ2 ) − ϑk2 |ˆ σ2 ) = σ 2 E tr(D) − [2(ˆ σ 2 /σ 2 ) − (ˆ σ 2 /σ 2 )2 ](p − 2)2 σ 2 κ(θ) σ 2 /σ 2 )2 ](p − 2)2 σ 2 κ(θ) σ 2 /σ 2 ) − E(ˆ = σ 2 tr(D) − [2E(ˆ = σ 2 tr(D) − [2tm − t2 m(m + 2)](p − 2)2 σ 2 κ(θ) ,

where θ = (ϑ, σ 2 ) and κ(θ) = E(kD−1 (X − c)k−2 ). Since 2tm − t2 m(m + 2) is maximized at t = 1/(m + 2), replacing t by 1/(m + 2) leads to Rδ˜c (θ) = σ 2 tr(D) − m(m + 2)−1 (p − 2)2 σ 2 E(kD−1 (X − c)k−2 ) .

Hence, the risk of the extended James-Stein estimator in (4.47) is smaller than that of X for any fixed θ, when p ≥ 3. Example 4.27. Consider the general linear model (3.25) with assumption A1, p ≥ 3, and a full rank Z, and the estimation of ϑ = β under the loss function (4.37). From Theorem 3.8, the LSE βˆ is from N (β, σ 2 D) with a ˆ and S 2 /σ 2 known matrix D = (Z τ Z)−1 ; S02 = SSR is independent of β; 0 2 has the chi-square distribution χn−p . Hence, from the previous discussion, the risk of the shrinkage estimator βˆ −

(p − 2)ˆ σ2 Z τ Z(βˆ − c) kZ τ Z(βˆ − c)k2

is smaller than that of βˆ for any β and σ 2 , where c ∈ Rp is fixed and σ ˆ 2 = SSR/(n − p + 2). From the previous discussion, the James-Stein estimators improve X substantially when we shrink the observations toward a vector c that is near

273

4.4. The Method of Maximum Likelihood

ϑ = EX. Of course, this cannot be done since ϑ is unknown. One may consider shrinking the observations toward the mean of the observations rather than a given point; that is, one may obtain a shrinkage estimator by ¯ p , where X ¯ = p−1 Pp Xi and Jp is replacing c in (4.39) or (4.47) by XJ i=1 the p-vector of ones. However, we have to replace the factor p − 2 in (4.39) or (4.47) by p − 3. This leads to shrinkage estimators p−3 ¯ ¯ p k2 (X − XJp ) kX − XJ

(4.48)

(p − 3)ˆ σ2 −1 ¯ ¯ p )k2 D (X − XJp ). kD−1 (X − XJ

(4.49)

X− and X−

These estimators are better than X (and, hence, are minimax) when p ≥ 4, under the loss function (4.37) (exercise). The results discussed in this section for the simultaneous estimation of a vector of normal means can be extended to a wide variety of cases where the loss functions are not given by (4.37) (Brown, 1966). The results have also been extended to exponential families and to general location parameter families. For example, Berger (1976) studied the inadmissibility of generalized Bayes estimators of a location vector; Berger (1980) considered simultaneous estimation of gamma scale parameters; and Tsui (1981) investigated simultaneous estimation of several Poisson parameters. See Lehmann (1983, pp. 320-330) for some further references.

4.4 The Method of Maximum Likelihood So far we have studied estimation methods in parametric families using the decision theory approach. The maximum likelihood method introduced next is the most popular method for deriving estimators in statistical inference that does not use any loss function.

4.4.1 The likelihood function and MLE’s To introduce the idea, let us consider an example. Example 4.28. Let X be a single observation taking values from {0, 1, 2} according to Pθ , where θ = θ0 or θ1 and the values of Pθj ({i}) are given by the following table: θ = θ0 θ = θ1

x=0 0.8 0.2

x=1 0.1 0.3

x=2 0.1 0.5

274

4. Estimation in Parametric Models

If X = 0 is observed, it is more plausible that it came from Pθ0 , since Pθ0 ({0}) is much larger than Pθ1 ({0}). We then estimate θ by θ0 . On the other hand, if X = 1 or 2, it is more plausible that it came from Pθ1 , although in this case the difference between the probabilities is not as large as that in the case of X = 0. This suggests the following estimator of θ: θ0 X=0 T (X) = X 6= 0. θ1 The idea in Example 4.28 can be easily extended to the case where Pθ is a discrete distribution and θ ∈ Θ ⊂ Rk . If X = x is observed, θ1 is more plausible than θ2 if and only if Pθ1 ({x}) > Pθ2 ({x}). We then estimate θ by a θˆ that maximizes Pθ ({x}) over θ ∈ Θ, if such a θˆ exists. The word plausible rather than probable is used because θ is considered to be nonrandom and Pθ is not a distribution of θ. Under the Bayesian approach with a prior that is the discrete uniform distribution on {θ1 , ..., θm }, Pθ ({x}) is proportional to the posterior probability and we can say that θ1 is more probable than θ2 if Pθ1 ({x}) > Pθ2 ({x}). Note that Pθ ({x}) in the previous discussion is the p.d.f. w.r.t. the counting measure. Hence, it is natural to extend the idea to the case of continuous (or arbitrary) X by using the p.d.f. of X w.r.t. some σ-finite measure on the range X of X. This leads to the following definition. Definition 4.3. Let X ∈ X be a sample with a p.d.f. fθ w.r.t. a σ-finite measure ν, where θ ∈ Θ ⊂ Rk . (i) For each x ∈ X, fθ (x) considered as a function of θ is called the likelihood function and denoted by ℓ(θ). ˆ = max ¯ ℓ(θ) is ¯ be the closure of Θ. A θˆ ∈ Θ ¯ satisfying ℓ(θ) (ii) Let Θ θ∈Θ ˆ called a maximum likelihood estimate (MLE) of θ. If θ is a Borel function of X a.e. ν, then θˆ is called a maximum likelihood estimator (MLE) of θ. (iii) Let g be a Borel function from Θ to Rp , p ≤ k. If θˆ is an MLE of θ, ˆ is defined to be an MLE of ϑ = g(θ). then ϑˆ = g(θ) ¯ instead of Θ is used in the definition of an MLE. This is Note that Θ because a maximum of ℓ(θ) may not exist when Θ is an open set (Examples 4.29 and 4.30). As an estimator, an MLE is defined a.e. ν. Part (iii) of Definition 4.3 is motivated by a fact given in Exercise 95 of §4.6. ¯ = Θ If the parameter space Θ contains finitely many points, then Θ and an MLE can always be obtained by comparing finitely many values ℓ(θ), θ ∈ Θ. If ℓ(θ) is differentiable on Θ◦ , the interior of Θ, then possible candidates for MLE’s are the values of θ ∈ Θ◦ satisfying ∂ℓ(θ) = 0, ∂θ

(4.50)

4.4. The Method of Maximum Likelihood

275

which is called the likelihood equation. Note that θ’s satisfying (4.50) may be local or global minima, local or global maxima, or simply stationary points. Also, extrema may occur at the boundary of Θ or when kθk → ∞. Furthermore, if ℓ(θ) is not always differentiable, then extrema may occur at nondifferentiable or discontinuity points of ℓ(θ). Hence, it is important to analyze the entire likelihood function to find its maxima. Since log x is a strictly increasing function and ℓ(θ) can be assumed to be positive without loss of generality, θˆ is an MLE if and only if it maximizes the log-likelihood function log ℓ(θ). It is often more convenient to work with log ℓ(θ) and the following analogue of (4.50) (which is called the log-likelihood equation or likelihood equation for simplicity): ∂ log ℓ(θ) = 0. ∂θ

(4.51)

Example 4.29. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ Θ = (0, 1). When (X1 , ..., Xn ) = (x1 , ..., xn ) is observed, the likelihood function is ℓ(p) =

n Y

i=1

Pn −1

pxi (1 − p)1−xi = pn¯x (1 − p)n(1−¯x) ,

◦ ¯ where x ¯=n i=1 xi . Note that Θ = [0, 1] and Θ = Θ. The likelihood equation (4.51) reduces to

n¯ x n(1 − x¯) − = 0. p 1−p If 0 < x ¯ < 1, then this equation has a unique solution x ¯. The second-order derivative of log ℓ(p) is n¯ x n(1 − x ¯) − 2 − , p (1 − p)2 which is always negative. Also, when p tends to 0 or 1 (the boundary of Θ), ℓ(p) → 0. Thus, x ¯ is the unique MLE of p. When x ¯ = 0, ℓ(p) = (1 − p)n is a strictly decreasing function of p and, therefore, its unique maximum is 0. Similarly, the MLE is 1 when x¯ = 1. Combining these results with the previous result, we conclude that the MLE of p is x ¯. When x¯ = 0 or 1, a maximum of ℓ(p) does not exist on Θ = (0, 1), although supp∈(0,1) ℓ(p) = 1; the MLE takes a value outside of Θ and, hence, is not a reasonable estimator. However, if p ∈ (0, 1), the probability that x ¯ = 0 or 1 tends to 0 quickly as n → ∞.

276

4. Estimation in Parametric Models

Example 4.29 indicates that, for small n, a maximum of ℓ(θ) may not exist on Θ and an MLE may be an unreasonable estimator; however, this is unlikely to occur when n is large. A rigorous result of this sort is given in §4.5.2, where we study asymptotic properties of MLE’s. Example 4.30. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with an unknown θ = (µ, σ 2 ), where n ≥ 2. Consider first the case where Θ = R × (0, ∞). When (X1 , ..., Xn ) = (x1 , ..., xn ) is observed, the log-likelihood function is log ℓ(θ) = −

n 1 X n n (xi − µ)2 − log σ 2 − log(2π). 2σ 2 i=1 2 2

The likelihood equation (4.51) becomes n 1 X (xi − µ) = 0 σ 2 i=1

and

n 1 X n (xi − µ)2 − 2 = 0. 4 σ i=1 σ

(4.52)

Solving ¯= Pn the first equation in (4.52) for µ, we obtain a unique solution x n−1 i=1 xi , and substituting x ¯ for µ in the second equation in (4.52), Pn we obtain a unique solution σ ˆ 2 = n−1 i=1 (xi − x ¯)2 . To show that θˆ = 2 (¯ x, σ ˆ ) is an MLE, first note that Θ is an open set and ℓ(θ) is differentiable everywhere; as θ tends to the boundary of Θ or kθk → ∞, ℓ(θ) tends to 0; and ! n 1 Pn ∂ 2 log ℓ(θ) i=1 (xi − µ) σ2 σ4 =− Pn 1 Pn ∂θ∂θτ (xi − µ) 16 (xi − µ)2 − n4 4 σ

i=1

σ

i=1

2σ

is negative definite when µ = x ¯ and σ = σ ˆ . Hence θˆ is the unique MLE. Sometimes we can avoid the calculation of the second-order derivatives. For instance, in this example we know that ℓ(θ) is bounded and ℓ(θ) → 0 as kθk → ∞ or θ tends to the boundary of Θ; hence the unique solution to (4.52) must be the MLE. Another way to show that θˆ is the MLE is indicated by the following discussion. Consider next the case where Θ = (0, ∞) × (0, ∞), i.e., µ is known to be positive. The likelihood function is differentiable on Θ◦ = Θ and ¯ = [0, ∞) × [0, ∞). If x Θ ¯ > 0, then the same argument for the previous ¯ ≤ 0, then the first case can be used to show that (¯ x, σ ˆ 2 ) is the MLE. If x equation in (4.52) does not have a solution in Θ. However, the function log ℓ(θ) = log ℓ(µ, σ 2 ) is strictly decreasing in µ for any fixed σ 2 . Hence, a maximum of log ℓ(µ, σ 2 ) is µ = 0, which does not depend on σ 2 . Then, the MLE is (0, σ ˜ 2 ), where σ ˜ 2 is the value maximizing log ℓ(0, σ 2 ) overPσ 2 ≥ 0. n Applying (4.51) to the function log ℓ(0, σ 2 ) leads to σ ˜ 2 = n−1 i=1 x2i . Thus, the MLE is x ¯>0 (¯ x, σ ˆ2 ) θˆ = x ¯ ≤ 0. (0, σ ˜2 ) 2

2

277

4.4. The Method of Maximum Likelihood

Again, the MLE in this case is not in Θ if x¯ ≤ 0. One can show that a maximum of ℓ(θ) does not exist on Θ when x ¯ ≤ 0. Example 4.31. Let X1 , ..., Xn be i.i.d. from the uniform distribution on an interval Iθ with an unknown θ. First, consider the case where Iθ = (0, θ) and θ > 0. The likelihood function is ℓ(θ) = θ−n I(x(n) ,∞) (θ), which is not always differentiable. In this case Θ◦ = (0, x(n) ) ∪ (x(n) , ∞). But, on (0, x(n) ), ℓ ≡ 0 and on (x(n) , ∞), ℓ′ (θ) = −nθn−1 < 0 for all θ. Hence, the method of using the likelihood equation is not applicable to this problem. Since ℓ(θ) is strictly decreasing on (x(n) , ∞) and is 0 on (0, x(n) ), a unique maximum of ℓ(θ) is x(n) , which is a discontinuity point of ℓ(θ). This shows that the MLE of θ is the largest order statistic X(n) . Next, consider the case where Iθ = (θ − 12 , θ + 12 ) with θ ∈ R. The likelihood function is ℓ(θ) = I(x(n) − 21 ,x(1) + 12 ) (θ). Again, the method of using the likelihood equation is not applicable. However, it follows from Definition 4.3 that any statistic T (X) satisfying x(n) − 21 ≤ T (x) ≤ x(1) + 12 is an MLE of θ. This example indicates that MLE’s may not be unique and can be unreasonable. Example 4.32. Let X be an observation from the hypergeometric distribution HG(r, n, θ − n) (Table 1.1, page 18) with known r, n, and an unknown θ = n + 1, n + 2, .... In this case, the likelihood function is defined on integers and the method of using the likelihood equation is certainly not applicable. Note that ℓ(θ) (θ − r)(θ − n) = , ℓ(θ − 1) θ(θ − n − r + x)

which is larger than 1 if and only if θ < rn/x and is smaller than 1 if and only if θ > rn/x. Thus, ℓ(θ) has a maximum θ = the integer part of rn/x, which is the MLE of θ. Example 4.33. Let X1 , ..., Xn be i.i.d. from the gamma distribution Γ(α, γ) with unknown α > 0 and γ > 0. The log-likelihood function is log ℓ(θ) = −nα log γ − n log Γ(α) + (α − 1) and the likelihood equation (4.51) becomes n

−n log γ − and −

n X i=1

log xi −

nΓ′ (α) X + log xi = 0 Γ(α) i=1

n 1 X nα + 2 xi = 0. γ γ i=1

n 1X xi γ i=1

278

4. Estimation in Parametric Models

The second equation yields γ = x ¯/α. Substituting γ = x ¯/α into the first equation we obtain that n

log α −

1X Γ′ (α) + log xi − log x ¯ = 0. Γ(α) n i=1

In this case, the likelihood equation does not have an explicit solution, although it can be shown (exercise) that a solution exists almost surely and it is the unique MLE. A numerical method has to be applied to compute the MLE for any given observations x1 , ..., xn . These examples indicate that we need to use various methods to derive MLE’s. In applications, MLE’s typically do not have analytic forms and some numerical methods have to be used to compute MLE’s. A commonly used numerical method is the Newton-Raphson iteration method, which repeatedly computes −1 2 ∂ log ℓ(θ) ∂ log ℓ(θ) (t+1) (t) ˆ ˆ θ =θ − ˆ(t) , ∂θ∂θτ θ=θˆ(t) ∂θ θ=θ

(4.53)

t = 0, 1, ..., where θˆ(0) is an initial value and ∂ 2 log ℓ(θ)/∂θ∂θτ is assumed of full rank for every θ ∈ Θ. If, at each iteration, we replace ∂ 2 log ℓ(θ)/∂θ∂θτ in (4.53) by its expected value E[∂ 2 log ℓ(θ)/∂θ∂θτ ], where the expectation is taken under Pθ , then the method is known as the Fisher-scoring method. If the iteration converges, then θˆ(∞) or θˆ(t) with a sufficiently large t is a numerical approximation to a solution of the likelihood equation (4.51). The following example shows that the MCMC methods discussed in §4.1.4 can also be useful in computing MLE’s. Example 4.34. Let X be a random k-vector from Pθ with the following p.d.f. w.r.t. a σ-finite measure ν: Z fθ (x) = fθ (x, y)dν(y), where fθ (x, y) is a joint p.d.f. w.r.t. ν × ν. This type of distribution is called a mixture distribution. Thus, the likelihood ℓ(θ) = fθ (x) involves a k-dimensional integral. In many cases this integral has to be computed in order to compute an MLE of θ. Let ℓ˜m (θ) be the MCMC approximation to ℓ(θ) based on one of the MCMC methods described in §4.1.4 and a Markov chain of length m. Under the conditions of Theorem 4.4, ℓ˜m (θ) →a.s. ℓ(θ) for every fixed θ and x. Suppose that, for each m, there exists θ˜m that maximizes ℓ˜m (θ) over θ ∈ Θ. Geyer (1994) studies the convergence of θ˜m to an MLE.

4.4. The Method of Maximum Likelihood

279

In terms of their mse’s, MLE’s are not necessarily better than UMVUE’s or Bayes estimators. Also, MLE’s are frequently inadmissible. This is not surprising, since MLE’s are not derived under any given loss function. The main theoretical justification for MLE’s is provided in the theory of asymptotic efficiency considered in §4.5.

4.4.2 MLE’s in generalized linear models Suppose that X has a distribution from a natural exponential family so that the likelihood function is ℓ(η) = exp{η τ T (x) − ζ(η)}h(x), where η ∈ Ξ is a vector of unknown parameters. The likelihood equation (4.51) is then ∂ζ(η) ∂ log ℓ(η) = T (x) − = 0, ∂η ∂η which has a unique solution T (x) = ∂ζ(η)/∂η, assuming that T (x) is in the range of ∂ζ(η)/∂η. Note that ∂ 2 log ℓ(η) ∂ 2 ζ(η) = − = −Var(T ) ∂η∂η τ ∂η∂η τ

(4.54)

(see the proof of Proposition 3.2). Since Var(T ) is positive definite, − log ℓ(η) is convex in η and T (x) is the unique MLE of the parameter µ(η) = ∂ζ(η)/∂η. By (4.54) again, the function µ(η) is one-to-one so that µ−1 exists. By Definition 4.3, the MLE of η is ηˆ = µ−1 (T (x)). If the distribution of X is in a general exponential family and the likelihood function is ℓ(θ) = exp{[η(θ)]τ T (x) − ξ(θ)}h(x), η ), if η −1 exists and ηˆ is in the range of η(θ). then the MLE of θ is θˆ = η −1 (ˆ Of course, θˆ is also the solution of the likelihood equation ∂η(θ) ∂ξ(θ) ∂ log ℓ(θ) = T (x) − = 0. ∂θ ∂θ ∂θ The results for exponential families lead to an estimation method in a class of models that have very wide applications. These models are generalizations of the normal linear model (model (3.25) with assumption A1) discussed in §3.3.1-§3.3.2 and, therefore, are named generalized linear models (GLM).

280

4. Estimation in Parametric Models

A GLM has the following structure. The sample X = (X1 , ..., Xn ) ∈ Rn has independent components and Xi has the p.d.f. o n i) h(xi , φi ), i = 1, ..., n, (4.55) exp ηi xi −ζ(η φi

w.r.t. a σ-finite measure ν, where ηi and φi are unknown, φi > 0, R ηi ∈ Ξ = η : 0 < h(x, φ)eηx/φ dν(x) < ∞ ⊂ R

for all i, ζ and h are known functions, and ζ ′′ (η) > 0 is assumed for all η ∈ Ξ◦ , the interior of Ξ. Note that the p.d.f. in (4.55) belongs to an exponential family if φi is known. As a consequence, E(Xi ) = ζ ′ (ηi ) and Var(Xi ) = φi ζ ′′ (ηi ),

i = 1, ..., n.

(4.56)

Define µ(η) = ζ ′ (η). It is assumed that ηi is related to Zi , the ith value of a p-vector of covariates (see (3.24)), through g(µ(ηi )) = β τ Zi ,

i = 1, ..., n,

(4.57)

where β is a p-vector of unknown parameters and g, called a link function, is a known one-to-one, third-order continuously differentiable function on {µ(η) : η ∈ Ξ◦ }. If µ = g −1 , then ηi = β τ Zi and g is called the canonical or d natural link function. If g is not canonical, we assume that dη (g ◦ µ)(η) 6= 0 for all η. In a GLM, the parameter of interest is β. We assume that the range of β is B = {β : (g ◦ µ)−1 (β τ z) ∈ Ξ◦ for all z ∈ Z}, where Z is the range of Zi ’s. φi ’s are called dispersion parameters and are considered to be nuisance parameters. It is often assumed that φi = φ/ti ,

i = 1, ..., n,

(4.58)

with an unknown φ > 0 and known positive ti ’s. As we discussed earlier, the linear model (3.24) with εi = N (0, φ) is a special GLM. One can verify this by taking g(µ) ≡ µ and ζ(η) = η 2 /2. The usefulness of the GLM is that it covers situations where the relationship between E(Xi ) and Zi is nonlinear and/or Xi ’s are discrete (in which case the linear model (3.24) is clearly not appropriate). The following is an example. Example 4.35. Let Xi ’s be independent discrete random variables taking values in {0, 1, ..., m}, where m is a known positive integer. First, suppose that Xi has the binomial distribution Bi(pi , m) with an unknown pi ∈ pi (0, 1), i = 1, ..., n. Let ηi = log 1−p and ζ(ηi ) = m log(1 + eηi ). Then the i p.d.f. of Xi (w.r.t. the counting measure) is given by (4.55) with φi = 1,

281

4.4. The Method of Maximum Likelihood

h(xi , φi ) = xmi , and Ξ = R. Under (4.57) and the logit link (canonical t link) g(t) = log m−t , τ

meηi meβ Zi E(Xi ) = mpi = = . 1 + eηi 1 + eβ τ Zi Another popular link in this problem is the probit link g(t) = Φ−1 (t/m), where Φ is the c.d.f. of the standard normal. Under the probit link, E(Xi ) = mΦ(β τ Zi ). The variance of Xi is mpi (1 − pi ) under the binomial distribution assumption. This assumption is often violated in applications, which results in an over-dispersion, i.e., the variance of Xi exceeds the nominal variance mpi (1 − pi ). Over-dispersion can arise in a number of ways, but the most common one is clustering in the population. Families, households, and litters Pm are common instances of clustering. For example, suppose that Xi = j=1 Xij , where Xij are binary random variables having a common distribution. If Xij ’s are independent, then Xi has a binomial distribution. However, if Xij ’s are from the same cluster (family or household), then they are often positively correlated. Suppose that the correlation coefficient (§1.3.2) between Xij and Xil , j 6= l, is ρi > 0. Then Var(Xi ) = mpi (1 − pi ) + m(m − 1)ρi pi (1 − pi ) = φi mpi (1 − pi ), where φi = 1 + (m − 1)ρi is the dispersion parameter. Of course, overdispersion can occur only if m > 1 in this case. This motivates the consideration of GLM (4.55)-(4.57) with dispersion parameters φi . If Xi has the p.d.f. (4.55) with ζ(ηi ) = m log(1 + eηi ), then E(Xi ) =

meηi 1 + eηi

and

Var(Xi ) = φi

meηi , (1 + eηi )2

which is exactly (4.56). Of course, the distribution of Xi is not binomial unless φi = 1. We now derive an MLE of β in a GLM under assumption (4.58). Let θ = (β, φ) and ψ = (g ◦ µ)−1 . Then the log-likelihood function is log ℓ(θ) =

n X i=1

ψ(β τ Zi )xi − ζ(ψ(β τ Zi )) log h(xi , φ/ti ) + φ/ti

and the likelihood equation is n

1X ∂ log ℓ(θ) = {[xi − µ(ψ(β τ Zi ))]ψ ′ (β τ Zi )ti Zi } = 0 ∂β φ i=1

(4.59)

282

4. Estimation in Parametric Models

and n

∂ log ℓ(θ) X = ∂φ i=1

∂ log h(xi , φ/ti ) ti [ψ(β τ Zi )xi − ζ(ψ(β τ Zi ))] − ∂φ φ2

= 0.

From the first equation, an MLE of β, if it exists, can be obtained without estimating φ. The second equation, however, is usually difficult to solve. Some other estimators of φ are suggested by various researchers; see, for example, McCullagh and Nelder (1989). Suppose that there is a solution βˆ ∈ B to equation (4.59). (The existence of βˆ is studied in §4.5.2.) We now study whether βˆ is an MLE of β. Let n X Mn (β) = [ψ ′ (β τ Zi )]2 ζ ′′ (ψ(β τ Zi ))ti Zi Ziτ (4.60) i=1

and

Rn (β) =

n X

[xi − µ(ψ(β τ Zi ))]ψ ′′ (β τ Zi )ti Zi Ziτ .

Var

i=1

Then

and

∂ log ℓ(θ) ∂β

= Mn (β)/φ

∂ 2 log ℓ(θ) = [Rn (β) − Mn (β)]/φ. ∂β∂β τ

(4.61)

(4.62)

(4.63)

Consider first the simple case of canonical g. Then ψ ′′ ≡ 0 and Rn ≡ 0. If Mn (β) is positive definite for all β, then − log ℓ(θ) is strictly convex in β for any fixed φ and, therefore, βˆ is the unique MLE of β. For the case of noncanonical g, Rn (β) 6= 0 and βˆ is not necessarily an MLE. If Rn (β) is dominated by Mn (β) (i.e., [Mn (β)]−1/2 Rn (β)[Mn (β)]−1/2 → 0 in some sense), then − log ℓ(θ) is convex and βˆ is an MLE for large n; see more details in the proof of Theorem 4.18 in §4.5.2. Example 4.36. Consider the GLM (4.55) with ζ(η) = η 2 /2, η ∈ R. If g in (4.57) is the canonical link, then the model is the same as (3.24) with independent εi ’s distributed as N (0, φi ). If (4.58) holds with ti ≡ 1, then (4.59) is exactly the same as equation (3.27). If Z is of full rank, then Mn (β) = Z τ Z is positive definite. Thus, we have shown that the LSE βˆ given by (3.28) is actually the unique MLE of β. Suppose now that g is noncanonical but (4.58) still holds with ti ≡ 1. Then the model reduces to the one with independent Xi ’s and Xi = N g −1 (β τ Zi ), φ , i = 1, ..., n. (4.64)

283

4.4. The Method of Maximum Likelihood

This type of model is called a nonlinear regression model (with normal errors) and an MLE of β under this model is also called a nonlinear LSE, since maximizing the log-likelihood is equivalent to minimizing the sum of P squares ni=1 [Xi −g −1 (β τ Zi )]2 . Under certain conditions the matrix Rn (β) is dominated by Mn (β) and an MLE of β exists. More details can be found in §4.5.2. Example 4.37 (The Poisson model). Consider the GLM (4.55) with ζ(η) = eη , η ∈ R. If φi ≡ 1, then Xi has the Poisson distribution with mean eηi . Assume that (4.58) holds. Under the canonical link g(t) = log t, Mn (β) =

n X

eβ

τ

Zi

ti Zi Ziτ ,

i=1

√ √ τ which is positive definite if inf i eβ Zi > 0 and the matrix ( t1 Z1 , ..., tn Zn ) is of full rank. There is one noncanonical link that deserves attention. Suppose that ′ 2 ′′ we Pn choose aτ link function so that [ψ (t)] ζ (ψ(t)) ≡ 1. Then Mn (β) ≡ i=1 ti Zi Zi does not depend on β. In §4.5.2 it is shown that the asymptotic variance of the MLE βˆ is φ[Mn (β)]−1 . The fact that Mn (β) does not depend on β makes the estimation of the asymptotic variance (and, thus, statistical inference) easy. Under the Poisson model, ζ ′′ (t) = et and, therefore, we need to solve the differential equation [ψ ′ (t)]2 eψ(t) = 1. A solution √ is ψ(t) = 2 log(t/2), which gives the link function g(µ) = 2 µ. In a GLM, an MLE βˆ usually does not have an analytic form. A numerical method such as the Newton-Raphson or the Fisher-scoring method has to be applied. Using the Newton-Raphson method, we have the following iteration procedure: βˆ(t+1) = βˆ(t) − [Rn (βˆ(t) ) − Mn (βˆ(t) )]−1 sn (βˆ(t) ),

t = 0, 1, ...,

where sn (β) = φ∂ log ℓ(θ)/∂β. Note that E[Rn (β)] = 0 if β is the true parameter value and xi is replaced by Xi . This means that the Fisherscoring method uses the following iteration procedure: βˆ(t+1) = βˆ(t) + [Mn (βˆ(t) )]−1 sn (βˆ(t) ),

t = 0, 1, ....

If the canonical link is used, then the two methods are identical.

4.4.3 Quasi-likelihoods and conditional likelihoods We now introduce two variations of the method of using likelihoods.

284

4. Estimation in Parametric Models

Consider a GLM (4.55)-(4.57). Assumption (4.58) is often unrealistic in applications. If there is no restriction on φi ’s, however, there are too many parameters and an MLE of β may not exist. (Note that assumption (4.58) reduces n nuisance parameters to one.) One way to solve this problem is to assume that φi = h ¯ (Zi , ξ) for some known function h ¯ and unknown parameter vector ξ (which may include β as a subvector). Let θ = (β, ξ). Then we can try to solve the likelihood equation ∂ log ℓ(θ)/∂θ = 0 to obtain an MLE of β and/or ξ. We omit the details, which can be found, for example, in Smyth (1989). Suppose that we do not impose any assumptions on φi ’s but still estimate β by solving s˜n (β) =

n X i=1

{[xi − µ(ψ(β τ Zi ))]ψ ′ (β τ Zi )ti Zi } = 0.

(4.65)

Note that (4.65) is not a likelihood equation unless (4.58) holds. In the special case of Example 4.36 where Xi = N (β τ Zi , φi ), i = 1, ..., n, a solution to (4.65) is simply an LSE of β whose properties are discussed at the end of §3.3.3. Estimating β by solving equation (4.65) is motivated by the following facts. First, if (4.58) does hold, then our estimate is an MLE. Second, if (4.58) is slightly violated, the performance of our estimate is still nearly the same as that of an MLE under assumption (4.58) (see the discussion of robustness at the end of §3.3.3). Finally, estimators obtained by solving (4.65) usually have good asymptotic properties. As a special case of a general result in §5.4, a solution to (4.65) is asymptotically normal under some regularity conditions. In general, an equation such as (4.65) is called a quasi-likelihood equation if and only if it is a likelihood equation when certain assumptions hold. The “likelihood” corresponding to a quasi-likelihood equation is called quasilikelihood and a maximum of the quasi-likelihood is then called a maximum quasi-likelihood estimate (MQLE). Thus, a solution to (4.65) is an MQLE. Note that (4.65) is a likelihood equation if and only if both (4.55) and (4.58) hold. The LSE (§3.3) without normality assumption on Xi ’s is a simple example of an MQLE without (4.55). Without assumption (4.55), the model under consideration is usually nonparametric and, therefore, the MQLE’s are studied in §5.4. While the quasi-likelihoods are used to relax some assumptions in our models, the conditional likelihoods discussed next are used mainly in cases where MLE’s are difficult to compute. We consider two cases. In the first case, θ = (θ1 , θ2 ), θ1 is the main parameter vector of interest, and θ2 is a nuisance parameter vector. Suppose that there is a statistic T2 (X) that is sufficient for θ2 for each fixed θ1 . By the sufficiency, the conditional distribution of X given T2 does not depend on θ2 . The likelihood function

285

4.4. The Method of Maximum Likelihood

corresponding to the conditional p.d.f. of X given T2 is called the conditional likelihood function. A conditional MLE of θ1 can then be obtained by maximizing the conditional likelihood function. This method can be applied to the case where the dimension of θ is considerably larger than the dimension of θ1 so that computing the unconditional MLE of θ is much more difficult than computing the conditional MLE of θ1 . Note that the conditional MLE’s are usually different from the unconditional MLE’s. As a more specific example, suppose that X has a p.d.f. in an exponential family: fθ (x) = exp{θ1τ T1 (x) + θ2τ T2 (x) − ζ(θ)}h(x). Then T2 is sufficient for θ2 for any given θ1 . Problems of this type are from comparisons of two binomial distributions or two Poisson distributions (Exercises 119-120). The second case is when our sample X = (X1 , ..., Xn ) follows a firstorder autoregressive time series model: Xt − µ = ρ(Xt−1 − µ) + εt ,

t = 2, ..., n,

where µ ∈ R and ρ ∈ (−1, 1) are unknown and εi ’s are i.i.d. from N (0, σ 2 ) with an unknown σ 2 > 0. This model is often a satisfactory representation of the error time series in economic models, and is one of the simplest and most heavily used models in time series analysis (Fuller, 1996). Let θ = (µ, ρ, σ2 ). The log-likelihood function is n n 1 log(2π) − log σ 2 + log(1 − ρ2 ) 2 ( 2 2 ) n X 1 2 2 2 − 2 (x1 − µ) (1 − ρ ) + [xt − µ − ρ(xt−1 − µ)] . 2σ t=2

log ℓ(θ) = −

The computation of the MLE is greatly simplified if we consider the conditional likelihood given X1 = x1 : log ℓ(θ|x1 ) = −

n n−1 n−1 1 X log(2π)− log σ 2 − 2 [xt −µ−ρ(xt−1 −µ)]2 . 2 2 2σ t=2

Let (¯ x−1 , x ¯0 ) = (n − 1)−1 ρˆ =

n X t=2

Pn

t=2 (xt−1 , xt ).

If

X n (xt − x¯0 )(xt−1 − x ¯−1 ) (xt−1 − x ¯−1 )2 t=2

is between −1 and 1, then it is the conditional MLE of ρ and the conditional MLE’s of µ and σ 2 are, respectively, ¯−1 )/(1 − ρˆ) µ ˆ = (¯ x0 − ρˆx

286

4. Estimation in Parametric Models

and

n

σ ˆ2 =

1 X [xt − x ¯0 − ρˆ(xt−1 − x ¯−1 )]2 . n − 1 t=2

Obviously, the result can be extended to the case where X follows a pth-order autoregressive time series model: Xt − µ = ρ1 (Xt−1 − µ) + · · · + ρp (Xt−p − µ) + εt ,

t = p + 1, ..., n, (4.66)

where ρj ’s are unknown parameters satisfying the constraint that the roots (which may be complex) of the polynomial xp − ρ1 xp−1 − · · · − ρp = 0 are less than one in absolute value (exercise). Some other likelihood based methods are introduced in §5.1.4. Although they can also be applied to parametric models, the methods in §5.1.4 are more useful in nonparametric models.

4.5 Asymptotically Efficient Estimation In this section, we consider asymptotic optimality of point estimators in parametric models. We use the asymptotic mean squared error (amse, see §2.5.2) or its multivariate generalization to assess the performance of an estimator. Reasons for considering asymptotics have been discussed in §2.5. We focus on estimators that are asymptotically normal, since this covers the majority of cases. Some cases of asymptotically nonnormal estimators are studied in Exercises 111-114 in §4.6.

4.5.1 Asymptotic optimality Let {θˆn } be a sequence of estimators of θ based on a sequence of samples {X = (X1 , ..., Xn ) : n = 1, 2, ...} whose distributions are in a parametric family indexed by θ. Suppose that as n → ∞, [Vn (θ)]−1/2 (θˆn − θ) →d Nk (0, Ik ),

(4.67)

where, for each n, Vn (θ) is a k × k positive definite matrix depending on θ. If θ is one-dimensional (k = 1), then Vn (θ) is the asymptotic variance as well as the amse of θˆn (§2.5.2). When k > 1, Vn (θ) is called the asymptotic covariance matrix of θˆn and can be used as a measure of asymptotic performance of estimators. If θˆjn satisfies (4.67) with asymptotic covariance matrix Vjn (θ), j = 1, 2, and V1n (θ) ≤ V2n (θ) (in the sense that V2n (θ) − V1n (θ) is nonnegative definite) for all θ ∈ Θ, then θˆ1n is said to be asymptotically more efficient than θˆ2n . Of course, some sequences of estimators are

4.5. Asymptotically Efficient Estimation

287

not comparable under this criterion. Also, since the asymptotic covariance matrices are unique only in the limiting sense, we have to make our comparison based on their limits. When Xi ’s are i.i.d., Vn (θ) is usually of the form n−δ V (θ) for some δ > 0 (= 1 in the majority of cases) and a positive definite matrix V (θ) that does not depend on n. Note that (4.67) implies that θˆn is an asymptotically unbiased estimator of θ. If Vn (θ) = Var(θˆn ), then, under some regularity conditions, it follows from Theorem 3.3 that Vn (θ) ≥ [In (θ)]−1 , (4.68) where, for every n, In (θ) is the Fisher information matrix (see (3.5)) for X of size n. (Note that (4.68) holds if and only if lτ Vn (θ)l ≥ lτ [In (θ)]−1 l for every l ∈ Rk .) Unfortunately, when Vn (θ) is an asymptotic covariance matrix, (4.68) may not hold (even in the limiting sense), even if the regularity conditions in Theorem 3.3 are satisfied.

Example 4.38 (Hodges). Let X1 , ..., Xn be i.i.d. from N (θ, 1), θ ∈ R. Then In (θ) = n. Define ¯ ¯ ≥ n−1/4 X |X| θˆn = ¯ ¯ tX |X| < n−1/4 , where t is a fixed constant. By Proposition 3.2, all conditions in Theorem 3.3 are satisfied. It can be shown (exercise) that (4.67) holds with Vn (θ) = V (θ)/n, where V (θ) = 1 if θ 6= 0 and V (θ) = t2 if θ = 0. If t2 < 1, (4.68) does not hold when θ = 0. However, the following result, due to Le Cam (1953), shows that (4.68) holds for i.i.d. Xi ’s except for θ in a set of Lebesgue measure 0. Theorem 4.16. Let X1 , ..., Xn be i.i.d. from a p.d.f. fθ w.r.t. a σ-finite measure ν on (R, B), where θ ∈ Θ and Θ is an open set in Rk . Suppose that for every x in the range of X1 , fθ (x) is twice continuously differentiable in θ and satisfies Z Z ∂ ∂ ψθ (x)dν = ψθ (x)dν ∂θ ∂θ for ψθ (x) = fθ (x) and = ∂fθ (x)/∂θ; the Fisher information matrix τ ∂ ∂ I1 (θ) = E log fθ (X1 ) log fθ (X1 ) ∂θ ∂θ is positive definite; and for any given θ ∈ Θ, there exists a positive number cθ and a positive function hθ such that E[hθ (X1 )] < ∞ and

2

∂ log fγ (x)

sup (4.69)

∂γ∂γ τ ≤ hθ (x) γ:kγ−θk [I1 (θ)]1/2 , Z Pθn (Kn (X, θ) ≤ t) = ℓ(θn )dν × · · · × dν Kn (x,θ)≤t

=

Z

Kn (x,θ)≤t

= e−I1 (θ)/2 =e

Z

ℓ(θn ) dPθ (x) ℓ(θ)

e[I1 (θ)]

Kn (x,θ)≤t Z t −I1 (θ)/2 [I1 (θ)]1/2 z

e

−∞ Z t

1/2

1/2

Kn (x,θ)

dPθ (x)

dHn (z)

e[I1 (θ)] z dΦ(z) + o(1) = e−I1 (θ)/2 −∞ = Φ t − [I1 (θ)]1/2 + o(1),

4.5. Asymptotically Efficient Estimation

289

where Hn denotes the distribution of Kn (X, θ) and the next to last equality follows from (4.70) and the dominated convergence theorem. This result and result (4.71) imply that there is a sequence {nj } such that for j = 1, 2, ..., Pθnj (θˆnj ≤ θnj ) < Pθnj (Knj (X, θ) ≤ t). (4.72) By the Neyman-Pearson lemma (Theorem 6.1 in §6.1.1), we conclude that (4.72) implies that for j = 1, 2, ..., Pθ (θˆnj ≤ θnj ) < Pθ (Knj (X, θ) ≤ t).

(4.73)

(The reader should come back to this after reading §6.1.1.) From (4.70) and (4.67) with Vn (θ) = V (θ)/n, (4.73) implies Φ [V (θ)]−1/2 ≤ Φ(t). Hence [V (θ)]−1/2 ≤ t. Since In (θ) = nI1 (θ) (Proposition 3.1(i)) and t is arbitrary but > [I1 (θ)]1/2 , we conclude that (4.68) holds.

Points at which (4.68) does not hold are called points of superefficiency. Motivated by the fact that the set of superefficiency points is of Lebesgue measure 0 under some regularity conditions, we have the following definition. Definition 4.4. Assume that the Fisher information matrix In (θ) is well defined and positive definite for every n. A sequence of estimators {θˆn } satisfying (4.67) is said to be asymptotically efficient or asymptotically optimal if and only if Vn (θ) = [In (θ)]−1 . Suppose that we are interested in estimating ϑ = g(θ), where g is a differentiable function from Θ to Rp , 1 ≤ p ≤ k. If θˆn satisfies (4.67), then, by Theorem 1.12(i), ϑˆn = g(θˆn ) is asymptotically distributed as Np (ϑ, [∇g(θ)]τ Vn (θ)∇g(θ)). Thus, inequality (4.68) becomes [∇g(θ)]τ Vn (θ)∇g(θ) ≥ [I˜n (ϑ)]−1 , where I˜n (ϑ) is the Fisher information matrix about ϑ contained in X. If p = k and g is one-to-one, then [I˜n (ϑ)]−1 = [∇g(θ)]τ [In (θ)]−1 ∇g(θ) and, therefore, ϑˆn is asymptotically efficient if and only if θˆn is asymptotically efficient. For this reason, in the case of p < k, ϑˆn is considered to be asymptotically efficient if and only if θˆn is asymptotically efficient, and we can focus on the estimation of θ only.

290

4. Estimation in Parametric Models

4.5.2 Asymptotic efficiency of MLE’s and RLE’s We now show that under some regularity conditions, a root of the likelihood equation (RLE), which is a candidate for an MLE, is asymptotically efficient. Theorem 4.17. Assume the conditions of Theorem 4.16. (i) There is a sequence of estimators {θˆn } such that and θˆn →p θ, P sn (θˆn ) = 0 → 1

(4.74)

where sn (γ) = ∂ log ℓ(γ)/∂γ. (ii) Any consistent sequence θ˜n of RLE’s is asymptotically efficient. Proof. (i) Let Bn (c) = {γ : k[In (θ)]1/2 (γ − θ)k ≤ c} for c > 0. Since Θ is open, for each c > 0, Bn (c) ⊂ Θ for sufficiently large n. Since Bn (c) shrinks to {θ} as n → ∞, the existence of θˆn satisfying (4.74) is implied by the fact that for any ǫ > 0, there exists c > 0 and n0 > 1 such that P log ℓ(γ) − log ℓ(θ) < 0 for all γ ∈ ∂Bn (c) ≥ 1 − ǫ, n ≥ n0 , (4.75)

where ∂Bn (c) is the boundary of Bn (c). (For a proof of the measurability of θˆn , see Serfling (1980, pp. 147-148).) For γ ∈ ∂Bn (c), the Taylor expansion gives log ℓ(γ) − log ℓ(θ) = cλτ [In (θ)]−1/2 sn (θ)

(4.76)

+ (c2 /2)λτ [In (θ)]−1/2 ∇sn (γ ∗ )[In (θ)]−1/2 λ,

where λ = [In (θ)]1/2 (γ − θ)/c satisfying kλk = 1, ∇sn (γ) = ∂sn (γ)/∂γ, and γ ∗ lies between γ and θ. Note that E

k∇sn (γ) − ∇sn (θ)k k∇sn (γ ∗ ) − ∇sn (θ)k ≤ E max n γ∈Bn (c) n

2

∂ log fγ (X1 ) ∂ 2 log fθ (X1 )

≤ E max −

γ∈Bn (c) ∂γ∂γ τ ∂θ∂θτ → 0, (4.77)

which follows from (a) ∂ 2 log fγ (x)/∂γ∂γ τ is continuous in a neighborhood of θ for any fixed x; (b) Bn (c) shrinks to {θ}; and (c) for sufficiently large n,

2

∂ log fγ (X1 ) ∂ 2 log fθ (X1 )

≤ 2hθ (X1 ) max −

γ∈Bn (c) ∂γ∂γ τ ∂θ∂θτ under condition (4.69). By the SLLN (Theorem 1.13) and Proposition 3.1, n−1 ∇sn (θ) →a.s. −I1 (θ) (i.e., kn−1 ∇sn (θ) + I1 (θ)k →a.s. 0). These results, together with (4.76), imply that log ℓ(γ) − log ℓ(θ) = cλτ [In (θ)]−1/2 sn (θ) − [1 + op (1)]c2 /2.

(4.78)

4.5. Asymptotically Efficient Estimation

291

Note that maxλ {λτ [In (θ)]−1/2 sn (θ)} = k[In (θ)]−1/2 sn (θ)k. Hence, (4.75) follows from (4.78) and P k[In (θ)]−1/2 sn (θ)k < c/4 ≥ 1 − (4/c)2 Ek[In (θ)]−1/2 sn (θ)k2 = 1 − k(4/c)2 ≥ 1−ǫ

by choosing c sufficiently large. This completes the proof of (i). (ii) Let Aǫ = {γ : kγ − θk ≤ ǫ} for ǫ > 0. Since Θ is open, Aǫ ⊂ Θ for sufficiently small ǫ. Let {θ˜n } be a sequence of consistent RLE’s, i.e., P (sn (θ˜n ) = 0 and θ˜n ∈ Aǫ ) → 1 for any ǫ > 0. Hence, we can focus on the set on which sn (θ˜n ) = 0 and θ˜n ∈ Aǫ . Using the mean-value theorem for vector-valued functions, we obtain that Z 1 −sn (θ) = ∇sn θ + t(θ˜n − θ) dt (θ˜n − θ). 0

Note that

Z 1

1 k∇sn (γ) − ∇sn (θ)k ˜

. ∇sn θ + t(θn − θ) dt − ∇sn (θ) ≤ max

γ∈A n 0 n ǫ

Using the argument in proving (4.77) and the fact that P (θ˜n ∈ Aǫ ) → 1 for arbitrary ǫ > 0, we obtain that

Z 1

1 ˜

→p 0. ∇s − θ) dt − ∇s (θ) θ + t( θ n n n

n 0 Since n−1 ∇sn (θ) →a.s. −I1 (θ) and In (θ) = nI1 (θ),

−sn (θ) = −In (θ)(θ˜n − θ) + op kIn (θ)(θ˜n − θ)k .

√ This and Slutsky’s theorem (Theorem 1.11) imply that n(θ˜n − θ) has the same asymptotic distribution as √ n[In (θ)]−1 sn (θ) = n−1/2 [I1 (θ)]−1 sn (θ) →d Nk 0, [I1 (θ)]−1

by the CLT (Corollary 1.2), since Var(sn (θ)) = In (θ).

Theorem 4.17(i) shows the asymptotic existence of a sequence of consistent RLE’s, and Theorem 4.17(ii) shows the asymptotic efficiency of any sequence of consistent RLE’s. However, for a given sequence of RLE’s, its consistency has to be checked unless the RLE’s are unique for sufficiently large n, in which case the consistency of the RLE’s is guaranteed by Theorem 4.17(i).

292

4. Estimation in Parametric Models

RLE’s are not necessarily MLE’s. We still have to use the techniques discussed in §4.4 to check whether an RLE is an MLE. However, according to Theorem 4.17, when a sequence of RLE’s is consistent, then it is asymptotically efficient and, therefore, we may not need to search for MLE’s, if asymptotic efficiency is the only criterion to select estimators. The method of estimating θ by solving sn (γ) = 0 over γ ∈ Θ is called scoring and the function sn (γ) is called the score function. Example 4.39. Suppose that Xi has a distribution in a natural exponential family, i.e., the p.d.f. of Xi is fη (xi ) = exp{η τ T (xi ) − ζ(η)}h(xi ).

(4.79)

Since ∂ 2 log fη (xi )/∂η∂η τ = −∂ 2 ζ(η)/∂η∂η τ , condition (4.69) is satisfied. From Proposition 3.2, other conditions in Theorem 4.16 are also satisfied. For i.i.d. Xi ’s, n X ∂ζ(η) . T (Xi ) − sn (η) = ∂η i=1

Pn If θˆn = n−1 i=1 T (Xi ) ∈ Θ, the range of θ = g(η) = ∂ζ(η)/∂η, then θˆn is a unique RLE of θ, which is also a unique MLE of θ since ∂ 2 ζ(η)/∂η∂η τ = Var(T (Xi )) is positive definite. Also, η = g −1 (θ) exists and a unique RLE (MLE) of η is ηˆn = g −1 (θˆn ). However, θˆn may not be in Θ and the previous argument fails (e.g., Example 4.29). What Theorem 4.17 tells us in this case is that as n → ∞, P (θˆn ∈ Θ) → 1 and, therefore, θˆn (or ηˆn ) is the unique asymptotically efficient RLE (MLE) of θ (or η) in the limiting sense. In an example like this we can directly show that P (θˆn ∈ Θ) → 1, using the fact that θˆn →a.s. E[T (X1 )] = g(η) (the SLLN).

The next theorem provides a similar result for the MLE or RLE in the GLM (§4.4.2). Theorem 4.18. Consider the GLM (4.55)-(4.58) with ti ’s in a fixed interval (t0 , t∞ ), 0 < t0 ≤ t∞ < ∞. Assume that the range of the unknown parameter β in (4.57) is an open subset of Rp ; at the true parameter value β, 0 < inf i ϕ(β τ Zi ) ≤ supi ϕ(β τ Zi ) < ∞, where ϕ(t) = [ψ ′ (t)]2 ζ ′′ (ψ(t)); as n → ∞, maxi≤n Ziτ (Z τ Z)−1 Zi → 0 and λ− [Z τ Z] → ∞, where Z is the n × p matrix whose ith row is the vector Zi and λ− [A] is the smallest eigenvalue of the matrix A. (i) There is a unique sequence of estimators {βˆn } such that P sn (βˆn ) = 0 → 1

and

βˆn →p β,

(4.80)

293

4.5. Asymptotically Efficient Estimation

where sn (γ) is the score function defined to be the left-hand side of (4.59) with γ = β. (ii) Let In (β) = Var(sn (β)). Then [In (β)]1/2 (βˆn − β) →d Np (0, Ip ).

(4.81)

(iii) If φ in (4.58) is known or the p.d.f. in (4.55) indexed by θ = (β, φ) satisfies the conditions for fθ in Theorem 4.16, then βˆn is asymptotically efficient. Proof. (i) The proof of the existence of βˆn satisfying (4.80) is the same as that of Theorem 4.17(i) with θ = β, except that we need to show

max [In (β)]−1/2 ∇sn (γ)[In (β)]−1/2 + Ip →p 0, γ∈Bn (c)

where Bn (c) = {γ : k[In (β)]1/2 (γ − β)k ≤ c}. From (4.62) and (4.63), In (β) = Mn (β)/φ and ∇sn (γ) = [Rn (γ) − Mn (γ)]/φ, where Mn (γ) and Rn (γ) are defined by (4.60)-(4.61) with γ = β. Hence, it suffices to show that for any c > 0,

max [Mn (β)]−1/2 [Mn (γ) − Mn (β)][Mn (β)]−1/2 → 0 (4.82) γ∈Bn (c)

and

max [Mn (β)]−1/2 Rn (γ)[Mn (β)]−1/2 →p 0.

γ∈Bn (c)

The left-hand side of (4.82) is bounded by √ 1 − ϕ(γ τ Zi )/ϕ(β τ Zi ) , p max γ∈Bn (c),i≤n

which converges to 0 since ϕ is continuous and, for γ ∈ Bn (c), |γ τ Zi − β τ Zi |2 = |(γ − β)τ [In (β)]1/2 [In (β)]−1/2 Zi |2 ≤ k[In (β)]1/2 (γ − β)k2 k[In (β)]−1/2 Zi k2 ≤ c2 max Ziτ [In (β)]−1 Zi i≤n

−1 ≤ c φ t0 inf ϕ(β τ Zi ) max Ziτ (Z τ Z)−1 Zi 2

i

i≤n

→0

under the assumed conditions. This proves (4.82). Let ei = Xi − µ(ψ(β τ Zi )), Un (γ) =

n X i=1

[µ(ψ(β τ Zi )) − µ(ψ(γ τ Zi ))]ψ ′′ (γ τ Zi )ti Zi Ziτ ,

(4.83)

294

4. Estimation in Parametric Models

Vn (γ) =

n X i=1

and

ei [ψ ′′ (γ τ Zi ) − ψ ′′ (β τ Zi )]ti Zi Ziτ ,

Wn (β) =

n X

ei ψ ′′ (β τ Zi )ti Zi Ziτ .

i=1

Then Rn (γ) = Un (γ) + Vn (γ) + Wn (β). Using the same argument as that in proving (4.82), we can show that

max [Mn (β)]−1/2 Un (γ)[Mn (β)]−1/2 → 0. γ∈Bn (c)

Note that [Mn (β)]−1/2 Vn (γ)[Mn (β)]−1/2 is bounded by the product of [Mn (β)]−1/2

n X i=1

|ei |ti Zi Ziτ [Mn (β)]−1/2 = Op (1)

and max

γ∈Bn (c),i≤n

′′ τ ψ (γ Zi ) − ψ ′′ (β τ Zi ) ,

which can be shown to be o(1) using the same argument as that in proving (4.82). Hence,

max [Mn (β)]−1/2 Vn (γ)[Mn (β)]−1/2 →p 0 γ∈Bn (c)

and (4.83) follows from

[Mn (β)]−1/2 Wn (β)[Mn (β)]−1/2 →p 0.

To show this result, we apply Theorem 1.14(ii). Since E(ei ) = 0 and ei ’s are independent, it suffices to show that n X 1+δ E ei ψ ′′ (β τ Zi )ti Ziτ [Mn (β)]−1 Zi →0

(4.84)

i=1

for some δ ∈ (0, 1). Note that supi E|ei |1+δ < ∞. Hence, there is a constant C > 0 such that the left-hand side of (4.84) is bounded by C

n X τ τ −1 1+δ Zi (Z Z) Zi ≤ pC max |Ziτ (Z τ Z)−1 Zi |δ → 0. i=1

i≤n

Hence, (4.84) follows from Theorem 1.14(ii). This proves (4.80). The uniqueness of βˆn follows from (4.83) and the fact that Mn (γ) is positive definite in a neighborhood of β. This completes the proof of (i).

4.5. Asymptotically Efficient Estimation

295

(ii) The proof of (ii) is very similar to that of Theorem 4.17(ii). Using the results in the proof of (i) and Taylor’s expansion, we can establish (exercise) that [In (β)]1/2 (βˆn − β) = [In (β)]−1/2 sn (β) + op (1). (4.85) Using the CLT (e.g., Corollary 1.3) and Theorem 1.9(iii), we can show (exercise) that [In (β)]−1/2 sn (β) →d Np (0, Ip ). (4.86) Result (4.81) follows from (4.85)-(4.86) and Slutsky’s theorem. (iii) The result is obvious if φ is known. When φ is unknown, it follows from (4.59) that ∂ ∂ log ℓ(θ) sn (β) =− . ∂φ ∂β φ Since E[sn (β)] = 0, the Fisher information about θ = (β, φ) is 2 ∂ log ℓ(θ) 0 In (β) , In (β, φ) = −E = 0 I˜n (φ) ∂θ∂θτ where I˜n (φ) is the Fisher information about φ. The result then follows from (4.81) and the discussion in the end of §4.5.1.

4.5.3 Other asymptotically efficient estimators To study other asymptotically efficient estimators, we start with MRIE’s in location-scale families. Since MLE’s and RLE’s are invariant (see Exercise 109 in §4.6), MRIE’s are often asymptotically efficient; see, for example, Stone (1974). Assume the conditions in Theorem 4.16 and let sn (γ) be the score func(0) tion. Let θˆn be an estimator of θ that may not be asymptotically efficient. The estimator θˆn(1) = θˆn(0) − [∇sn (θˆn(0) )]−1 sn (θˆn(0) ) (4.87) is the first iteration in computing an MLE (or RLE) using the Newton(0) Raphson iteration method with θˆn as the initial value (see (4.53)) and, (1) therefore, is called the one-step MLE. Without any further iteration, θˆn (1) can be used as a numerical approximation to an MLE or RLE; and θˆn is asymptotically efficient under some conditions, as the following result shows. Theorem 4.19. Assume that the conditions in Theorem 4.16 hold and √ (0) that θˆn is n-consistent for θ (Definition 2.10). (1) (i) The one-step MLE θˆn is asymptotically efficient. (ii) The one-step MLE obtained by replacing ∇sn (γ) in (4.87) with its

296

4. Estimation in Parametric Models

expected value, −In (γ) (the Fisher-scoring method), is asymptotically efficient. √ (0) (0) Proof. Since θˆn is n-consistent, we can focus on the event θˆn ∈ Aǫ = {γ : kγ − θk ≤ ǫ} for a sufficiently small ǫ such that Aǫ ⊂ Θ. From the mean-value theorem, Z 1 sn (θˆn(0) ) = sn (θ) + ∇sn θ + t(θˆn(0) − θ) dt (θˆn(0) − θ). 0

Substituting this into (4.87) we obtain that

θˆn(1) − θ = −[∇sn (θˆn(0) )]−1 sn (θ) + [Ik − Gn (θˆn(0) )](θˆn(0) − θ), where Gn (θˆn(0) ) = [∇sn (θˆn(0) )]−1

Z

1

0

∇sn θ + t(θˆn(0) − θ) dt.

(0) From (4.77), k[In (θ)] [∇sn (θˆn )]−1 [In (θ)]1/2 + Ik k →p 0. Using an argument similar to those in the proofs of (4.77) and (4.82), we can show that √ (0) (0) kGn (θˆn ) − Ik k →p 0. These results and the fact that n(θˆn − θ) = Op (1) imply √ √ (1) n(θˆn − θ) = n[In (θ)]−1 sn (θ) + op (1). 1/2

This proves (i). The proof for (ii) is similar. Example 4.40. Let X1 , ..., Xn be i.i.d. from the Weibull distribution W (θ, 1), where θ > 0 is unknown. Note that n

sn (θ) =

n

X n X + log Xi − Xiθ log Xi θ i=1 i=1

and

n

∇sn (θ) = −

X n − Xiθ (log Xi )2 . 2 θ i=1

Hence, the one-step MLE of θ is " # (0) ˆn(0) (Pn log Xi − Pn X θˆn log Xi ) n + θ i=1 i=1 i . θˆn(1) = θˆn(0) 1 + Pn ˆ(0) (0) n + (θˆn )2 i=1 Xiθn (log Xi )2

Usually one can use a moment estimator (§3.5.2) as the initial estimator (0) ¯ = θˆn . In this example, a moment estimator of θ is the solution of X −1 Γ(θ + 1). Results similar to that in Theorem 4.19 can be obtained in non-i.i.d. cases, for example, the GLM discussed in §4.4.2 (exercise); see also §5.4.

297

4.5. Asymptotically Efficient Estimation

As we discussed in §4.1.3, Bayes estimators are usually consistent. The next result, due to Bickel and Yahav (1969) and Ibragimov and Has’minskii (1981), states that Bayes estimators are asymptotically efficient when Xi ’s are i.i.d. Theorem 4.20. Assume the conditions of Theorem 4.16. Let π(γ) be a prior p.d.f. (which may be improper) w.r.t. the Lebesgue measure on Θ and pn (γ) be the posterior p.d.f., given X1 , ..., Xn , n = 1, 2, .... Assume that Rthere exists an n0 such R that pn0 (γ) is continuous and positive for all γ ∈ Θ, pn0 (γ)dγ = 1 and kγkpn0 (γ)dγ < ∞. Suppose further that, for any ǫ > 0, there exists a δ > 0 such that ! log ℓ(γ) − log ℓ(θ) lim P > −δ = 0 (4.88) sup n→∞ n kγ−θk≥ǫ and lim P

n→∞

k∇sn (γ) − ∇sn (θ)k ≥ǫ sup n kγ−θk≤δ

!

= 0,

(4.89)

where ℓ(γ) is the likelihood function and s√ n (γ) is the score function. (i) Let p∗n (γ) be the posterior p.d.f. of n(γ − Tn ), where Tn = θ + [In (θ)]−1 sn (θ) and θ is the true parameter value, and let ψ(γ) be the p.d.f. of Nk (0, [I1 (θ)]−1 ). Then Z (1 + kγk) p∗n (γ) − ψ(γ) dγ →p 0. (4.90)

(ii) The Bayes estimator of θ under the squared error loss is asymptotically efficient.

The proof of Theorem 4.20 is lengthy and is omitted; see Lehmann (1983, §6.7) for a proof of the case of univariate θ. A number of conclusions can be drawn from Theorem 4.20. First, result (4.90) shows that the posterior p.d.f. is approximately normal with mean θ + [In (θ)]−1 sn (θ) and covariance matrix [In (θ)]−1 . This result is useful in Bayesian computation; see Berger (1985, §4.9.3). Second, (4.90) shows that the posterior distribution and its first-order moments converge to the degenerate distribution at θ and its first-order moments, which implies the consistency and asymptotic unbiasedness of Bayes estimators such as the posterior means. Third, the Bayes estimator under the squared error loss is asymptotically efficient, which provides an additional support for the early suggestion that the Bayesian approach is a useful method for generating estimators. Finally, the results hold regardless of the prior being used, indicating that the effect of the prior declines as n increases.

298

4. Estimation in Parametric Models

In addition to the regularity conditions in Theorem 4.16, Theorem 4.20 requires two more nontrivial regularity conditions, (4.88) and (4.89). Let us verify these conditions for natural exponential families (Example 4.39), i.e., Xi ’s are i.i.d. with p.d.f. (4.79). Since ∇sn (η) = −n∂ 2 ζ(η)/∂η∂η τ , (4.89) follows from the continuity of the second-order derivatives of ζ. To show (4.88), consider first the case of univariate η. Without loss of generality, we assume that γ > η. Note that log ℓ(γ) − log ℓ(η) ζ(γ) − ζ(η) ′ ′ ¯ = T − ζ (η) + ζ (η) − (γ − η), (4.91) n γ−η

where T¯ is the average of T (Xi )’s. Since ζ(γ) is strictly convex, γ > η implies ζ ′ (η) < [ζ(γ) − ζ(η)]/(γ − η). Also, T¯ →a.s. ζ ′ (η). Hence, with probability tending to 1, the factor in front of (γ − η) on the right-hand side of (4.91) is negative. Then (4.88) holds with ǫ ζ(γ) − ζ(η) δ= inf − ζ ′ (η) . 2 γ≥η+ǫ γ −η To show how to extend this to multivariate η, consider the case of bivariate η. Let ηj , γj , and ξj be the jth components of η, γ, and T¯ − ∇ζ(η), respectively. Assume γ1 > η1 and γ2 > η2 . Let ζj′ be the derivative of ζ w.r.t. the jth component of η. Then the left-hand side of (4.91) is the sum of (γ1 − η1 )ξ1 − [ζ(η1 , γ2 ) − ζ(η1 , η2 ) − (γ2 − η2 )ζ2′ (η1 , η2 )] and

(γ2 − η2 )ξ2 − [ζ(γ1 , γ2 ) − ζ(η1 , γ2 ) − (γ1 − η1 )ζ1′ (η1 , η2 )],

where the last quantity is bounded by

(γ2 − η2 )ξ2 − [ζ(γ1 , γ2 ) − ζ(η1 , γ2 ) − (γ1 − η1 )ζ1′ (η1 , γ2 )], since ζ1′ (η1 , η2 ) ≤ ζ1′ (η1 , γ2 ). The rest of the proof is the same as the case of univariate η. When Bayes estimators have explicit forms under a specific prior, it is usually easy to prove the asymptotic efficiency of the Bayes estimators directly. For instance, in Example 4.7, the Bayes estimator of θ is −1 ¯ + γ −1 ¯ nX ¯ + γ − (α − 1)X = X ¯ + O 1 a.s., =X n+α−1 n+α−1 n

¯ is the MLE of θ. Hence the Bayes estimator is asymptotically where X efficient by Slutsky’s theorem. A similar result can be obtained for the Bayes estimator δt (X) in Example 4.7. Theorem 4.20, however, is useful in cases where Bayes estimators do not have explicit forms and/or the prior is not specified clearly. One such example is the problem in Example 4.40 (Exercises 153 and 154).

299

4.6. Exercises

4.6 Exercises 1. Show that the priors in the following cases are conjugate priors: (a) X1 , ..., Xn are i.i.d. from Nk (θ, Ik ), θ ∈ Rk , and Π = Nk (µ0 , Σ0 ) (Normal family); (b) X1 , ..., Xn are i.i.d. from the binomial distribution Bi(θ, k), θ ∈ (0, 1), and Π = B(α, β) (Beta family); (c) X1 , ..., Xn are i.i.d. from the uniform distribution U (0, θ), θ > 0, and Π = P a(a, b) (Pareto family); (d) X1 , ..., Xn are i.i.d. from the exponential distribution E(0, θ), θ > 0, Π = the inverse gamma distribution Γ−1 (α, γ) (a random variable Y has the inverse gamma distribution Γ−1 (α, γ) if and only if Y −1 has the gamma distribution Γ(α, γ)). (e) X1 , ..., Xn are i.i.d. from the exponential distribution E(θ, 1), θ ∈ R, and Π has a Lebesgue p.d.f. b−1 e−a/b eθ/b I(−∞,a) (θ), a ∈ R, b > 0. 2. In Exercise 1, find the posterior mean and variance for each case. 3. Let X1 , ..., Xn be i.i.d. from the N (θ, 1) distribution and let the prior be the double exponential distribution DE(0, 1). Obtain the posterior mean. 4. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (0, θ), where θ > 0 is unknown. Let the prior of θ be the log-normal distribution LN (µ0 , σ02 ), where µ0 ∈ R and σ0 > 0 are known constants. (a) Find the posterior p.d.f. of ϑ = log θ. (b) Find the rth posterior moment of θ. (c) Find a value that maximizes the posterior p.d.f. of θ. 5. Show that if T (X) is a sufficient statistic for θ ∈ Θ, then the Bayes action δ(x) in (4.3) is a function of T (x). ¯ be the sample mean of n i.i.d. observations from N (θ, σ 2 ) with 6. Let X a known σ > 0 and an unknown θ ∈ R. Let π(θ) be a prior p.d.f. w.r.t. a σ-finite measure on R. ¯ = x, is of the form (a) Show that the posterior mean of θ, given X σ 2 d log(p(x)) , n dx ¯ unconditional on θ. where p(x) is the marginal p.d.f. of X, ¯ = x) as a function (b) Express the posterior variance of θ (given X of the first two derivatives of log(p(x)) w.r.t. x. (c) Find explicit expressions for p(x) and δ(x) in (a) when the prior is N (µ0 , σ02 ) with probability 1 − ǫ and a point mass at µ1 with probability ǫ, where µ0 , µ1 , and σ02 are known constants. δ(x) = x +

300

4. Estimation in Parametric Models

7. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). Find the Bayes action w.r.t. the uniform prior on [0, 1] in the problem of estimating p under the loss L(p, a) = (p−a)2 /[p(1−p)]. 8. Consider the estimation of θ in Exercise 41 of §2.6 under the squared error loss. Suppose that the prior of θ is the uniform distribution U (0, 1), the prior of j is P (j = 1) = P (j = 2) = 21 , and the joint prior of (θ, j) is the product probability of the two marginal priors. Show that the Bayes action is δ(x) =

H(x)B(t + 1) + G(t + 1) , H(x)B(t) + G(t)

where x = (x1 , ..., xn ) is the vector of observations, t = x1 + · · · + xn , R1 R1 B(t) = 0 θt (1 − θ)n−t dθ, G(t) = 0 θt e−nθ dθ, and H(x) is a function of x with range {0, 1}. 9. Consider the estimation problem in Example 4.1 R with the loss function L(θ, a) = w(θ)[g(θ) − a]2 , where w(θ) ≥ 0 and Θ w(θ)[g(θ)]2 dΠ < ∞. Show that the Bayes action is R w(θ)g(θ)fθ (x)dΠ . δ(x) = ΘR Θ w(θ)fθ (x)dΠ

10. Let X be a sample from Pθ , θ ∈ Θ ⊂ R. Consider the estimation of θ under the loss L(|θ − a|), where L is an increasing function on [0, ∞). Let π(θ|x) be the posterior p.d.f. of θ given X = x. Suppose that π(θ|x) is symmetric about δ(x) ∈ Θ and that π(θ|x) is nondecreasing for θ ≤ δ(x) and nonincreasing for θ ≥ δ(x). Show that δ(x) is a Bayes action, assuming that all integrals involved are finite. 11. Let X be a sample of size 1 from the geometric distribution G(p) with an unknown p ∈ (0, 1]. Consider the estimation of p with A = [0, 1] and the loss function L(p, a) = (p − a)2 /p. (a) RShow that δ is aR Bayes action w.r.t. Π if and only if δ(x) = 1 − (1 − p)x dΠ(p)/ (1 − p)x−1 dΠ(p), x = 1, 2, .... (b) Let δ0 be a rule such that δ0 (1) = 1/2 and δ0 (x) = 0 for all x > 1. Show that δ0 is a limit of Bayes actions. (c) Let δ0 be a rule such that δ0 (x) = 0 for all x > 1 and δ0 (1) is arbitrary. Show that δ0 is a generalized Bayes action. 12. Let X be a single observation from N (µ, σ 2 ) with a known σ 2 and an unknown µ > 0. Consider the estimation of µ under the squared error loss and the noninformative prior Π = the Lebesgue measure on (0, ∞). Show that the generalized Bayes action when X = x is δ(x) = x + σΦ′ (x/σ)/[1 − Φ(−x/σ)], where Φ is the c.d.f. of the standard normal distribution and Φ′ is its derivative.

301

4.6. Exercises

13. Let X be a sample from Pθ having the p.d.f. h(x) exp{θτ x − ζ(θ)} w.r.t. ν. Let Π be the Lebesgue measure on Θ = Rp . Show that the generalized Bayes action under the loss L(θ, a) = kE(X) − ak2 is δ(x) = x when X = x. 14. Let , ..., Xn be i.i.d. random variables with the Lebesgue p.d.f. p X1−(x−θ) 2 /2 2/πe I(θ,∞) (x), where θ ∈ R is unknown. Find the generalized Bayes action for estimating θ under the squared error loss, when the (improper) prior of θ is the Lebesgue measure on R.

15. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) and π(µ, σ 2 ) = σ −2 I(0,∞) (σ 2 ) be an improper prior for (µ, σ 2 ) w.r.t. the Lebesgue measure on R2 . (a) Show that the posterior p.d.f. of (µ, σ2 ) given x = (x1 , ..., xn ) is π(µ, σ 2 |x) = π1 (µ|σ 2 , x)π2 (σ 2 |x), where π1 (µ|σ 2 , x) is the p.d.f. of 2 N (¯ x, σ 2 /n) and π2 (σP |x) is the p.d.f. of the inverse gamma distribun −1 tion Γ ((n − 1)/2, [ i=1 (xi − x ¯)2 /2]−1 ) (see Exercise 1(d)). x (b) Show that the marginal posterior p.d.f. of µ given x is f µ−¯ , τ P n where τ 2 = ¯)2 /[n(n − 1)] and f is the p.d.f. of the ti=1 (xi − x distribution tn−1 . (c) Obtain the generalized Bayes action for estimating µ/σ under the squared error loss. 16. Consider Example 3.13. Under the squared error loss and the prior 2 with the improper Lebesgue density π(µ1 , ..., µm , σP ) = σ −2 , obtain m −2 the generalized P Bayes action for estimating θ = σ ¯)2 , i=1 ni (µi − µ m −1 where µ ¯=n i=1 ni µi . 17. Let X be a single observation from the Lebesgue p.d.f. e−x+θ I(θ,∞) (x), where θ > 0 is an unknown parameter. Consider the estimation of j θ ∈ (j − 1, j], j = 1, 2, 3, ϑ= 4 θ>3 under the loss L(i, j), 1 ≤ i, j 0 1 1 3

≤ 4, given by the following matrix: 1 1 2 0 2 2 . 2 0 2 3

3 0

When X = 4, find the Bayes action w.r.t. the prior with the Lebesgue p.d.f. e−θ I(0,∞) (θ).

18. (Bayesian hypothesis testing). Let X be a sample from Pθ , where θ ∈ Θ. Let Θ0 ⊂ Θ and Θ1 = Θc0 , the complement of Θ0 . Consider the problem of testing H0 : θ ∈ Θ0 versus H1 : θ ∈ Θ1 under the loss 0 θ ∈ Θi L(θ, ai ) = Ci θ 6∈ Θi ,

302

4. Estimation in Parametric Models

where Ci > 0 are known constants and {a0 , a1 } is the action space. Let Πθ|x be the posterior distribution of θ w.r.t. a prior distribution Π, given X = x. Show that the Bayes action δ(x) = a1 if and only if Πθ|x (Θ1 ) ≥ C1 /(C0 + C1 ). 19. In (b)-(d) of Exercise 1, assume that the parameters in priors are unknown. Using the method of moments, find empirical Bayes actions under the squared error loss. 20. In Example 4.5, assume that both µ0 and σ02 in the prior for µ are unknown. Let the second-stage joint prior for (µ0 , σ02 ) be the product of N (a, v 2 ) and the Lebesgue measure on (0, ∞), where a and v are known. Under the squared error loss, obtain a formula for the hierarchical Bayes action in terms of a one-dimensional integral. 21. Let X1 , ..., Xn be i.i.d. random variables from the uniform distribution U (0, θ), where θ > 0 is unknown. Let π(θ) = bab θ−(b+1) I(a,∞) (θ) be a prior p.d.f. w.r.t. the Lebesgue measure, where b > 1 is known but a > 0 is an unknown hyperparameter. Consider the estimation of θ under the squared error loss. (a) Show that the empirical Bayes method using the method of moments produces the empirical Bayes action δ(ˆ a), where δ(a) = 2(b−1) Pn b+n ˆ = bn i=1 Xi , and X(n) is the largest orb+n−1 max{a, X(n) }, a der statistic. (b) Let h(a) = a−1 I(0,∞) (a) be an improper Lebesgue prior density for a. Obtain explicitly the hierarchical generalized Bayes action. 22. Let X be a sample and δ(X) with any fixed X R = x ∈ A be a Bayes action, where δ is a measurable function and Θ Pθ (A)dΠ = 1. Show that δ(X) is a Bayes rule as defined in §2.3.2. 23. Let X1 , ..., random variables with the Lebesgue p.d.f. p Xn be i.i.d. 2 fθ (x) = 2θ/πe−θx /2 I[0,∞) (x), where θ > 0 is unknown. Let the prior of θ be the gamma distribution Γ(α, γ) with known α and γ. Find the Bayes estimator of fθ (0) and its Bayes risk under the loss function L(θ, a) = (a − θ)2 /θ. 24. Let X be a single observation from N (θ, θ2 ) and consider a prior p.d.f. −1 2 2 πξ (θ) = c(α, µ, τ )|θ|−α e−(θ −µ) /(2τ ) w.r.t. the Lebesgue measure, where ξ = (α, µ, τ ) is a vector of hyperparameters and c(α, µ, τ ) ensures that πξ (θ) is a p.d.f. (a) Identify the constraints on the hyperparameters for πξ (θ) to be a proper prior. (b) Show that the posterior p.d.f. is πξ∗ (θ) for given X = x and identify ξ∗ .

303

4.6. Exercises

(c) Express the Bayes estimator of |θ| and its Bayes risk in terms of the function c and ξ∗ and state any additional constraints needed on the hyperparameters. 25. Let X1 , X2 , ... be i.i.d. from the exponential distribution E(0, 1). Suppose that we observe T = X1 + · · · + Xθ , where θ is an unknown integer ≥ 1. Consider the estimation of θ under the loss function L(θ, a) = (θ − a)2 /θ and the geometric distribution G(p) as the prior for θ, where p ∈ (0, 1) is known. (a) Show that the posterior expected loss is E[L(θ, a)|T = t] = 1 + ξ − 2a + (1 − e−ξ )a2 /ξ, where ξ = (1 − p)t. (b) Find theP Bayes estimator of θ and show that its posterior expected −mξ loss is 1 − ξ ∞ . m=1 e (c) Find the marginal distribution of (1 − p)T , unconditional on θ. (d) Obtain an explicit expression for the Bayes risk of the Bayes estimator in part (b). 26. Prove (ii) and (iii) of Theorem 4.2. 27. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). ¯ is an admissible estimator of p under the loss function (a) Show that X 2 (a − p) /[p(1 − p)]. ¯ is an admissible estimator of p under the squared (b) Show that X error loss. 28. Let X be a sample (of size 1) from N (µ, 1). Consider the estimation of µ under the loss function L(µ, a) = |µ − a|. Show that X is an admissible estimator. 29. In Exercise 1, consider the posterior mean to be the Bayes estimator of the corresponding parameter in each case. (a) Show that the bias of the Bayes estimator converges to 0 if n → ∞. (b) Show that the Bayes estimator is consistent. (c) Discuss whether the Bayes estimator is admissible. 30. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). (a) Obtain the Bayes estimator of p(1 − p) w.r.t. Π = the beta distribution B(α, β) with known α and β, under the squared error loss. (b) Compare the Bayes estimator in part (a) with the UMVUE of p(1 − p). (c) Discuss the bias, consistency, and admissibility of the Bayes estimator in (a).

304

4. Estimation in Parametric Models (d) Let π(p) = [p(1 − p)]−1 I(0,1) (p) be an improper Lebesgue prior density for p. Show that the posterior of p given Xi ’s is a p.d.f. pro¯ ∈ (0, 1). vided that the sample mean X (e) Under the squared error loss, find the generalized Bayes estimator of p(1 − p) w.r.t. the improper prior in (d).

31. Let X be an observation from the negative binomial distribution N B(p, r) with a known r and an unknown p ∈ (0, 1). (a) Under the squared error loss, find Bayes estimators of p and p−1 w.r.t. Π = the beta distribution B(α, β) with known α and β. (b) Show that the Bayes estimators in (a) are consistent as r → ∞. 32. In Example 4.7, show that ¯ is the generalized Bayes estimator of θ w.r.t. the improper (a) X prior dΠ dω = I(0,∞) (ω) and is a limit of Bayes estimators (as α → 1 and γ → ∞); (b) under the squared error loss for estimating θ, the Bayes estimator ¯ +γ −1 )/(n+α−1) is admissible, but the limit of Bayes estimators, (nX ¯ nX/(n + α − 1) with an α 6= 2, is inadmissible. ¯ is a generalized 33. Consider Example 4.8. Show that the sample mean X ¯ is admissible Bayes estimator of µ under the squared error loss and X using (a) Theorem 4.3 and (b) the result in Example 4.6. 34. Let X be an observation from the gamma distribution Γ(α, θ) with a known α and an unknown θ > 0. Show that X/(α+1) is an admissible estimator of θ under the squared error loss, using Theorem 4.3. 35. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (θ, θ + 1), θ ∈ R. Consider the estimation of θ under the squared error loss. (a) Let π(θ) be a continuous and positive Lebesgue p.d.f. on R. Derive the Bayes estimator w.r.t. the prior π and show that it is a consistent estimator of θ. (b) Show that (X(1) + X(n) − 1)/2 is an admissible estimator of θ and obtain its risk, where X(j) is the jth order statistic. 36. Consider the normal linear model X = Nn (Zβ, σ 2 In ), where Z is an n × p known matrix of full rank, p < n, β ∈ Rp , and σ 2 > 0. (a) Assume that σ 2 is known. Derive the posterior distribution of β when the prior distribution for β is Np (β0 , σ 2 V ), where β0 ∈ Rp is known and V is a known positive definite matrix, and find the Bayes estimator of lτ β under the squared error loss, where l ∈ Rp is known. (b) Show that the Bayes estimator in (a) is admissible and consistent as n → ∞, assuming that the minimum eigenvalue of Z τ Z → ∞. (c) Repeat (a) and (b) when σ 2 is unknown and has the inverse gamma distribution Γ−1 (α, γ) (see Exercise 1(d)), where α and γ are known.

305

4.6. Exercises

(d) In part (c), obtain Bayes estimators of σ 2 and lτ β/σ under the squared error loss and show that they are consistent under the condition in (b). 37. In Example 4.9, suppose that εij has the Lebesgue p.d.f. o n κ(δ)σi−1 exp −c(δ)|x/σi |2/(1+δ) , where

c(δ) =

Γ

3(1+δ) 2 Γ 1+δ 2

(

)

1 1+δ

,

κ(δ) =

Γ

3(1+δ) 2

1/2

3/2

(1+δ)[Γ( 1+δ 2 )]

,

−1 < δ ≤ 1 and σi > 0. −2/(1+δ) (a) Assume that δ is known. Let ωi = c(δ)σi . Under the squared error loss and the same prior in Example 4.9, show that the Bayes estimator of σi2 is 1+δ Z ni X 1 qi (δ) + |xij − β τ Zi |2/(1+δ) f (β|x, δ)dβ, γ j=1

where qi (δ) = [c(δ)]1+δ Γ

1+δ 2 ni

+α−δ

Γ

1+δ 2 ni

+ α + 1 and

−(α+1+ 1+δ 2 ni ) ni k Y X 1 + f (β|x, δ) ∝ π(β) |xij − β τ Zi |2/(1+δ) . γ i=1 j=1

(b) Assume that δ has a prior p.d.f. f (δ) and that given δ, ωi still has the same prior in (a). Derive a formula (similar to that in (a)) for the Bayes estimator of σi2 . 38. Suppose that we have observations Xij = µi + εij ,

i = 1, ..., k, j = 1, ..., m,

where εij ’s are i.i.d. from N (0, σε2 ), µi ’s are i.i.d. from N (µ, σµ2 ), and εij ’s and µi ’s are independent. Suppose that the distribution for σε2 is the inverse gamma distribution Γ−1 (α1 , β1 ) (see Exercise 1(d)); the distribution for σµ2 is the inverse gamma distribution Γ−1 (α2 , β2 ); the distribution for µ is N (µ0 , σ02 ); and σε , σµ , and µ are independent. Describe a Gibbs sampler and obtain explicit forms of (a) the distribution of µ, given Xij ’s, µi ’s, σε2 , and σµ2 ; (b) the distribution of µi , given Xij ’s, µ, σε2 , and σµ2 ; (c) the distribution of σε2 , given Xij ’s, µi ’s, µ, and σµ2 ; (d) the distribution of σµ2 , given Xij ’s, µi ’s, µ, and σε2 . 39. Prove (4.16).

306

4. Estimation in Parametric Models

40. Consider a Lebesgue p.d.f. p(y) ∝ (2+y)125 (1−y)38 y 34 I(0,1) (y). Generate Markov chains of length 10,000 and compute approximations to R yp(y)dy, using the Metropolis kernel with q(y, z) being the p.d.f. of N (y, r2 ), given y, where (a) r = 0.001; (b) r = 0.05; (c) r = 0.12. 41. Prove Proposition 4.4 for the cases of variance and risk. 42. In the proof of Theorem 4.5, show that if L is (strictly) convex and not monotone, then E[L(T0 (x) − a)|D = d] is (strictly) convex and not monotone in a. 43. Prove part (iii) of Theorem 4.5. 44. Under the conditions of Theorem 4.5 and the loss function L(µ, a) = |µ − a|, show that u∗ (d) in Theorem 4.5 is any median (Exercise 92 in §2.6) of T0 (X) under the conditional distribution of X given D = d when µ = 0. 45. Show that if there is a location invariant estimator T0 of µ with finite mean, then E0 [T (X)|D = d] is finite a.s. P for any location invariant estimator T . 46. Show (4.21) under the squared error loss. 47. In Exercise 14, find the MRIE of θ under the squared error loss. 48. In Example 4.12, (a) show that X(1) − θ log 2/n is an MRIE of µ under the absolute error loss L(µ − a) = |µ − a|; (b) show that X(1) − t is an MRIE under the loss function L(µ − a) = I(t,∞) (|µ − a|). 49. In Example 4.13, show that T∗ is also an MRIE of µ if the loss function is convex and even. (Hint: the distribution of T∗ (X) given D depends only on X(n) − X(1) and is symmetric about 0 when µ = 0.) 50. Let X1 , ..., Xn be i.i.d. from the double exponential distribution DE(µ, 1) with an unknown µ ∈ R. Under the squared error loss, find the MRIE of µ. (Hint: for x1 < · · · < xn and xk < t < xk+1 , Pn Pk Pn i=1 |xi − t| = i=k+1 xi − i=1 xi + (2k − n)t.) 51. In Example 4.11, find the MRIE of µ under the loss function −α(µ − a) µ 0 satisfying Z c Z ∞ xdPx|z = xdPx|z , 0

c

where Px|z is the conditional distribution of X given Z = z when σ = 1. −1

56. In Example 4.15, show that the MRIE is 2(n+1) X(n) when the loss is given by (4.24) with p = 1. 57. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(0, θ) with an unknown θ > 0. (a) Find the MRIE of θ under the loss (4.24) with p = 2. (b) Find the MRIE of θ under the loss (4.24) with p = 1. (c) Find the MRIE of θ2 under the loss (4.24) with p = 2. 58. Let X1 , ..., Xn be i.i.d. with a Lebesgue p.d.f. (2/σ)[1−(x/σ)]I(0,σ) (x), where σ > 0 is an unknown scale parameter. Find Pitman’s estimator of σ h for n = 2, 3, and 4. 59. Let X1 , ..., Xn be i.i.d. from the Pareto distribution P a(σ, α), where σ > 0 is an unknown parameter and α > 2 is known. Find the MRIE of σ under the loss function (4.24) with p = 2. 60. Assume that the sample X has a joint Lebesgue p.d.f. given by (4.25). Show that a loss function for the estimation of µ is invariant under the location-scale transformations gc,r (X) = (rX1 + c, ..., rXn + c), r > 0, c ∈ R, if and only if it is of the form L a−µ . σ 61. Prove Proposition 4.6, Theorem 4.9, and Corollary 4.2.

62. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(µ, σ), where µ ∈ R and σ > 0 are unknown. (a) Find the MRIE of σ under the loss (4.24) with p = 1 or 2. (b) Under the loss function (a − µ)2 /σ 2 , find the MRIE of µ. (c) Compute the bias of the MRIE of µ in (b).

308

4. Estimation in Parametric Models

63. Suppose that X and Y are two samples with p.d.f. given by (4.30). (a) Suppose that µx = µy = 0 and consider the estimation of η = (σy /σx )h with a fixed h 6= 0 under the loss L(a/η). Show that the problem is invariant under the transformations g(X, Y ) = (rX, r′ Y ), r > 0, r′ > 0. Generalize Proposition 4.5, Theorem 4.8, and Corollary 4.1 to the present problem. (b) Generalize the result in (a) to the case of unknown µx and µy under the transformations in (4.31). 64. Under the conditions of part (a) of the previous exercise and the loss function (a − η)2 /η 2 , determine the MRIE of η in the following cases: (a) m = n = 1, X and Y are independent, X has the gamma distribution Γ(αx , γ) with a known αx and an unknown γ = σx > 0, and Y has the gamma distribution Γ(αy , γ) with a known αy and an unknown γ = σy > 0; (b) X is Nm (0, σx2 Im ), Y is Nn (0, σy2 In ), and X and Y are independent; (c) X and Y are independent, the components of X are i.i.d. from the uniform distribution U (0, σx ), and the components of Y are i.i.d. from the uniform distribution U (0, σy ). samples, where Xi ’s 65. Let X1 , ..., Xm and Y1 , ..., Yn betwo independent x−µx −1 are i.i.d. having the p.d.f. σx f with µx ∈ R and σx > 0, and σx x−µy with µy ∈ R and σy > 0. Yi ’s are i.i.d. having the p.d.f. σy−1 f σy

Under the loss function (a − η)2 /η 2 and the transformations in (4.31), obtain the MRIE of η = σy /σx when (a) f is the p.d.f. of N (0, 1); (b) f is the p.d.f. of the exponential distribution E(0, 1); (c) f is the p.d.f. of the uniform distribution U − 12 , 12 ; (d) In (a)-(c), find the MRIE of ∆ = µy − µx under the assumption that σx = σy = σ and under the loss function (a − ∆)2 /σ 2 .

66. Consider the general linear model (3.25) under the assumption that εi ’s are i.i.d. with the p.d.f. σ −1 f (x/σ), where f is a known Lebesgue p.d.f. (a) Show that the family of populations is invariant under the transformations in (4.32). τ (b) Show that the estimation τ of l β with l ∈ R(Z) is invariant under the loss function L a−lσ β . (c) Show that the LSE lτ βˆ is an invariant estimator of lτ β, l ∈ R(Z). (d) Prove Theorem 4.10. 67. In Example 4.18, let T be a randomized estimator of p with probabil¯ and probability 1/(n + 1) being 1 . Show that ity n/(n + 1) being X 2

309

4.6. Exercises

¯ T has a constant risk that is smaller than the maximum risk of X. 68. Let X be a single sample from the geometric distribution G(p) with an unknown p ∈ (0, 1). Show that I{1} (X) is a minimax estimator of p under the loss function (a − p)2 /[p(1 − p)]. ¯ is a minimax estimator of µ under the 69. In Example 4.19, show that X 2 2 loss function (a − µ) /σ when Θ = R × (0, ∞). 70. Let T be a minimax (or admissible) estimator of ϑ under the squared error loss. Show that c1 T + c0 is a minimax (or admissible) estimator of c1 ϑ+c0 under the squared error loss, where c1 and c0 are constants. 71. Let X be a sample from Pθ with an unknown θ = (θ1 , θ2 ), where θj ∈ Θj , j = 1, 2, and let Π2 be a probability R measure on Θ2 . Suppose that an estimator T0 minimizesRsupθ1 ∈Θ1 RT (θ)dΠ2 (θ2 ) over all estimators T and that supθ1 ∈Θ1 RT0 (θ)dΠ2 (θ2 ) = supθ1 ∈Θ1 ,θ2 ∈Θ2 RT0 (θ). Show that T0 is a minimax estimator. 72. Let X1 , ..., Xm be i.i.d. from N (µx , σx2 ) and Y1 , ..., Yn be i.i.d. from N (µy , σy2 ). Assume that Xi ’s and Yj ’s are independent. Consider the estimation of ∆ = µy − µx under the squared error loss. ¯ is a minimax estimator of ∆ when σx and σy (a) Show that Y¯ − X ¯ are known, where X and Y¯ are the sample means based on Xi ’s and Yi ’s, respectively. ¯ is a minimax estimator of ∆ when σx ∈ (0, cx ] (b) Show that Y¯ − X and σy ∈ (0, cy ], where cx and cy are constants. 73. Consider the general linear model (3.25) with assumption A1 and the estimation of lτ β under the squared error loss, where l ∈ R(Z). Show that the LSE lτ βˆ is minimax if σ 2 ∈ (0, c] with a constant c. 74. Let X be a random variable having the hypergeometric distribution HG(r, θ, N − θ) (Table 1.1, page 18) with known N and r but an unknown θ. Consider the estimation of θ/N under the squared error loss. (a) Show that the p risk function of T (X) = αX/r + β is constant, where α = {1 + (N − r)/[r(N − 1)]}−1 and β = (1 − α)/2. (b) Show that T in (a) is the minimax estimator of θ/N and the Bayes estimator w.r.t. the prior Π({θ}) =

Γ(2c) [Γ(c)]2

Z

0

1

N θ+c−1 t (1 − t)N −θ+c−1 dt, θ = 1, ..., N, θ

where c = β/(α/r − 1/N ).

310

4. Estimation in Parametric Models

75. Let X be a single observation from N (µ, 1) and let µ have the improper Lebesgue prior density π(µ) = eµ . Under the squared error loss, show that the generalized Bayes estimator of µ is X + 1, which is neither minimax nor admissible. 76. Let X be a random variable having the Poisson distribution P (θ) with an unknown θ > 0. Consider the estimation of θ under the squared error loss. (a) Show that supθ RT (θ) = ∞ for any estimator T = T (X). (b) Let ℑ = {aX + b : a ∈ R, b ∈ R}. Show that 0 is a ℑ-admissible estimator of θ. 77. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(a, θ) with a known θ and an unknown a ∈ R. Under the squared error loss, show that X(1) − θ/n is the unique minimax estimator of a. 78. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (µ − 12 , µ + 12 ) with an unknown µ ∈ R. Under the squared error loss, show that (X(1) + X(n) )/2 is the unique minimax estimator of µ. 79. Let X1 , ..., Xn be i.i.d. from the double exponential distribution DE(µ, 1) with an unknown µ ∈ R. Under the squared error loss, find a minimax estimator of µ. ¯ + b)/(n + 1) is an admissi80. Consider Example 4.7. Show that (nX ble estimator of θ under the squared error loss for any b ≥ 0 and ¯ that nX/(n + 1) is a minimax estimator of θ under the loss function L(θ, a) = (a − θ)2 /θ2 . 81. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). Consider the estimation of p under the squared error loss. ¯ and (X ¯ + γλ)/(1 + λ) with λ > 0 Using Theorem 4.14, show that X and 0 ≤ γ ≤ 1 are admissible. 82. Let X be a single observation. Using Theorem 4.14, find values of α and β such that αX + β are admissible for estimating EX under the squared error loss when (a) X has the Poisson distribution P (θ) with an unknown θ > 0; (b) X has the negative binomial distribution N B(p, r) with a known r and an unknown p ∈ (0, 1). 83. Let X be a single observation having the Lebesgue p.d.f. 21 c(θ)eθx−|x| , |θ| < 1. (a) Show that c(θ) = 1 − θ2 . (b) Show that if 0 ≤ α ≤ 21 , then αX + β is admissible for estimating E(X) under the squared error loss.

311

4.6. Exercises

84. Let X be a single observation from the discrete p.d.f. fθ (x) = [x!(1 − e−θ )]−1 θx e−θ I{1,2,...} (x), where θ > 0 is unknown. Consider the estimation of ϑ = θ/(1 − e−θ ) under the squared error loss. (a) Show that the estimator X is admissible. (b) Show that X is not minimax unless supθ RT (θ) = ∞ for any estimator T = T (X). (c) Find a loss function under which X is minimax and admissible. 85. In Example 4.23, find the UMVUE of θ = (p1 , ..., pk ) under the loss function (4.37). 86. Let X be a sample from Pθ , θ ∈ Θ ⊂ Rp . Consider the estimation of θ under the loss (θ − a)τ Q(θ − a), where a ∈ A = Θ and Q is a known positive definite matrix. Show that the Bayes action is the posterior mean E(θ|X = x), assuming that all integrals involved are finite. 87. In Example 4.24, show that X is the MRIE of θ under the loss function (4.37), if Qp (a) f (x − θ) = j=1 fj (xj − θj ), where each fj is a known Lebesgue p.d.f. with mean 0; R (b) f (x − θ) = f (kx − θk) with xf (kxk)dx = 0.

88. Prove that X in Example 4.25 is a minimax estimator of θ under the loss function (4.37). 89. Let X1 , ..., Xk be independent random variables, where Xi has the binomial distribution Bi(pi , ni ) with an unknown pi ∈ (0, 1) and a known ni . For estimating θ = (p1 , ..., pk ) under the loss (4.37), find a minimax estimator of θ and determine whether it is admissible. 90. Show that the risk function in (4.42) tends to p as kθk → ∞.

91. Suppose that X is Np (θ, Ip ). Consider the estimation of θ under the loss (a − θ)τ Q(a − θ) with a positive definite p × p matrix Q. Show that the risk of the estimator Q δc,r =X−

r(p − 2) Q−1 (X − c) kQ−1/2 (X − c)k2

is equal to tr(Q) − (2r − r2 )(p − 2)2 E(kQ−1/2 (X − c)k−2 ). 92. Show that under the loss (4.37), the risk of δ˜c,r in (4.45) is given by (4.46).

312

4. Estimation in Parametric Models

93. Suppose that X is Np (θ, V ) with p ≥ 4. Consider the estimation of θ under the loss function (4.37). (a) When V = Ip , show that the risk of the estimator in (4.48) is ¯ p k−2 ). p − (p − 3)2 E(kX − XJ 2 (b) When V = σ D with an unknown σ 2 > 0 and a known matrix D, show that the risk function of the estimator in (4.49) is smaller than that of X for any θ and σ 2 . 94. Let X be a sample from a p.d.f. fθ and T (X) be a sufficient statistic for θ. Show that if an MLE exists, it is a function of T but it may not be sufficient for θ. 95. Let {fθ : θ ∈ Θ} be a family of p.d.f.’s w.r.t. a σ-finite measure, where Θ ⊂ Rk ; h be a Borel function from Θ onto Λ ⊂ Rp , 1 ≤ p ≤ k; and ˜ let ℓ(λ) = supθ:h(θ)=λ ℓ(θ) be the induced likelihood function for the transformed parameter λ. Show that if θˆ ∈ Θ is an MLE of θ, then ˆ = h(θ) ˆ maximizes ˜l(λ). λ 96. Let X1 , ..., Xn be i.i.d. with a p.d.f. fθ . Find an MLE of θ in each of the following cases. (a) fθ (x) = θ−1 I{1,...,θ}(x), θ is an integer between 1 and θ0 . (b) fθ (x) = e−(x−θ)I(θ,∞) (x), θ > 0. (c) fθ (x) = θ(1 − x)θ−1 I(0,1) (x), θ > 1. θ x(2θ−1)/(1−θ) I(0,1) (x), θ ∈ ( 12 , 1). (d) fθ (x) = 1−θ −1 −|x−θ| (e) fθ (x) = 2 e , θ ∈ R. (f) fθ (x) = θx−2 I(θ,∞) (x), θ > 0. (g) fθ (x) = θx (1 − θ)1−x I{0,1} (x), θ ∈ [ 21 , 34 ]. (h) fθ (x) is the p.d.f. of N (θ, θ2 ), θ ∈ R, θ 6= 0. (i) fθ (x) is the p.d.f. of the exponential distribution E(µ, σ), θ = (µ, σ) ∈ R × (0, ∞). (j) fθ (x) is the p.d.f. of the log-normal distribution LN (µ, σ 2 ), θ = (µ, σ 2 ) ∈ R × (0, ∞). √ (k) fθ (x) = I(0,1) (x) if θ = 0 and fθ (x) = (2 x)−1 I(0,1) (x) if θ = 1. (l) fθ (x) = β −α αxα−1 I(0,β) (x), θ = (α, β) ∈ (0, ∞) × (0, ∞). (m) fθ (x) = xθ px (1 − p)θ−xI{0,1,...,θ} (x), θ = 1, 2, ..., where p ∈ (0, 1) is known. (n) fθ (x) = 12 (1 − θ2 )eθx−|x|, θ ∈ (−1, 1). 97. In Exercise 14, obtain an MLE of θ when (a) θ ∈ R and (b) θ ≤ 0. 98. Suppose that n observations are taken from N (µ, 1) with an unknown µ. Instead of recording all the observations, one records only whether the observation is less than 0. Find an MLE of µ. 99. Find an MLE of θ in Exercise 43 of §2.6.

4.6. Exercises

313

100. Let (Y1 , Z1 ), ..., (Yn , Zn ) be i.i.d. random 2-vectors such that Y1 and Z1 are independently distributed as the exponential distributions E(0, λ) and E(0, µ), respectively, where λ > 0 and µ > 0. (a) Find the MLE of (λ, µ). (b) Suppose that we only observe Xi = min{Yi , Zi } and ∆i = 1 if Xi = Yi and ∆i = 0 if Xi = Zi . Find the MLE of (λ, µ). 101. In Example 4.33, show that almost surely the likelihood equation has a unique solution that is the MLE of θ = (α, γ). Obtain iteration equation (4.53) for this example. Discuss how to apply the Fisherscoring method in this example. 102. Let X1 , ..., Xn be i.i.d. from the discrete p.d.f. in Exercise 84 with an unknown θ > 0. Show that the likelihood equation has a unique root when the sample mean > 1. Show whether this root is an MLE of θ. 103. Let X1 , ..., Xn be i.i.d. from the logistic distribution LG(µ, σ) (Table 1.2, page 20). (a) Show how to find an MLE of µ when µ ∈ R and σ is known. (b) Show how to find an MLE of σ when σ > 0 and µ is known. 104. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. from a two-dimensional normal distribution with E(X1 ) = E(Y1 ) = 0, Var(X1 ) = Var(Y1 ) = 1, and an unknown correlation coefficient ρ ∈ (−1, 1). Show that the likelihood equation is a cubic in ρ and the probability that it has a unique root tends to 1 as n → ∞. 105. Let X1 , ..., Xn be i.i.d. from the Weibull distribution W (α, θ) (Table 1.2, page 20) with unknown α > 0 and θ > P 0. Show that n −1 the likelihood equation is equivalent to h(α) = n i=1 log xi and Pn Pn Pn −1 α α −1 α −1 θ=n x , where h(α) = ( x ) x , i=1 i i=1 i i=1 i log xi − α and that the likelihood equation has a unique solution. 106. Consider the random effects model in Example 3.17. Assume that µ = 0 and ni = n0 for all i. Provide a condition on Xij ’s under which a unique MLE of (σa2 , σ 2 ) exists and find this MLE. 107. Let X1 , ..., Xn be i.i.d. with the p.d.f. θf (θx), where f is a Lebesgue p.d.f. on (0, ∞) or symmetric about 0, and θ > 0 is an unknown parameter. Show that the likelihood equation has a unique root if xf ′ (x)/f (x) is continuous in x and strictly decreasing for x > 0. Verify that this condition is satisfied if f is the p.d.f. of the Cauchy distribution C(0, 1). 108. Let X1 , ..., Xn be i.i.d. with the Lebesgue p.d.f. fθ (x) = θf1 (x) + (1 − θ)f2 (x), where fj ’s are two different known Lebesgue p.d.f.’s and

314

4. Estimation in Parametric Models

θ ∈ (0, 1) is unknown. (a) Provide a necessary and sufficient condition for the likelihood equation to have a unique solution and show that if there is a solution, it is the MLE of θ. (b) Derive the MLE of θ when the likelihood equation has no solution. 109. Consider the location family in §4.2.1 and the scale family in §4.2.2. In each case, show that an MLE or an RLE (root of the likelihood equation) of the parameter, if it exists, is invariant. 110. Let X be a sample from Pθ , θ ∈ R. Suppose that Pθ ’s have p.d.f.’s fθ w.r.t. a common σ-finite measure and that {x : fθ (x) > 0} does not depend on θ. Assume further that an estimator θˆ of θ attains the Cram´er-Rao lower bound and that the conditions in Theorem 3.3 ˆ Show that θˆ is a unique MLE of θ. hold for θ. 111. Let Xij , j = 1, ..., r > 1, i = 1, ..., n, be independently distributed as N (µi , σ 2 ). Find the MLE of (µ1 , ..., µn , σ 2 ). Show that the MLE of σ 2 is not a consistent estimator (as n → ∞). 112. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (0, θ), where θ > 0 is unknown. Let θˆ be the MLE of θ and T be the UMVUE. (a) Obtain the ratio mseT (θ)/mseθˆ(θ) and show that the MLE is inadmissible when n ≥ 2. (b) Let Za,θ be a random variable having the exponential distribution ˆ →d Z0,θ and n(θ − T ) →d Z−θ,θ . Obtain the E(a, θ). Prove n(θ − θ) asymptotic relative efficiency of θˆ w.r.t. T . 113. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(a, θ) with unknown a and θ. Obtain the asymptotic relative efficiency of the MLE of a (or θ) w.r.t. the UMVUE of a (or θ). 114. Let X1 , ..., Xn be i.i.d. from the Pareto distribution P a(a, θ) with unknown a and θ. (a) Find the MLE of (a, θ). (b) Find the asymptotic relative efficiency of the MLE of a w.r.t. the UMVUE of a. 115. In Exercises 40 and 41 of §2.6, (a) obtain an MLE of (θ, j); (b) show whether the MLE of j in part (a) is consistent; (c) show that the MLE of θ is consistent and derive its nondegenerated asymptotic distribution. 116. In Example 4.36, obtain the MLE of β under the canonical link and assumption (4.58) but ti 6≡ 1.

4.6. Exercises

315

117. Consider the GLM P in Example 4.35 with φi ≡ 1 and the canonical n link. Assume that i=1 Zi Ziτ is positive definite for n ≥ n0 . Show that the likelihood equation has at most one solution when n ≥ n0 and a solution exists with probability tending to 1. 118. Consider the linear model (3.25) with ε = Nn (0, V ), where V is an unknown positive definite matrix. Show that the LSE βˆ defined by (3.29) is an MQLE and that βˆ is an MLE if and only if one of (a)-(e) in Theorem 3.10 holds. 119. Let Xj be a random variable having the binomial distribution Bi(pj , nj ) with a known nj and an unknown pj ∈ (0, 1), j = 1, 2. Assume that Xj ’s are independent. Obtain a conditional likelihood p1 p2 function of the odds ratio θ = 1−p 1−p2 , given X1 + X2 . 1 120. Let X1 and X2 be independent from Poisson distributions P (µ1 ) and P (µ2 ), respectively. Suppose that we are interested in θ1 = µ1 /µ2 . Derive a conditional likelihood function of θ1 , using (a) θ2 = µ1 ; (b) θ2 = µ1 + µ2 ; and (c) θ2 = µ1 µ2 . 121. Assume model (4.66) with p = 2 and normally distributed i.i.d. εt ’s. Obtain the conditional likelihood given (X1 , X2 ) = (x1 , x2 ). 122. Prove the claim in Example 4.38. 123. Prove (4.70). (Hint: Show, using the argument in proving (4.77), that ∂2 ∂2 n−1 | ∂θ 2 log ℓ(ξn ) − ∂θ 2 log ℓ(θ)| = op (1) for any random variable ξn satisfying |ξn − θ| ≤ |θ − θn |.) 124. Let X1 , ..., Xn be i.i.d. from N (µ, 1) truncated at two known points α < β, i.e., the Lebesgue p.d.f. of Xi is √ 2 { 2π[Φ(β − µ) − Φ(α − µ)]}−1 e−(x−µ) /2 I(α,β) (x). ¯ is asymptotically efficient for esti(a) Show that the sample mean X mating θ = EX1 . ¯ is the unique MLE of θ. (b) Show that X 125. Let X1 , ..., Xn be i.i.d. from the discrete p.d.f. x m−x fθ (x) = [1 − (1 − θ)m ]−1 m I{1,2,...,m} (x), x θ (1 − θ)

where θ ∈ (0, 1) is unknown and m ≥ 2 is a known integer. ¯ = m, show that X/m ¯ (a) When the sample mean X is an MLE of θ. ¯ (b) When 1 < X < m, show that the likelihood equation has at least one solution. (c) Show that the regularity conditions of Theorem 4.16 are satisfied and find the asymptotic variance of a consistent RLE of θ.

316

4. Estimation in Parametric Models

126. In Exercise 96, check whether the regularity conditions of Theorem 4.16 are satisfied for cases (b), (c), (d), (e), (g), (h), (j) and (n). Obtain nondegenerated asymptotic distributions of RLE’s for cases in which Theorem 4.17 can be applied. 127. Let X1 , ..., Xn be i.i.d. random variables such that log Xi is N (θ, θ) with an unknown θ > 0. (a) Obtain the likelihood equation and show that one of the solutions of the likelihood equation is the unique MLE of θ. (b) Using Theorem 4.17, obtain the asymptotic distribution of the MLE of θ. 128. In Exercise 107 of §3.6, find the MLE’s of α and β and obtain their nondegenerated asymptotic joint distribution. 129. In Example 4.30, show that the MLE (or RLE) of θ is asymptotically efficient by (a) applying Theorem 4.17 and (b) directly deriving the asymptotic distribution of the MLE. 130. In Example 4.23, show that there is a unique asymptotically efficient RLE of θ = (p1 , ..., pk ). Discuss whether this RLE is the MLE. 131. Let X1 , ..., Xn be i.i.d. with P (X1 = 0) = 6θ2 − 4θ + 1, P (X1 = 1) = θ − 2θ2 , and P (X1 = 2) = 3θ − 4θ2 , where θ ∈ (0, 21 ) is unknown. Apply Theorem 4.17 to obtain the asymptotic distribution of an RLE of θ. 132. Let X1 , ..., Xn be i.i.d. random variables from N (µ, 1), where µ ∈ R is unknown. Let θ = P (X1 ≤ c), where c is a known constant. Find the asymptotic relative efficiency ofP the MLE of θ w.r.t. (a) the UMVUE of θ and (b) the estimator n−1 ni=1 I(−∞,c] (Xi ).

133. In Exercise 19 of §3.6, find the MLE’s of θ and ϑ = P (Y1 > 1) and find the asymptotic relative efficiency of the MLE of ϑ w.r.t. the UMVUE of ϑ in part (b). 134. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors. Suppose that both X1 and Y1 are binary, P (X1 = 1) = 21 , P (Y1 = 1|X1 = 0) = e−aθ , and P (Y1 = 1|X1 = 0) = e−bθ , where θ > 0 is unknown and a > 0 and b > 0 are known constants. (a) Suppose that (Xi , Yi ), i = 1, ..., n, are observed. Find the MLE of θ and its nondegenerated asymptotic distribution. (b) Suppose that only Y1 , ..., Yn are observed. Find the MLE of θ and its nondegenerated asymptotic distribution. (c) Calculate the asymptotic relative efficiency of the MLE in (a) w.r.t. the MLE in (b). How much efficiency is lost in the special case of a = b?

317

4.6. Exercises

135. In Exercise 110 of §3.6, derive (a) the MLE of (θ1 , θ2 ); (b) a nondegenerated asymptotic distribution of the MLE of (θ1 , θ2 ); (c) the asymptotic relative efficiencies of the MLE’s w.r.t. the moment estimators in Exercise 110 of §3.6. 136. In Exercise 104, show that the RLE of ρ is asymptotically distributed as N ρ, (1 − ρ2 )2 /[n(1 + ρ2 )] .

137. In Exercise 107, obtain a nondegenerated asymptotic distribution of the RLE of θ when f is the p.d.f. of the Cauchy distribution C(0, 1). 138. Let X1 , ..., Xn be i.i.d. from the logistic distribution LG(µ, σ) with unknown µ ∈ R and σ > 0. Obtain a nondegenerated asymptotic distribution of the RLE of (µ, σ). 139. In Exercise 105, show that the conditions of Theorem 4.16 are satisfied. 140. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p, where p ∈ (0, 1) is unknown. Let ϑˆn be the MLE of ϑ = p(1 − p). (a) Show that ϑˆn is asymptotically normal when p 6= 21 . (b) When p = 12 , derive a nondegenerated asymptotic distribution of ϑˆn with an appropriate normalization. 141. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors satisfying 0 ≤ X1 ≤ 1, 0 ≤ Y1 ≤ 1, and P (X1 > x, Y1 > y) = (1 − x)(1 − y)(1 − max{x, y})θ for 0 ≤ x ≤ 1, 0 ≤ y ≤ 1, where θ ≥ 0 is unknown. (a) Obtain the likelihood function and the likelihood equation. (b) Show that an RLE of θ is asymptotically normal and derive its amse. 142. Assume the conditions in Theorem 4.16. Suppose that θ = (θ1 , ..., θk ) and there is a positive integer p < k such that ∂ log ℓ(θ)/∂θi and ∂ log ℓ(θ)/∂θj are uncorrelated whenever i ≤ p < j. Show that the asymptotic distribution of the RLE of (θ1 , ..., θp ) is unaffected by whether θp+1 , ..., θk are known. 143. Let X1 , ..., Xn be i.i.d. random p-vectors from Np (µ, Σ) with unknown µ and Σ. Find the MLE’s of µ and Σ and derive their nondegenerated asymptotic distributions. 144. Let X1 , ..., Xn be i.i.d. bivariate normal random vectors with mean 0 and an unknown covariance matrix whose diagonal elements are

318

4. Estimation in Parametric Models σ12 and σ22 and off-diagonal element is σ1 σ2 ρ. Let θ = (σ12 , σ22 , ρ). Obtain In (θ) and [In (θ)]−1 and derive a nondegenerated asymptotic distribution of the MLE of θ.

145. Let X1 , ..., Xn be i.i.d. each with probability p as N (µ, σ 2 ) and probability 1 − p as N (η, τ 2 ), where θ = (µ, η, σ 2 , τ 2 , p) is unknown. (a) Show that the conditions in Theorem 4.16 are satisfied. (b) Show that the likelihood function is unbounded. (c) Show that an MLE may be inconsistent. 146. Let X1 , ..., Xn and Y1 , ..., Yn be independently distributed as N (µ, σ 2 ) and N (µ, τ 2 ), respectively, with unknown θ = (µ, σ 2 , τ 2 ). Find the MLE of θ and show that it is asymptotically efficient. 147. Find a nondegenerated asymptotic distribution of the MLE of (σa2 , σ 2 ) in Exercise 106. 148. Under the conditions in Theorem 4.18, prove (4.85) and (4.86). 149. Assume linear model (3.25) with ε = Nn (0, σ 2 In ) and a full rank Z. Apply Theorem 4.18 to show that the LSE βˆ is asymptotically efficient. Compare this result with that in Theorem 3.12. 150. Apply Theorem 4.18 to obtain the asymptotic distribution of the RLE of β in (a) Example 4.35 and (b) Example 4.37. 151. Let X1 , ..., Xn be i.i.d. from the logistic distribution LG(µ, σ), µ ∈ R, σ > 0. Using Newton-Raphson and Fisher-scoring methods, find (a) one-step MLE’s of µ when σ is known; (b) one-step MLE’s of σ when µ is known; (c) one-step MLE’s of (µ, σ); √ (d) n-consistent initial estimators in (a)-(c). 152. Under the GLM (4.55)-(4.58), (a) show how to obtain a one-step MLE of β, if an initial estimator (0) βˆn is available; (b) show that under the conditions in Theorem 4.18, the one-step (0) MLE satisfies (4.81) if k[In (β)]1/2 (βˆn − β)k = Op (1). 153. In Example 4.40, show that the conditions in Theorem 4.20 concerning the likelihood function are satisfied. 154. Let X1 , ..., Xn be i.i.d. from the logistic distribution LG(µ, σ) with unknown µ ∈ R and σ > 0. Show that the conditions in Theorem 4.20 concerning the likelihood function are satisfied.

Chapter 5

Estimation in Nonparametric Models Estimation methods studied in this chapter are useful for nonparametric models as well as for parametric models in which the parametric model assumptions might be violated (so that robust estimators are required) or the number of unknown parameters is exceptionally large. Some such methods have been introduced in Chapter 3; for example, the methods that produce UMVUE’s in nonparametric models, the U- and V-statistics, the LSE’s and BLUE’s, the Horvitz-Thompson estimators, and the sample (central) moments. The theoretical justification for estimators in nonparametric models, however, relies more on asymptotics than that in parametric models. This means that applications of nonparametric methods usually require large sample sizes. Also, estimators derived using parametric methods are asymptotically more efficient than those based on nonparametric methods when the parametric models are correct. Thus, to choose between a parametric method and a nonparametric method, we need to balance the advantage of requiring weaker model assumptions (robustness) against the drawback of losing efficiency, which results in requiring a larger sample size. It is assumed in this chapter that a sample X = (X1 , ..., Xn ) is from a population in a nonparametric family, where Xi ’s are random vectors.

5.1 Distribution Estimators In many applications the c.d.f.’s of Xi ’s are determined by a single c.d.f. F on Rd ; for example, Xi ’s are i.i.d. random d-vectors. In this section, we 319

320

5. Estimation in Nonparametric Models

consider the estimation of F or F (t) for several t’s, under a nonparametric model in which very little is assumed about F .

5.1.1 Empirical c.d.f.’s in i.i.d. cases For i.i.d. random variables X1 , ..., Xn , the empirical c.d.f. Fn is defined in (2.28). The definition of the empirical c.d.f. based on X = (X1 , ..., Xn ) in the case of Xi ∈ Rd is analogously given by n

Fn (t) =

1X I(−∞,t] (Xi ), n i=1

t ∈ Rd ,

(5.1)

where (−∞, a] denotes the set (−∞, a1 ] × · · · × (−∞, ad ] for any a = (a1 , ..., ad ) ∈ Rd . Similar to the case of d = 1 (Example 2.26), Fn (t) as an estimator of F (t) has the following properties. For any t ∈ Rd , nFn (t) has the binomial distribution Bi(F (t), n); Fn (t) is unbiased with variance F (t)[1 − F (t)]/n; √Fn (t) is the UMVUE under some nonparametric models; and Fn (t) is n-consistent for F (t). For any m fixed distinct points t1 , ..., tm in Rd , it follows from the multivariate CLT (Corollary 1.2) and (5.1) that as n → ∞, √ n Fn (t1 ), ..., Fn (tm ) − F (t1 ), ..., F (tm ) →d Nm (0, Σ), (5.2) where Σ is the m × m matrix whose (i, j)th element is P X1 ∈ (−∞, ti ] ∩ (−∞, tj ] − F (ti )F (tj ).

Note that these results hold without any assumption on F . Considered as a function of t, Fn is a random element taking values in √ F, the collection of all c.d.f.’s on Rd . As n → ∞, n(Fn − F ) converges in some sense to a random element defined on some probability space. A detailed discussion of such a result is beyond our scope and can be found, for example, in Shorack and Wellner (1986). To discuss some global properties of Fn as an estimator of F ∈ F, we need to define a closeness measure between the elements (c.d.f.’s) in F. Definition 5.1. Let F0 be a collection of c.d.f.’s on Rd . (i) A function ̺ from F0 × F0 to [0, ∞) is called a distance or metric on F0 if and only if for any Gj in F0 , (a) ̺(G1 , G2 ) = 0 if and only if G1 = G2 ; (b) ̺(G1 , G2 ) = ̺(G2 , G1 ); and (c) ̺(G1 , G2 ) ≤ ̺(G1 , G3 ) + ̺(G3 , G2 ). (ii) Let D = {c(G1 − G2 ) : c ∈ R, Gj ∈ F0 , j = 1, 2}. A function k · k from D to [0, ∞) is called a norm on D if and only if (a) k∆k = 0 if and only if ∆ = 0; (b) kc∆k = |c|k∆k for any ∆ ∈ D and c ∈ R; and (c) k∆1 + ∆2 k ≤ k∆1 k + k∆2 k for any ∆j ∈ D, j = 1, 2.

321

5.1. Distribution Estimators

Any norm k·k on D induces a distance given by ̺(G1 , G2 ) = kG1 −G2 k. The most commonly used distance is the sup-norm distance ̺∞ , i.e., the distance induced by the sup-norm kG1 − G2 k∞ = sup |G1 (t) − G2 (t)|, t∈Rd

Gj ∈ F.

(5.3)

The following result concerning the sup-norm distance between Fn and F is due to Dvoretzky, Kiefer, and Wolfowitz (1956). Lemma 5.1. (DKW’s inequality). Let Fn be the empirical c.d.f. based on i.i.d. X1 , ..., Xn from a c.d.f. F on Rd . (i) When d = 1, there exists a positive constant C (not depending on F ) such that 2 P ̺∞ (Fn , F ) > z ≤ Ce−2nz , z > 0, n = 1, 2, ....

(ii) When d ≥ 2, for any ǫ > 0, there exists a positive constant Cǫ,d (not depending on F ) such that 2 P ̺∞ (Fn , F ) > z ≤ Cǫ,d e−(2−ǫ)nz , z > 0, n = 1, 2, .... The proof of this lemma is omitted. The following results useful in statistics are direct consequences of Lemma 5.1. Theorem 5.1. Let Fn be the empirical c.d.f. based on i.i.d. X1 , ..., Xn from a c.d.f. F on Rd . Then (i) ̺∞ √ (Fn , F ) →a.s. 0 as n → ∞; (ii) E[ n̺∞ (Fn , F )]s = O(1) for any s > 0. Proof. (i) From DKW’s inequality, ∞ X

n=1

P ̺∞ (Fn , F ) > z < ∞.

Hence, the result follows from Theorem 1.8(v). √ (ii) Using DKW’s inequality with z = y 1/s / n and the result in Exercise 55 of §1.6, we obtain that Z ∞ √ √ E[ n̺∞ (Fn , F )]s = P n̺∞ (Fn , F ) > y 1/s dy 0 Z ∞ 2/s ≤ Cǫ,d e−(2−ǫ)y dy 0

= O(1)

as long as 2 − ǫ > 0.

322

5. Estimation in Nonparametric Models

Theorem 5.1(i) means that Fn (t) →a.s. F (t) uniformly in t ∈ Rd , a result stronger than√the strong consistency of Fn (t) for every t. Theorem √ 5.1(ii) implies that n̺∞ (Fn , F ) = Op (1), a result stronger than the nconsistency of Fn (t). These results hold without any condition on F . R Let p ≥ 1 and Fp = {G ∈ F : ktkp dG < ∞}, which is the subset of c.d.f.’s in F having finite pth moments. Mallows’ distance between G1 and G2 in Fp is defined to be ̺Mp (G1 , G2 ) = inf(EkY1 − Y2 kp )1/p ,

(5.4)

where the infimum is taken over all pairs of Y1 and Y2 having c.d.f.’s G1 and G2 , respectively. Let {Gj : j = 0, 1, 2, ...} ⊂ Fp . Then ̺Mp (Gj , G0 ) → 0 as R R j → ∞ if and only if ktkp dGj → ktkp dG0 and Gj (t) → G0 (t) for every t ∈ Rd at which G0 is continuous. It follows from Theorem 5.1 and the SLLN (Theorem 1.13) that ̺Mp (Fn , F ) →a.s. 0 if F ∈ Fp . When d = 1, another useful distance for measuring the closeness between Fn and F is the Lp distance ̺Lp induced by the Lp -norm (p ≥ 1) Z 1/p p |G1 (t) − G2 (t)| dt , Gj ∈ F1 . (5.5) kG1 − G2 kLp = A result similar to Theorem 5.1 is given as follows. Theorem 5.2. Let Fn be the empirical c.d.f. based on i.i.d. random variables X1 , ..., Xn from a c.d.f. F ∈ F1 . Then (i) ̺Lp (Fn , F ) →a.s. 0; R √ (ii) E[ n̺Lp (Fn , F )] = O(1) if 1 ≤ p < 2 and {F (t)[1 − F (t)]}p/2 dt < ∞, or p ≥ 2. Proof. (i) Since [̺Lp (Fn , F )]p ≤ [̺∞ (Fn , F )]p−1 [̺L1 (Fn , F )] and, by Theorem 5.1, ̺∞ (Fn , F ) →a.s. 0, it suffices to show the result for p = 1. Let R0 Yi = −∞ [I(−∞,t] (Xi ) − F (t)]dt. Then Y1 , ..., Yn are i.i.d. and Z Z E|Yi | ≤ E|I(−∞,t] (Xi ) − F (t)|dt = 2 F (t)[1 − F (t)]dt,

which is finite under the condition that F ∈ F1 . By the SLLN, Z 0 n 1X [Fn (t) − F (t)]dt = Yi →a.s. E(Y1 ) = 0. (5.6) n i=1 −∞ R0 Since [Fn (t) − F (t)]− ≤ F (t) and −∞ F (t)dt < ∞ (Exercise 55 in §1.6), it R 0follows from Theorem 5.1 and the dominated convergence theorem that −∞ [Fn (t) − F (t)]− dt →a.s. 0, which with (5.6) implies Z 0 |Fn (t) − F (t)|dt →a.s. 0. (5.7) −∞

323

5.1. Distribution Estimators

The result follows since we can similarly show that (5.7) holds with R∞ replaced by 0 . (ii) When 1 ≤ p < 2, the result follows from E[̺Lp (Fn , F )] ≤ ≤

Z

Z

=n

p

E|Fn (t) − F (t)| dt

Z

−∞

1/p

[E|Fn (t) − F (t)|2 ]p/2 dt

−1/2

R0

{F (t)[1 − F (t)]}

1/p

p/2

dt

= O(n−1/2 ),

1/p

where the two inequalities follow from Jensen’s inequality. When p ≥ 2, o n E[̺Lp (Fn , F )] ≤ E [̺∞ (Fn , F )]1−2/p [̺L2 (Fn , F )]2/p n o1/q 1/p ≤ E[̺∞ (Fn , F )](1−2/p)q E[̺L2 (Fn , F )]2 1/p n o1/q Z −(1−2/p)q/2 2 E |Fn (t) − F (t)| dt = O(n ) = O(n

−(1−2/p)/2

= O(n−1/2 ),

)

1 n

Z

F (t)[1 − F (t)]dt

1/p

where 1q + 1p = 1, the second inequality follows from H¨older’s inequality (see (1.40) in §1.3.2), and the first equality follows from Theorem 5.1(ii).

5.1.2 Empirical likelihoods In §4.4 and §4.5, we have shown that the method of using likelihoods provides some asymptotically efficient estimators. We now introduce some likelihoods in nonparametric models. This not only provides another justification for the use of the empirical c.d.f. in (5.1), but also leads to a useful method of deriving estimators in various (possibly non-i.i.d.) cases, some of which are discussed later in this chapter. Let X1 , ..., Xn be i.i.d. with F ∈ F and PG be the probability measure corresponding to G ∈ F. Given X1 = x1 , ..., Xn = xn , the nonparametric likelihood function is defined to be the following functional from F to [0, ∞): ℓ(G) =

n Y

i=1

PG ({xi }),

G ∈ F.

(5.8)

324

5. Estimation in Nonparametric Models

Apparently, ℓ(G) = 0 if PG ({xi }) = 0 for at least one i. The following result, due to Kiefer and Wolfowitz (1956), shows that the empirical c.d.f. Fn is a nonparametric maximum likelihood estimator of F . Theorem 5.3. Let X1 , ..., Xn be i.i.d. with F ∈ F and ℓ(G) be defined by (5.8). Then Fn maximizes ℓ(G) over G ∈ F. Proof. We only need to consider G ∈ F such that ℓ(G) > 0. Let c ∈ (0, 1] and F(c) be the subset Pn of F containing G’s satisfying pi = PG ({xi }) > 0, i = 1, ..., n, and i=1 pi = c. We now apply the Lagrange multiplier method to solve the problem of maximizing ℓ(G) over G ∈ F(c). Define ! n n Y X H(p1 , ..., pn , λ) = pi + λ pi − c , i=1

i=1

where λ is the Lagrange multiplier. Set n

X ∂H = pi − c = 0, ∂λ i=1

n Y ∂H = p−1 pi + λ = 0, j ∂pj i=1

j = 1, ..., n.

The solution is pi = c/n, i = 1, ..., n, λ = −(c/n)n−1 . It can be shown (exercise) that Pn this solution is a maximum of H(p1 , ..., pn , λ) over pi > 0, i = 1, ..., n, i=1 pi = c. This shows that max ℓ(G) = (c/n)n ,

G∈F(c)

which is maximized at c = 1 for any fixed n. The result follows from PFn ({xi }) = n−1 for given Xi = xi , i = 1, ..., n. From the proof of Theorem 5.3, P Fn maximizes the likelihood ℓ(G) in (5.8) over pi > 0, i = 1, ..., n, and ni=1 pi = 1, where pi = PG ({xi }). This method of deriving an estimator of F can be extended to various situations with some modifications of (5.8) and/or constraints on pi ’s. Modifications of the likelihood in (5.8) are called empirical likelihoods (Owen, 1988, 2001; Qin and Lawless, 1994). An estimator obtained by maximizing an empirical likelihood is then called a maximum empirical likelihood estimator (MELE). We now discuss several applications of the method of empirical likelihoods. Consider first the estimation of F with auxiliary information about F (and i.i.d. X1 , ..., Xn ). For instance, suppose that there is a known Borel function u from Rd to Rs such that Z u(x)dF = 0 (5.9) (e.g., some components of the mean of F are 0). It is thenR reasonable to expect that any estimate Fˆ of F has property (5.9), i.e., u(x)dFˆ = 0,

325

5.1. Distribution Estimators

which is not true for the empirical c.d.f. Fn in (5.1), since Z n 1X u(x)dFn = u(Xi ) 6= 0 n i=1

even if E[u(X1 )] = 0. Using the method of empirical likelihoods, a natural solution is to put another constraint in the process of maximizing the likelihood. That is, we maximize ℓ(G) in (5.8) subject to pi > 0,

i = 1, ..., n,

n X

pi = 1,

and

i=1

n X

pi u(xi ) = 0,

(5.10)

i=1

where pi = PG ({xi }). Using the Lagrange multiplier method and an argument similar to the proof of Theorem 5.3, it can be shown (exercise) that an MELE of F is n X Fˆ (t) = pˆi I(−∞,t] (Xi ), (5.11) i=1

where the notation (−∞, t] is the same as that in (5.1), pˆi = n−1 [1 + λτn u(Xi )]−1 ,

i = 1, ..., n,

(5.12)

and λn ∈ Rs is the Lagrange multiplier satisfying n X i=1

n

pˆi u(Xi ) =

u(Xi ) 1X = 0. n i=1 1 + λτn u(Xi )

(5.13)

Note that Fˆ reduces to Fn if u ≡ 0. To see that (5.13) has a solution asymptotically, note that " n # n ∂ 1X u(Xi ) 1X τ log 1 + λ u(Xi ) = ∂λ n i=1 n i=1 1 + λτ u(Xi )

and

∂2 ∂λ∂λτ

"

# n n 1X 1 X u(Xi )[u(Xi )]τ τ log 1 + λ u(Xi ) = − , n i=1 n i=1 [1 + λτ u(Xi )]2

which is negative definite if Var(u(X1 )) is positive definite. Also, " n ) # ( ∂ 1X = E[u(X1 )] = 0. log 1 + λτ u(Xi ) E ∂λ n i=1 λ=0

Hence, using the same argument as in the proof of Theorem 4.18, we can show that there exists a unique sequence {λn (X)} such that as n → ∞, ! n u(Xi ) 1X =0 →1 and λn →p 0. (5.14) P n i=1 1 + λτn u(Xi )

326

5. Estimation in Nonparametric Models

Theorem 5.4. Let X1 , ..., Xn be i.i.d. with F ∈ F, u be a Borel function on Rd satisfying (5.9), and Fˆ be given by (5.11)-(5.13). Suppose that U = Var(u(X1 )) is positive definite. Then, for any m fixed distinct t1 , ..., tm in Rd , √ n[ Fˆ (t1 ), ..., Fˆ (tm ) − F (t1 ), ..., F (tm ) ] →d Nm (0, Σu ), (5.15)

where

Σu = Σ − W τ U −1 W, Σ is given in (5.2), W = W (t1 ), ..., W (tm ) , W (tj ) = E[u(X1 )I(−∞,tj ] (X1 )], and the notation (−∞, t] is the same as that in (5.1). Proof. We prove the case Pn of m = 1. The case of m ≥ 2 is left as an exercise. Let u ¯ = n−1 i=1 u(Xi ). It follows from (5.13), (5.14), and Taylor’s expansion that n

u ¯=

1X u(Xi )[u(Xi )]τ λn [1 + op (1)]. n i=1

By the SLLN and CLT,

U −1 u ¯ = λn + op (n−1/2 ). Using Taylor’s expansion and the SLLN again, we have n n 1X 1X 1 − 1 I(−∞,t] (Xi )(nˆ pi − 1) = I(−∞,t] (Xi ) n i=1 n i=1 1 + λτn u(Xi ) n

=−

1X I(−∞,t] (Xi )λτn u(Xi ) + op (n−1/2 ) n i=1

= −λτn W (t) + op (n−1/2 )

= −¯ uτ U −1 W (t) + op (n−1/2 ).

Thus, n

1X Fˆ (t) − F (t) = Fn (t) − F (t) + I(−∞,t] (Xi )(nˆ pi − 1) n i=1

¯τ U −1 W (t) + op (n−1/2 ) = Fn (t) − F (t) − u n 1 X I(−∞,t] (Xi )−F (t)−[u(Xi )]τ U −1 W (t) + op (n−1/2 ). = n i=1

The result follows from the CLT and the fact that Var [W (t)]τ U −1 u(Xi ) = [W (t)]τ U −1 U U −1 W (t) = [W (t)]τ U −1 W (t)

= E{[W (t)]τ U −1 u(Xi )I(−∞,t] (Xi )}

= Cov I(−∞,t] (Xi ), [W (t)]τ U −1 u(Xi ) .

327

5.1. Distribution Estimators

Comparing (5.15) with (5.2), we conclude that Fˆ is asymptotically more efficient than Fn . Example 5.1 (Survey problems). An example of situations in which we have auxiliary information expressed as (5.9) is a survey problem (Example 2.3) where the population P = {y1 , ..., yN } consists of two-dimensional yj ’s, PN yj = (y1j , y2j ), and the population mean Y¯2 = N −1 j=1 y2j is known. For example, suppose that y1j is the current year’s income of unit j in the population and y2j is the last year’s income. In many applications the population total or mean of y2j ’s is known, for example, from tax return records. Let X1 , ..., Xn be a simple random sample (see Example 2.3) selected from P with replacement. Then Xi ’s are i.i.d. bivariate random vectors whose c.d.f. is F (t) =

N 1 X I(−∞,t] (yj ), N j=1

(5.16)

where the notation (−∞, t] is the same as that in (5.1). If Y¯2 is known, then it can be expressed as (5.9) with u(x1 , x2 ) = x2 − Y¯2 . In survey problems Xi ’s are usually sampled without replacement so that X1 , ..., Xn are not i.i.d. However, for a simple random sample without replacement, (5.8) can still be treated as an empirical likelihood, given Xi ’s. Note that F in (5.16) is the c.d.f. of Xi , regardless of whether Xi ’s are sampled with replacement. If X = (X1 , ..., Xn ) is not a simple random sample, then the likelihood (5.8) has to be modified. Suppose that πi is the probability that the ith unit is selected (see Theorem 3.15). Given X = {yi , i ∈ s}, an empirical likelihood is Y Y 1/π ℓ(G) = (5.17) [PG ({yi })]1/πi = pi i , i∈s i∈s where pi = PG ({yi }). With the auxiliary information (5.9), an MELE of F in (5.16) can be obtained by maximizing ℓ(G) in (5.17) subject to (5.10). In this case F may not be the c.d.f. of Xi , but the c.d.f.’s of Xi ’s are determined by F and πi ’s. It can be shown (exercise) that an MELE is given by (5.11) with X 1 1 pˆi = (5.18) πi [1 + λτn u(yi )] i∈s πi and

X

u(yi ) = 0. π [1 + λτn u(yi )] i∈s i

(5.19)

If πi = a constant, then the MELE reduces to that in (5.11)-(5.13). If

328

5. Estimation in Nonparametric Models

u(x) = 0 (no auxiliary information), then the MELE is X X 1 1 ˆ I(−∞,t] (yi ) , F (t) = π π i∈s i i∈s i which is a ratio of two Horvitz-Thompson estimators (§3.4.2). Some asymptotic properties of the MELE Fˆ can be found in Chen and Qin (1993). The second part of Example 5.1 shows how to use empirical likelihoods in a non-i.i.d. problem. Applications of empirical likelihoods in non-i.i.d. problems are usually straightforward extensions of those in i.i.d. cases. The following is another example. Example 5.2 (Biased sampling). Biased sampling is often used in applications. Suppose that n = n1 + · · · + nk , k ≥ 2; Xi ’s are independent random variables; X1 , ..., Xn1 are i.i.d. with F ; and Xn1 +···+nj +1 , ..., Xn1 +···+nj+1 are i.i.d. with the c.d.f. Z ∞ Z t wj+1 (s)dF (s) wj+1 (s)dF (s), −∞

−∞

j = 1, ..., k − 1, where wj ’s are some nonnegative Borel functions. A simple example is that X1 , ..., Xn1 are sampled from F and Xn1 +1 , ..., Xn1 +n2 are sampled from F but conditional on the fact that each sampled value exceeds a given value x0 (i.e., w2 (s) = I(x0 ,∞) (s)). For instance, Xi ’s are blood pressure measurements; X1 , ..., Xn1 are sampled from ordinary people and Xn1 +1 , ..., Xn1 +n2 are sampled from patients whose blood pressures are higher than x0 . The name biased sampling comes from the fact that there is a bias in the selection of samples. For simplicity we consider the case of k = 2, since the extension to k ≥ 3 is straightforward. Denote w2 by w. An empirical likelihood is ℓ(G) =

n1 Y

i=1

=

"

PG ({xi })

n X i=1

n Y

i=n1

w(xi )PG ({xi }) R w(s)dG(s) +1

#−n2

pi w(xi )

n Y

i=1

pi

n Y

w(xi ),

(5.20)

i=n1 +1

the where pi = PG ({xi }). An MELE of F can be obtained by maximizing Pn empirical likelihood (5.20) subject to pi > 0, i = 1, ..., n, and i=1 pi = 1. Using the Lagrange multiplier method we can show (exercise) that an MELE Fˆ is given by (5.11) with pˆi = [n1 + n2 w(Xi )/w] ˆ −1 ,

i = 1, ..., n,

(5.21)

329

5.1. Distribution Estimators

where w ˆ satisfies w ˆ=

n X i=1

w(Xi ) . n1 + n2 w(Xi )/w ˆ

An asymptotic result similar to that in Theorem 5.4 can be established (Vardi, 1985; Qin, 1993). If the function w depends on an unknown parameter vector θ, then the method of profile empirical likelihood (see §5.1.4) can be applied. Our last example concerns an important application in survival analysis. Example 5.3 (Censored data). Let T1 , ..., Tn be survival times that are i.i.d. nonnegative random variables from a c.d.f. F , and C1 , ..., Cn be i.i.d. nonnegative random variables independent of Ti ’s. In a variety of applications in biostatistics and life-time testing, we are only able to observe the smaller of Ti and Ci and an indicator of which variable is smaller: Xi = min{Ti , Ci },

δi = I(0,Ci ) (Ti ),

i = 1, ..., n.

This is called a random censorship model and Ci ’s are called censoring times. We consider the estimation of the survival distribution F ; see Kalbfleisch and Prentice (1980) for other problems involving censored data. An MELE of F can be derived as follows. Let x(1) ≤ · · · ≤ x(n) be ordered values of Xi ’s and δ(i) be the δ-value associated with x(i) . Consider a c.d.f. G that assigns its mass to the points x(1) , ..., x(n) and the interval (x(n) , ∞). Let pi = PG ({x(i) }), i = 1, ..., n, and pn+1 = 1 − G(x(n) ). An MELE of F is then obtained by maximizing 1−δ(i) n n+1 Y X δ(i) pi pj (5.22) ℓ(G) = i=1

j=i+1

subject to

pi ≥ 0,

i = 1, ..., n + 1,

n+1 X

pi = 1.

(5.23)

i=1

It can be shown (exercise) that an MELE is Fˆ (t) =

n+1 X

pˆi I(0,t] (X(i) ),

(5.24)

i=1

where X(0) = 0, X(n+1) = ∞, X(1) ≤ · · · ≤ X(n) are order statistics, and pˆi =

δ(i) n−i+1

i−1 Y

j=1

1−

δ(j) n−j+1

,

i = 1, ..., n,

pˆn+1 = 1 −

n X j=1

pˆj .

330

5. Estimation in Nonparametric Models

The Fˆ in (5.24) can also be written as (exercise) Y δ(i) 1 − n−i+1 , Fˆ (t) = 1 −

(5.25)

X(i) ≤t

which is the well-known Kaplan-Meier (1958) product-limit estimator. Some asymptotic results for Fˆ in (5.25) can be found, for example, in Shorack and Wellner (1986).

5.1.3 Density estimation Suppose that X1 , ..., Xn are i.i.d. random variables from F and that F is unknown but has a Lebesgue p.d.f. f . Estimation of F can be done by estimating f , which is called density estimation. Note that estimators of F derived in §5.1.1 and §5.1.2 do not have Lebesgue p.d.f.’s. Since f (t) = F ′ (t) a.e., a simple estimator of f (t) is the difference quotient Fn (t + λn ) − Fn (t − λn ) fn (t) = , t ∈ R, (5.26) 2λn where Fn is the empirical c.d.f. given by (2.28) or (5.1) with d = 1, and {λn } is a sequence of positive constants. Since 2nλn fn (t) has the binomial distribution Bi(F (t + λn ) − F (t − λn ), n), E[fn (t)] → f (t) and

Var fn (t) → 0

if λn → 0 as n → ∞ if λn → 0 and nλn → ∞.

Thus, we should choose λn converging to 0 slower than n−1 . If we assume that λn → 0, nλn → ∞, and f is continuously differentiable at t, then it can be shown (exercise) that f (t) 1 + O(λ2n ) msefn (t) (F ) = +o (5.27) 2nλn nλn and, under the additional condition that nλ3n → 0, p nλn [fn (t) − f (t)] →d N 0, 12 f (t) .

(5.28)

A useful class of estimators is the class of kernel density estimators of the form n 1 X t−Xi fˆ(t) = (5.29) w λn , nλn i=1

331

5.1. Distribution Estimators

where w is a known Lebesgue p.d.f. on R and is called the kernel. If we ˆ in (5.29) is essentially the same as the choose w(t) = 12 I[−1,1] (t), then f(t) so-called histogram. The bias of fˆ(t) in (5.29) is Z 1 E[fˆ(t)] − f (t) = f (z)dz − f (t) w t−z λ n λn Z = w(y)[f (t − λn y) − f (t)]dy. If f is bounded and continuous at t, then, by the dominated convergence theorem (Theorem 1.1(iii)), the bias of fˆ(t) converges to 0 as λn → 0; if f ′ R ˆ is bounded and continuous at t and |t|w(t)dt < ∞, then the bias of f(t) ˆ is O(λn ). The variance of f(t) is 1 t−X1 Var fˆ(t) = Var w λn nλ2n Z h i2 1 t−z = w f (z)dz λn nλ2n 2 Z 1 1 f (z)dz − w t−z λn n λn Z 1 1 [w(y)]2 f (t − λn y)dy + O = nλn n 1 w0 f (t) +o = nλn nλn R if f is bounded and continuous at t and w0 = [w(t)]2 dt < ∞. Hence, if λn → 0, nλn → ∞, and f ′ is bounded and continuous at t, then msefˆ(t) (F ) =

w0 f (t) + O(λ2n ). nλn

Using the CLT (Theorem 1.15), one can show (exercise) that if λn → 0, nλn → ∞, and f is bounded and continuous at t, then p ˆ − E[fˆ(t)]} →d N 0, w0 f (t) . nλn {f(t) (5.30) R Furthermore, if f ′ is bounded and continuous at t, |t|w(t)dt < ∞, and nλ3n → 0, then p p ˆ nλn {E[f(t)] − f (t)} = O nλn λn → 0

and, therefore, (5.30) holds with E[fˆ(t)] replaced by f (t). Similar to the estimation of a c.d.f., we can also study global properties of fn or fˆ as an estimator of the density curve f , using a suitably defined

332

5. Estimation in Nonparametric Models

0.0

0.1

0.2

f(t)

0.3

0.4

0.5

True p.d.f. Estimator (5.26) Estimator (5.29)

-2

-1

0

1

2

t

Figure 5.1: Density estimates in Example 5.4 distance between f and its density estimator. R For example, we may study the convergence of supt∈R |fˆ(t) − f (t)| or |fˆ(t) − f (t)|2 dt. More details can be found, for example, in Silverman (1986). Example 5.4. An i.i.d. sample of size n = 200 was generated from N (0, 1). Density curve estimates (5.26) and (5.29) are plotted in Figure 5.1 with the curve of the true p.d.f. For the kernel density estimator (5.29), w(t) = 21 e−|t| is used and λn = 0.4. From Figure 5.1, it seems that the kernel estimate (5.29) is much better than the estimate (5.26). There are many other density estimation methods, for example, the nearest neighbor method (Stone, 1977), the smoothing splines (Wahba, 1990), and the method of empirical likelihoods described in §5.1.2 (see, e.g., Jones (1991)), which produces estimators of the form n 1 X i fˆ(t) = . pˆi w t−X λn λn i=1

333

5.1. Distribution Estimators

5.1.4 Semi-parametric methods Suppose that the sample X is from a population in a family indexed by (θ, ξ), where θ is a parameter vector, i.e., θ ∈ Θ ⊂ Rk with a fixed positive integer k, but ξ is not vector-valued, e.g., ξ is a c.d.f. Such a model is often called a semi-parametric model, although it is nonparametric according to our definition in §2.1.2. A semi-parametric method refers to a statistical inference method that combines a parametric method and a nonparametric method in making an inference about the parametric component θ and the nonparametric component ξ. In the following, we consider two important examples of semi-parametric methods. Partial likelihoods and proportional hazards models The idea of partial likelihood (Cox, 1972) is similar to that of conditional likelihood introduced in §4.4.3. To illustrate this idea, we assume that X has a p.d.f. fθ,ξ and ξ is also a vector-valued parameter. Suppose that X can be transformed into a sequence of pairs (V1 , U1 ), ..., (Vm , Um ) such that "m #" m # Y Y fθ,ξ (x) = gθ (ui |v1 , u1 , ..., ui−1 , vi ) hθ,ξ (vi |v1 , u1 , ..., vi−1 , ui−1 ) , i=1

i=1

where gθ (·|v1 , u1 , ..., ui−1 , vi ) is the conditional p.d.f. of Ui given V1 = v1 , U1 = u1 , ..., Ui−1 = ui−1 , Vi = vi , which does not depend on ξ, and hθ,ξ (·|v1 , u1 , ..., vi−1 , ui−1 ) is the conditional p.d.f. of Vi given V1 = v1 , U1 = u1 , ..., Vi−1 = vi−1 , Ui−1 = ui−1 . The first product in the previous expression for fθ,ξ (x) is called the partial likelihood for θ. When ξ is a nonparametric component, the partial likelihood for θ can be similarly defined, in which case the full likelihood fθ,ξ (x) should be replaced by a nonparametric likelihood or an empirical likelihood. As long as the conditional distributions of Ui given V1 , U1 , ..., Ui−1 , Vi , i = 1, ..., m, are in a parametric family (indexed by θ), the partial likelihood is parametric. A semi-parametric estimation method consists of a parametric method (typically the maximum likelihood method in §4.4) for estimating θ and a nonparametric method for estimating ξ. To illustrate the application of the method of partial likelihoods, we consider the estimation of the c.d.f. of survival data in the random censorship model described in Example 5.3. Following the notation in Example 5.3, we assume that {T1 , ..., Tn } (survival times) and {C1 , ..., Cn } (censoring times) are two sets of independent nonnegative random variables and that Xi = min{Ti , Ci } and δi = I(0,Ci ) (Ti ), i = 1, ..., n, are independent observations. In addition, we assume that there is a p-vector Zi of covariate values associated with Xi and δi . The situation considered in Example 5.3

334

5. Estimation in Nonparametric Models

can be viewed as a special homogeneous case with Zi ≡ a constant. The survival function when the covariate vector is equal to z is defined to be Sz (t) = 1 − Fz (t), where Fz is the c.d.f. of the survival time T having the same distribution as Ti . Assume that fz (t) = Fz′ (t) exists for all t > 0. The function λz (t) = fz (t)/Sz (t) is called the hazard function and the Rt function Λz (t) = 0 λz (s)ds is called the cumulative hazard function, when the covariate vector is equal to z. A commonly adopted model for λz is the following proportional hazards model: λz (t) = λ0 (t)φ(β τ z),

(5.31)

where φ is a known function (typically φ(x) = ex ), z is a value of the pvector of covariates, β ∈ Rp is an unknown parameter vector, and λ0 (t) is the unknown hazard function when the covariate vector is 0 and is referred to as the baseline hazard function. Under model (5.31), 1 − Fz (t) = exp{−Λz (t)} = exp{−φ(β τ z)Λ0 (t)}. Thus, the estimation of the c.d.f. Fz or the survival function Sz can be done through the estimation of β, the parametric component of model (5.31), and Λ0 , the nonparametric component of model (5.31). Consider first the estimation of β using the method of partial likelihoods. Suppose that there are l observed failures at times T(1) < · · · < T(l) , where (i) is the label for the ith failure ordered according to the time to failure. (Note that a failure occurs when δi = 1.) Suppose that there are mi items censored at or after T(i) but before T(i+1) at times T(i,1) , ..., T(i,mi ) (setting T(0) = 0). Let Ui = (i) and Vi = (T(i) , T(i−1,1) , ..., T(i−1,mi−1 ) ), i = 1, ..., l. Then the partial likelihood is l Y P (Ui = (i)|V1 , U1 , ..., Ui−1 , Vi ). i=1

Since λz (t) = lim∆>0,∆→0 ∆−1 Pz (t ≤ T < t + ∆|T > t), where Pz denotes the probability measure of T when the covariate is equal to z, P (Ui = (i)|V1 , U1 , ..., Ui−1 , Vi ) = P

λZ(i) (ti ) φ(β τ Z(i) ) = P , τ j∈Ri λZj (ti ) j∈Ri φ(β Zj )

where ti is the observed value of T(i) , Ri = {j : Xj ≥ ti } is called the risk set, and the last equality follows from assumption (5.31). This leads to the partial likelihood #δi " l n Y Y φ(β τ Z(i) ) φ(β τ Zi ) P P ℓ(β) = = , τ τ j∈Ri φ(β Zj ) j∈Ri φ(β Zj ) i=1 i=1 which is a function of the parameter β, given the observed data. The maximum likelihood method introduced for parametric models in §4.4 can

335

5.1. Distribution Estimators

be applied to obtain a maximum partial likelihood estimator βˆ of β. It is shown in Tsiatis (1981) that βˆ is consistent for β and is asymptotically normal under some regularity conditions. We now consider the estimation of Λ0 . First, assume that the covariate vector Zi is random, (Ti , Ci , Zi ) are i.i.d., and Ti and Ci are conditionally independent given Zi . Let (T, C, Z) be the random vector having the same distribution as (Ti , Ci , Zi ), X = min{T, C}, and δ = I(0,C) (T ). Under assumption (5.31), it can be shown (exercise) that ZZ ∞ λ0 (s)φ(β τ z)H(s|z)dsdG(z), (5.32) Q(t) = P (X > t, δ = 1) = t

where H(s|z) = P (X > s|Z = z) and G is the c.d.f. of Z. Then Z dQ(t) = −λ0 (t) φ(β τ z)H(t|z)dG(z) dt and λ0 (t) = −

dQ(t) 1 , dt K(t)

(5.33)

(5.34)

where K(t) = E[φ(β τ Z)I(t,∞) (X)] (exercise). Consequently, Z t Z t dQ(s) . Λ0 (t) = λ0 (s)ds = − 0 0 K(s) An estimator of Λ0 can then be obtained by substituting Q and K in the previous expression by their estimators n

X ˆ = 1 Q(t) I{Xi >t,δi =1} n i=1 and

n

1 X ˆτ ˆ φ(β Zi )I(t,∞) (Xi ). K(t) = n i=1

(5.35)

This estimator is known as Breslow’s estimator. When Z1 , ..., Zn are nonrandom, we can still use Breslow’s estimator. Its asymptotic properties can be found, for example, in Fleming and Harrington (1991). Profile likelihoods Let ℓ(θ, ξ) be a likelihood (or empirical likelihood), where θ and ξ are not necessarily vector-valued. It may be difficult to maximize the likelihood ℓ(θ, ξ) simultaneously over θ and ξ. For each fixed θ, let ξ(θ) satisfy ℓ(θ, ξ(θ)) = sup ℓ(θ, ξ). ξ

336

5. Estimation in Nonparametric Models

The function ℓP (θ) = ℓ(θ, ξ(θ)) is called a profile likelihood function for θ. Suppose that θˆP maximizes ℓP (θ). Then θˆP is called a maximum profile likelihood estimator of θ. Note that θˆP may be different from an MLE of θ. Although this idea can be applied to parametric models, it is more useful in nonparametric models, especially when θ is a parametric component. For example, consider the empirical likelihood in (5.8) subject to the constraints in (5.10). Sometimes it is more convenient to allow the function u in (5.10) to depend on an unknown parameter vector θ ∈ Rk , where k ≤ s. This leads to the empirical likelihood ℓ(G) in (5.8) subject to (5.10) with u(x) replaced by ψ(x, θ), where ψ is a known function from Rd × Rk to Rs . Maximizing this empirical likelihood is equivalent to maximizing ! n n n Y X X ℓ(p1 , ..., pn , ω, λ, θ) = pi + ω 1 − pi + pi λτ ψ(xi , θ), i=1

i=1

i=1

where ω and λ are Lagrange multipliers. It follows from (5.12) and (5.13) that ω = n, p˜i (θ) = n−1 {1 + [λn (θ)]τ ψ(xi , θ)}−1 with a λn (θ) satisfying n

ψ(xi , θ) 1X =0 n i=1 1 + [λn (θ)]τ ψ(xi , θ)

P maximize ℓ(p1 , ...pn , ω, λ, θ) for any fixed θ. Substituting p˜i with ni=1 p˜i = 1 into ℓ(p1 , ...pn , ω, λ, θ) leads to the following profile empirical likelihood for θ: n Y 1 ℓP (θ) = . (5.36) τ ψ(x , θ)} n{1 + [λ (θ)] n i i=1

If θˆ is a maximum of ℓP (θ) in (5.36), then θˆ is a maximum profile empirical ˆ A likelihood estimator of θ and the corresponding estimator of pi is p˜i (θ). result similar to Theorem 5.4 and a result on asymptotic normality of θˆ are established in Qin and Lawless (1994), under some conditions on ψ. Another example is the empirical likelihood (5.20) in the problem of biased sampling with a function w(x) = wθ (x) depending on an unknown θ ∈ Rk . The profile empirical likelihood for θ is then ℓP (θ) = w ˆθ−n2

n Y

1 n + n w ˆθ 2 θ (xi )/w i=1 1

where w ˆθ satisfies w ˆθ =

n X i=1

n Y

i=n1 +1

wθ (xi ) . n1 + n2 wθ (xi )/w ˆθ

wθ (xi ),

337

5.1. Distribution Estimators

Finally, we consider the problem of missing data. Assume that X1, ...,Xn are i.i.d. random variables from an unknown c.d.f. F and some Xi ’s are missing. Let δi = 1 if Xi is observed and δi = 0 if Xi is missing. Suppose that (Xi , δi ) are i.i.d. Let π(x) = P (δi = 1|Xi = x). If Xi and δi are independent, i.e., π(x) ≡ π does not depend on x, then the empirical c.d.f. based on observed data, i.e., the c.d.f. putting mass r−1 to each observed Xi , where r is the number of observed Xi ’s, is an unbiased and consistent estimator of F , provided that π > 0. On the other hand, if π(x) depends on x, then the empirical c.d.f. based on observed data is a biased and inconsistent estimator of F . In fact, it can be shown (exercise) that the empirical c.d.f. based on observed data is an unbiased estimator of P (Xi ≤ x|δi = 1), which is generally different from the unconditional probability F (x) = P (Xi ≤ x). If both π and F are in parametric models, then we can apply the method of maximum likelihood. For example, if π(x) = πθ (x) and F (x) = Fϑ (x) has a p.d.f. fϑ , where θ and ϑ are vectors of unknown parameters, then a parametric likelihood of (θ, ϑ) is ℓ(θ, ϑ) =

n Y

[πθ (xi )fϑ (xi )]δi (1 − π)1−δi ,

i=1

R

where π = πθ (x)dF (x). Suppose now that π(x) = πθ (x) is the parametric component and F is the nonparametric component. Then an empirical likelihood can be defined as ℓ(θ, G) =

n Y

[πθ (xi )PG ({xi })]δi (1 − π)1−δi

i=1

Pn

Pn subject to pi ≥ 0, i=1 δi pi = 1, i=1 δi pi [πθ (xi ) − π] = 0, where pi = PG ({xi }), i = 1, ..., n. It can be shown (exercise) that the logarithm of the profile empirical likelihood for (θ, π) (with a Lagrange multiplier) is n X i=1

δi log πθ (xi ) +(1−δi ) log(1−π)−δi log 1+λ[πθ (xi )−π] . (5.37)

Under some regularity conditions, Qin, Leung, and Shao (2002) show that ˆ π ˆ obtained by maximizing the likelihood in (5.37) the estimators θ, ˆ , and λ are consistent and asymptotically normal and that the empirical c.d.f. ˆ ˆ(Xi ) − π putting mass pˆi = r−1 {1 + λ[π ˆ ]}−1 to each observed Xi is conθ sistent for F . The results are also extended to the case where a covariate vector Zi associated with Xi is observed for all i.

338

5. Estimation in Nonparametric Models

5.2 Statistical Functionals In many nonparametric problems, we are interested in estimating some characteristics (parameters) of the unknown population, not the entire population. We assume in this section that Xi ’s are i.i.d. from an unknown c.d.f. F on Rd . Most characteristics of F can be written as T(F ), where T is a functional from F to Rs . If we estimate F by the empirical c.d.f. Fn in (5.1), then a natural estimator of T(F ) is T(Fn ), which is called a statistical functional. Many commonly used statistics can be written as T(F R n ) for some T. Two simple examples are given as follows. Let T(F ) = ψ(x)dF R P (x) with an integrable function ψ, and T(Fn ) = ψ(x)dFn (x) = n−1 ni=1 ψ(Xi ). The sample moments discussed in §3.5.2 are particular examples of this kind of statistical functional. For d = 1, let T(F ) = F −1 (p) = inf{x : F (x) ≥ p}, where p ∈ (0, 1) is a fixed constant. F −1 (p) is called the pth quantile of F . The statistical functional T(Fn ) = Fn−1 (p) is called the pth sample quantile. More examples of statistical functionals are provided in §5.2.1 and §5.2.2. In this section, we study asymptotic distributions of T(Fn ). We focus on the case of real-valued T (s = 1), since the extension to the case of s ≥ 2 is straightforward.

5.2.1 Differentiability and asymptotic normality Note that T(Fn ) is a function of the “statistic” Fn . In Theorem 1.12 (and §3.5.1) we have studied how to use Taylor’s expansion to establish asymptotic normality of differentiable functions of statistics that are asymptotically normal. This leads to the approach of establishing asymptotic normality of T(Fn ) by using some generalized Taylor expansions for functionals and using asymptotic properties of Fn given in §5.1.1. First, we need a suitably defined differential of T. Several versions of differentials are given in the following definition. Definition 5.2. Let T be a functional on F0 , a collection of c.d.f.’s on Rd , and let D = {c(G1 − G2 ) : c ∈ R, Gj ∈ F0 , j = 1, 2}. (i) A functional T on F0 is Gˆ ateaux differentiable at G ∈ F0 if and only if there is a linear functional LG on D (i.e., LG (c1 ∆1 + c2 ∆2 ) = c1 LG (∆1 ) + c2 LG (∆2 ) for any ∆j ∈ D and cj ∈ R) such that ∆ ∈ D and G + t∆ ∈ F0 imply T(G + t∆) − T(G) − LG (∆) = 0. lim t→0 t (ii) Let ̺ be a distance on F0 induced by a norm k · k on D. A functional T on F0 is ̺-Hadamard differentiable at G ∈ F0 if and only if there is a

339

5.2. Statistical Functionals

linear functional LG on D such that for any sequence of numbers tj → 0 and {∆, ∆j , j = 1, 2, ...} ⊂ D satisfying k∆j − ∆k → 0 and G + tj ∆j ∈ F0 , T(G + tj ∆j ) − T(G) − LG (∆j ) = 0. lim j→∞ tj (iii) Let ̺ be a distance on F0 . A functional T on F0 is ̺-Fr´echet differentiable at G ∈ F0 if and only if there is a linear functional LG on D such that for any sequence {Gj } satisfying Gj ∈ F0 and ̺(Gj , G) → 0, lim

j→∞

T(Gj ) − T(G) − LG (Gj − G) = 0. ̺(Gj , G)

The functional LG is called the differential of T at G. If we define h(t) = T(G + t∆), then the Gˆ ateaux differentiability is equivalent to the differentiability of the function h(t) at t = 0, and LG (∆) is simply h′ (0). Let διx denote the d-dimensional c.d.f. degenerated at the point x and φG (x) = LG (διx − G). Then φF (x) is called the influence function of T at F , which is an important tool in robust statistics (see Hampel (1974)). If T is Gˆateaux differentiable √ at F , then we have the following expansion (taking t = n−1/2 and ∆ = n(Fn − F )): √ √ n[T(Fn ) − T(F )] = LF n(Fn − F ) + Rn . (5.38) Since LF is linear, LF

n √ 1 X n(Fn − F ) = √ φF (Xi ) →d N (0, σF2 ) n i=1

(5.39)

by the CLT, provided that E[φF (X1 )] = 0

and

σF2 = E[φF (X1 )]2 < ∞

(5.40)

(which is usually true when φF is bounded or when F has some finite moments). By Slutsky’s theorem and (5.39), √ n[T(Fn ) − T(F )] →d N (0, σF2 ) (5.41) if Rn in (5.38) is op (1). Unfortunately, Gˆ ateaux differentiability is too weak to be useful in establishing Rn = op (1) (or (5.41)). This is why we need other types of differentiability. Hadamard differentiability, which is also referred to as compact differentiability, is clearly stronger than Gˆateaux differentiability but weaker than Fr´echet differentiability (exercise). For a given functional

340

5. Estimation in Nonparametric Models

T, we can first find LG by differentiating h(t) = T(G + t∆) at t = 0 and then check whether T is ̺-Hadamard (or ̺-Fr´echet) differentiable with a given ̺. The most commonly used distances on F0 are the sup-norm distance ̺∞ and the Lp distance ̺Lp . Their corresponding norms are given by (5.3) and (5.5), respectively. Theorem 5.5. Let X1 , ..., Xn be i.i.d. from a c.d.f. F on Rd . (i) If T is ̺∞ -Hadamard differentiable at F , then Rn in (5.38) is op (1). (ii) If T is ̺-Fr´echet differentiable at F with a distance ̺ satisfying √ n̺(Fn , F ) = Op (1), (5.42) then Rn in (5.38) is op (1). (iii) In either (i) or (ii), if (5.40) is also satisfied, then (5.41) holds. Proof. Part (iii) follows directly from (i) or (ii). The proof of (i) involves some high-level mathematics and is omitted; see, for example, Fernholz (1983). We now prove (ii).√From Definition 5.2(iii), for any ǫ > 0, there is a δ > 0 such that |Rn | < ǫ n̺(Fn , F ) whenever ̺(Fn , F ) < δ. Then √ n̺(Fn , F ) > η/ǫ + P (̺(Fn , F ) ≥ δ) P (|Rn | > η) ≤ P for any η > 0, which implies

lim sup P (|Rn | > η) ≤ lim sup P n

n

√ n̺(Fn , F ) > η/ǫ .

The result follows from (5.42) and the fact that ǫ can be made arbitrarily small. Since ̺-Fr´echet differentiability implies ̺-Hadamard differentiability, Theorem 5.5(ii) is useful when ̺ is not the sup-norm distance. There are functionals that are not ̺∞ -Hadamard differentiable (and hence R not ̺∞ -Fr´echet differentiable). For example, if d = 1 and T(G) = g( xdG) with a differentiable function g, then T is not necessarily ̺∞ -Hadamard differentiable, but is ̺L1 -Fr´echet differentiable (exercise). From Theorem 5.2, condition (5.42) holds for ̺Lp under the moment conditions on F given in Theorem 5.2. Note that if ̺ and ̺˜ are two distances on F0 satisfying ̺˜(G1 , G2 ) ≤ c̺(G1 , G2 ) for a constant c and all Gj ∈ F0 , then ̺˜-Hadamard (Fr´echet) differentiability implies ̺-Hadamard (Fr´echet) differentiability. This suggests the use of the distance ̺∞+p = ̺∞ + ̺Lp , which also satisfies (5.42) under the moment conditions in Theorem 5.2. The distance ̺∞+p is useful in some cases (Theorem 5.6). A ̺∞ -Hadamard differentiable T having a bounded and continuous influence function φF is robust in Hampel’s sense (see, e.g., Huber (1981)).

5.2. Statistical Functionals

341

This is motivated by the fact that the asymptotic behavior of T(Fn ) is determined by that of LF (Fn − F ), and a small change in the sample, i.e., small changes in all xi ’s (rounding, grouping) or large changes in a few xi ’s (gross errors, blunders), will result in a small change of T(Fn ) if and only if φF is bounded and continuous. We now consider some examples. For the sample moments related to R functionals of the form T(G) = ψ(x)dG(x), it is clear that T is a linear functional. Any linear functional is trivially ̺-Fr´echet differentiable for any ̺. Next, if F is one-dimensional and F ′ (x) > 0 for all x, then the quantile functional T(G) = G−1 (p) is ̺∞ -Hadamard differentiable at F (Fernholz, 1983). Hence, Theorem 5.5 applies to these functionals. But the asymptotic normality of sample quantiles can be established under weaker conditions, which are studied in §5.3.1. Example 5.5 (Convolution functionals). Suppose that F is on R and for a fixed z ∈ R, Z T(G) = G(z − y)dG(y), G ∈ F. If X1 and X2 are i.i.d. with c.d.f. G, then T(G) is the c.d.f. of X1 + X2 (Exercise 47 in §1.6), and is also called the convolution of G evaluated at z. For tj → 0 and k∆j − ∆k∞ → 0, Z Z T(G + tj ∆j ) − T(G) = 2tj ∆j (z − y)dG(y) + t2j ∆j (z − y)d∆j (y) (for ∆ = c1 G1 + c2 G2 , Gj ∈ F0 , and cj ∈ R, d∆ denotes c1 dG1 + c2 dG2 ). Using Lemma 5.2, one can show (exercise) that Z ∆j (z − y)d∆j (y) = O(1). (5.43) Hence T is ̺∞ -Hadamard differentiable at any G R∈ F with LG (∆) = R 2 ∆(z−y)dG(y). The influence function, φF (x) = 2 (διx −F )(z−y)dF (y), is a bounded function and clearly satisfies (5.40). Thus, (5.41) holds. If F is continuous, then T is robust in Hampel’s sense (exercise). Three important classes of statistical functionals, i.e., L-estimators, Mestimators, and rank statistics and R-estimators, are considered in §5.2.2. Lemma 5.2. Let ∆ ∈ D and h be a continuous function on R such that R h(x)d∆(x) is finite. Then Z h(x)d∆(x) ≤ khkV k∆k∞ ,

342

5. Estimation in Nonparametric Models

where khkV is the variation norm defined by m X sup khkV = lim |h(xj ) − h(xj−1 )| a→−∞,b→∞

j=1

with the supremum being taken over all partitions a = x0 < · · · < xm = b of the interval [a, b]. The proof of Lemma 5.2 can be found in Natanson (1961, p. 232). The differentials in Definition 5.2 are first-order differentials. For some functionals, we can also consider their second-order differentials, which provides a way of defining the order of the asymptotic biases via expansion (2.37). Definition 5.3. Let T be a functional on F0 and ̺ be a distance on F0 . (i) T is second-order ̺-Hadamard differentiable at G ∈ F0 if and only if there is a functional QG on D such that for any sequence of numbers tj → 0 and {∆, ∆j , j = 1, 2, ...} ⊂ D satisfying k∆j − ∆k → 0 and G + tj ∆j ∈ F0 , lim

j→∞

T(G + tj ∆j ) − T(G) − QG (tj ∆j ) = 0, t2j

RR ψG (x, y)d(GR + where QG (∆) = R ∆)(x)d(G + ∆)(y) for a function ψG satisfying ψG (x, y) = ψG (y, x), ψG (x, y)dG(x)dG(y) = 0, and D and k · k are the same as those in Definition 5.2(ii). (ii) T is second-order ̺-Fr´echet differentiable at G ∈ F0 if and only if, for any sequence {Gj } satisfying Gj ∈ F0 and ̺(Gj , G) → 0, lim

j→∞

T (Gj ) − T (G) − QG (Gj − G) = 0, [̺(Gj , G)]2

where QG is the same as that in (i). For a second-order differentiable T, we have the following expansion: n[T(Fn ) − T(F )] = nVn + Rn ,

(5.44)

where Vn = QF (Fn − F ) =

Z Z

ψF (x, y)dFn (x)dFn (y) =

n n 1 XX ψF (Xi , Xj ) n2 j=1 i=1

is a “V-statistic” (§3.5.3) whose asymptotic properties are given by Theorem 3.16. If Rn in (5.44) is op (1), then the asymptotic behavior of T(Fn ) − T(F ) is the same as that of Vn .

343

5.2. Statistical Functionals

Proposition 5.1. Let X1 , ..., Xn be i.i.d. from F . (i) If T is second-order ̺∞ -Hadamard differentiable at F , then Rn in (5.44) is op (1). (ii) If T is second-order ̺-Fr´echet differentiable at F with a distance ̺ satisfying (5.42), then Rn in (5.44) is op (1). Combining Proposition 5.1 with Theorem 3.16, we conclude that if Z ψF (X1 , y)dF (y) > 0, ζ1 = Var then (5.41) holds with σF2 = 4ζ1 and amseT (Fn ) (P ) = σF2 /n; if ζ1 = 0, then n[T(Fn ) − T(F )] →d

∞ X

λj χ21j

j=1

and amseT (Fn ) (P ) = {2Var(ψF (X1 , X2 )) + [EψF (X1 , X1 )]2 }/n2 . In any case, expansion (2.37) holds and the n−1 order asymptotic bias of T (Fn ) is EψF (X1 , X1 )/n. If T is also first-order differentiable, then it can be shown (exercise) that Z φF (x) = 2 ψF (x, y)dF (y). (5.45)

Then ζ1 = 4−1 Var(φF (X1 )) and ζ1 = 0 corresponds to the case of φF (x) ≡ 0. However, second-order ̺-Hadamard (Fr´echet) differentiability does not imply first-order ̺-Hadamard (Fr´echet) differentiability (exercise). The technique in this section can be applied to non-i.i.d. Xi ’s when the c.d.f.’s of Xi ’s are determined by an unknown c.d.f. F , provided that results similar to (5.39) and (5.42) (with Fn replaced by some other estimator Fˆ ) can be established.

5.2.2 L-, M-, and R-estimators and rank statistics Three large classes of statistical functionals based on i.i.d. Xi ’s are studied in this section. L-estimators Let J(t) be a Borel function on [0, 1]. An L-functional is defined as Z (5.46) T(G) = xJ(G(x))dG(x), G ∈ F0 , where F0 contains all c.d.f.’s on R for which T is well defined. For X1 , ..., Xn i.i.d. from F ∈ F0 , T(Fn ) is called an L-estimator of T(F ).

344

5. Estimation in Nonparametric Models

Example 5.6. The following are some examples of commonly used Lestimators. ¯ the sample mean. (i) When J ≡ 1, T(Fn ) = X, (ii) When J(t) = 4t − 2, T(Fn ) is proportional to Gini’s mean difference. (iii) When J(t) = (β − α)−1 I(α,β) (t) for some constants α < β, T(Fn ) is called the trimmed sample mean. For an L-functional T, it can be shown (exercise) that Z T(G) − T(F ) = φF (x)d(G − F )(x) + R(G, F ), where φF (x) = − R(G, F ) = − and WG (x) =

(

[G(x) − F (x)]−1 0

Z

Z

(διx − F )(y)J(F (y))dy,

(5.47)

(5.48)

WG (x)[G(x) − F (x)]dx,

R G(x) F (x)

J(t)dt − J(F (x))

G(x) 6= F (x) G(x) = F (x).

A sufficient condition for (5.40) in this case is that J is bounded and F has a finite variance (exercise). However, (5.40) is also satisfied if φF is bounded. The differentiability of T can be verified under some conditions on J. Theorem 5.6. Let T be an L-functional defined by (5.46). (i) Suppose that J is bounded, J(t) = 0 when t ∈ [0, α] ∪ [β, 1] for some constants α < β, and that the set D = {x : J is discontinuous at F (x)} has Lebesgue measure 0. Then T is ̺∞ -Fr´echet differentiable at F with the influence function φF given by (5.48), and φF is bounded and continuous and satisfies (5.40). (ii) Suppose that J is bounded, the set D in (i) has Lebesgue measure 0, and J is continuous on [0, α] ∪ [β, 1] for some constants α < β. Then T is ̺∞+1 -Fr´echet differentiable at F . (iii) Suppose that |J(t) − J(s)| ≤ C|t − s|p−1 , where C > 0 and p > 1 are some constants. Then T is ̺Lp -Fr´echet differentiable at F . (iv) If, in addition to the conditions in part (i), J ′ is continuous on [α, β], then T is second-order ̺∞ -Fr´echet differentiable at F with Z ψF (x, y) = φF (x) + φF (y) − (διx − F )(z)(διy − F )(z)J ′ (F (z))dz. (v) Suppose that J ′ is continuous on [0, 1]. Then T is second-order ̺L2 Fr´echet differentiable at F with the same ψF given in (iv).

345

5.2. Statistical Functionals

Proof. We prove (i)-(iii). The proofs for (iv) and (v) are similar and are left to the reader. (i) Let Gj ∈ F and ̺∞ (Gj , F ) → 0. Let c and d be two constants such that F (c) > β and F (d) < α. Then, for sufficiently large j, Gj (x) ∈ [0, α] ∪ [β, 1] if x > c or x < d. Hence, for sufficiently large j, Z c |R(Gj , F )| = WGj (x)(Gj − F )(x)dx d Z c ≤ ̺∞ (Gj , F ) |WGj (x)|dx. d

Since J is continuous at F (x) when x 6∈ D and D has Lebesgue measure R0,c WGj (x) → 0 a.e. Lebesgue. By the dominated convergence theorem, |WGj (x)|dx → 0. This proves that T is ̺∞ -Fr´echet differentiable. The d assertions on φF can be proved by noting that Z c (διx − F )(y)J(F (y))dy. φF (x) = − d

(ii) From the proof of (i), we only need to show that Z ̺∞+1 (Gj , F ) → 0, W (x)(G − F )(x)dx Gj j

(5.49)

A

where A = {x : F (x) ≤ α or F (x) > β}. The quantity on the left-hand side of (5.49) is bounded by supx∈A |WGj (x)|, which converges to 0 under the continuity assumption of J on [0, α] ∪ [β, 1]. Hence (5.49) follows. (iii) The result follows from Z |R(G, F )| ≤ C |G(x) − F (x)|p dx = O [̺Lp (G, F )]p and the fact that p > 1.

An L-estimator with J(t) = 0 when t ∈ [0, α] ∪ [β, 1] is called a trimmed L-estimator. Theorem 5.6(i) shows that trimmed L-estimators satisfy (5.41) and are robust in Hampel’s sense. In cases (ii) and (iii) of Theorem 5.6, (5.41) holds if Var(X1 ) < ∞, but T(Fn ) may not be robust in Hampel’s sense. It can be shown (exercise) that one or several of (i)-(v) of Theorem 5.6 can be applied to each of the L-estimators in Example 5.6. M-estimators Let ρ(x, t) be a Borel function on Rd × R and Θ be an open subset of R. An M-functional is defined to be a solution of Z Z ρ(x, T(G))dG(x) = min ρ(x, t)dG(x), G ∈ F0 , (5.50) t∈Θ

346

5. Estimation in Nonparametric Models

where F0 contains all c.d.f.’s on Rd for which the integrals in (5.50) are well defined. For X1 , ..., Xn i.i.d. from F ∈ F0 , T(Fn ) is called an M-estimator of T(F ). Assume that ψ(x, t) = ∂ρ(x, t)/∂t exists a.e. and Z Z ∂ λG (t) = ψ(x, t)dG(x) = ρ(x, t)dG(x). (5.51) ∂t Then λG (T(G)) = 0. Example 5.7. The following are some examples of M-estimators. R (i) If ρ(x, t) = (x − t)2 /2, then ψ(x, t) = t − x; T(G) = xdG(x) is the ¯ is the sample mean. mean functional; and T(Fn ) = X (ii) If ρ(x, t) = |x − t|p /p, where p ∈ [1, 2), then x≤t |x − t|p−1 ψ(x, t) = x > t. −|x − t|p−1 When p = 1, T(Fn ) is the sample median. When 1 < p < 2, T (Fn ) is called the pth least absolute deviations estimator or the minimum Lp distance estimator. (iii) Let F0 = {fθ : θ ∈ Θ} be a parametric family of p.d.f.’s with Θ ⊂ R and ρ(x, t) = − log ft (x). Then T(Fn ) is an MLE. This indicates that Mestimators are extensions of MLE’s in parametric models. (iv) Let C > 0 be a constant. Huber (1964) considers ( 1 2 |x − t| ≤ C 2 (x − t) ρ(x, t) = 1 2 |x − t| > C 2C with

t−x |x − t| ≤ C 0 |x − t| > C. The corresponding T(Fn ) is a type of trimmed sample mean. (v) Let C > 0 be a constant. Huber (1964) considers 1 2 |x − t| ≤ C 2 (x − t) ρ(x, t) = |x − t| > C C|x − t| − 12 C 2 ψ(x, t) =

t−x >C C ψ(x, t) = t−x |x − t| ≤ C −C t − x < −C. The corresponding T(Fn ) is a type of Winsorized sample mean. (vi) Hampel (1974) considers ψ(x, t) = ψ0 (t − x) with ψ0 (s) = −ψ0 (−s) and s 0≤s≤a a a<s≤b ψ0 (s) = a(c−s) b <s≤c c−b 0 s > c,

with

347

5.2. Statistical Functionals

where 0 < a < b < c are constants. A smoothed version of ψ0 is sin(as) 0 ≤ s < π/a ψ1 (s) = 0 s > π/a. For bounded and continuous ψ, the following result shows that T is ̺∞ Hadamard differentiable with a bounded and continuous influence function and, hence, T(Fn ) satisfies (5.41) and is robust in Hampel’s sense. Theorem 5.7. Let T be an M-functional defined by (5.50). Assume that ψ is a bounded and continuous function on Rd × R and that λF (t) is continuously differentiable at T(F ) and λ′F (T(F )) 6= 0. Then T is ̺∞ Hadamard differentiable at F with φF (x) = −ψ(x, T(F ))/λ′F (T(F )). Proof. Let tj → 0, ∆j ∈ D, k∆j − ∆k∞ → 0, and Gj = F + tj ∆j ∈ F. Since λG (T(G)) = 0, Z |λF (T(Gj )) − λF (T(F ))| = tj ψ(x, T(Gj ))d∆j (x) → 0

by k∆j − ∆k∞ → 0 and the boundedness of ψ. Note that λ′F (T(F )) 6= 0. Hence, the inverse of λF (t) exists and is continuous in a neighborhood of 0 = λF (T(F )). Therefore, T(Gj ) − T(F ) → 0.

(5.52)

Let hF (T(F )) = λ′F (T(F )), hF (t) = [λF (t) − λF (T(F ))]/[t − T(F )] if t 6= T(F ), Z 1 1 − , R1j = ψ(x, T(F ))d∆j (x) ′ λF (T(F )) hF (T(Gj )) Z 1 [ψ(x, T(Gj )) − ψ(x, T(F ))]d∆j (x), R2j = hF (T(Gj )) and LF (∆) = −

1 ′ λF (T(F ))

Z

ψ(x, T(F ))d∆(x),

∆ ∈ D.

Then T(Gj ) − T(F ) = −LF (tj ∆j ) + tj (R1j − R2j ). By (5.52), k∆j − ∆k∞ → 0, and the boundedness of ψ, Rj1 → 0. The result then follows from R2j → 0, which follows from k∆j − ∆k∞ → 0 and the boundedness and continuity of ψ (exercise).

348

5. Estimation in Nonparametric Models

Some ψ functions in Example 5.7 satisfy the conditions in Theorem 5.7 (exercise). Under more conditions on ψ, it can be shown that an Mfunctional is ̺∞ -Fr´echet differentiable at F (Clarke, 1986; Shao, 1993). Some M-estimators that satisfy (5.41) but are not differentiable functionals are studied in §5.4. Rank statistics and R-estimators Assume that X1 , ..., Xn are i.i.d. from a c.d.f. F on R. The rank of Xi among X1 , ..., Xn , denoted by Ri , is defined to be the number of Xj ’s satisfying Xj ≤ Xi , i = 1, ..., n. The rank of |Xi | among |X1 |, ..., |Xn | is ˜ i . A statistic that is a function of Ri ’s similarly defined and denoted by R ˜ i ’s is called a rank statistic. For G ∈ F, let or R ˜ G(x) = G(x) − G (−x)− , x > 0,

where g(x−) denotes the left limit of the function g at x. Define a functional T by Z ∞ ˜ T(G) = J(G(x))dG(x), G ∈ F, (5.53) 0

where J is a function on [0, 1] with a bounded derivative J ′ . Then Z ∞ n 1 X R˜ i T(Fn ) = J(F˜n (x))dFn (x) = J n I(0,∞) (Xi ) n i=1 0

is a (one-sample) signed rank statistic. If J(t) = t, then T(Fn ) is the wellknown Wilcoxon signed rank test statistic (§6.5.1). Statistics based on ranks (or signed ranks) are robust against changes in values of xi ’s, but may not provide efficient inference procedures, since the values of xi ’s are discarded after ranks (or signed ranks) are determined. It can be shown (exercise) that T in (5.53) is ̺∞ -Hadamard differentiable at F with the differential Z ∞ Z ∞ ˜ LF (∆) = J ′ (F˜ (x))∆(x)dF (x) + J(F˜ (x))d∆(x), (5.54) 0

0

˜ where ∆ ∈ D and ∆(x) = ∆(x) − ∆((−x)−). These results can be extended to the case where X1 , ..., Xn are i.i.d. from a c.d.f. F on R2 . For any c.d.f. G on R2 , let J be a function on [0, 1] with J(1 − t) = −J(t) and a bounded J ′ , ¯ G(y) = [G(y, ∞) + G(∞, y)]/2,

and T(G) =

Z

¯ J(G(y))dG(y, ∞).

y ∈ R, (5.55)

349

5.2. Statistical Functionals

Let Xi = (Yi , Zi ), Ri be the rank of Yi , and Ui be the number of Zj ’s satisfying Zj ≤ Yi , i = 1, ..., n. Then Z n 1X +Ui ¯ T(Fn ) = J(Fn (y))dFn (y, ∞) = J Ri2n n i=1 is called a two-sample linear rank statistic. It can be shown (exercise) that T in (5.55) is ̺∞ -Hadamard differentiable at F with the differential Z Z ¯ LF (∆) = J ′ (F¯ (y))∆(y)dF (y, ∞) + J(F¯ (y))d∆(y, ∞), (5.56)

¯ where ∆(y) = [∆(y, ∞) + ∆(∞, y)]/2. Rank statistics (one-sample or two-sample) are asymptotically normal and robust in Hampel’s sense (exercise). These results are useful in testing hypotheses (§6.5). Let F be a continuous c.d.f. on R symmetric about an unknown parameter θ ∈ R. An estimator of θ closely related to a rank statistic can be derived as follows. Let Xi be i.i.d. from F and Wi = (Xi , 2t − Xi ) with a fixed t ∈ R. The functional T in (5.55) evaluated at the c.d.f. of Wi is equal to Z (2t−x) λF (t) = J F (x)+1−F dF (x). (5.57) 2

If J is strictly increasing and F is strictly increasing in a neighborhood of θ, then λF (t) = 0 if and only if t = θ (exercise). For G ∈ F, define T(G) to be a solution of Z J G(x)+1−G(2T(G)−x) dG(x) = 0. (5.58) 2 T(Fn ) is called an R-estimator of T(F ) = θ. When J(t) = t − 12 (which is related to the Wilcoxon signed rank test), T(Fn ) is the well-known HodgesLehmann estimator and is equal to any value between the two middle points of the values (Xi + Xj )/2, i = 1, ..., n, j = 1, ..., n.

Theorem 5.8. Let T be the functional defined by (5.58). Suppose that F is continuous and symmetric about θ, the derivatives F ′ and J ′ exist, and J ′ is bounded. Then T is ̺∞ -Hadamard differentiable at F with the influence function φF (x) = R

J(F (x)) J ′ (F (x))F ′ (x)dF (x)

.

Proof. Since F is symmetric about θ, F (x) + RF (2θ − x) = 1. Under the assumed conditions, λF (t) is continuous and J ′ (F (x))F ′ (x)dF (x) = −λ′F (θ) 6= 0 (exercise). Hence, the inverse of λF exists and is continuous

350

5. Estimation in Nonparametric Models

at 0 = λF (θ). Suppose that tj → 0, ∆j ∈ D, k∆j − ∆k∞ → 0, and Gj = F + tj ∆j ∈ F. Then Z [J(Gj (x, t)) − J(F (x, t))]dGj (x) → 0 uniformly in t, where G(x, t) = [G(x) + 1 − G(2t − x)]/2, and Z Z J(F (x, t))d(Gj − F )(x) = (F − Gj )(x)J ′ (F (x, t))dF (x, t) → 0 uniformly in t. Let λG (t) be defined by (5.57) with F replaced by G. Then λGj (t) − λF (t) → 0 uniformly in t. Thus, λF (T(Gj )) → 0, which implies T(Gj ) → T(F ) = θ.

R

(5.59)

Let ξG (t) = J(F (x, t))dG(x), hF (t) = [λ R F (t) − λF (θ)]/(t − θ) if t 6= θ, and hF (θ) = λ′F (θ). Then T(Gj ) − T(F ) − φF (x)d(Gj − F )(x) is equal to λF (T(Gj )) − ξGj (θ) 1 1 − + . (5.60) ξGj (θ) ′ λF (θ) hF (T(Gj )) hF (T(Gj )) Note that ξGj (θ) =

Z

J(F (x))dGj (x) = tj

Z

J(F (x))d∆j (x).

By (5.59), Lemma 5.2, and k∆j − ∆k∞ → 0, the first term in (5.60) is o(tj ). The second term in (5.60) is the sum of Z tj − [J(F (x, T(Gj ))) − J(F (x))]d∆j (x) (5.61) hF (T(Gj )) and 1 hF (T(Gj ))

Z

[J(F (x, T(Gj ))) − J(Gj (x, T(Gj )))]dGj (x).

(5.62)

From the continuity of J and F , the quantity in (5.61) is o(tj ). Similarly, the quantity in (5.62) is equal to Z 1 [J(F (x, T(Gj ))) − J(Gj (x, T(Gj )))]dF (x) + o(tj ). (5.63) hF (T(Gj )) From Taylor’s expansion, (5.59), and k∆j − ∆k∞ → 0, the quantity in (5.63) is equal to Z tj J ′ (F (x))∆(x, θ)dF (x) + o(tj ). (5.64) hF (T(Gj ))

351

5.3. Linear Functions of Order Statistics

Since J(1 − t) = −J(t), the integral in (5.64) is 0. This proves that the second term in (5.60) is o(tj ) and thus the result. It is clear that the influence function φF for an R-estimator is bounded and continuous if J and F are continuous. Thus, R-estimators satisfy (5.41) and are robust in Hampel’s sense. Example 5.8. Let J(t) = t − 12 . Then T(Fn ) is the Hodges-Lehmann estiR mator. From Theorem 5.8, φF (x) = [F (x)− 21 ]/γ, where γ = F ′ (x)dF (x). Since F (X1 ) has a uniform √ distribution on [0, 1], φF (X1 ) has mean 0 and variance (12γ 2 )−1 . Thus, n[T(Fn ) − T(F )] →d N (0, (12γ 2)−1 ).

5.3 Linear Functions of Order Statistics In this section, we study statistics that are linear functions of order statistics X(1) ≤ · · · ≤ X(n) based on independent random variables X1 , ..., Xn (in §5.3.1 and §5.3.2, X1 , ..., Xn are assumed i.i.d.). Order statistics, first introduced in Example 2.9, are usually sufficient and often complete (or minimal sufficient) for nonparametric families (Examples 2.12 and 2.14). L-estimators defined in §5.2.2 are in fact linear functions of order statistics. If T is given by (5.46), then T(Fn ) =

Z

n

xJ(Fn (x))dFn (x) =

1X J n i=1

i n

X(i) ,

(5.65)

since Fn (X(i) ) = i/n, i = 1, ..., n. If J is a smooth function, such as those given in Example 5.6 or those satisfying the conditions in Theorem 5.6, the corresponding L-estimator is often called a smooth L-estimator. Asymptotic properties of smooth L-estimators can be obtained using Theorem 5.6 and the results in §5.2.1. Results on L-estimators that are slightly different from that in (5.65) can be found in Serfling (1980, Chapter 8). In §5.3.1, we consider another useful class of linear functions of order statistics, the sample quantiles described in the beginning of §5.2. In §5.3.2, we study robust linear functions of order statistics (in Hampel’s sense) ¯ an efficient but and their relative efficiencies w.r.t. the sample mean X, nonrobust estimator. In §5.3.3, extensions to linear models are discussed.

5.3.1 Sample quantiles Recall that G−1 (p) is defined to be inf{x : G(x) ≥ p} for any c.d.f. G on R, where p ∈ (0, 1) is a fixed constant. For i.i.d. X1 , ..., Xn from F , let θp = F −1 (p) and θˆp = Fn−1 (p) denote the pth quantile of F and the pth

352

5. Estimation in Nonparametric Models

sample quantile, respectively. Then θˆp = cnp X(mp ) + (1 − cnp )X(mp +1) ,

(5.66)

where mp is the integer part of np, cnp = 1 if np is an integer, and cnp = 0 if np is not an integer. Thus, θˆp is a linear function of order statistics. Note that F (θp −) ≤ p ≤ F (θp ) and F is not flat in a neighborhood of θp if and only if p < F (θp + ǫ) for any ǫ > 0. Theorem 5.9. Let X1 , ..., Xn be i.i.d. random variables from a c.d.f. F satisfying p < F (θp +ǫ) for any ǫ > 0. Then, for every ǫ > 0 and n = 1, 2,..., 2 (5.67) P |θˆp − θp | > ǫ ≤ 2Ce−2nδǫ ,

where δǫ is the smaller of F (θp + ǫ) − p and p − F (θp − ǫ) and C is the same constant in Lemma 5.1(i). Proof. Let ǫ > 0 be fixed. Note that G(x) ≥ t if and only if x ≥ G−1 (t) for any c.d.f. G on R (exercise). Hence P θˆp > θp + ǫ = P p > Fn (θp + ǫ) = P F (θp + ǫ) − Fn (θp + ǫ) > F (θp + ǫ) − p ≤ P ̺∞ (Fn , F ) > δǫ 2

≤ Ce−2nδǫ ,

where the last inequality follows from DKW’s inequality (Lemma 5.1(i)). Similarly, 2 P θˆp < θp − ǫ ≤ Ce−2nδǫ . This proves (5.67).

Result (5.67) implies that θˆp is strongly consistent for θp (exercise) and √ that θˆp is n-consistent for θp if F ′ (θp −) and F ′ (θp +) (the left and right derivatives of F at θp ) exist (exercise). The exact distribution of θˆp can be obtained as follows. Since nFn (t) has the binomial distribution Bi(F (t), n) for any t ∈ R, P θˆp ≤ t = P Fn (t) ≥ p n X n [F (t)]i [1 − F (t)]n−i , (5.68) = i i=lp

where lp = np if np is an integer and lp = 1+ the integer part of np if np is not an integer. If F has a Lebesgue p.d.f. f , then θˆp has the Lebesgue p.d.f. n−1 ϕn (t) = n [F (t)]lp −1 [1 − F (t)]n−lp f (t). (5.69) lp − 1

353

5.3. Linear Functions of Order Statistics

The following result provides an asymptotic distribution for

√ ˆ n(θp −θp ).

Theorem 5.10. Let X1 , ..., √ Xn be i.i.d. random variables from F . (i) If F (θp ) = p, then P ( n(θˆp − θp ) ≤ 0) → Φ(0) = 12 , where Φ is the c.d.f. of the standard normal. (ii) If F is continuous at θp and there exists F ′ (θp −) > 0, then P

√ n(θˆp − θp ) ≤ t → Φ(t/σ − ), F

t < 0,

p where σF− = p(1 − p)/F ′ (θp −). (iii) If F is continuous at θp and there exists F ′ (θp +) > 0, then P

√ n(θˆp − θp ) ≤ t → Φ(t/σF+ ),

t > 0,

p where σF+ = p(1 − p)/F ′ (θp +). (iv) If F ′ (θp ) exists and is positive, then

√ n(θˆp − θp ) →d N (0, σF2 ),

(5.70) p where σF = p(1 − p)/F ′ (θp ). Proof. The proof of (i) is left as an exercise. Part (iv) is a direct consequence of (i)-(iii) and the proofs of (ii) and (iii) are similar. Thus, we only give a proof for (iii). p √ Let t > 0, pnt = F (θp +p tσF+ n−1/2 ), cnt = n(pnt − p)/ pnt (1 − pnt ), and Znt = [Bn (pnt )−npnt ]/ npnt (1 − pnt ), where Bn (q) denotes a random variable having the binomial distribution Bi(q, n). Then P θˆp ≤ θp + tσF+ n−1/2 = P p ≤ Fn (θp + tσF+ n−1/2 ) = P Znt ≥ −cnt .

Under the assumed conditions on F , pnt → p and cnt → t. Hence, the result follows from P Znt < −cnt − Φ(−cnt ) → 0.

But this follows from the CLT (Example 1.33) and P´olya’s theorem (Proposition 1.16). If both F ′ (θp −) and F ′ (θp +) exist and are positive, but F ′ (θp −) 6= √ F (θp +), then the asymptotic distribution of n(θˆp − θp ) has the c.d.f. Φ(t/σF− )I(−∞,0) (t) + Φ(t/σF+ )I[0,∞) (t), a mixture of two normal distributions. An example of such a case when p = 12 is ′

F (x) = xI[0, 21 ) (x) + (2x − 12 )I[ 12 , 34 ) (x) + I[ 34 ,∞) (x).

354

5. Estimation in Nonparametric Models

When F ′ (θp −) = F ′ (θp +) = F ′ (θp ) > 0, (5.70) shows that the asymptotic √ √ distribution of n(θˆp −θp ) is the same as that of n[Fn (θp )−F (θp )]/F ′ (θp ) (see (5.2)). The following result reveals a stronger relationship between sample quantiles and the empirical c.d.f. Theorem 5.11 (Bahadur’s representation). Let X1 , ..., Xn be i.i.d. random variables from F . Suppose that F ′ (θp ) exists and is positive. Then F (θp ) − Fn (θp ) √1 + o θˆp = θp + . p n F ′ (θp )

(5.71)

Zn (t) − Zn (0) = op (1).

(5.72)

√ Proof. Let t ∈ R, θnt = θp +tn−1/2 , Zn (t) = n[F (θnt )−Fn (θnt )]/F ′ (θp ), √ and Un (t) = n[F (θnt ) − Fn (θˆp )]/F ′ (θp ). It can be shown (exercise) that

Note that |p − Fn (θˆp )| ≤ n−1 . Then Un (t) = =

√

√

→ t.

n[F (θnt ) − p + p − Fn (θˆp )]/F ′ (θp ) n[F (θnt ) − p]/F ′ (θp ) + O(n−1/2 )

(5.73)

√ ˆ n(θp − θp ). Then, for any t ∈ R and ǫ > 0, P ξn ≤ t, Zn (0) ≥ t + ǫ = P Zn (t) ≤ Un (t), Zn (0) ≥ t + ǫ (5.74) ≤ P |Zn (t) − Zn (0)| ≥ ǫ/2 + P |Un (t) − t| ≥ ǫ/2

Let ξn =

→0

by (5.72) and (5.73). Similarly, P ξn ≥ t + ǫ, Zn (0) ≤ t → 0.

(5.75)

It follows from the result in Exercise 128 of §1.6 that ξn − Zn (0) = op (1), which is the same as (5.71). If F has a positive Lebesgue p.d.f., then θˆp viewed as a statistical functional (§5.2) is ̺∞ -Hadamard differentiable at F (Fernholz, 1983) with the influence function φF (x) = [F (θp ) − I(−∞,θp ] (x)]/F ′ (θp ).

355

5.3. Linear Functions of Order Statistics

This implies result (5.71). Note that φF is bounded and is continuous except when x = θp . Corollary 5.1. Let X1 , ..., Xn be i.i.d. random variables from F having positive derivatives at θpj , where 0 < p1 < · · · < pm < 1 are fixed constants. Then √ n[(θˆp1 , ..., θˆpm ) − (θp1 , ..., θpm )] →d Nm (0, D), where D is the m × m symmetric matrix whose (i, j)th element is pi (1 − pj )/[F ′ (θpi )F ′ (θpj )],

i ≤ j.

The proof of this corollary is left to the reader. Example 5.9 (Interquartile range). One application of Corollary 5.1 is the derivation of the asymptotic distribution of the interquartile range θˆ0.75 − θˆ0.25 . The interquartile range is used as a measure of the variability among Xi ’s. It can be shown (exercise) that √ n[(θˆ0.75 − θˆ0.25 ) − (θ0.75 − θ0.25 )] →d N (0, σF2 ) with σF2 =

3 3 1 . + − 16[F ′ (θ0.75 )]2 16[F ′ (θ0.25 )]2 8F ′ (θ0.75 )F ′ (θ0.25 )

There are some applications of using extreme order statistics such as X(1) and X(n) . One example is given in Example 2.34. Some other examples and references can be found in Serfling (1980, pp. 89-91).

5.3.2 Robustness and efficiency Let F be a c.d.f. on R symmetric about θ ∈ R with F ′ (θ) > 0. Then θ = θ0.5 and is called the median of F . If F has a finite mean, then θ is also equal to the mean. In this section, we consider the estimation of θ based on i.i.d. Xi ’s from F . If F is normal, it has been shown in previous chapters that the sample ¯ is the UMVUE, MRIE, and MLE of θ, and is asymptotically mean X efficient. On the other hand, if F is the c.d.f. of the Cauchy distribution ¯ has the same distribution C(θ, 1), it follows from Exercise 78 in §1.6 that X ¯ as X1 , i.e., X is as variable as X1 , and is inconsistent as an estimator of θ. ¯ perform so differently? An important difference between Why does X the normal and Cauchy p.d.f.’s is that the former tends to 0 at the rate

356

5. Estimation in Nonparametric Models 2

the latter tends to 0 at the much slower rate e−x /2 as |x| → ∞, whereas R ¯ in the x−2 , which results in |x|dF (x) = ∞. The poor performance of X Cauchy case is due to the high probability of getting extreme observations ¯ is sensitive to large changes in a few of the Xi ’s. (Note and the fact that X R ¯ that X is not robust in Hampel’s sense, since the functional xdG(x) has an unbounded influence function at F .) This suggests the use of a robust estimator that discards some extreme observations. The sample median, which is defined to be the 50%th sample quantile θˆ0.5 described in §5.3.1, is insensitive to the behavior of F as |x| → ∞. Since both the sample mean and the sample median can be used to estimate θ, a natural question is when is one better than the other, using a criterion such as the amse. Unfortunately, a general answer does not exist, since the asymptotic relative efficiency between these two estimators depends on the unknown distribution F . If F does not have a finite vari¯ = ∞ and X ¯ may be inconsistent. In such a case the ance, then Var(X) sample median is certainly preferred, since θˆ0.5 is consistent and asymptotically normal as long as F ′ (θ) > 0, and may have a finite variance (Exercise 60). The following example, which compares the sample mean and median in some cases, shows that the sample median can be better even if Var(X1 ) < ∞. Example 5.10. Suppose that Var(X1 ) < ∞. Then, by the CLT, √ ¯ − θ) →d N (0, Var(X1 )). n(X By Theorem 5.10(iv), √ ˆ n(θ0.5 − θ) →d N (0, [2F ′ (θ)]−2 ). ¯ is Hence, the asymptotic relative efficiency of θˆ0.5 w.r.t. X e(F ) = 4[F ′ (θ)]2 Var(X1 ). √ (i) If F is the c.d.f. of N (θ, σ 2 ), then Var(X1 ) = σ 2 , F ′ (θ) = ( 2πσ)−1 , and e(F ) = 2/π = 0.637. (ii) If F is the c.d.f. of the logistic distribution LG(θ, σ), then Var(X1 ) = σ 2 π 2 /3, F ′ (θ) = (4σ)−1 , and e(F ) = π 2 /12 = 0.822. (iii) If F (x) = F0 (x − θ) and F0 is the c.d.f. of the√t-distribution tν with ν ν ≥ 3, then Var(X1 ) = ν/(ν − 2), F ′ (θ) = Γ( ν+1 2 )/[ νπΓ( 2 )], e(F ) = 1.62 when ν = 3, e(F ) = 1.12 when ν = 4, and e(F ) = 0.96 when ν = 5. (iv) If F is the c.d.f. of the double exponential distribution DE(θ, σ), then F ′ (θ) = (2σ)−1 and e(F ) = 2. (v) Consider the Tukey model F (x) = (1 − ǫ)Φ x−θ + ǫΦ x−θ (5.76) σ τσ ,

357

5.3. Linear Functions of Order Statistics

2 2 2 where σ > 0, τ > 0, and √ 0 < ǫ < 1. Then Var(X1 ) = 2(1 − ǫ)σ + ǫτ 2 σ , ′ F (θ) = (1 − ǫ + ǫ/τ )/( 2πσ), and e(F ) = 2(1 − ǫ + ǫτ )(1 − ǫ + ǫ/τ ) /π. Note that limǫ→0 e(F ) = 2/π and limτ →∞ e(F ) = ∞.

Since the sample median uses at most two actual values of xi ’s, it may go too far in discarding observations, which results in a possible loss of efficiency. The trimmed sample mean introduced in Example 5.6(iii) is a natural compromise between the sample mean and median. Since F is symmetric, we consider β = 1 − α in the trimmed mean, which results in the following L-estimator: ¯α = X

n−m Xα 1 X(j) , (1 − 2α)n j=m +1

(5.77)

α

where mα is the integer part of nα and α ∈ (0, 12 ). The estimator in (5.77) is called the α-trimmed sample mean. It discards the mα smallest and mα largest observations. The sample mean and median can be viewed as two ¯ α as α → 0 and 1 , respectively. extreme cases of X 2 It follows from Theorem 5.6 that if F (x) = F0 (x − θ), where F0 is symmetric about 0 and has a Lebesgue p.d.f. positive in the range of X1 , then √ ¯ α − θ) →d N (0, σ 2 ), n(X (5.78) α where

σα2

2 = (1 − 2α)2

(Z

0

F0−1 (1−α)

2

x dF0 (x) +

α[F0−1 (1

2

− α)]

)

.

Lehmann (1983, §5.4) provides various values of the asymptotic relative efficiency eX¯ α ,X¯ (F ) = Var(X1 )/σα2 . For instance, when F (x) = F0 (x − θ) and F0 is the c.d.f. of the t-distribution t3 , eX¯ α ,X¯ (F ) = 1.70, 1.91, and 1.97 for α = 0.05, 0.125, and 0.25, respectively; when F is given by (5.76) with τ = 3 and ǫ = 0.05, eX¯ α ,X¯ (F ) = 1.20, 1.19, and 1.09 for α = 0.05, 0.125, and 0.25, respectively; when F is given by (5.76) with τ = 3 and ǫ = 0.01, eX¯ α ,X¯ (F ) = 1.04, 0.98, and 0.89 for α = 0.05, 0.125, and 0.25, respectively. Robustness and efficiency of other L-estimators can be discussed similarly. For an L-estimator T(Fn ) with T given by (5.46), if the conditions in one of (i)-(iii) of Theorem 5.6 are satisfied, then (5.41) holds with Z ∞Z ∞ σF2 = J(F (x))J(F (y))[F (min{x, y}) − F (x)F (y)]dxdy, (5.79) −∞

−∞

provided that σF2 < ∞ (exercise). If F is symmetric about θ, J is symmetric R1 about 21 , and 0 J(t)dt = 1, then T(F ) = θ (exercise) and, therefore, the ¯ is Var(X1 )/σ 2 . asymptotic relative efficiency of T(Fn ) w.r.t. X F

358

5. Estimation in Nonparametric Models

5.3.3 L-estimators in linear models In this section, we extend L-estimators to the following linear model: Xi = β τ Z i + ε i ,

i = 1, ..., n,

(5.80)

with i.i.d. εi ’s having an unknown c.d.f. F0 and a full rank Z whose ith row is the vector Zi . Note that the c.d.f. of Xi is F0 (x − β τ Zi ). Instead of assuming E(εi ) = 0 (as we did in Chapter 3), we assume that Z (5.81) xJ(F0 (x))dF0 (x) = 0, where J is a Borel function on [0, 1] (the same as that in (5.46)). Note that (5.81) may hold without any assumption on the existence of E(εi ). For instance, (5.81) holds if F0 is symmetric about 0, J is symmetric about 21 , R1 and 0 J(t)dt = 1 (Exercise 69). Since Xi ’s are not identically distributed, the use of the order statistics and the empirical c.d.f. based on X1 , ..., Xn may not be appropriate. Inˆ i = 1, ..., n, stead, we consider the ordered values of residuals ri = Xi −Ziτ β, τ ˆ and some empirical c.d.f.’s based on residuals, where β = (Z Z)−1 Z τ X is the LSE of β (§3.3.1). To illustrate the idea, let us start with the case where β and Zi are univariate. First, assume thatP Zi ≥ 0 for all i (or Zi ≤ 0 for all i). Let Fˆ0 be the c.d.f. putting mass Zi / ni=1 Zi at ri , i = 1, ..., n. An L-estimator of β is defined to be X Z n n X ˆ ˆ ˆ ˆ βL = β + xJ(F0 (x))dF0 (x) Zi Zi2 . i=1

i=1

When J(t) = (1 − 2α)−1 I(α,1−α) (t) with an α ∈ (0, 21 ), βˆL is similar to the α-trimmed sample mean in the i.i.d. case. If not all Zi ’s have the same sign, we can define L-estimators as follows. Let Zi+ = P max{Zi , 0} and Zi− = Zi+ − Zi . Let Fˆ0± be the c.d.f. putting ± mass Zi / ni=1 Zi± at ri , i = 1, ..., n. An L-estimator of β is defined to be βˆL = βˆ + −

Z

Z

xJ(Fˆ0+ (x))dFˆ0+ (x)

xJ(Fˆ0− (x))dFˆ0− (x)

n X

Zi+

i=1

n X i=1

Zi−

X n

Zi2

i=1

X n

Zi2 .

i=1

For a general p-vector Zi , let zij be the jth component of Zi , j = 1, ..., p. + − + ± = max{zij , 0}, zij = zij − zij , and Fˆ0j be the c.d.f. putting mass Let zij

359

5.4. Generalized Estimating Equations

± Pn ± / i=1 zij at ri , i = 1, ..., n. For any j, if zij ≥ 0 for all i (or zij ≤ 0 for zij + − all i), then we set Fˆ0j ≡ 0 (or Fˆ0j ≡ 0). An L-estimator of β is defined to be βˆL = βˆ + (Z τ Z)−1 (A+ − A− ), (5.82)

where

A± =

Z

± ± xJ(Fˆ01 (x))dFˆ01 (x)

n X i=1

± zi1 , ...,

Z

± ± xJ(Fˆ0p (x))dFˆ0p (x)

n X i=1

± zip

!

.

Obviously, βˆL in (5.82) reduces to the previously defined βˆL when β and Zi are univariate. Theorem 5.12. Assume model (5.80) with i.i.d. εi ’s from a c.d.f. F0 satisfying (5.81) for a given J. Suppose that F0 has a uniformly continuous, positive, and bounded derivative on the range of ε1 . Suppose further that the conditions on Zi ’s in Theorem 3.12 are satisfied. (i) If the function J is continuous on (α1 , α2 ) and equals 0 on [0, α1 ]∪[α2 , 1], where 0 < α1 < α2 < 1 are constants, then σF−1 (Z τ Z)1/2 (βˆL − β) →d Np (0, Ip ), 0

(5.83)

where σF2 0 is given by (5.79) with F = F0 . (ii) Result (5.83) also holds if J ′ is bounded on [0, 1], E|ε1 | < ∞, and σF2 0 is finite. The proof of this theorem can be found in Bickel (1973). Robustness and efficiency comparisons between the LSE βˆ and L-estimators βˆL can be made in a way similar to those in §5.3.2.

5.4 Generalized Estimating Equations The method of generalized estimating equations (GEE) is a powerful and general method of deriving point estimators, which includes many previously described methods as special cases. In §5.4.1, we begin with a description of this method and, to motivate the idea, we discuss its relationship with other methods that have been studied. Consistency and asymptotic normality of estimators derived from generalized estimating equations are studied in §5.4.2 and §5.4.3. Throughout this section, we assume that X1 , ..., Xn are independent (not necessarily identically distributed) random vectors, where the dimension of Xi is di , i = 1, ..., n (supi di < ∞), and that we are interested in estimating θ, a k-vector of unknown parameters related to the unknown population.

360

5. Estimation in Nonparametric Models

5.4.1 The GEE method and its relationship with others The sample mean and, more generally, the LSE in linear models are solutions of equations of the form n X i=1

(Xi − γ τ Zi )Zi = 0.

Also, MLE’s (or RLE’s) in §4.4 and, more generally, M-estimators in §5.2.2 are solutions to equations of the form n X

ψ(Xi , γ) = 0.

n X

ψi (Xi , γ),

i=1

This leads to the following general estimation method. Let Θ ⊂ Rk be the range of θ, ψi be a Borel function from Rdi × Θ to Rk , i = 1, ..., n, and sn (γ) =

i=1

γ ∈ Θ.

(5.84)

ˆ = 0, then θˆ is called a GEE If θ is estimated by θˆ ∈ Θ satisfying sn (θ) estimator. The equation sn (γ) = 0 is called a GEE. Apparently, the LSE’s, RLE’s, MQLE’s, and M-estimators are special cases of GEE estimators. Usually GEE’s are chosen so that E[sn (θ)] =

n X

E[ψi (Xi , θ)] = 0,

(5.85)

i=1

where the expectation E may be replaced by an asymptotic expectation defined in §2.5.2 if the exact expectation does not exist. If this is true, ˆ = 0 is a sample analogue of then θˆ is motivated by the fact that sn (θ) E[sn (θ)] = 0. To motivate the idea, let us study the relationship between the GEE method and other methods that have been introduced. M-estimators The M-estimators defined in §5.2.2 for univariate θ = T(F ) in the i.i.d. case are special cases of GEE estimators. Huber (1981) also considers regression M-estimators in the linear model (5.80). A regression M-estimator of β is defined as a solution to the GEE n X ψ(Xi − γ τ Zi )Zi = 0, i=1

where ψ is one of the functions given in Example 5.7.

361

5.4. Generalized Estimating Equations

LSE’s in linear and nonlinear regression models Suppose that Xi = f (Zi , θ) + εi ,

i = 1, ..., n,

(5.86)

where Zi ’s are the same as those in (5.80), θ is an unknown k-vector of parameters, f is a known function, and εi ’s are independent random variables. Model (5.86) is the same as model (5.80) if f is linear in θ and is called a nonlinear regression model otherwise. Note that model (4.64) is a special case of P model (5.86). The LSE under model (5.86) is any point in n Θ minimizing i=1 [Xi − f (Zi , γ)]2 over γ ∈ Θ. If f is differentiable, then the LSE is a solution to the GEE n X i=1

[Xi − f (Zi , γ)]

∂f (Zi , γ) = 0. ∂γ

Quasi-likelihoods This is a continuation of the discussion of the quasi-likelihoods introduced in §4.4.3. Assume first that Xi ’s are univariate (di ≡ 1). If Xi ’s follow a GLM, i.e., Xi has the p.d.f. in (4.55) and (4.57) holds, and if (4.58) holds, then the likelihood equation (4.59) can be written as n X xi − µi (γ) i=1

vi (γ)

Gi (γ) = 0,

(5.87)

where µi (γ) = µ(ψ(γ τ Zi )), Gi (γ) = ∂µi (γ)/∂γ, vi (γ) = Var(Xi )/φ, and we have used the following fact: ψ ′ (t) = (µ−1 )′ (g −1 (t))(g −1 )′ (t) = (g −1 )′ (t)/ζ ′′ (ψ(t)). Equation (5.87) is a quasi-likelihood equation if either Xi does not have the p.d.f. in (4.55) or (4.58) does not hold. Note that this generalizes the discussion in §4.4.3. If Xi does not have the p.d.f. in (4.55), then the problem is often nonparametric. Let sn (γ) be the left-hand side of (5.87). Then sn (γ) = 0 is a GEE and E[sn (β)] = 0 is satisfied as long as the first condition in (4.56), E(Xi ) = µi (β), is satisfied. For general di ’s, let Xi = (Xi1 , ..., Xidi ), i = 1, ..., n, where each Xit satisfies (4.56) and (4.57), i.e., E(Xit ) = µ(ηit ) = g −1 (β τ Zit ) and Var(Xit ) = φi µ′ (ηit ), and Zit ’s are k-vector values of covariates. In biostatistics and life-time testing problems, components of Xi are repeated measurements at different times from subject i and are called longitudinal data. Although Xi ’s are

362

5. Estimation in Nonparametric Models

assumed independent, Xit ’s are likely to be dependent for each i. Let Ri be the di × di correlation matrix whose (t, l)th element is the correlation coefficient between Xit and Xil . Then Var(Xi ) = φi [Di (β)]1/2 Ri [Di (β)]1/2 ,

(5.88)

where Di (γ) is the di × di diagonal matrix with the tth diagonal element (g −1 )′ (γ τ Zit ). If Ri ’s in (5.88) are known, then an extension of (5.87) to the multivariate xi ’s is n X i=1

Gi (γ){[Di (γ)]1/2 Ri [Di (γ)]1/2 }−1 [xi − µi (γ)] = 0,

(5.89)

where µi (γ) = (µ(ψ(γ τ Zi1 )), ..., µ(ψ(γ τ Zidi ))) and Gi (γ) = ∂µi (γ)/∂γ. In ˜ i be a most applications, Ri is unknown and its form is hard to model. Let R known correlation matrix (called a working correlation matrix). Replacing ˜ i leads to the quasi-likelihood equation Ri in (5.89) by R n X i=1

˜ i [Di (γ)]1/2 }−1 [xi − µi (γ)] = 0. Gi (γ){[Di (γ)]1/2 R

(5.90)

For example, we may assume that the components of Xi are independent ˜ i = Idi . Although the working correlation matrix R ˜ i may not be and take R the same as the true unknown correlation matrix Ri , an MQLE obtained from (5.90) is still consistent and asymptotically normal (§5.4.2 and §5.4.3). ˜ i is closer to Ri . Of course, MQLE’s are asymptotically more efficient if R ˜ Even if Ri = Ri and φi ≡ φ, (5.90) is still a quasi-likelihood equation, since the covariance matrix of Xi cannot determine the distribution of Xi unless Xi is normal. ˜ i closer to Ri results in a better MQLE, sometimes it is Since an R ˜ i in (5.90) by R ˆ i , an estimator of Ri (Liang and suggested to replace R Zeger, 1986). The resulting equation is called a pseudo-likelihood p equation. ˆ i − Ui k →p 0 as n → ∞, where kAk = tr(Aτ A) for As long as maxi≤n kR a matrix A and Ui is a correlation matrix (not necessarily the same as Ri ), i = 1, ..., n, MQLE’s are consistent and asymptotically normal. Empirical likelihoods The previous discussion shows that the GEE method coincides with the method of deriving M-estimators, LSE’s, MLE’s, or MQLE’s. The following discussion indicates that the GEE method is also closely related to the method of empirical likelihoods introduced in §5.1.4. Assume that Xi ’s are i.i.d. from a c.d.f. F on Rd and ψi = ψ for all i. Then condition (5.85) reduces to E[ψ(X1 , θ)] = 0. Hence, we can consider

363

5.4. Generalized Estimating Equations

the empirical likelihood ℓ(G) =

n Y

i=1

subject to pi ≥ 0,

n X

PG ({xi }),

pi = 1,

i=1

and

G∈F n X

pi ψ(xi , θ) = 0,

(5.91)

i=1

where pi = PG ({xi }). However, in this case the dimension of the function ψ is the same as the dimension of the parameter θ and, hence, the last equation in (5.91) does not impose any restriction on pi ’s. Then, it follows from Theorem 5.3 that (p1 , ..., pn ) = (n−1 , ..., n−1 ) maximizes ℓ(G) for any fixed θ. Substituting pi = n−1 into the last equation in (5.91) leads to n

1X ψ(xi , θ) = 0. n i=1

That is, any MELE θˆ of θ is a GEE estimator.

5.4.2 Consistency of GEE estimators We now study under what conditions (besides (5.85)) GEE estimators are consistent. For each n, let θˆn be a GEE estimator, i.e., sn (θˆn ) = 0, where sn (γ) is defined by (5.84). First, Theorem 5.7 and its proof can be extended to multivariate T in a straightforward manner. Hence, we have the following result. Proposition 5.2. Suppose that X1 , ..., Xn are i.i.d. from F and ψi ≡ d k Rψ, a bounded and continuous function from R × Θ to R . Let Ψ(t) = ψ(x, t)dF (x). Suppose that Ψ(θ) = 0 and ∂Ψ(t)/∂t exists and is of full rank at t = θ. Then θˆn →p θ. For unbounded ψ in the i.i.d. case, the following result and its proof can be found in Qin and Lawless (1994). Proposition 5.3. Suppose that X1 , ..., Xn are i.i.d. from F and ψi ≡ ψ. Assume that ϕ(x, γ) = ∂ψ(x, γ)/∂γ exists in Nθ , a neighborhood of θ, and is continuous at θ; there is a function h(x) such that supγ∈Nθ kϕ(x, γ)k ≤ h(x), supγ∈Nθ kψ(x, γ)k3 ≤ h(x), and E[h(X1 )] < ∞; E[ϕ(X1 , θ)] is of full rank; E{ψ(X1 , θ)[ψ(X1 , θ)]τ } is positive definite; and (5.85) holds. Then, there exists a sequence of random vectors {θˆn } such that and θˆn →p θ. (5.92) P sn (θˆn ) = 0 → 1

364

5. Estimation in Nonparametric Models

Next, we consider non-i.i.d. Xi ’s. Proposition 5.4. Suppose that X1 , ..., Xn are independent and θ is univariate. Assume that ψi (x, γ) is real-valued and nonincreasing in γ for all i; there is a δ > 0 such that supi E|ψi (Xi , γ)|1+δ < ∞ for any γ in Nθ , a neighborhood of θ (this condition can be replaced by E|ψ(X1 , γ)| < ∞ for any γ in Nθ when Xi ’s are i.i.d. and ψi ≡ ψ); ψi (x, γ) are continuous in Nθ ; (5.85) holds; and lim sup E[Ψn (θ + ǫ)] < 0 < lim inf E[Ψn (θ − ǫ)] n

n

(5.93)

for any ǫ > 0, where Ψn (γ) = n−1 sn (γ). Then, there exists a sequence of random variables {θˆn } such that (5.92) holds. Furthermore, any sequence {θˆn } satisfying sn (θˆn ) = 0 satisfies (5.92). Proof. Since ψi ’s are nonincreasing, the functions Ψn (γ) and E[Ψn (γ)] are nonincreasing. Let ǫ > 0 be fixed so that θ ± ǫ ∈ Nθ . Under the assumed conditions, Ψn (θ ± ǫ) − E[Ψn (θ ± ǫ)] →p 0 (Theorem 1.14(ii)). By condition (5.93), P Ψn (θ + ǫ) < 0 < Ψn (θ − ǫ) → 1.

The rest of the proof is left as an exercise.

To establish the next result, we need the following lemma. First, we need the following concept. A sequence of functions {gi } from Rk to Rk is called equicontinuous on an open set O ⊂ Rk if and only if, for any ǫ > 0, there is a δǫ > 0 such that supi kgi (t) − gi (s)k < ǫ whenever t ∈ O, s ∈ O, and kt − sk < δǫ . Since a continuous function on a compact set is uniformly continuous, functions such as gi (γ) = g(ti , γ) form an equicontinuous sequence on O if ti ’s vary in a compact set containing O and g(t, γ) is a continuous function in (t, γ). Lemma 5.3. Suppose that Θ is a compact subset of Rk . Let hi (Xi ) = supγ∈Θ kψi (Xi , γ)k, i = 1, 2,.... Suppose that supi E|hi (Xi )|1+δ < ∞ and supi EkXi kδ < ∞ for some δ > 0 (this condition can be replaced by E|h(X1 )| < ∞ when Xi ’s are i.i.d. and ψi ≡ ψ). Suppose further that for any c > 0 and sequence {xi } satisfying kxi k ≤ c, the sequence of functions {gi (γ) = ψi (xi , γ)} is equicontinuous on any open subset of Θ. Then

n

1 X

{ψi (Xi , γ) − E[ψi (Xi , γ)]} sup

→p 0. n γ∈Θ i=1

365

5.4. Generalized Estimating Equations

Proof. Since we only need to consider components of ψi ’s, without loss of generality we can assume that ψi ’s are functions from Rdi × Θ to R. For any c > 0, " n # 1X sup E hi (Xi )I(c,∞) (kXi k) ≤ sup E[hi (Xi )I(c,∞) (kXi k)]. n i=1 n i Let c0 = supi E|hi (Xi )|1+δ and c1 = supi EkXi kδ . By H¨older’s inequality, 1/(1+δ) δ/(1+δ) [P (kXi k > c)] E[hi (Xi )I(c,∞) (kXi k)] ≤ E|hi (Xi )|1+δ 1/(1+δ) δ/(1+δ) −δ 2 /(1+δ) c1 c

≤ c0

1/(1+δ) δ/(1+δ)

2

for all i. For ǫ > 0 and ǫ˜ > 0, choose a c such that c0 c1 c−δ /(1+δ) < ǫ˜ ǫ/4. Then, for any O ⊂ Θ, the probability ! ( ) n ǫ 1X (5.94) sup ψi (Xi , γ) − inf ψi (Xi , γ) I(c,∞) (kXi k) > P γ∈O n i=1 γ∈O 2 is bounded by ǫ˜ (exercise). From the equicontinuity of {ψi (xi , γ)}, there is a δǫ > 0 such that ( ) n 1X ǫ sup ψi (Xi , γ) − inf ψi (Xi , γ) I[0,c] (kXi k) < γ∈Oǫ n i=1 γ∈Oǫ 2 for sufficiently large n, where Oǫ denotes any open ball in Rk with radius less than δǫ . These results, together with Theorem 1.14(ii) and the fact that kψi (Xi , γ)k ≤ hi (Xi ), imply that ( ! ) n 1X sup ψi (Xi , γ) − E inf ψi (Xi , γ) > ǫ → 0. (5.95) P γ∈Oǫ n i=1 γ∈Oǫ Let Hn (γ) = n−1

Pn

i=1 {ψi (Xi , γ) n

1X sup Hn (γ) ≤ n i=1 γ∈Oǫ

(

− E[ψi (Xi , γ)]}. Then

sup ψi (Xi , γ) − E

γ∈Oǫ

) inf ψi (Xi , γ) ,

γ∈Oǫ

which with (5.95) implies that

P Hn (γ) > ǫ for all γ ∈ Oǫ = P

sup Hn (γ) > ǫ γ∈Oǫ

Similarly we can show that P Hn (γ) < −ǫ for all γ ∈ Oǫ → 0.

!

→ 0.

366

5. Estimation in Nonparametric Models

Since Θ is compact, there exists mǫ open balls Oǫ,j such that Θ ⊂ ∪Oǫ,j . Then, the result follows from ! X mǫ P sup |Hn (γ)| > ǫ ≤ P sup |Hn (γ)| > ǫ → 0. γ∈Θ

j=1

γ∈Oǫ,j

˜i} Example 5.11. Consider the quasi-likelihood equation (5.90). Let {R be a sequence of working correlation matrices and ˜ i [Di (γ)]1/2 }−1 [xi − µi (γ)]. ψi (xi , γ) = Gi (γ){[Di (γ)]1/2 R

(5.96)

It can be shown (exercise) that ψi ’s satisfy the conditions of Lemma 5.3 if Θ is compact and supi kZi k < ∞. Proposition 5.5. Assume (5.85) and the conditions in Lemma 5.3 (with Θ replaced by any compact subset of the parameter space). Suppose that the functions ∆n (γ) = E[n−1 sn (γ)] have the property that limn→∞ ∆n (γ) = 0 if and only if γ = θ. (If ∆n converges to a function ∆, then this condition and (5.85) imply that ∆ has a unique 0 at θ.) Suppose that {θˆn } is a sequence of GEE estimators and that θˆn = Op (1). Then θˆn →p θ. Proof. First, assume that Θ is a compact subset of Rk . From Lemma 5.3 and sn (θˆn ) = 0, ∆n (θˆn ) →p 0. By Theorem 1.8(vi), there is a subsequence {ni } such that (5.97) ∆ni (θˆni ) →a.s. 0.

Let x1 , x2 , ... be a fixed sequence such that (5.97) holds and let θ0 be a limit point of {θˆn }. Since Θ is compact, θ0 ∈ Θ and there is a subsequence {mj } ⊂ {ni } such that θˆmj → θ0 . Using the argument in the proof of Lemma 5.3, it can be shown (exercise) that {∆n (γ)} is equicontinuous on any open subset of Θ. Then ∆mj (θˆmj ) − ∆mj (θ0 ) → 0, which with (5.97) implies ∆mj (θ0 ) → 0. Under the assumed condition, θ0 = θ. Since this is true for any limit point of {θˆn }, θˆn →p θ. Next, consider a general Θ. For any ǫ > 0, there is an Mǫ > 0 such that P (kθˆn k ≤ Mǫ ) > 1 − ǫ. The result follows from the previous proof by considering the closure of Θ ∩ {γ : kγk ≤ Mǫ } as the parameter space.

Condition θˆn = Op (1) in Proposition 5.5 is obviously necessary for the consistency of θˆn . It has to be checked in any particular problem. If a GEE is a likelihood equation under some conditions, then we can often show, using an argument similar to the proof of Theorem 4.17 or 4.18, that there exists a consistent sequence of GEE estimators.

5.4. Generalized Estimating Equations

367

Proposition 5.6. Suppose that sn (γ) = ∂ log ℓn (γ)/∂γ for some function ℓn ; Dn (θ) = Var(sn (θ)) → 0; ϕi (x, γ) = ∂ψi (x, γ)/∂γ exists and the sequence of functions {ϕij , i = 1, 2, ...} satisfies the conditions in Lemma 5.3 with Θ replaced by a compact neighborhood of θ, where ϕij is the jth row of ϕi , j = 1, ..., k; − lim inf n [Dn (θ)]1/2 E[∇sn (θ)][Dn (θ)]1/2 is positive definite, where ∇sn (γ) = ∂sn (γ)/∂γ; and (5.85) holds. Then, there exists a sequence of estimators {θˆn } satisfying (5.92). The proof of Proposition 5.6 is similar to that of Theorem 4.17 or Theorem 4.18 and is left as an exercise. ˜i = Example 5.12. Consider the quasi-likelihood equation (5.90) with R Idi for all i. Then the GEE is a likelihood equation under a GLM (§4.4.2) assumption. It can be shown (exercise) that the conditions of Proposition 5.6 are satisfied if supi kZi k < ∞.

5.4.3 Asymptotic normality of GEE estimators Asymptotic normality of a consistent sequence of GEE estimators can be established under some conditions. We first consider the special case where θ is univariate and X1 , ..., Xn are i.i.d. Theorem 5.13. LetR X1 , ..., Xn be i.i.d. from F , ψi ≡ ψ, and θ ∈ R. Suppose that Ψ(γ) = ψ(x, γ)dF (x) = 0 if and only if γ = θ, Ψ′ (θ) exists and Ψ′ (θ) 6= 0. R (i) Assume that ψ(x, γ) is nonincreasing in γ and that [ψ(x, γ)]2 dF (x) is finite for γ in a neighborhood of θ and is continuous at θ. Then, any sequence of GEE estimators (M-estimators) {θˆn } satisfies √ n(θˆn − θ) →d N (0, σF2 ), (5.98) where σF2

=

Z

[ψ(x, θ)]2 dF (x)/[Ψ′ (θ)]2 .

R (ii) Assume that [ψ(x, θ)]2 dF (x) < ∞, ψ(x, γ) is continuous in x, and limγ→θ kψ(·, γ) − ψ(·, θ)kV = 0, where k · kV is the variation norm defined in Lemma 5.2. Then, any consistent sequence of GEE estimators {θˆn } satisfies (5.98). Proof. (i) Let Ψn (γ) = n−1 sn (γ). Since Ψn is nonincreasing, P (Ψn (t) < 0) ≤ P (θˆn ≤ t) ≤ P (Ψn (t) ≤ 0) for any t ∈ R. Then, (5.98) follows from lim P Ψn (tn ) < 0 = lim P Ψn (tn ) ≤ 0 = Φ(t) n→∞

n→∞

368

5. Estimation in Nonparametric Models

for all t ∈ R, where tn = θ + tσF n−1/2 . Let s2t,n = Var(ψ(X1 , tn )) and Yni = [ψ(Xi , tn ) − Ψ(tn )]/st,n . Then, it suffices to show that ! √ n nΨ(tn ) 1 X = Φ(t) lim P √ Yni ≤ − n→∞ st,n n i=1

√ for all t. Under the assumed conditions, nΨ(tn ) → Ψ′ (θ)tσF and st,n → −Ψ′ (θ)σF . Hence, it suffices to show that n

1 X √ Yni →d N (0, 1). n i=1 Note that Yn1 , ..., Ynn are i.i.d. random variables. Hence we can apply Lindeberg’s CLT (Theorem 1.15). In this case, Lindeberg’s condition (1.92) is implied by Z lim [ψ(x, tn )]2 dF (x) = 0 √ n→∞

|ψ(x,tn )|> nǫ

for any ǫ > 0. For any η > 0, ψ(x, θ + η) ≤ ψ(x, tn ) ≤ ψ(x, θ − η) for all x and sufficiently large n. Let u(x) = max{|ψ(x, θ − η)|, |ψ(x, θ + η)|}. Then Z Z 2 [ψ(x, t )] dF (x) ≤ [u(x)]2 dF (x), n √ √ |ψ(x,tn )|> nǫ

u(x)> nǫ

R

which converges to 0 since [ψ(x, γ)]2 dF (x) is finite for γ in a neighborhood of θ. This proves (i). (ii) Let φF (x) = −ψ(x, θ)/Ψ′ (θ). Following the proof of Theorem 5.7, we have n √ 1 X n(θˆn − θ) = √ φF (Xi ) + R1n − R2n , n i=1 where

" # n 1 1 X 1 − ψ(Xi , θ) , R1n = √ n i=1 Ψ′ (θ) hF (θˆn ) √ Z n [ψ(x, θˆn ) − ψ(x, θ)]d(Fn − F )(x), R2n = hF (θˆn )

and hF is defined in the proof of Theorem 5.7 with Ψ = λF . By the CLT and the consistency of θˆn , R1n = op (1). Hence, the result follows if we can show that R2n = op (1). By Lemma 5.2, √ |R2n | ≤ n|hF (θˆn )|−1 ̺∞ (Fn , F )kψ(·, θˆn ) − ψ(·, θ)kV . The result follows from the assumed condition on ψ and the fact that √ n̺∞ (Fn , F ) = Op (1) (Theorem 5.1).

369

5.4. Generalized Estimating Equations

Note that the result in Theorem 5.13 coincides with the result in Theorem 5.7 and (5.41). Example 5.13. Consider the M-estimators given in Example 5.7 based on i.i.d. random variables X1 , ..., Xn . If ψ is bounded and continuous, then Theorem 5.7 applies and (5.98) holds. For case (ii), ψ(x, γ) is not bounded but is nondecreasing in γ (−ψ(x, γ) is nonincreasing in γ). Hence Theorem 5.13 can be applied to this case. Consider Huber’s ψ given in Example 5.7(v). Assume that F is continuous at θ − C and θ + C. Then Z γ+C Ψ(γ) = (γ − x)dF (x) + CF (γ − C) − C[1 − F (γ + C)] γ−C

is differentiable at θ (exercise); Ψ(θ) = 0 if F is symmetric about θ (exercise); and Z

2

[ψ(x, γ)] dF (x) =

Z

γ+C

γ−C

(γ − x)2 dF (x) + C 2 F (γ − C) + C 2 [1 − F (γ + C)]

is continuous at θ (exercise). Therefore, (5.98) holds with σF2

=

R θ+C θ−C

(θ − x)2 dF (x) + C 2 F (θ − C) + C 2 [1 − F (θ + C)] [F (θ + C) − F (θ − C)]2

(exercise). Note that Huber’s M-estimator is robust in Hampel’s sense. ¯ can be obAsymptotic relative efficiency of θˆn w.r.t. the sample mean X tained (exercise). The next result is for general θ and independent Xi ’s. Theorem 5.14. Suppose that ϕi (x, γ) = ∂ψi (x, γ)/∂γ exists and the sequence of functions {ϕij , i = 1, 2, ...} satisfies the conditions in Lemma 5.3 with Θ replaced by a compact neighborhood of θ, where ϕij is the jth row of ϕi ; supi Ekψi (Xi , θ)k2+δ < ∞ for some δ > 0 (this condition can be replaced by Ekψ(X1 , θ)k2 < ∞ if Xi ’s are i.i.d. and ψi ≡ ψ); E[ψi (Xi , θ)] = 0; lim inf n λ− [n−1 Var(sn (θ))] > 0 and lim inf n λ− [n−1 Mn (θ)] > 0, where Mn (θ) = −E[∇sn (θ)] and λ− [A] is the smallest eigenvalue of the matrix A. If {θˆn } is a consistent sequence of GEE estimators, then Vn−1/2 (θˆn − θ) →d Nk (0, Ik ),

(5.99)

Vn = [Mn (θ)]−1 Var(sn (θ))[Mn (θ)]−1 .

(5.100)

where

370

5. Estimation in Nonparametric Models

Proof. The proof is similar to that of Theorem 4.17. By the consistency of θˆn , we can focus on the event {θˆn ∈ Aǫ }, where Aǫ = {γ : kγ − θk ≤ ǫ} with a given ǫ > 0. For sufficiently small ǫ, it can be shown (exercise) that max

γ∈Aǫ

k∇sn (γ) − ∇sn (θ)k = op (1), n

(5.101)

using an argument similar to the proof of Lemma 5.3. From the mean-value theorem and sn (θˆn ) = 0, Z 1 ∇sn θ + t(θˆn − θ) dt (θˆn − θ). −sn (θ) = 0

It follows from (5.101) that

Z 1

1 ˆ

= op (1). θ + t( θ ∇s − θ) dt − ∇s (θ) n n n

n 0 Also, by Theorem 1.14(ii),

n−1 k∇sn (θ) + Mn (θ)k = op (1). This and lim inf n λ− [n−1 Mn (θ)] > 0 imply [Mn (θ)]−1 sn (θ) = [1 + op (1)](θˆn − θ). The result follows if we can show that Vn−1/2 [Mn (θ)]−1 sn (θ) →d Nk (0, Ik ).

(5.102)

For any nonzero l ∈ Rk , n X 1 E|lτ [Mn (θ)]−1 ψi (Xi , θ)|2+δ → 0, (lτ Vn l)1+δ/2 i=1

(5.103)

since lim inf n λ− [n−1 Var(sn (θ))] > 0 and supi Ekψi (Xi , θ)k2+δ < ∞ (exercise). Applying the CLT (Theorem 1.15) with Liapounov’s condition (5.103), we obtain that p lτ [Mn (θ)]−1 sn (θ)/ lτ Vn l →d N (0, 1) (5.104) for any l, which implies (5.102) (exercise).

Asymptotic normality of GEE estimators can be established under various other conditions; see, for example, Serfling (1980, Chapter 7) and He and Shao (1996).

371

5.5. Variance Estimation

If Xi ’s are i.i.d. and ψi ≡ ψ, the asymptotic covariance matrix in (5.100) reduces to Vn = n−1 {E[ϕ(X1 , θ)]}−1 E{ψ(X1 , θ)[ψ(X1 , θ)]τ }{E[ϕ(X1 , θ)]}−1 , where ϕ(x, γ) = ∂ψ(x, γ)/∂γ. When θ is univariate, Vn further reduces to Vn = n−1 E[ψ(X1 , θ)]2 /{E[ϕ(X1 , θ)]}2 . Under the conditions of Theorem 5.14, Z Z ∂ ∂ψ(x, θ) dF (x) = ψ(x, θ)dF (x). E[ϕ(X1 , θ)] = ∂θ ∂θ Hence, the result in Theorem 5.14 coincides with that in Theorem 5.13. Example 5.14. Consider the quasi-likelihood equation in (5.90) and ψi in (5.96). If supi kZi k < ∞, then ψi satisfies the conditions in Theorem 5.14 ˜ i [Di (γ)]1/2 . Then (exercise). Let V˜n (γ) = [Di (γ)]1/2 R Var(sn (θ)) =

n X

Gi (θ)[V˜n (θ)]−1 Var(Xi )[V˜n (θ)]−1 [Gi (θ)]τ

i=1

and Mn (θ) =

n X

Gi (θ)[V˜n (θ)]−1 [Gi (θ)]τ .

i=1

˜ i = Ri (the true correlation matrix) for all i, then If R Var(sn (θ)) =

n X

φi Gi (θ)[V˜n (θ)]−1 [Gi (θ)]τ .

i=1

If, in addition, φi ≡ φ, then Vn = [Mn (θ)]−1 Var(sn (θ))[Mn (θ)]−1 = φ[Mn (θ)]−1 .

5.5 Variance Estimation In statistical inference the accuracy of a point estimator is usually assessed by its mse or amse. If the bias or asymptotic bias of an estimator is (asymptotically) negligible w.r.t. its mse or amse, then assessing the mse or amse is equivalent to assessing variance or asymptotic variance. Since variances and asymptotic variances usually depend on the unknown population, we have to estimate them in order to report accuracies of point estimators. Variance estimation is an important part of statistical inference, not only for

372

5. Estimation in Nonparametric Models

assessing accuracy, but also for constructing inference procedures studied in Chapters 6 and 7. See also the discussion at the end of §2.5.1. Let θ be a parameter of interest and θˆn be its estimator. Suppose that, as the sample size n → ∞, Vn−1/2 (θˆn − θ) →d Nk (0, Ik ),

(5.105)

where Vn is the covariance matrix or an asymptotic covariance matrix of θˆn . An essential asymptotic requirement in variance estimation is the consistency of variance estimators according to the following definition. See also (3.60) and Exercise 116 in §3.6. Definition 5.4. Let {Vn } be a sequence of k × k positive definite matrices and Vˆn be a positive definite matrix estimator of Vn for each n. Then {Vˆn } or Vˆn is said to be consistent for Vn (or strongly consistent for Vn ) if and only if kVn−1/2 Vˆn Vn−1/2 − Ik k →p 0 (5.106) (or (5.106) holds with →p replaced by →a.s. ). Note that (5.106) is different from kVˆn −Vn k →p 0, because kVn k → 0 in most applications. It can be shown (Exercise 93) that (5.106) holds if and only if lnτ Vˆn ln /lnτ Vn ln →p 1 for any sequence of nonzero vectors {ln } ⊂ Rk . If (5.105) and (5.106) hold, then Vˆn−1/2 (θˆn − θ) →d Nk (0, Ik ) (exercise), a result useful for asymptotic inference discussed in Chapters 6 and 7. If the unknown population is in a parametric family indexed by θ, then Vn is a function of θ, say Vn = Vn (θ), and it is natural to estimate Vn (θ) by Vn (θˆn ). Consistency of Vn (θˆn ) according to Definition 5.4 can usually be directly established. Thus, variance estimation in parametric problems is usually simple. In a nonparametric problem, Vn may depend on unknown quantities other than θ and, thus, variance estimation is much more complex. We introduce three commonly used variance estimation methods in this section, the substitution method, the jackknife, and the bootstrap.

5.5.1 The substitution method Suppose that we can obtain a formula for the covariance or asymptotic covariance matrix Vn in (5.105). Then a direct method of variance estimation is to substitute unknown quantities in the variance formula by some

373

5.5. Variance Estimation

estimators. To illustrate, consider the simplest case where X1 , ..., Xn are ¯ i.i.d. random d-vectors with EkX1 k2 < ∞, θ = g(µ), µ = EX1 , θˆn = g(X), d k and g is a function from R to R . Suppose that g is differentiable at µ. Then, by the CLT and Theorem 1.12(i), (5.105) holds with Vn = [∇g(µ)]τ Var(X1 )∇g(µ)/n,

(5.107)

which depends on unknown quantities µ and Var(X1 ). A substitution estimator of Vn is ¯ τ S 2 ∇g(X)/n, ¯ Vˆn = [∇g(X)] (5.108) where

n

S2 =

1 X ¯ ¯ τ (Xi − X)(X i − X) n − 1 i=1

is the sample covariance matrix, an extension of the sample variance to the multivariate Xi ’s. ¯ →a.s. µ and S 2 →a.s. Var(X1 ). Hence, Vˆn in (5.108) By the SLLN, X is strongly consistent for Vn in (5.107), provided that ∇g(µ) 6= 0 and ∇g is continuous at µ. Example 5.15. Let Y1 , ..., Yn be i.i.d. random variables with finite µy = EY1 , σy2 = Var(Y1 ), γy = EY13 , and κy = EY14 . Consider the estimation Pn ¯ σ of θ = (µy , σy2 ). Let θˆn = (X, ˆy2 ), where σ ˆy2 = n−1 i=1 (Yi − Y¯ )2 . If ¯ with g(x) = (x1 , x2 − x2 ). Hence, (5.105) Xi = (Yi , Yi2 ), then θˆn = g(X) 1 holds with γy − µy (σy2 + µ2y ) σy2 Var(X1 ) = γy − µy (σy2 + µ2y ) κy − (σy2 + µ2y )2 and ∇g(x) =

1 −2x1

0 1

.

The estimator Vˆn in (5.108) is strongly consistent, since ∇g(x) is obviously a continuous function. Similar results can be obtained for problems in Examples 3.21 and 3.23 and Exercises 100 and 101 in §3.6. A key step in the previous discussion is the derivation of formula (5.107) ¯ via Taylor’s expansion for the asymptotic covariance matrix of θˆn = g(X) (Theorem 1.12) and the CLT. Thus, the idea can be applied to the case where θˆn = T(Fn ), a differentiable statistical functional. We still consider i.i.d. random d-vectors X1 , ..., Xn from F . Suppose that T is a vector-valued functional whose components are ̺-Hadamard

374

5. Estimation in Nonparametric Models

differentiable at F , where ̺ is either ̺∞ or a distance satisfying (5.42). Let φF be the vector of influence functions of components of T. If the components of φF satisfy (5.40), then (5.105) holds with θ = T(F ), θˆn = T(Fn ), Fn = the empirical c.d.f. in (5.1), and Z 1 Var(φF (X1 )) = φF (x)[φF (x)]τ dF (x). Vn = (5.109) n n Formula (5.109) leads to a natural substitution variance estimator 1 Vˆn = n

Z

φFn (x)[φFn (x)]τ dFn (x) =

n 1 X φF (Xi )[φFn (Xi )]τ , (5.110) n2 i=1 n

provided that φFn (x) is well defined, i.e., the components of T are Gˆateaux differentiable at Fn for sufficiently large n. Under some more conditions on φFn we can establish the consistency of Vˆn in (5.110). Theorem 5.15. Let X1 , ..., Xn be i.i.d. random d-vectors from F , T be a vector-valued functional whose components are Gˆateaux differentiable at F and Fn , and φF be the vector of influence functions of components of T. Suppose that supkxk≤c kφFn (x) − φF (x)k = op (1) for any c > 0 and that there exist a constant c0 > 0 and a function h(x) ≥ 0 such that E[h(X1 )] < ∞ and P kφFn (x)k2 ≤ h(x) for all kxk ≥ c0 → 1. Then Vˆn in (5.110) is consistent for Vn in (5.109). Proof. Let ζ(x) = φF (x)[φF (x)]τ and ζn (x) = φFn (x)[φFn (x)]τ . By the SLLN, Z n 1X ζ(Xi ) →a.s. ζ(x)dF (x). n i=1

Hence the result follows from

n

1 X

[ζn (Xi ) − ζ(Xi )] = op (1).

n i=1

Using the assumed conditions and the argument in the proof of Lemma 5.3, we can show that for any ǫ > 0, there is a c > 0 such that ! n 1X ǫ P ≤ǫ kζn (Xi ) − ζ(Xi )kI(c,∞) (kXi k) > n i=1 2

and

n

P

1X ǫ kζn (Xi ) − ζ(Xi )kI[0,c] (kXi k) > n i=1 2

for sufficiently large n. This completes the proof.

!

≤ǫ

375

5.5. Variance Estimation

Example 5.16. Consider the L-functional defined in (5.46) and the Lestimator θˆn = T(Fn ). Theorem 5.6 shows that T is Hadamard differentiable at F under some conditions on J. It can be shown (exercise) that T is Gˆateaux differentiable at Fn with φFn (x) given by (5.48) (with F replaced by Fn ). Then the difference φFn (x) − φF (x) is equal to Z Z (Fn − F )(y)J(Fn (y))dy + (F − διx )(y)[J(Fn (y)) − J(F (y))]dy. One can show (exercise) that the conditions in Theorem 5.15 are satisfied if the conditions in Theorem 5.6(i) or (ii) (with E|X1 | < ∞) hold. Substitution variance estimators for M-estimators and U-statistics can also be derived (exercises). The substitution method can clearly be applied to non-i.i.d. cases. For example, the LSE βˆ in linear model (3.25) with a full rank Z and i.i.d. εi ’s ˆ = σ 2 (Z τ Z)−1 , where σ 2 = Var(ε1 ). A consistent substitution has Var(β) ˆ ˆ can be obtained by replacing σ 2 in the formula of Var(β) estimator of Var(β) by a consistent estimator of σ 2 such as SSR/(n − p) (see (3.35)). We now consider variance estimation for the GEE estimators described in §5.4.1. By Theorem 5.14, the asymptotic covariance matrix of the GEE estimator θˆn is given by (5.100), where Var(sn (θ)) =

n X i=1

Mn (θ) =

E{ψi (Xi , θ)[ψi (Xi , θ)]τ }, n X

E[ϕi (Xi , θ)],

i=1

and ϕi (x, γ) = ∂ψi (x, γ)/∂γ. Substituting θ by θˆn and the expectations by their empirical analogues, we obtain the substitution estimator Vˆn = ˆ −1 Var(s ˆ −1 , where d n )M M n n and

d n) = Var(s

n X

ψi (Xi , θˆn )[ψi (Xi , θˆn )]τ

i=1

ˆn = M

n X

ϕi (Xi , θˆn ).

i=1

The proof of the following result is left as an exercise. Theorem 5.16. Let X1 , ..., Xn be independent and {θˆn } be a consistent sequence of GEE estimators. Assume the conditions in Theorem 5.14. Suppose further that the sequence of functions {hij , i = 1, 2, ...} satisfies the

376

5. Estimation in Nonparametric Models

conditions in Lemma 5.3 with Θ replaced by a compact neighborhood of θ, where hij (x, γ) is the jth row of ψi (x, γ)[ψi (x, γ)]τ , j = 1, ..., k. Let Vn be ˆ −1 Var(s ˆ −1 is consistent for Vn . d n )M given by (5.100). Then Vˆn = M n n

5.5.2 The jackknife

Applying the substitution method requires the derivation of a formula for the covariance matrix or asymptotic covariance matrix of a point estimator. There are variance estimation methods that can be used without actually deriving such a formula (only the existence of the covariance matrix or asymptotic covariance matrix is assumed), at the expense of requiring a large number of computations. These methods are called resampling methods, replication methods, or data reuse methods. The jackknife method introduced here and the bootstrap method in §5.5.3 are the most popular resampling methods. The jackknife method was proposed by Quenouille (1949) and Tukey (1958). Let θˆn be a vector-valued estimator based on independent Xi ’s, where each Xi is a random di -vector and supi di < ∞. Let θˆ−i be the same estimator but based on X1 , ..., Xi−1 , Xi+1 , ..., Xn , i = 1, ..., n. Note that θˆ−i also depends on n but the subscript n is omitted for simplicity. Since θˆn and θˆ−1 , ..., θˆ−n are estimators of the same quantity, the “sample covariance matrix” τ 1 X ˆ θ−i − θ¯n θˆ−i − θ¯n n − 1 i=1 n

(5.111)

can be used as a measure of the variation of θˆn , where θ¯n is the average of θˆ−i ’s. There are two major differences between the quantity in (5.111) and the sample covariance matrix S 2 previously discussed. First, θˆ−i ’s are not independent. Second, θˆ−i − θˆ−j usually converges to 0 at a fast rate (such as n−1 ). Hence, to estimate the asymptotic covariance matrix of θˆn , the ¯ quantity in (5.111) should be multiplied by a correction factor cn . If θˆn = X −1 ¯ ˆ ¯ (di ≡ d), then θ−i − θn = (n − 1) (X − Xi ) and the quantity in (5.111) reduces to n X 1 1 ¯ Xi − X ¯ τ = S2, Xi − X (n − 1)3 i=1 (n − 1)2

where S 2 is the sample covariance matrix. Thus, the correction factor cn ¯ since, by the SLLN, S 2 /n is strongly is (n − 1)2 /n for the case of θˆn = X ¯ consistent for Var(X).

377

5.5. Variance Estimation

It turns out that the same correction factor works for many other estimators. This leads to the following jackknife variance estimator for θˆn : τ n − 1 X ˆ θ−i − θ¯n θˆ−i − θ¯n . VˆJ = n i=1 n

(5.112)

Theorem 5.17. Let X1 , ..., Xn be i.i.d. random d-vectors from F with ¯ Suppose that ∇g is finite µ = E(X1 ) and Var(X1 ), and let θˆn = g(X). continuous at µ and ∇g(µ) 6= 0. Then the jackknife variance estimator VˆJ in (5.112) is strongly consistent for Vn in (5.107). Proof. We prove the case where g is real-valued. The proof of the gen¯ −i be the sample mean based on eral case is left to the reader. Let X X1 , ..., Xi−1 , Xi+1 , ..., Xn . From the mean-value theorem, we have ¯ −i ) − g(X) ¯ θˆ−i − θˆn = g(X τ ¯ ¯ = [∇g(ξn,i )] (X−i − X) ¯ −i − X) ¯ + Rn,i , ¯ τ (X = [∇g(X)]

¯ τ (X ¯ −i − X) ¯ and ξn,i is a point on the where Rn,i = ∇g(ξn,i ) − ∇g(X) ¯ ¯ ¯ ¯ = (n − 1)−1 (X ¯ − Xi ), it line segmentP between X−i and X. From X−i − X n ¯ ¯ follows that i=1 (X−i − X) = 0 and n

n

1X ˆ 1X ¯n. (θ−i − θˆn ) = Rn,i = R n i=1 n i=1

From the definition of the jackknife estimator in (5.112), VˆJ = An + Bn + 2Cn , where n

An =

X n−1 ¯ τ ¯ −i − X)( ¯ X ¯ −i − X) ¯ τ ∇g(X), ¯ [∇g(X)] (X n i=1 n

Bn = and

n−1 X ¯ n )2 , (Rn,i − R n i=1

n

Cn =

n−1X ¯ n )[∇g(X)] ¯ τ (X ¯ −i − X). ¯ (Rn,i − R n i=1

¯ −i − X ¯ = (n − 1)−1 (X ¯ − Xi ), the SLLN, and the continuity of ∇g at By X µ, An /Vn →a.s. 1.

378

5. Estimation in Nonparametric Models

Also, (n − 1) Hence

n X i=1

n

¯ −i − Xk ¯ 2= kX

1 X ¯ 2 = O(1) a.s. kXi − Xk n − 1 i=1

(5.113)

¯ −i − Xk ¯ 2 →a.s. 0, max kX i≤n

¯ ≤ kX ¯ −i − Xk, ¯ which, together with the continuity of ∇g at µ and kξn,i − Xk implies that ¯ →a.s. 0. un = max k∇g(ξn,i ) − ∇g(X)k i≤n P ¯ −i − Xk ¯ 2 /Vn = O(1) a.s. Hence From (5.107) and (5.113), ni=1 kX n n Bn n−1X 2 un X ¯ ¯ 2 →a.s. 0. ≤ Rn,i ≤ kX−i − Xk Vn Vn n i=1 Vn i=1

By the Cauchy-Schwarz inequality, (Cn /Vn )2 ≤ (An /Vn )(Bn /Vn ) →a.s. 0. This proves the result. A key step in the proof of Theorem 5.17 is that θˆ−i − θˆn can be approx¯ τ (X ¯ −i − X) ¯ and the contributions of the remainders, imated by [∇g(X)] Rn,1 , ..., Rn,n , are sufficiently small, i.e., Bn /Vn →a.s. 0. This indicates that the jackknife estimator (5.112) is consistent for θˆn that can be well approximated by some linear statistic. In fact, the jackknife estimator (5.112) has been shown to be consistent when θˆn is a U-statistic (Arvesen, 1969) or a statistical functional that is Hadamard differentiable and continuously Gˆateaux differentiable at F (which includes certain types of L-estimators and M-estimators). More details can be found in Shao and Tu (1995, Chapter 2). The jackknife method can be applied to non-i.i.d. problems. A detailed discussion of the use of the jackknife method in survey problems can be found in Shao and Tu (1995, Chapter 6). We now consider the jackknife variance estimator for the LSE βˆ in linear model (3.25). For simplicity, assume that Z is of full rank. Assume also that εi ’s are independent with E(εi ) = 0 and Var(εi ) = σi2 . Then ˆ = (Z τ Z)−1 Var(β)

n X

σi2 Zi Ziτ (Z τ Z)−1 .

i=1

Let βˆ−i be the LSE of β based on the data with the ith pair (Xi , Zi ) deleted. Using the fact that (A + ccτ )−1 = A−1 − A−1 ccτ A−1 /(1 + cτ A−1 c) for a matrix A and a vector c, we can show that (exercise) βˆ−i = βˆ − ri Zi /(1 − hi ),

(5.114)

379

5.5. Variance Estimation

where ri = Xi − Ziτ βˆ is the ith residual and hi = Ziτ (Z τ Z)−1 Zi . Hence " n # n n τ X r 2 Zi Z τ X X r Z r Z n − 1 1 i i i i i i (Z τ Z)−1 (Z τ Z)−1 . − VˆJ = 2 n (1 − h ) n 1 − h 1 − h i i i i=1 i=1 i=1 Wu (1986) proposed the following weighted jackknife variance estimator that improves VˆJ : VˆW J =

n τ X ri2 Zi Ziτ τ −1 (1 − hi ) βˆ−i − βˆ βˆ−i − βˆ = (Z τ Z)−1 (Z Z) . 1 − hi i=1 i=1

n X

Theorem 5.18. Assume the conditions in Theorem 3.12 and that εi ’s are ˆ independent. Then both VˆJ and VˆW J are consistent for Var(β). p Proof. Let ln ∈ R , n = 1, 2, ..., be nonzero vectors and li = lnτ (Z τ Z)−1 Zi . Since maxi≤n hi → 0, the result for VˆW J follows from X n n X 2 2 li ri li2 σi2 →p 1 (5.115) i=1

i=1

(see Exercise 93). By the WLLN (Theorem 1.14(ii)) and maxi≤n hi → 0, X n n X 2 2 li εi li2 σi2 →p 1. i=1

i=1

ˆ and Note that ri = εi + Ziτ (β − β) ˆ 2 ≤ kZ(β − β)k ˆ 2 max hi = op (1). max[Ziτ (β − β)] i≤n

i≤n

Hence (5.115) holds. The consistency of VˆJ follows from (5.115) and !2 n n X n − 1 X li ri li2 σi2 = op (1). n2 1 − h i i=1 i=1

(5.116)

The proof of (5.116) is left as an exercise. Finally, let us consider the jackknife estimators for GEE estimators in §5.4.1. Under the conditions of Proposition 5.5 or 5.6, it can be shown that ˆ = op (1), max kθˆ−i − θk i≤n

where θˆ−i is a root of sni (γ) = 0 and X ψj (Xj , γ). sni (γ) = j6=i,j≤n

(5.117)

380

5. Estimation in Nonparametric Models

Assume that ψi (x, γ) is continuously differentiable w.r.t. γ in a neighborhood of θ. Using Taylor’s expansion and the fact that sni (θˆ−i ) = 0 and sn (θˆn ) = 0, we obtain that Z 1 ∇sn θˆn + t(θˆ−i − θˆn ) dt (θˆ−i − θˆn ). ψi (Xi , θˆ−i ) = 0

Following the proof of Theorem 5.14, we obtain that VˆJ = [Mn (θ)]−1

n X i=1

ψi (Xi , θˆ−i )[ψi (Xi , θˆ−i )]τ [Mn (θ)]−1 + Rn ,

−1/2

−1/2

where Rn satisfies kVn Rn Vn k = op (1) for Vn in (5.100). Under the conditions of Theorem 5.16, it follows from (5.117) that VˆJ is consistent. If θˆn is computed using an iteration method, then the computation of ˆ VJ requires n additional iteration processes. We may use the idea of a one-step MLE to reduce the amount of computation. For each i, let θˆ−i = θˆn − [∇sni (θˆn )]−1 sni (θˆn ),

(5.118)

which is the result from the first iteration when the Newton-Raphson method is applied in computing a root of sni (γ) = 0 and θˆn is used as the initial point. Note that θˆ−i ’s in (5.118) satisfy (5.117) (exercise). If the jackknife variance estimator is based on θˆ−i ’s in (5.118), then VˆJ = [Mn (θ)]−1

n X

˜n, ψi (Xi , θˆn )[ψi (Xi , θˆn )]τ [Mn (θ)]−1 + R

i=1

˜ n satisfies kVn−1/2 R ˜ n Vn−1/2 k = op (1). These results are summarized where R in the following theorem. Theorem 5.19. Assume the conditions in Theorems 5.14 and 5.16. Assume further that θˆ−i ’s are given by (5.118) or GEE estimators satisfying (5.117). Then the jackknife variance estimator VˆJ is consistent for Vn given in (5.100).

5.5.3 The bootstrap The basic idea of the bootstrap method can be described as follows. Suppose that P is a population or model that generates the sample X and that ˆ where θˆ = θ(X) ˆ we need to estimate Var(θ), is an estimator, a statistic based on X. Suppose further that the unknown population P is estimated by Pˆ , based on the sample X. Let X ∗ be a sample (called a bootstrap

381

5.5. Variance Estimation

sample) taken from the estimated population Pˆ using the same or a similar ˆ ∗ ), which is the sampling procedure used to obtain X, and let θˆ∗ = θ(X ∗ ˆ same as θ but with X replaced by X . If we believe that P = Pˆ (i.e., ˆ = Var∗ (θˆ∗ ), we have a perfect estimate of the population), then Var(θ) where Var∗ is the conditional variance w.r.t. the randomness in generating ˆ 6= Var∗ (θˆ∗ ). But X ∗ , given X. In general, P 6= Pˆ and, therefore, Var(θ) ˆ and can be used as an VˆB = Var∗ (θˆ∗ ) is an empirical analogue of Var(θ) ˆ estimate of Var(θ). In a few cases, an explicit form of VˆB = Var∗ (θˆ∗ ) can be obtained. First, consider i.i.d. X1 , ..., Xn from a c.d.f. F on Rd . The population is determined by F . Suppose that we estimate F by the empirical c.d.f. Fn ¯ its bootstrap in (5.1) and that X1∗ , ..., Xn∗ are i.i.d. from Fn . For θˆ = X, ∗ ∗ ∗ ˆ ¯ analogue is θ = X , the average of Xi ’s. Then n X n−1 2 ¯ ∗) = 1 ¯ ¯ τ (Xi − X)(X S , VˆB = Var∗ (X i − X) = n2 i=1 n2

¯ ∗ ) is where S 2 is the sample covariance matrix. In this case VˆB = Var∗ (X ¯ a strongly consistent estimator for Var(X). Next, consider i.i.d. random variables X1 , ..., Xn from a c.d.f. F on R and θˆ = Fn−1 ( 12 ), the sample median. Suppose that n = 2l − 1 for an integer l. Let X1∗ , ..., Xn∗ be i.i.d. from Fn and θˆ∗ be the sample median based on X1∗ , ..., Xn∗ . Then !2 n n X X ∗ ˆ pj X(j) − pi X(i) , VˆB = Var∗ (θ ) = j=1

i=1

where X(1) ≤ · · · ≤ X(n) are order statistics and pj = P (θˆ∗ = X(j) |X). It can be shown (exercise) that l−1 X n (j − 1)t (n − j + 1)n−t − j t (n − j)n−t pj = . (5.119) nn t t=0 However, in most cases VˆB does not have a simple explicit form. When P is known, the Monte Carlo method described in §4.1.4 can be used to ˆ That is, we draw repeatedly new data sets from P and approximate Var(θ). then use the sample covariance matrix based on the values of θˆ computed ˆ This idea from new data sets as a numerical approximation to Var(θ). can be used to approximate VˆB , since Pˆ is a known population. That is, we can draw m bootstrap data sets X ∗1 , ..., X ∗m independently from Pˆ ˆ ∗j ), j = 1, ..., m, and approximate (conditioned on X), compute θˆ∗j = θ(X VˆB by m 1 X ˆ∗j ¯∗ ˆ∗j ¯∗ τ VˆBm = θ −θ θ −θ , m j=1

382

5. Estimation in Nonparametric Models

where θ¯∗ is the average of θˆ∗j ’s. Since each X ∗j is a data set generated from Pˆ , VˆBm is a resampling estimator. From the SLLN, as m → ∞, VˆBm →a.s. VˆB , conditioned on X. Both VˆB and its Monte Carlo approximation VˆBm are ˆ Vˆ m is more useful in practical called bootstrap variance estimators for θ. B applications, whereas in theoretical studies, we usually focus on VˆB . The consistency of the bootstrap variance estimator VˆB is a much more complicated problem than that of the jackknife variance estimator in §5.5.2. Some examples can be found in Shao and Tu (1995, §3.2.2). The bootstrap method can also be applied to estimate quantities other ˆ For example, let K(t) = P (θˆ ≤ t) be the c.d.f. of a real-valued than Var(θ). ˆ From the previous discussion, a bootstrap estimator of K(t) estimator θ. is the conditional probability P (θˆ∗ ≤ t|X), which can be approximated Pm by the Monte Carlo approximation m−1 j=1 I(−∞,t] (θˆ∗j ). An important application of bootstrap distribution estimators in problems of constructing confidence sets is studied in §7.4. Here, we study the use of a bootstrap distribution estimator to form a consistent estimator of the asymptotic ˆ variance of a real-valued estimator θ. Suppose that √ n(θˆ − θ) →d N (0, v), (5.120) √ ˆ where v is unknown. Let Hn (t) be the c.d.f. of n(θ − θ) and √ ˆ ≤ t|X) ˆ B (t) = P ( n(θˆ∗ − θ) H

(5.121)

be a bootstrap estimator of Hn (t). If ˆ B (t) − Hn (t) →p 0 H for any t, then, by (5.120), √ ˆ B (t) − Φ t/ v →p 0, H

which implies (Exercise 112) that

ˆ −1 (α) →p H B

√ vzα

for any α ∈ (0, 1), where zα = Φ−1 (α). Then, for α 6= 12 , ˆ −1 (α) →p ˆ −1 (1 − α) − H H B B

√ v(z1−α − zα ).

ˆ is Therefore, a consistent estimator of v/n, the asymptotic variance of θ, " #2 ˆ −1 (α) ˆ −1 (1 − α) − H H 1 B B . V˜B = n z1−α − zα

383

5.6. Exercises

ˆ B (t)−Hn (t) →p 0. The following result gives some conditions under which H The proof of part (i) is omitted. The proof of part (ii) is given in Exercises 113-115 in §5.6. Theorem 5.20. Suppose that X1 , ..., Xn are i.i.d. from a c.d.f. F on Rd . Let θˆ = T(Fn ), where T is a real-valued functional, θˆ∗ = T(Fn∗ ), where Fn∗ is the empirical c.d.f. based on a bootstrap sample X1∗ , ..., Xn∗ i.i.d. from Fn , ˆ B be given by (5.121). and let H (i) If T is ̺∞ -Hadamard differentiable at F and (5.40) holds, then ˆ B , Hn ) →p 0. ̺∞ (H

(5.122) R (ii) If d = 1 and T is ̺Lp -Fr´echet differentiable at F ( {F (t)[1 − F (t)]}p/2 dt < ∞ if 1 ≤ p < 2) and (5.40) holds, then (5.122) holds. Applications of the bootstrap method to non-i.i.d. cases can be found, for example, in Efron and Tibshirani (1993), Hall (1992), and Shao and Tu (1995).

5.6 Exercises 1. Let ̺∞ be the sup-norm distance. Find an example of a sequence {Gn } of c.d.f.’s satisfying Gn →w G for a c.d.f. G, but ̺∞ (Gn , G) does not converge to 0. 2. Let X1 , ..., Xn be i.i.d. random d-vectors with c.d.f. F and Fn be the empirical c.d.f. defined by (5.1). Show that for any t > 0 and ǫ > 0, there is a Cǫ,d such that for all n = 1, 2, ..., P

sup ̺∞ (Fm , F ) > t

m≥n

2

≤

Cǫ,d e−(2−ǫ)t n . 1 − e−(2−ǫ)t2

3. Show that ̺Mp defined by (5.4) is a distance on Fp , p ≥ 1. 4. Show that k · kLp in (5.5) is a norm for any p ≥ 1. 5. Let F1 be the collection of c.d.f.’s finite means. R 1 on R with −1 −1 (a) Show that ̺M1 (G1 , G2 ) = 0 |G−1 (z) 1 (z) − G2 (z)|dz, where G = inf{t : G(t) ≥ z} for any G ∈ F. (b) Show that ̺M1 (G1 , G2 ) = ̺L1 (G1 , G2 ). 6. Find an example of a sequence {Gj } ⊂ F for which (a) limj→∞ ̺∞ (Gj , G0 ) = 0 but ̺M2 (Gj , G0 ) does not converge to 0; (b) limj→∞ ̺M2 (Gj , G0 ) = 0 but ̺∞ (Gj , G0 ) does not converge to 0.

384

5. Estimation in Nonparametric Models

7. Repeat the previous exercise with ̺M2 replaced by ̺L2 . 8. Let X be a random variable having c.d.f. F . Show that R (a) E|X|2 < ∞ implies {F (t)[1 − F (t)]}p/2Rdt < ∞ for p ∈ (1, 2); (b) E|X|2+δ < ∞ with some δ > 0 implies {F (t)[1 − F (t)]}1/2 dt < ∞. 9. For Gj ∈ F1 , j = 1, 2, show that ̺L1 (G1 , G2 ) ≥ R any one-dimensional R | xdG1 − xdG2 |.

10. In the proof of Theorem 5.3, show that pi = c/n, i = 1, ..., n, λ = −(c/n)n−1 is Pan maximum of the function H(p1 , ..., pn , λ) over pi > 0, i = 1, ..., n, i=1 pi = c.

11. Show that (5.11)-(5.13) is a solution to the problem of maximizing ℓ(G) in (5.8) subject to (5.10). 12. In the proof of Theorem 5.4, prove the case of m ≥ 2. 13. Show that a maximum of ℓ(G) in (5.17) subject to (5.10) is given by (5.11) with pˆi defined by (5.18) and (5.19). 14. In Example 5.2, show that an MELE is given by (5.11) with pˆi ’s given by (5.21). 15. In Example 5.3, show that (a) maximizing (5.22) subject to (5.23) is equivalent to maximizing n Y

i=1

Pn+1

δ

qi (i) (1 − qi )n−i+1−δ(i) ,

where qi = pi / j=i pj , i = 1, ..., n; (b) Fˆ given by (5.24) maximizes (5.22) subject to (5.23); (Hint: use Qi−1 part (a) and the fact that pi = qi j=1 (1 − qj ).) (c) Fˆ given by (5.25) is the same as that in (5.24); (d) if δi = 1 for all i (no censoring), then Fˆ in (5.25) is the same as the empirical c.d.f. in (5.1). 16. Let fn be given by (5.26). (a) Show that fn is a Lebesgue p.d.f. on R. (b) Suppose that f is continuously differentiable at t, λn → 0, and nλn → ∞. Show that (5.27) holds. (c) Under nλ3n → 0 and the conditions of (b), show that (5.28) holds. (d) Suppose that f is continuous on [a, b], −∞ < a < b < ∞, λn → 0, Rb Rb and nλn → ∞. Show that a fn (t)dt →p a f (t)dt.

5.6. Exercises

385

17. Let fˆ be given by (5.29). (a) Show that fˆ is a Lebesgue p.d.f. on R. (b) Prove (5.30) under the condition that λn → 0, nλn → ∞, and R f is bounded and continuous at t and [w(t)]2 dt < ∞. (Hint: check Lindeberg’s condition and apply Theorem 1.15.) (c) Assume that λn → 0, nλn → ∞, w is bounded, and f is bounded Rb ˆ and continuous on [a, b], −∞ < a < b < ∞. Show that a f(t)dt →p Rb a f (t)dt.

18. Prove (5.32)-(5.34) under the conditions described in §5.1.4.

ˆ 19. Show that K(t) in (5.35) is a consistent estimator of K(t) in (5.34), ˆ assuming that β →p β, φ is a continuous function on R, (Xi , Zi )’s are i.i.d., and kZi k ≤ c for a constant c > 0. 20. Let ℓ(θ, ξ) be a likelihood. Show that a maximum profile likelihood estimator θˆ of θ is an MLE if ξ(θ), the maximum of supξ ℓ(θ, ξ) for a fixed θ, does not depend on θ. 21. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ). Derive the profile likelihood function for µ or σ 2 . Discuss in each case whether the maximum profile likelihood estimator is the same as the MLE.

22. Derive the profile empirical likelihoods in (5.36) and (5.37). 23. Let X1 , ..., Xn be i.i.d. random variables from a c.d.f. F and let π(x) = P (δi = 1|Xi = x), where δi = 1R if Xi is observed and δi = 0 if Xi is missing. Assume that 0 < π = π(x)dF (x) < 1. (a) Let F1 (x) = P (Xi ≤ x|δi = 1). Show that F and F1 are the same if and only if π(x) ≡ π. (b) Let Fˆ be the c.d.f. putting mass r−1 to each observed Xi , where r is the number of observed Xi ’s. Show that Fˆ (x) is unbiased and consistent for F1 (x), x ∈ R. (c) When π(x) ≡ π, show that Fˆ (x) in part (b) is unbiased and consistent for F (x), x ∈ R. When π(x) is not constant, show that Fˆ (x) is biased and inconsistent for F (x) for some x ∈ R. 24. Show that ̺-Fr´echet differentiability implies ̺-Hadamard differentiability. 25. Suppose that a functional T is Gˆ ateaux differentiable at F with a continuous differential LF in the sense that ̺∞ (∆j , ∆) → 0 implies LF (∆j ) → LF (∆). Show that φF is bounded. 26. Suppose that a functional T is Gˆ ateaux differentiable at F with a bounded and continuous influence function φF . Show that the differential LF is continuous in the sense described in the previous exercise.

386

5. Estimation in Nonparametric Models

R 27. Let T(G) = g( xdG) be a functional defined on F1 , the collection of one-dimensional c.d.f.’s with finite means. (a) Find a differentiable function g for which the functional T is not ̺∞ -Hadamard differentiable at F . (b) Show that if g is a differentiable function, then T is ̺L1 -Fr´echet differentiable at F . (Hint: use the result in Exercise 9.) 28. In Example 5.5, show that (5.43) holds. (Hint: for ∆ = c(G1 − G2 ), show that k∆kV ≤ |c|(kG1 kV + kG2 kV ) = 2|c|.) 29. In Example 5.5, show that φF is continuous if F is continuous. 30. In Example 5.5, show that T is not ̺∞ -Fr´echet differentiable at F . 31. Prove Proposition 5.1(ii). 32. Suppose that T is first-order and second-order ̺-Hadamard differentiable at F . Prove (5.45). 33. Find an example of a second-order ̺-Fr´echet differentiable functional T that is not first-order ̺-Hadamard differentiable. 34. Prove (5.47) and that (5.40) is satisfied for an L-functional if J is bounded and F has a finite variance. 35. Prove (iv) and (v) of Theorem 5.6. 36. Discuss which of (i)-(v) in Theorem 5.6 can be applied to each of the L-estimators in Example 5.6. 37. Obtain explicit forms of the influence functions for L-estimators in Example 5.6. Discuss which of them are bounded and continuous. 38. Provide an example in which the L-functional T given by (5.46) is not ̺∞ -Hadamard differentiable at F . (Hint: consider an untrimmed J.) 39. Discuss which M-functionals defined in (i)-(vi) of Example 5.7 satisfy the conditions of Theorem 5.7. 40. In the proof of Theorem 5.7, show that R2j → 0. 41. Show that the second equality in (5.51) holds when ψ is Borel and bounded. 42. Show that the functional T in (5.53) is ̺∞ -Hadamard differentiable at F with the differential given by (5.54). Obtain the influence function φF and show that it is bounded and continuous if F is continuous.

387

5.6. Exercises

43. Show that the functional T in (5.55) is ̺∞ -Hadamard differentiable at F with the differential given by (5.56). Obtain the influence function φF and show that it is bounded and continuous if F (y, ∞) and F (∞, z) are continuous. 44. Let F be a continuous c.d.f. on R. Suppose that F is symmetric about θ and is strictly increasing in a neighborhood of θ. Show that λF (t) = 0 if and only if t = θ, where λF (t) is defined by (5.57) with a strictly increasing J satisfying J(1 − t) = −J(t). 45. Show in (5.57) is differentiable at θ and λ′F (θ) is equal to R ′that λF (t) ′ − J (F (x))F (x)dF (x).

46. Let T(Fn ) be an R-estimator satisfying the conditions in Theorem 5.8. Show that (5.41) holds with σF2 =

Z

0

1

[J(t)]2 dt

Z

∞

−∞

2 J ′ (F (x))F ′ (x)dF (x) .

47. Calculate the asymptotic relative efficiency of the Hodges-Lehmann estimator in Example 5.8 w.r.t. the sample mean based on an i.i.d. sample from F when (a) F is the c.d.f. of N (µ, σ 2 ); (b) F is the c.d.f. of the logistic distribution LG(µ, σ); (c) F is the c.d.f. of the double exponential distribution DE(µ, σ); (d) F (x) = F0 (x − θ), where F0 (x) is the c.d.f. of the t-distribution tν with ν ≥ 3. 48. Let G be a c.d.f. on R. Show that G(x) ≥ t if and only if x ≥ G−1 (t). 49. Show that (5.67) implies that θˆp is strongly consistent for θp and is √ n-consistent for θp if F ′ (θp −) and F ′ (θp +) exist and are positive. 2

50. Under the condition of Theorem 5.9, show that, for ρǫ = e−2δǫ , 2Cρnǫ , n = 1, 2, .... P sup |θˆp − θp | > ǫ ≤ 1 − ρǫ m≥n 51. Prove that ϕn (t) in (5.69) is the Lebesgue p.d.f. of the pth sample quantile θˆp when F has the Lebesgue p.d.f. f by (a) differentiating the c.d.f. of θˆp in (5.68); (b) using result (5.66) and the result in Example 2.9. 52. Let X1 , ..., Xn be i.i.d. random variables from F with a finite mean. Show that θˆp has a finite jth moment for sufficiently large n, j = 1, 2,....

388

5. Estimation in Nonparametric Models

53. Prove Theorem 5.10(i). 54. Suppose that a c.d.f. F has a Lebesgue p.d.f. f that is continuous at the pth quantile of F , p ∈ (0, 1). Using the p.d.f. in (5.69) and Scheff´e’s theorem (Proposition 1.18), prove part (iv) of Theorem 5.10. 55. Let {kn } be a sequence of integers satisfying kn /n = p + o(n−1/2 ) with p ∈ (0, 1), and let X1 , ..., Xn be i.i.d. random variables from a c.d.f. F with F ′ (θp ) > 0. Show that √ n(X(kn ) − θp ) →d N (0, p(1 − p)/[F ′ (θp )]2 ). 56. In the proof of Theorem 5.11, prove (5.72), (5.75), and inequality (5.74). 57. Prove Corollary 5.1. 58. Prove the claim in Example 5.9. 59. Let T (G) = G−1 (p) be the pth quantile functional. Suppose that F has a positive derivative F ′ in a neighborhood of θ = F −1 (p). Show that T is Gˆateaux differentiable at F and obtain the influence function. 60. Let X1 , ..., Xn be i.i.d. from the Cauchy distribution C(0, 1). (a) Show that E(X(j) )2 < ∞ if and only if 3 ≤ j ≤ n − 2. (b) Show that E(θˆ0.5 )2 < ∞ for n ≥ 5. 61. Suppose that F is the c.d.f. of the uniform distribution U (θ− 21 , θ+ 12 ), θ ∈ R. Obtain the asymptotic relative efficiency of the sample median w.r.t. the sample mean, based on an i.i.d. sample of size n from F . 62. Suppose that F (x) = F0 (x − θ) and F0 is the c.d.f. of the Cauchy distribution C(0, 1) truncated R c at c and −c, i.e., F0 has the Lebesgue p.d.f. (1 + x2 )−1 I(−c,c) (x)/ −c (1 + x2 )−1 dt. Obtain the asymptotic relative efficiency of the sample median w.r.t. the sample mean, based on an i.i.d. sample of size n from F . 63. Let X1 , ..., Xn be i.i.d. with the c.d.f. (1−ǫ)Φ x−µ +ǫD x−µ , where σ σ ǫ ∈ (0, 1) is a known constant, Φ is the c.d.f. of the standard normal distribution, D is the c.d.f. of the double exponential distribution D(0, 1), and µ ∈ R and σ > 0 are unknown parameters. Consider the estimation of µ. Obtain the asymptotic relative efficiency of the sample mean w.r.t. the sample median. 64. Let X1 , ..., Xn be i.i.d. with the Lebesgue p.d.f. 2−1 (1 − θ2 )eθx−|x|, where θ ∈ (−1, 1) is unknown. (a) Show that the median of the distribution of X1 is given by m(θ) =

5.6. Exercises

389

(1 − θ)−1 log(1 + θ) when θ ≥ 0 and m(θ) = −m(−θ) when θ < 0. (b) Show that the mean of the distribution of X1 is µ(θ) = 2θ/(1−θ2 ). (c) Show that the inverse functions of m(θ) and µ(θ) exist. Obtain ¯ where m the asymptotic relative efficiency of m−1 (m) ˆ w.r.t. µ−1 (X), ˆ ¯ is the sample median and X is the sample mean. ¯ in (d) asymptotically efficient in estimating θ? (e) Is µ−1 (X) ¯ α in (5.77) is the L-estimator corresponding to the J 65. Show that X function given in Example 5.6(iii) with β = 1 − α. 66. Let X1 , ..., Xn be i.i.d. random variables from F , where F is symmetric about θ. (a) Show that P X(j) − θ and θ − X(n−j+1) have the same distribution. n (b) Show that j=1 wj X(j) has a c.d.f. symmetric about θ, if wi ’s are Pn constants satisfying i=1 wi = 1 and wj = wn−j+1 for all j. ¯ α has a c.d.f. symmetric (c) Show that the trimmed sample mean X about θ. 67. Under the conditions in one of (i)-(iii) of Theorem 5.6, show that (5.41) holds for T(Fn ) with σF2 given by (5.79), if σF2 < ∞. 68. Prove (5.78) under the assumed conditions. 69. For the functional T given by (5.46), show that T(F ) = θ if F is R1 symmetric about θ, J is symmetric about 12 , and 0 J(t)dt = 1.

70. Obtain the asymptotic relative efficiency of the trimmed sample mean ¯ α w.r.t. the sample mean, based on an i.i.d. sample of size n from the X double exponential distribution DE(θ, 1), where θ ∈ R is unknown. 71. Obtain the asymptotic relative efficiency of the trimmed sample mean ¯ α w.r.t. the sample median, based on an i.i.d. sample of size n from X the Cauchy distribution C(θ, 1), where θ ∈ R is unknown. 72. Consider the α-trimmed sample mean defined in (5.77). Show that σα2 in (5.78) is the same as σF2 in (5.79) with J(t) = (1−2α)−1 I(α,1−α) (t), when F (x) = F0 (x − θ) and F0 is symmetric about 0. 73. For σα2 in (5.78), show that (a) if F0′ (0) exists and is positive, then limα→ 12 σα2 = 1/[2F0′ (0)]2 ; R (b) if σ 2 = x2 dF0 (x) < ∞, then limα→0 σα2 = σ 2 .

74. Show that if J ≡ 1, then σF2 in (5.79) is equal to the variance of the c.d.f. F .

75. Calculate σF2 in (5.79) with J(t) = 4t − 2 and F being the double exponential distribution DE(θ, 1), θ ∈ R.

390

5. Estimation in Nonparametric Models

76. Consider the simple linear model in Example 3.12 with positive ti ’s. Derive the L-estimator of β defined by (5.82) with a J symmetric about 12 and compare it with the LSE of β. 77. Consider the one-way ANOVA model in Example 3.13. Derive the L-estimator of β defined by (5.82) when (a) J is symmetric about 12 and (b) J(t) = (1 − 2α)−1 I(α,1−α) (t). Compare these L-estimators with the LSE of β. 78. Show that the method of moments in §3.5.2 is a special case of the GEE method. 79. Complete the proof of Proposition 5.4. 80. In the proof of Lemma 5.3, show that the probability in (5.94) is bounded by ǫ. 81. In Example 5.11, show that ψi ’s satisfy the conditions of Lemma 5.3 if Θ is compact and supi kZi k < ∞. 82. In the proof of Proposition 5.5, show that {∆n (γ)} is equicontinuous on any open subset of Θ. 83. Prove Proposition 5.6. 84. Prove the claim in Example 5.12. 85. Prove the claims in Example 5.13. 86. For Huber’s M-estimator discussed in Example 5.13, obtain a formula ¯ when F is for e(F ), the asymptotic relative efficiency of θˆn w.r.t. X, given by (5.76). Show that limτ →∞ e(F ) = ∞. Find the value of e(F ) when ǫ = 0, σ = 1, and C = 1.5. 87. Consider the ψ function in Example 5.7(ii). Show that under some conditions on F , ψ satisfies the conditions given in Theorem 5.13(i) or (ii). Obtain σF2 in (5.98) in this case. 88. In the proof of Theorem 5.14, show that (a) (5.101) holds; (b) (5.103) holds; (c) (5.104) implies (5.102). (Hint: use Theorem 1.9(iii).) 89. Prove the claim in Example 5.14, assuming some necessary moment conditions. 90. Derive the asymptotic distribution of the MQLE (the GEE estimator based on (5.90)), assuming that Xi = (Xi1 , ..., Xidi ), E(Xit ) = meηi /(1 + eηi ), Var(Xit ) = mφi eηi /(1 + eηi )2 , and (4.57) holds with t g(t) = log 1−t .

5.6. Exercises

391

ηi 91. Repeat the previous exercise under the assumption that E(Xit ) = √e , ηi Var(Xit ) = φi e , and (4.57) holds with g(t) = log t or g(t) = 2 t.

˜ i is replaced 92. In Theorem 5.14, show that result (5.99) still holds if R ˆ i satisfying maxi≤n kR ˆ i − Ui k = op (1), where Ui ’s by an estimator R are correlation matrices. 93. Show that (5.106) holds if and only if one of the following holds: (a) λ− →p 1 and λ+ →p 1, where λ− and λ+ are respectively the −1/2 ˆ −1/2 smallest and largest eigenvalues of Vn . Vn Vn (b) lnτ Vˆn ln /lnτ Vn ln →p 1, where {ln } is any sequence of nonzero vectors in Rk . −1/2 ˆ (θn − θ) →d Nk (0, Ik ). 94. Show that (5.105) and (5.106) imply Vˆn

95. Suppose that X1 , ..., Xn are independent (not necessarily identically distributed) random d-vectors with E(Xi ) = µ for all i. Suppose also that supi EkXi k2+δ < ∞ for some δ > 0. Let µ = E(X1 ), θ = g(µ), ¯ Show that and θˆn = g(X). Pn (a) (5.105) holds with Vn = n−2 [∇g(µ)]τ i=1 Var(Xi )∇g(µ); (b) Vˆn in (5.108) is consistent for Vn in part (a). 96. Consider the ratio estimator in Example 3.21. Derive the estimator Vˆn given by (5.108) and show that Vˆn is consistent for the asymptotic variance of the ratio estimator. ˆ 97. Derive a consistent variance estimator for R(t) in Example 3.23. 98. Prove the claims in Example 5.16. 99. Let σF2 n be given by (5.79) with F replaced by the empirical c.d.f. Fn . (a) Show that σF2 n /n is the same as Vˆn in (5.110) for an L-estimator with influence function φF . (b) Show directly (without using Theorem 5.15) σF2 n →a.s. σF2 in (5.79), under the conditions in Theorem 5.6(i) or (ii) (with EX12 < ∞). 100. Derive a consistent variance estimator for a U-statistic satisfying the conditions in Theorem 3.5(i). 101. Derive a consistent variance estimator for Huber’s M-estimator discussed in Example 5.13. 102. Assume the conditions in Theorem 5.8. Let r ∈ (0, 21 ). (a) Show that nr λF (T(Fn ) + n−r ) →p λF (T(F )). (b) Show that nr [λFn (T(Fn ) + n−r ) − λF (T(Fn ) + n−r )] →p 0. (c) Derive a consistent estimator of the asymptotic variance of T(Fn ), using the results in (a) and (b).

392

5. Estimation in Nonparametric Models

103. Prove Theorem 5.16. ¯ 2 . Show that the 104. Let X1 , ..., Xn be random variables and θˆ = X ¯ 2 cˆ2 ¯ cˆ3 cˆ4 −ˆ c22 4X 4X jackknife estimator in (5.112) equals n−1 − (n−1)2 + (n−1) 3 , where cˆj ’s are the sample central moments defined by (3.52). 105. Prove Theorem 5.17 for the case where g is from Rd to Rk and k ≥ 2. 106. Prove (5.114). 107. In the proof of Theorem 5.18, prove (5.116). 108. Show that θˆ−i ’s in (5.118) satisfy (5.117), under the conditions of Theorem 5.14. 109. Prove Theorem 5.19. 110. Prove (5.119). ¯ 2 . Show that the 111. Let X1 , ..., Xn be random variables and θˆ = X ∗ bootstrap variance estimator based on i.i.d. Xi ’s from Fn is equal to ¯2 ¯ cˆ3 cˆ4 −ˆ c22 ˆj ’s are the sample central moments VˆB = 4Xn cˆ2 + 4X n2 + n3 , where c defined by (3.52). 112. Let G, G1 , G2 ,..., be c.d.f.’s on R. Suppose that ̺∞ (Gj , G) → 0 as j → ∞ and G′ (x) exists and is positive for all x ∈ R. Show that −1 G−1 (p) for any p ∈ (0, 1). j (p) → G 113. Let X1 , ..., Xn be i.i.d. from a c.d.f. F on Rd with a finite Var(X1 ). Let X1∗ , ..., Xn∗ be i.i.d. from the empirical √ ¯ ∗c.d.f.¯ Fn . Show that for almost all given sequences X1 , X2 , ..., n(X − X) →d Nd (0, Var(X1 )). (Hint: verify Lindeberg’s condition.) 114. Let X1 , ..., Xn be i.i.d. from a c.d.f. F on Rd , X1∗ , ..., Xn∗ be i.i.d. from the empirical c.d.f. Fn , and let Fn∗ be the empirical c.d.f. based on Xi∗ ’s. Using DKW’s inequality (Lemma 5.1), show that (a) ̺∞ (Fn∗ , F ) →a.s. 0; (b) ̺∞ (Fn∗ , F ) = Op (n−1/2 ); (c) ̺Lp (Fn∗ , F ) = Op (n−1/2 ), under the condition in Theorem 5.20(ii). 115. Using the results from the previous two exercises, prove Theorem 5.20(ii). 116. Under the conditions in Theorem 5.11, establish a Bahadur’s representation for the bootstrap sample quantile θˆp∗ .

Chapter 6

Hypothesis Tests A general theory of testing hypotheses is presented in this chapter. Let X be a sample from a population P in P, a family of populations. Based on the observed X, we test a given hypothesis H0 : P ∈ P0 versus H1 : P ∈ P1 , where P0 and P1 are two disjoint subsets of P and P0 ∪ P1 = P. Notational conventions and basic concepts (such as two types of errors, significance levels, and sizes) given in Example 2.20 and §2.4.2 are used in this chapter.

6.1 UMP Tests A test for a hypothesis is a statistic T (X) taking values in [0, 1]. When X = x is observed, we reject H0 with probability T (x) and accept H0 with probability 1−T (x). If T (X) = 1 or 0 a.s. P, then T (X) is a nonrandomized test. Otherwise T (X) is a randomized test. For a given test T (X), the power function of T (X) is defined to be βT (P ) = E[T (X)],

P ∈ P,

(6.1)

which is the type I error probability of T (X) when P ∈ P0 and one minus the type II error probability of T (X) when P ∈ P1 . As we discussed in §2.4.2, with a sample of a fixed size, we are not able to minimize two error probabilities simultaneously. Our approach involves maximizing the power βT (P ) over all P ∈ P1 (i.e., minimizing the type II error probability) and over all tests T satisfying sup βT (P ) ≤ α,

(6.2)

P ∈P0

where α ∈ [0, 1] is a given level of significance. Recall that the left-hand side of (6.2) is defined to be the size of T . 393

394

6. Hypothesis Tests

Definition 6.1. A test T∗ of size α is a uniformly most powerful (UMP) test if and only if βT∗ (P ) ≥ βT (P ) for all P ∈ P1 and T of level α. If U (X) is a sufficient statistic for P ∈ P, then for any test T (X), E(T |U ) has the same power function as T and, therefore, to find a UMP test we may consider tests that are functions of U only. The existence and characteristics of UMP tests are studied in this section.

6.1.1 The Neyman-Pearson lemma A hypothesis H0 (or H1 ) is said to be simple if and only if P0 (or P1 ) contains exactly one population. The following useful result, which has already been used once in the proof of Theorem 4.16, provides the form of UMP tests when both H0 and H1 are simple. Theorem 6.1 (Neyman-Pearson lemma). Suppose that P0 = {P0 } and P1 = {P1 }. Let fj be the p.d.f. of Pj w.r.t. a σ-finite measure ν (e.g., ν = P0 + P1 ), j = 0, 1. (i) (Existence of a UMP test). For every α, there exists a UMP test of size α, which is equal to f1 (X) > cf0 (X) 1 T∗ (X) = (6.3) γ f1 (X) = cf0 (X) 0 f1 (X) < cf0 (X),

where γ ∈ [0, 1] and c ≥ 0 are some constants chosen so that E[T∗ (X)] = α when P = P0 (c = ∞ is allowed). (ii) (Uniqueness). If T∗∗ is a UMP test of size α, then 1 f1 (X) > cf0 (X) T∗∗ (X) = a.s. P. (6.4) 0 f1 (X) < cf0 (X) Proof. The proof for the case of α = 0 or 1 is left as an exercise. Assume now that 0 < α < 1. (i) We first show that there exist γ and c such that E0 [T∗ (X)] = α, where Ej is the expectation w.r.t. Pj . Let γ(t) = P0 (f1 (X) > tf0 (X)). Then γ(t) is nonincreasing, γ(0) = 1, and γ(∞) = 0 (why?). Thus, there exists a c ∈ (0, ∞) such that γ(c) ≤ α ≤ γ(c−). Set ( α−γ(c) γ(c−) 6= γ(c) γ(c−)−γ(c) γ= 0 γ(c−) = γ(c). Note that γ(c−) − γ(c) = P (f1 (X) = cf0 (X)). Then E0 [T∗ (X)] = P0 f1 (X) > cf0 (X) + γP0 f1 (X) = cf0 (X) = α.

395

6.1. UMP Tests

Next, we show that T∗ in (6.3) is a UMP test. Suppose that T (X) is a test satisfying E0 [T (X)] ≤ α. If T∗ (x) − T (x) > 0, then T∗ (x) > 0 and, therefore, f1 (x) ≥ cf0 (x). If T∗ (x) − T (x) < 0, then T∗ (x) < 1 and, therefore, f1 (x) ≤ cf0 (x). In any case, [T∗ (x) − T (x)][f1 (x) − cf0 (x)] ≥ 0 and, therefore, Z [T∗ (x) − T (x)][f1 (x) − cf0 (x)]dν ≥ 0, i.e.,

Z

[T∗ (x) − T (x)]f1 (x)dν ≥ c

Z

[T∗ (x) − T (x)]f0 (x)dν.

(6.5)

The left-hand side of (6.5) is E1 [T∗ (X)] − E1 [T (X)] and the right-hand side of (6.5) is c{E0 [T∗ (X)] − E0 [T (X)]} = c{α − E0 [T (X)]} ≥ 0. This proves the result in (i). (ii) Let T∗∗ (X) be a UMP test of size α. Define A = {x : T∗ (x) 6= T∗∗ (x), f1 (x) 6= cf0 (x)}. Then [T∗ (x)−T∗∗ (x)][f1 (x)−cf0 (x)] > 0 when x ∈ A and = 0 when x ∈ Ac , and Z [T∗ (x) − T∗∗ (x)][f1 (x) − cf0 (x)]dν = 0,

since both T∗ and T∗∗ are UMP tests of size α. By Proposition 1.6(ii), ν(A) = 0. This proves (6.4).

Theorem 6.1 shows that when both H0 and H1 are simple, there exists a UMP test that can be determined by (6.4) uniquely (a.s. P) except on the set B = {x : f1 (x) = cf0 (x)}. If ν(B) = 0, then we have a unique nonrandomized UMP test; otherwise UMP tests are randomized on the set B and the randomization is necessary for UMP tests to have the given size α; furthermore, we can always choose a UMP test that is constant on B. Example 6.1. Suppose that X is a sample of size 1, P0 = {P0 }, and P1 = {P1 }, where P0 is N (0, 1) and P1 is the double exponential distribution DE(0, 2) with the p.d.f. 4−1 e−|x|/2. Since P (f1 (X) = cf0 (X)) = 0, there is a unique nonrandomized UMP test. From (6.3), the UMP test T∗ (x) = 1 2 if and only if π8 ex −|x| > c2 for some c > 0, which is equivalent to |x| > t or |x| < 1 − t for some t > 21 . Suppose that α < 13 . To determine t, we use α = E0 [T∗ (X)] = P0 (|X| > t) + P0 (|X| < 1 − t).

(6.6)

If t ≤ 1, then P0 (|X| > t) ≥ P0 (|X| > 1) = 0.3374 > α. Hence t should be larger than 1 and (6.6) becomes α = P0 (|X| > t) = Φ(−t) + 1 − Φ(t).

396

6. Hypothesis Tests

Thus, t = Φ−1 (1 − α/2) and T∗ (X) = I(t,∞) (|X|). Note that it is not necessary to find out what c is. Intuitively, the reason why the UMP test in this example rejects H0 when |X| is large is that the probability of getting a large |X| is much higher under H1 (i.e., P is the double exponential distribution DE(0, 2)). The power of T∗ when P ∈ P1 is Z 1 t −|x|/2 e dx = e−t/2 . E1 [T∗ (X)] = P1 (|X| > t) = 1 − 4 −t Example 6.2. Let X1 , ..., Xn be i.i.d. binary random variables with p = P (X1 = 1). Suppose that H0 : p = p0 and H1 : p = p1 , where 0 < p0 < p1 < 1. By Theorem 6.1, a UMP test of size α is λ(Y ) > c 1 T∗ (Y ) = γ λ(Y ) = c 0 λ(Y ) < c, Pn where Y = i=1 Xi and n−Y Y 1 − p1 p1 . λ(Y ) = p0 1 − p0 Since λ(Y ) is increasing in Y , there is an integer m > 0 such that Y >m 1 T∗ (Y ) = γ Y =m 0 Y < m,

where m and γ satisfy α = E0 [T∗ (Y )] = P0 (Y > m) + γP0 (Y = m). Since Y has the binomial distribution Bi(p, n), we can determine m and γ from n X n j n m α= p0 (1 − p0 )n−j + γ p0 (1 − p0 )n−m . (6.7) j m j=m+1 Unless α=

n X n j p0 (1 − p0 )n−j j j=m+1

for some integer m, in which case we can choose γ = 0, the UMP test T∗ is a randomized test. An interesting phenomenon in Example 6.2 is that the UMP test T∗ does not depend on p1 . In such a case, T∗ is in fact a UMP test for testing H0 : p = p0 versus H1 : p > p0 .

6.1. UMP Tests

397

Lemma 6.1. Suppose that there is a test T∗ of size α such that for every P1 ∈ P1 , T∗ is UMP for testing H0 versus the hypothesis P = P1 . Then T∗ is UMP for testing H0 versus H1 . Proof. For any test T of level α, T is also of level α for testing H0 versus the hypothesis P = P1 with any P1 ∈ P1 . Hence βT ∗ (P1 ) ≥ βT (P1 ). We conclude this section with the following generalized Neyman-Pearson lemma. Its proof is left to the reader. Other extensions of the NeymanPearson lemma can be found in Exercises 8 and 9 in §6.6. Proposition 6.1. Let f1 , ..., fm+1 be Borel functions on Rp that are integrable w.r.t. a σ-finite measure ν. For given constants t1 , ..., tm , let T be the class of Borel functions φ (from Rp to [0, 1]) satisfying Z φfi dν ≤ ti , i = 1, ..., m, (6.8) and T0 be the set of φ’s in T satisfying (6.8) with all inequalities replaced by equalities. If there are constants c1 , ..., cm such that 1 fm+1 (x) > c1 f1 (x) + · · · + cm fm (x) φ∗ (x) = (6.9) 0 fm+1 (x) < c1 f1 (x) + · · · + cm fm (x) R is a member of T0 , then φR∗ maximizes φfm+1 dν over φ ∈ T0 . If ci ≥ 0 for all i, then φ∗ maximizes φfm+1 dν over φ ∈ T .

The existence of constants ci ’s in (6.9) is considered in the following lemma whose proof can be found in Lehmann (1986, pp. 97-99).

Lemma 6.2. R Let f1 , ...,Rfm and ν be given by Proposition 6.1. Then the set M = ( φf1 dν, ..., φfm dν) : φ is from Rp to [0, 1] is convex and closed. If (t1 , ..., tm ) is an interior point of M , then there exist constants c1 , ..., cm such that the function defined by (6.9) is in T0 .

6.1.2 Monotone likelihood ratio The case where both H0 and H1 are simple is mainly of theoretical interest. If a hypothesis is not simple, it is called composite. As we discussed in §6.1.1, UMP tests for composite H1 exist in the problem discussed in Example 6.2. We now extend this result to a class of parametric problems in which the likelihood functions have a special property. Definition 6.2. Suppose that the distribution of X is in P = {Pθ : θ ∈ Θ}, a parametric family indexed by a real-valued θ, and that P is dominated by a σ-finite measure ν. Let fθ = dPθ /dν. The family P is said to have

398

6. Hypothesis Tests

monotone likelihood ratio in Y (X) (a real-valued statistic) if and only if, for any θ1 < θ2 , fθ2 (x)/fθ1 (x) is a nondecreasing function of Y (x) for values x at which at least one of fθ1 (x) and fθ2 (x) is positive. The following lemma states a useful result for a family with monotone likelihood ratio. Lemma 6.3. Suppose that the distribution of X is in a parametric family P indexed by a real-valued θ and that P has monotone likelihood ratio in Y (X). If ψ is a nondecreasing function of Y , then g(θ) = E[ψ(Y )] is a nondecreasing function of θ. Proof. Let θ1 < θ2 , A = {x : fθ1 (x) > fθ2 (x)}, a = supx∈A ψ(Y (x)), B = {x : fθ1 (x) < fθ2 (x)}, and b = inf x∈B ψ(Y (x)). Since P has monotone likelihood ratio in Y (X) and ψ is nondecreasing in Y , b ≥ a. Then the result follows from Z g(θ2 ) − g(θ1 ) = ψ(Y (x))(fθ2 − fθ1 )(x)dν Z Z ≥ a (fθ2 − fθ1 )(x)dν + b (fθ2 − fθ1 )(x)dν A B Z = (b − a) (fθ2 − fθ1 )(x)dν B

≥ 0.

Before discussing UMP tests in families with monotone likelihood ratio, let us consider some examples of such families. Example 6.3. Let θ be real-valued and η(θ) be a nondecreasing function of θ. Then the one-parameter exponential family with fθ (x) = exp{η(θ)Y (x) − ξ(θ)}h(x)

(6.10)

has monotone likelihood ratio in Y (X). From Tables 1.1-1.2 (§1.3.1), this includes the binomial family {Bi(θ, r)}, the Poisson family {P (θ)}, the negative binomial family {N B(θ, r)}, the log-distribution family {L(θ)}, the normal family {N (θ, c2 )} or {N (c, θ)}, the exponential family {E(c, θ)}, the gamma family {Γ(θ, c)} or {Γ(c, θ)}, the beta family {B(θ, c)} or {B(c, θ)}, and the double exponential family {DE(c, θ)}, where r or c is known. Example 6.4. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (0, θ), where θ > 0. The Lebesgue p.d.f. of X = (X1 , ..., Xn ) is fθ (x) = θ−n I(0,θ) (x(n) ), where x(n) is the value of the largest order statistic X(n) . For θ1 < θ2 , fθ2 (x) θn I(0,θ2 ) (x(n) ) = 1n , fθ1 (x) θ2 I(0,θ1 ) (x(n) )

6.1. UMP Tests

399

which is a nondecreasing function of x(n) for x’s at which at least one of fθ1 (x) and fθ2 (x) is positive, i.e., x(n) < θ2 . Hence the family of distributions of X has monotone likelihood ratio in X(n) . Example 6.5. The following families have monotone likelihood ratio: (a) the double exponential distribution family {DE(θ, c)} with a known c; (b) the exponential distribution family {E(θ, c)} with a known c; (c) the logistic distribution family {LG(θ, c)} with a known c; (d) the uniform distribution family {U (θ, θ + 1)}; (e) the hypergeometric distribution family {HG(r, θ, N − θ)} with known r and N (Table 1.1, page 18). An example of a family that does not have monotone likelihood ratio is the Cauchy distribution family {C(θ, c)} with a known c. Hypotheses of the form H0 : θ ≤ θ0 (or H0 : θ ≥ θ0 ) versus H1 : θ > θ0 (or H1 : θ < θ0 ) are called one-sided hypotheses for any given constant θ0 . The following result provides UMP tests for testing one-sided hypotheses when the distribution of X is in a parametric family with monotone likelihood ratio. Theorem 6.2. Suppose that X has a distribution in P = {Pθ : θ ∈ Θ} (Θ ⊂ R) that has monotone likelihood ratio in Y (X). Consider the problem of testing H0 : θ ≤ θ0 versus H1 : θ > θ0 , where θ0 is a given constant. (i) There exists a UMP test of size α, which is given by Y (X) > c 1 T∗ (X) = (6.11) γ Y (X) = c 0 Y (X) < c,

where c and γ are determined by βT∗ (θ0 ) = α, and βT (θ) = E[T (X)] is the power function of a test T . (ii) βT∗ (θ) is strictly increasing for all θ’s for which 0 < βT∗ (θ) < 1. (iii) For any θ < θ0 , T∗ minimizes βT (θ) (the type I error probability of T ) among all tests T satisfying βT (θ0 ) = α. (iv) Assume that Pθ (fθ (X) = cfθ0 (X)) = 0 for any θ > θ0 and c ≥ 0, where fθ is the p.d.f. of Pθ . If T is a test with βT (θ0 ) = βT∗ (θ0 ), then for any θ > θ0 , either βT (θ) < βT∗ (θ) or T = T∗ a.s. Pθ . (v) For any fixed θ1 , T∗ is UMP for testing H0 : θ ≤ θ1 versus H1 : θ > θ1 , with size βT∗ (θ1 ). Proof. (i) Consider the hypotheses θ = θ0 versus θ = θ1 with any θ1 > θ0 . From Theorem 6.1, a UMP test is given by (6.3) with fj = the p.d.f. of Pθj , j = 0, 1. Since P has monotone likelihood ratio in Y (X), this UMP test can be chosen to be the same as T∗ in (6.11) with possibly different c and γ satisfying βT∗ (θ0 ) = α. Since T∗ does not depend on θ1 , it follows from

400

6. Hypothesis Tests

Lemma 6.1 that T∗ is UMP for testing the hypothesis θ = θ0 versus H1 . Note that if T∗ is UMP for testing θ = θ0 versus H1 , then it is UMP for testing H0 versus H1 , provided that βT∗ (θ) ≤ α for all θ ≤ θ0 , i.e., the size of T∗ is α. But this follows from Lemma 6.3, i.e., βT∗ (θ) is nondecreasing in θ. This proves (i). (ii) See Exercise 2 in §6.6. (iii) The result can be proved using Theorem 6.1 with all inequalities reversed. (iv) The proof for (iv) is left as an exercise. (v) The proof for (v) is similar to that of (i). By reversing inequalities throughout, we can obtain UMP tests for testing H0 : θ ≥ θ0 versus H1 : θ < θ0 . A major application of Theorem 6.2 is to problems with one-parameter exponential families. Corollary 6.1. Suppose that X has the p.d.f. given by (6.10) w.r.t. a σ-finite measure, where η is a strictly monotone function of θ. If η is increasing, then T∗ given by (6.11) is UMP for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 , where γ and c are determined by βT∗ (θ0 ) = α. If η is decreasing or H0 : θ ≥ θ0 (H1 : θ < θ0 ), the result is still valid by reversing inequalities in (6.11). Example 6.6. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 . Consider H0 : µ ≤ µ0 versus H1 : µ > µ0 , where µ0 is a fixed constant. The p.d.f. of X = (X1 , ..., Xn ) is of ¯ and η(µ) = nµ/σ 2 . By Corollary 6.1 and the form (6.10) with Y (X) = X 2 ¯ is N (µ, σ /n), the UMP test is T∗ (X) = I(c ,∞) (X), ¯ where the fact that √ X α cα = σz1−α / n + µ0 and za = Φ−1 (a) (see also Example 2.28). To derive a UMP test for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 when X has the p.d.f. (6.10), it is essential to know the distribution of Y (X). Typically, a nonrandomized test can be obtained if the distribution of Y is continuous; otherwise UMP tests are randomized. Example 6.7. Let X1 , ..., Xn be i.i.d. binary random variables with p = P 1 = 1). The p.d.f. of X = (X1 , ..., Xn ) is of the form (6.10) with Y = P(X n p i=1 Xi and η(p) = log 1−p . Note that η(p) is a strictly increasing function of p. By Corollary 6.1, a UMP test for H0 : p ≤ p0 versus H1 : p > p0 is given by (6.11), where c and γ are determined by (6.7) with c = m. Example 6.8. Let X1 , ..., Xn be i.i.d. random variables from the Poisson distribution P (θ) with an unknown θ > 0. The p.d.f. of X = (X1 , ..., Xn )

401

6.1. UMP Tests

P is of the form (6.10) with Y (X) = ni=1 Xi and η(θ) = log θ. Note that Y has the Poisson distribution P (nθ). By Corollary 6.1, a UMP test for H0 : θ ≤ θ0 versus H1 : θ > θ0 is given by (6.11) with c and γ satisfying α=

∞ X enθ0 (nθ0 )j enθ0 (nθ0 )c +γ . j! c! j=c+1

Example 6.9. Let X1 , ..., Xn be i.i.d. random variables from the uniform distribution U (0, θ), θ > 0. Consider the hypotheses H0 : θ ≤ θ0 and H1 : θ > θ0 . Since the p.d.f. of X = (X1 , ..., Xn ) is in a family with monotone likelihood ratio in Y (X) = X(n) (Example 6.4), by Theorem 6.2, a UMP test is of the form (6.11). Since X(n) has the Lebesgue p.d.f. nθ−n xn−1 I(0,θ) (x), the UMP test in (6.11) is nonrandomized and Z θ0 n cn xn−1 dx = 1 − n . α = βT∗ (θ0 ) = n θ0 c θ0 Hence c = θ0 (1 − α)1/n . The power function of T∗ when θ > θ0 is Z θ n θn (1 − α) βT∗ (θ) = n xn−1 dx = 1 − 0 n . θ c θ In this problem, however, UMP tests are not unique. (Note that the condition Pθ (fθ (X) = cfθ0 (X)) = 0 in Theorem 6.2(iv) is not satisfied.) It can be shown (exercise) that the following test is also UMP with size α: 1 X(n) > θ0 T (X) = α X(n) ≤ θ0 .

6.1.3 UMP tests for two-sided hypotheses The following hypotheses are called two-sided hypotheses: H0 : θ ≤ θ1 or θ ≥ θ2 H0 : θ 1 ≤ θ ≤ θ 2

H0 : θ = θ 0

versus H1 : θ1 < θ < θ2 ,

(6.12)

versus H1 : θ < θ1 or θ > θ2 ,

(6.13)

versus H1 : θ 6= θ0 ,

(6.14)

where θ0 , θ1 , and θ2 are given constants and θ1 < θ2 .

Theorem 6.3. Suppose that X has the p.d.f. given by (6.10) w.r.t. a σfinite measure, where η is a strictly increasing function of θ. (i) For testing hypotheses (6.12), a UMP test of size α is c1 < Y (X) < c2 1 T∗ (X) = (6.15) γ Y (X) = ci , i = 1, 2 i 0 Y (X) < c1 or Y (X) > c2 ,

402

6. Hypothesis Tests

where ci ’s and γi ’s are determined by βT∗ (θ1 ) = βT∗ (θ2 ) = α.

(6.16)

(ii) The test defined by (6.15) minimizes βT (θ) over all θ < θ1 , θ > θ2 , and T satisfying βT (θ1 ) = βT (θ2 ) = α. (iii) If T∗ and T∗∗ are two tests satisfying (6.15) and βT∗ (θ1 ) = βT∗∗ (θ1 ) and if the region {T∗∗ = 1} is to the right of {T∗ = 1}, then βT∗ (θ) < βT∗∗ (θ) for θ > θ1 and βT∗ (θ) > βT∗∗ (θ) for θ < θ1 . If both T∗ and T∗∗ satisfy (6.15) and (6.16), then T∗ = T∗∗ a.s. P. Proof. (i) The distribution of Y has a p.d.f. gθ (y) = exp{η(θ)y − ξ(θ)}

(6.17)

(Theorem 2.1). Since Y is sufficient for θ, we only need to consider tests of the form T (Y ). Let θ1 < θ3 < θ2 . Consider the problem of testing θ = θ1 or θ = θ2 versus θ = θ3 . Clearly, (α, α) is an interior point of the set of all points (βT (θ1 ), βT (θ2 )) as T ranges over all tests of the form T (Y ). By (6.17) and Lemma 6.2, there are constants c˜1 and c˜2 such that 1 a1 eb1 Y + a2 eb2 Y < 1 T∗ (Y ) = 0 a1 eb1 Y + a2 eb2 Y > 1 satisfies (6.16), where ai = c˜i eξ(θ3 )−ξ(θi ) and bi = η(θi ) − η(θ3 ), i = 1, 2. Clearly ai ’s cannot both be ≤ 0. If one of the ai ’s is ≤ 0 and the other is > 0, then a1 eb1 Y + a2 eb2 Y is strictly monotone (since b1 < 0 < b2 ) and T∗ or 1 − T∗ is of the form (6.11), which has a strictly monotone power function (Theorem 6.2) and, therefore, cannot satisfy (6.16). Thus, both ai ’s are positive. Then, T∗ is of the form (6.15) (since b1 < 0 < b2 ) and it follows from Proposition 6.1 that T∗ is UMP for testing θ = θ1 or θ = θ2 versus θ = θ3 . Since T∗ does not depend on θ3 , it follows from Lemma 6.1 that T∗ is UMP for testing θ = θ1 or θ = θ2 versus H1 . To show that T∗ is a UMP test of size α for testing H0 versus H1 , it remains to show that βT∗ (θ) ≤ α for θ ≤ θ1 or θ ≥ θ2 . But this follows from part (ii) of the theorem by comparing T∗ with the test T (Y ) ≡ α. (ii) The proof is similar to that in (i) and is left as an exercise. (iii) The first claim in (iii) follows from Lemma 6.4, since the function T∗∗ − T∗ has a single change of sign. The second claim in (iii) follows from the first claim. Lemma 6.4. Suppose that X has a p.d.f. in {fθ (x) : θ ∈ Θ}, a parametric family of p.d.f.’s w.r.t. a single σ-finite measure ν on R, where Θ ⊂ R. Suppose that this family has monotone likelihood ratio in X. Let ψ be a function with a single change of sign. (i) There exists θ0 ∈ Θ such that Eθ [ψ(X)] ≤ 0 for θ < θ0 and Eθ [ψ(X)] ≥ 0

403

6.1. UMP Tests

for θ > θ0 , where Eθ is the expectation w.r.t. fθ . (ii) Suppose that fθ (x) > 0 for all x and θ, that fθ1 (x)/fθ (x) is strictly increasing in x for θ < θ1 , and that ν({x : ψ(x) 6= 0}) > 0. If Eθ0 [ψ(X)] = 0, then Eθ [ψ(X)] < 0 for θ < θ0 and Eθ [ψ(X)] > 0 for θ > θ0 . Proof. (i) Suppose that there is an x0 ∈ R such that ψ(x) ≤ 0 for x < x0 and ψ(x) ≥ 0 for x > x0 . Let θ1 < θ2 . We first show that Eθ1 [ψ(X)] > 0 implies Eθ2 [ψ(X)] ≥ 0. If fθ2 (x0 )/fθ1 (x0 ) = ∞, then fθ1 (x) = 0 for x ≥ x0 and, therefore, Eθ1 [ψ(X)] ≤ 0. Hence fθ2 (x0 )/fθ1 (x0 ) = c < ∞. Then ψ(x) ≥ 0 on the set A = {x : fθ1 (x) = 0 and fθ2 (x) > 0}. Thus, Z fθ Eθ2 [ψ(X)] ≥ ψ 2 fθ1 dν fθ1 c Z ZA cψfθ1 dν + cψfθ1 dν (6.18) ≥ x<x0

x≥x0

= cEθ1 [ψ(X)].

The result follows by letting θ0 = inf{θ : Eθ [ψ(X)] > 0}. (ii) Under the assumed conditions, fθ2 (x0 )/fθ1 (x0 ) = c < ∞. The result follows from the proof in (i) with θ1 replaced by θ0 and the fact that ≥ should be replaced by > in (6.18) under the assumed conditions. Part (iii) of Theorem 6.3 shows that the ci ’s and γi ’s are uniquely determined by (6.15) and (6.16). It also indicates how to select the ci ’s and (0) (0) (0) (0) γi ’s. One can start with some trial values c1 and γ1 , find c2 and γ2 such that βT∗ (θ1 ) = α, and compute βT∗ (θ2 ). If βT∗ (θ2 ) < α, by Theorem 6.3(iii), the correct rejection region {T∗ = 1} is to the right of the one (1) (0) (1) (0) (1) (0) chosen so that one should try c1 > c1 or c1 = c1 and γ1 < γ1 ; the converse holds if βT∗ (θ2 ) > α. Example 6.10. Let X1 , ..., Xn be i.i.d. from N (θ, 1). By Theorem 6.3, a ¯ where ci ’s are deterUMP test for testing (6.12) is T∗ (X) = I(c1 ,c2 ) (X), mined by √ √ Φ n(c2 − θ1 ) − Φ n(c1 − θ1 ) = α and

Φ

√ √ n(c2 − θ2 ) − Φ n(c1 − θ2 ) = α.

When the distribution of X is not given by (6.10), UMP tests for hypotheses (6.12) exist in some cases (see Exercises 17 and 26). Unfortunately, a UMP test does not exist in general for testing hypotheses (6.13) or (6.14) (Exercises 28 and 29). A key reason for this phenomenon is that UMP tests for testing one-sided hypotheses do not have level α for testing (6.12); but they are of level α for testing (6.13) or (6.14) and there does not exist a single test more powerful than all tests that are UMP for testing one-sided hypotheses.

404

6. Hypothesis Tests

6.2 UMP Unbiased Tests When a UMP test does not exist, we may use the same approach used in estimation problems, i.e., imposing a reasonable restriction on the tests to be considered and finding optimal tests within the class of tests under the restriction. Two such types of restrictions in estimation problems are unbiasedness and invariance. We consider unbiased tests in this section. The class of invariant tests is studied in §6.3.

6.2.1 Unbiasedness, similarity, and Neyman structure A UMP test T of size α has the property that βT (P ) ≤ α,

P ∈ P0

and

βT (P ) ≥ α,

P ∈ P1 .

(6.19)

This means that T is at least as good as the silly test T ≡ α. Thus, we have the following definition. Definition 6.3. Let α be a given level of significance. A test T for H0 : P ∈ P0 versus H1 : P ∈ P1 is said to be unbiased of level α if and only if (6.19) holds. A test of size α is called a uniformly most powerful unbiased (UMPU) test if and only if it is UMP within the class of unbiased tests of level α. Since a UMP test is UMPU, the discussion of unbiasedness of tests is useful only when a UMP test does not exist. In a large class of problems for which a UMP test does not exist, there do exist UMPU tests. Suppose that U is a sufficient statistic for P ∈ P. Then, similar to the search for a UMP test, we need to consider functions of U only in order to find a UMPU test, since, for any unbiased test T (X), E(T |U ) is unbiased and has the same power function as T . Throughout this section, we consider the following hypotheses: H0 : θ ∈ Θ 0

versus

H1 : θ ∈ Θ 1 ,

(6.20)

where θ = θ(P ) is a functional from P onto Θ and Θ0 and Θ1 are two disjoint Borel sets with Θ0 ∪ Θ1 = Θ. Note that Pj = {P : θ ∈ Θj }, j = 0, 1. For instance, X1 , ..., Xn are i.i.d. from F but we are interested in testing H0 : θ ≤ 0 versus H1 : θ > 0, where θ = EX1 or the median of F . Definition 6.4. Consider the hypotheses specified by (6.20). Let α be a ¯ 01 be the common boundary of Θ0 and given level of significance and let Θ Θ1 , i.e., the set of points θ that are points or limit points of both Θ0 and ¯ 01 if and only if Θ1 . A test T is similar on Θ ¯ 01 . βT (P ) = α, θ∈Θ (6.21)

405

6.2. UMP Unbiased Tests

It is more convenient to work with (6.21) than to work with (6.19) when the hypotheses are given by (6.20). Thus, the following lemma is useful. For a given test T , the power function βT (P ) is said to be continuous in θ if and only if for any {θj : j = 0, 1, 2, ...} ⊂ Θ, θj → θ0 implies βT (Pj ) → βT (P0 ), where Pj ∈ P satisfying θ(Pj ) = θj , j = 0, 1,.... Note that if βT is a function of θ, then this continuity property is simply the continuity of βT (θ). Lemma 6.5. Consider hypotheses (6.20). Suppose that, for every T , βT (P ) is continuous in θ. If T∗ is uniformly most powerful among all tests satisfying (6.21) and has size α, then T∗ is a UMPU test. Proof. Under the continuity assumption on βT , the class of tests satisfying (6.21) contains the class of tests satisfying (6.19). Since T∗ is uniformly at least as powerful as the test T ≡ α, T∗ is unbiased. Hence, T∗ is a UMPU test. Using Lemma 6.5, we can derive a UMPU test for testing hypotheses given by (6.13) or (6.14), when X has the p.d.f. (6.10) in a one-parameter exponential family. (Note that a UMP test does not exist in these cases.) We do not provide the details here, since the results for one-parameter exponential families are special cases of those in §6.2.2 for multiparameter exponential families. To prepare for the discussion in §6.2.2, we introduce the following result that simplifies (6.21) when there is a statistic sufficient ¯ 01 }. and complete for P ∈ P¯ = {P : θ(P ) ∈ Θ Let U (X) be a sufficient statistic for P ∈ P¯ and let P¯U be the family of ¯ If T is a test satisfying distributions of U as P ranges over P. E[T (X)|U ] = α

a.s. P¯U ,

(6.22)

then

¯ E[T (X)] = E{E[T (X)|U ]} = α P ∈ P, ¯ 01 . A test satisfying (6.22) is said to have Neyman i.e., T is similar on Θ ¯ 01 have Neyman structure w.r.t. structure w.r.t. U . If all tests similar on Θ U , then working with (6.21) is the same as working with (6.22). ¯ Then a necLemma 6.6. Let U (X) be a sufficient statistic for P ∈ P. ¯ essary and sufficient condition for all tests similar on Θ01 to have Neyman ¯ structure w.r.t. U is that U is boundedly complete for P ∈ P. ¯ Let Proof. (i) Suppose first that U is boundedly complete for P ∈ P. ¯ ¯ T (X) be a test similar on Θ01 . Then E[T (X) − α] = 0 for all P ∈ P. From the boundedness of T (X), E[T (X)|U ] is bounded (Proposition 1.10). Since ¯ (6.22) holds. E{E[T (X)|U ] − α} = E[T (X) − α] = 0 for all P ∈ P, ¯ Then (ii) Suppose now that U is not boundedly complete for P ∈ P. ¯ and there is a function h such that |h(u)| ≤ C, E[h(U )] = 0 for all P ∈ P, ¯ h(U ) 6= 0 with positive probability for some P ∈ P. Let T (X) = α + ch(U ),

406

6. Hypothesis Tests

where c = min{α, 1 − α}/C. The result follows from the fact that T is a ¯ 01 but does not have Neyman structure w.r.t. U . test similar on Θ

6.2.2 UMPU tests in exponential families Suppose that the distribution of X is in a multiparameter natural exponential family (§2.1.3) with the following p.d.f. w.r.t. a σ-finite measure: fθ,ϕ(x) = exp {θY (x) + ϕτ U (x) − ζ(θ, ϕ)} ,

(6.23)

where θ is a real-valued parameter, ϕ is a vector-valued parameter, and Y (real-valued) and U (vector-valued) are statistics. It follows from Theorem 2.1(i) that the p.d.f. of (Y, U ) (w.r.t. a σ-finite measure) is in a natural exponential family of the form exp {θy + ϕτ u − ζ(θ, ϕ)} and, given U = u, the p.d.f. of the conditional distribution of Y (w.r.t. a σ-finite measure νu ) is in a natural exponential family of the form exp {θy − ζu (θ)}. Theorem 6.4. Suppose that the distribution of X is in a multiparameter natural exponential family given by (6.23). (i) For testing H0 : θ ≤ θ0 versus H1 : θ > θ0 , a UMPU test of size α is Y > c(U ) 1 (6.24) T∗ (Y, U ) = γ(U ) Y = c(U ) 0 Y < c(U ),

where c(u) and γ(u) are Borel functions determined by Eθ0 [T∗ (Y, U )|U = u] = α

for every u, and Eθ0 is the expectation w.r.t. fθ0 ,ϕ . (ii) For testing hypotheses (6.12), a UMPU test of size α is c1 (U ) < Y < c2 (U ) 1 T∗ (Y, U ) = Y = ci (U ), i = 1, 2, γi (U ) 0 Y < c1 (U ) or Y > c2 (U ),

(6.25)

(6.26)

where ci (u)’s and γi (u)’s are Borel functions determined by Eθ1 [T∗ (Y, U )|U = u] = Eθ2 [T∗ (Y, U )|U = u] = α

for every u. (iii) For testing hypotheses (6.13), a UMPU test of size α is Y < c1 (U ) or Y > c2 (U ) 1 T∗ (Y, U ) = Y = ci (U ), i = 1, 2, γ (U ) i 0 c1 (U ) < Y < c2 (U ),

(6.27)

(6.28)

407

6.2. UMP Unbiased Tests

where ci (u)’s and γi (u)’s are Borel functions determined by (6.27) for every u. (iv) For testing hypotheses (6.14), a UMPU test of size α is given by (6.28), where ci (u)’s and γi (u)’s are Borel functions determined by (6.25) and Eθ0 [T∗ (Y, U )Y |U = u] = αEθ0 (Y |U = u)

(6.29)

for every u. Proof. Since (Y, U ) is sufficient for (θ, ϕ), we only need to consider tests that are functions of (Y, U ). Hypotheses in (i)-(iv) are of the form (6.20) ¯ 01 = {(θ, ϕ) : θ = θ0 } or = {(θ, ϕ) : θ = θi , i = 1, 2}. In case (i) or with Θ (iv), U is sufficient and complete for P ∈ P¯ and, hence, Lemma 6.6 applies. In case (ii) or (iii), applying Lemma 6.6 to each {(θ, ϕ) : θ = θi } also shows that working with (6.21) is the same as working with (6.22). By Theorem 2.1, the power functions of all tests are continuous and, hence, Lemma 6.5 applies. Thus, for (i)-(iii), we only need to show that T∗ is UMP among all tests T satisfying (6.25) (for part (i)) or (6.27) (for part (ii) or (iii)) with T∗ replaced by T . For (iv), any unbiased T should satisfy (6.25) with T∗ replaced by T and ∂ Eθ,ϕ [T (Y, U )] = 0, ∂θ

¯ 01 . θ∈Θ

(6.30)

By Theorem 2.1, the differentiation can be carried out under the expectation sign. Hence, one can show (exercise) that (6.30) is equivalent to Eθ,ϕ [T (Y, U )Y − αY ] = 0,

¯ 01 . θ∈Θ

(6.31)

Using the argument in the proof of Lemma 6.6, one can show (exercise) that (6.31) is equivalent to (6.29) with T∗ replaced by T . Hence, to prove (iv) we only need to show that T∗ is UMP among all tests T satisfying (6.25) and (6.29) with T∗ replaced by T . Note that the power function of any test T (Y, U ) is Z Z βT (θ, ϕ) = T (y, u)dPY |U=u (y) dPU (u). Thus, it suffices to show that for every fixed u and θ ∈ Θ1 , T∗ maximizes Z T (y, u)dPY |U =u (y) over all T subject to the given side conditions. Since PY |U=u is in a one-parameter exponential family, the results in (i) and (ii) follow from Corollary 6.1 and Theorem 6.3, respectively. The result in (iii) follows from Theorem 6.3(ii) by considering 1 − T∗ with T∗ given by (6.15). To

408

6. Hypothesis Tests

prove the result in (iv), it suffices to show that if Y has the p.d.f. given by (6.10) and if U is treated as a constant in (6.25), (6.28), and (6.29), T∗ in (6.28) is UMP subject to conditions (6.25) and (6.29). We now omit U in the following proof for (iv), which is very similar to the proof of Theorem 6.3. First, (α, αEθ0 (Y )) is an interior point of the set of points (Eθ0 [T (Y )], Eθ0 [T (Y )Y ]) as T ranges over all tests of the form T (Y ) (exercise). By Lemma 6.2 and Proposition 6.1, for testing θ = θ0 versus θ = θ1 , the UMP test is equal to 1 when (k1 + k2 y)eθ0 y < C(θ0 , θ1 )eθ1 y ,

(6.32)

where ki ’s and C(θ0 , θ1 ) are constants. Note that (6.32) is equivalent to a1 + a2 y < eby for some constants a1 , a2 , and b. This region is either one-sided or the outside of an interval. By Theorem 6.2(ii), a one-sided test has a strictly monotone power function and therefore cannot satisfy (6.29). Thus, this test must have the form (6.28). Since T∗ in (6.28) does not depend on θ1 , by Lemma 6.1, it is UMP over all tests satisfying (6.25) and (6.29); in particular, the test ≡ α. Thus, T∗ is UMPU. Finally, it can be shown that all the c- and γ-functions in (i)-(iv) are Borel functions (see Lehmann (1986, p. 149)). Example 6.11. A problem arising in many different contexts is the comparison of two treatments. If the observations are integer-valued, the problem often reduces to testing the equality of two Poisson distributions (e.g., a comparison of the radioactivity of two substances or the car accident rate in two cities) or two binomial distributions (when the observation is the number of successes in a sequence of trials for each treatment). Consider first the Poisson problem in which X1 and X2 are independently distributed as the Poisson distributions P (λ1 ) and P (λ2 ), respectively. The p.d.f. of X = (X1 , X2 ) is e−(λ1 +λ2 ) exp {x2 log(λ2 /λ1 ) + (x1 + x2 ) log λ1 } x1 !x2 !

(6.33)

w.r.t. the counting measure on {(i, j) : i = 0, 1, 2, ..., j = 0, 1, 2, ...}. Let θ = log(λ2 /λ1 ). Then hypotheses such as λ1 = λ2 and λ1 ≥ λ2 are equivalent to θ = 0 and θ ≤ 0, respectively. The p.d.f. in (6.33) is of the form (6.23) with ϕ = log λ1 , Y = X2 , and U = X1 + X2 . Thus, Theorem 6.4 applies. To obtain various tests in Theorem 6.4, it is enough to derive the conditional distribution of Y = X2 given U = X1 + X2 = u. Using the fact that X1 + X2 has the Poisson distribution P (λ1 + λ2 ), one can show that u y P (Y = y|U = u) = p (1 − p)u−y I{0,1,...,u} (y), u = 0, 1, 2, ..., y

409

6.2. UMP Unbiased Tests

where p = λ2 /(λ1 + λ2 ) = eθ /(1 + eθ ). This is the binomial distribu¯ 01 , θ = θj (a known value) and the tion Bi(p, u). On the boundary set Θ distribution PY |U =u is known. The previous result can obviously be extended to the case where two independent samples, Xi1 , ..., Xini , i = 1, 2, are i.i.d. from the Poisson distributions P (λi ), i = 1, 2, respectively. Consider next the binomial problem in which Xj , j = 1, 2, are independently distributed as the binomial distributions Bi(pj , nj ), j = 1, 2, respectively, where nj ’s are known but pj ’s are unknown. The p.d.f. of X = (X1 , X2 ) is o n n2 n1 p1 1) (1 − p1 )n1 (1 − p2 )n2 exp x2 log pp21 (1−p + (x1 + x2 ) log (1−p (1−p ) ) 2 1 x1 x2 w.r.t. the counting measure on {(i, j) : i = 0, 1, ..., n1 , j = 0, 1, ..., n2 }. This 1) p.d.f. is of the form (6.23) with θ = log pp21 (1−p (1−p2 ) , Y = X2 , and U = X1 +X2 . Thus, Theorem 6.4 applies. Note that hypotheses such as p1 = p2 and p1 ≥ p2 are equivalent to θ = 0 and θ ≤ 0, respectively. Using the joint distribution of (X1 , X2 ), one can show (exercise) that n2 θy n1 e IA (y), P (Y = y|U = u) = Ku (θ) u−y y

u = 0, 1, ..., n1 + n2 ,

where A = {y : y = 0, 1, ..., min{u, n2 }, u − y ≤ n1 } and −1 X n1 n2 eθy . Ku (θ) = u−y y

(6.34)

y∈A

If θ = 0, this distribution reduces to a known distribution: the hypergeometric distribution HG(u, n2 , n1 ) (Table 1.1, page 18). Example 6.12 (2 × 2 contingency tables). Let A and B be two different events in a probability space related to a random experiment. Suppose that n independent trials of the experiment are carried out and that we observe the frequencies of the occurrence of the events A ∩ B, A ∩ B c , Ac ∩ B, and Ac ∩ B c . The results can be summarized in the following 2 × 2 contingency table:

B Bc Total

A X11 X21 m1

Ac X12 X22 m2

Total n1 n2 n

410

6. Hypothesis Tests

The distribution of X = (X11 , X12 , X21 , X22 ) is multinomial (Example 2.7) with probabilities p11 , p12 , p21 , and p22 , where pij = E(Xij )/n. Thus, the p.d.f. of X is o n n! pn22 exp x11 log pp11 + x12 log pp12 + x21 log pp21 22 22 22 x11 !x12 !x21 !x22 ! w.r.t. the counting measure on the range of X. This p.d.f. is clearly of the form (6.23). By Theorem 6.4, we can derive UMPU tests for any parameter of the form θ = a0 log pp11 + a1 log pp12 + a2 log pp21 , 22 22 22 where ai ’s are given constants. In particular, testing independence of A and B is equivalent to the hypotheses H0 : θ = 0 versus H1 : θ 6= 0 when a0 = 1 and a1 = a2 = −1 (exercise). For hypotheses concerning θ with a0 = 1 and a1 = a2 = −1, the p.d.f. of X can be written as (6.23) with Y = X11 and U = (X11 + X12 , X11 + X21 ). A direct calculation shows that P (Y = y|X11 + X12 = n1 , X11 + X21 = m1 ) is equal to n2 n1 eθ(m1 −y) IA (y), Km1 (θ) y m1 − y where A = {y : y = 0, 1, ..., min{m1 , n1 }, m1 − y ≤ n2 } and Ku (θ) is given by (6.34). This distribution is known when θ = θj is known. In particular, for testing independence of A and B, θ = 0 implies that PY |U =u is the hypergeometric distribution HG(m1 , n1 , n2 ), and the UMPU test in Theorem 6.4(iv) is also known as Fisher’s exact test. Suppose that Xij ’s in the 2 × 2 contingency table are from two binomial distributions, i.e., Xi1 is from the binomial distribution Bi(pi , ni ), Xi2 = ni − Xi1 , i = 1, 2, and that Xi1 ’s are independent. Then the UMPU test for independence of A and B previously derived is exactly the same as the UMPU test for p1 = p2 given in Example 6.11. The only difference is that ni ’s are fixed for testing the equality of two binomial distributions, whereas ni ’s are random for testing independence of A and B. This is also true for the general r × c contingency tables considered in §6.4.3.

6.2.3 UMPU tests in normal families An important application of Theorem 6.4 to problems with continuous distributions in exponential families is the derivation of UMPU tests in normal families. The results presented here are the basic justifications for tests in elementary textbooks concerning parameters in normal families. We start with the following lemma, which is useful especially when X is from a population in a normal family.

6.2. UMP Unbiased Tests

411

Lemma 6.7. Suppose that X has the p.d.f. (6.23) and that V (Y, U ) is a statistic independent of U when θ = θj , where θj ’s are known values given in the hypotheses in (i)-(iv) of Theorem 6.4. (i) If V (y, u) is increasing in y for each u, then the UMPU tests in (i)-(iii) of Theorem 6.4 are equivalent to those given by (6.24)-(6.28) with Y and (Y, U ) replaced by V and with ci (U ) and γi (U ) replaced by constants ci and γi , respectively. (ii) If there are Borel functions a(u) > 0 and b(u) such that V (y, u) = a(u)y + b(u), then the UMPU test in Theorem 6.4(iv) is equivalent to that given by (6.25), (6.28), and (6.29) with Y and (Y, U ) replaced by V and with ci (U ) and γi (U ) replaced by constants ci and γi , respectively. Proof. (i) Since V is increasing in y, Y > ci (u) is equivalent to V > di (u) for some di . The result follows from the fact that V is independent of U so that di ’s and γi ’s do not depend on u when Y is replaced by V . (ii) Since V = a(U )Y + b(U ), the UMPU test in Theorem 6.4(iv) is the same as V < c1 (U ) or V > c2 (U ) 1 T∗ (V, U ) = (6.35) V = ci (U ), i = 1, 2, γi (U ) 0 c1 (U ) < V < c2 (U ), subject to Eθ0 [T∗ (V, U )|U = u] = α and V − b(U ) V − b(U ) U = αEθ0 U . Eθ0 T∗ (V, U ) a(U ) a(U )

(6.36)

Under Eθ0 [T∗ (V, U )|U = u] = α, (6.36) is the same as Eθ0 [T∗ (V, U )V |U ] = αEθ0 (V |U ). Since V and U are independent when θ = θ0 , ci (u)’s and γi (u)’s do not depend on u and, therefore, T∗ in (6.35) does not depend on U. If the conditions of Lemma 6.7 are satisfied, then UMPU tests can be derived by working with the distribution of V instead of PY |U=u . In exponential families, a V (Y, U ) independent of U can often be found by applying Basu’s theorem (Theorem 2.4). When we consider normal families, γi ’s can be chosen to be 0 since the c.d.f. of Y given U = u or the c.d.f. of V is continuous. One-sample problems Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ ∈ R and σ 2 > 0, where n ≥ 2. The joint p.d.f. of X = (X1 , ..., Xn ) is ) ( n n 1 X 2 µ X nµ2 1 exp − 2 x + 2 xi − 2 . 2σ i=1 i σ i=1 2σ (2πσ 2 )n/2

412

6. Hypothesis Tests

Consider first hypotheses concerning σ 2 . The Pnp.d.f. of X has the¯ form (6.23) with θ = −(2σ 2 )−1 , ϕ = nµ/σ 2 , Y = i=1 Xi2 , and U = X. By ¯ (Example 2.18), Basu’s theorem, V = (n − 1)S 2 is independent of U = X where S 2 is the sample variance. Also, n X i=1

2

¯ 2, Xi2 = (n − 1)S 2 + nX

i.e., V = Y − nU . Hence the conditions of Lemma 6.7 are satisfied. Since V /σ 2 has the chi-square distribution χ2n−1 (Example 2.18), values of ci ’s for hypotheses in (i)-(iii) of Theorem 6.4 are related to quantiles of χ2n−1 . For testing H0 : θ = θ0 versus H1 : θ 6= θ0 (which is equivalent to testing H0 : σ 2 = σ02 versus H1 : σ 2 6= σ02 ), di = ci /σ02 , i = 1, 2, are determined by Z d2 Z d2 fn−1 (v)dv = 1 − α and vfn−1 (v)dv = (n − 1)(1 − α), d1

d1

where fm is the Lebesgue p.d.f. of the chi-square distribution χ2m . Since vfn−1 (v) = (n − 1)fn+1 (v), d1 and d2 are determined by Z d2 Z d2 fn−1 (v)dv = fn+1 (v)dv = 1 − α. d1

d1

If n − 1 ≈ n + 1, then d1 and d2 are nearly the (α/2)th and (1 − α/2)th quantiles of χ2n−1 , respectively, in which case the UMPU test in Theorem 6.4(iv) is the same as the “equal-tailed” chi-square test for H0 in elementary textbooks. Consider next hypotheses concerning µ. The p.d.f. of X has the form ¯ U = Pn (Xi − µ0 )2 , θ = n(µ − µ0 )/σ 2 , and ϕ = (6.23) with Y = X, i=1 −(2σ 2 )−1 . For testing √ ¯ hypotheses H0 : µ ≤ µ0 versus H1 : µ > µ0 , we take V to be t(X) = n(X − µ0 )/S. By Basu’s theorem, t(X) is independent of U when µ = µ0 . Hence it satisfies the conditions in Lemma 6.7(i). From Examples 1.16 and 2.18, t(X) has the t-distribution tn−1 when µ = µ0 . Thus, c(U ) in Theorem 6.4(i) is the (1 − α)th quantile of tn−1 . For the two-sided√hypotheses H0 : µ = µ0 versus H1 : µ 6= µ0 , the statistic V = ¯ (X−µ 0 )/ U satisfies the conditions in Lemma 6.7(ii) and has a distribution symmetric about 0 when µ = µ0 . Then the UMPU test in Theorem 6.4(iv) rejects H0 when |V | > d, where d satisfies P (|V | > d) = α when µ = µ0 . Since p p t(X) = (n − 1)nV (X) 1 − n[V (X)]2 , the UMPU test rejects H0 if and only if |t(X)| > tn−1,α/2 , where tn−1,α is the (1 − α)th quantile of the t-distribution tn−1 . The UMPU tests derived here are the so-called one-sample t-tests in elementary textbooks. The power function of a one-sample t-test is related to the noncentral t-distribution introduced in §1.3.1 (see Exercise 36).

413

6.2. UMP Unbiased Tests

Two-sample problems The problem of comparing the parameters of two normal distributions arises in the comparison of two treatments, products, and so on (see also Example 6.11). Suppose that we have two independent samples, Xi1 , ..., Xini , i = 1, 2, i.i.d. from N (µi , σi2 ), i = 1, 2, respectively, where ni ≥ 2. The joint p.d.f. of Xij ’s is ni 2 2 X X X 1 ni µi 2 C(µ1 , µ2 , σ12 , σ22 ) exp − x + x ¯i , ij 2σi2 σi2 i=1

j=1

i=1

where x¯i is the sample mean based on xi1 , ..., xini and C(·) is a known function. Consider first the hypothesis H0 : σ22 /σ12 ≤ ∆0 or H0 : σ22 /σ12 = ∆0 . The p.d.f. of Xij ’s is of the form (6.23) with 1 1 1 n1 µ1 n2 µ2 , θ= − , ϕ = − , , 2∆0 σ12 2σ22 2σ12 σ12 σ22 n2 n2 n1 X X X 1 2 2 2 ¯1, X ¯2 . Y = X2j , U = X1j + X2j , X ∆ 0 j=1 j=1 j=1

To apply Lemma 6.7, consider V =

(Y − n2 U3 )/∆0 (n2 − 1)S22 /∆0 = , 2 2 (n1 − 1)S1 + (n2 − 1)S2 /∆0 U1 − n1 U2 − n2 U3 /∆0

where Si2 is the sample variance based on Xi1 , ..., Xini and Uj is the jth component of U . By Basu’s theorem, V and U are independent when θ = 0 (σ22 = ∆0 σ12 ). Since V is increasing and linear in Y , the conditions of Lemma 6.7 are satisfied. Thus, a UMPU test rejects H0 : θ ≤ 0 (which is equivalent to H0 : σ22 /σ12 ≤ ∆0 ) when V > c0 , where c0 satisfies P (V > c0 ) = α when θ = 0; and a UMPU test rejects H0 : θ = 0 (which is equivalent to H0 : σ22 /σ12 = ∆0 ) when V < c1 or V > c2 , where ci ’s satisfy P (c1 < V < c2 ) = 1 − α and E[V T∗ (V )] = αE(V ) when θ = 0. Note that V =

(n2 − 1)F n1 − 1 + (n2 − 1)F

with

F=

S22 /∆0 . S12

It follows from Example 1.16 that F has the F-distribution Fn2 −1,n1 −1 (Table 1.2, page 20) when θ = 0. Since V is a strictly increasing function of F, a UMPU test rejects H0 : θ ≤ 0 when F > Fn2 −1,n1 −1,α , where Fa,b,α is the (1 − α)th quantile of the F-distribution Fa,b . This is the F-test in elementary textbooks.

414

6. Hypothesis Tests

When θ = 0, V has the beta distribution B((n2 − 1)/2, (n1 − 1)/2) and E(V ) = (n2 − 1)/(n1 + n2 − 2) (Table 1.2). Then, E[V T∗ (V )] = αE(V ) when θ = 0 is the same as Z c2 (1 − α)(n2 − 1) = vf(n2 −1)/2,(n1 −1)/2 (v)dv, n1 + n2 − 2 c1 where fa,b is the p.d.f. of the beta distribution B(a, b). Using the fact that vf(n2 −1)/2,(n1 −1)/2 (v) = (n1 + n2 − 2)−1 (n2 − 1)f(n2 +1)/2,(n1 −1)/2 (v), we conclude that a UMPU test rejects H0 : θ = 0 when V < c1 or V > c2 , where c1 and c2 are determined by Z c2 Z c2 1−α= f(n2 −1)/2,(n1 −1)/2 (v)dv = f(n2 +1)/2,(n1 −1)/2 (v)dv. c1

c1

If n2 − 1 ≈ n2 + 1 (i.e., n2 is large), then this UMPU test can be approximated by the F-test that rejects H0 : θ = 0 if and only if F < Fn2 −1,n1 −1,1−α/2 or F > Fn2 −1,n1 −1,α/2 . Consider next the hypothesis H0 : µ1 ≥ µ2 or H0 : µ1 = µ2 . If σ12 6= σ22 , the problem is the so-called Behrens-Fisher problem and is not accessible by the method introduced in this section. We now assume that σ12 = σ22 = σ 2 but σ 2 is unknown. The p.d.f. of Xij ’s is then ni 2 X 1 X n1 µ1 n2 µ2 x2ij + 2 x ¯1 + 2 x ¯2 , C(µ1 , µ2 , σ 2 ) exp − 2 2σ σ σ i=1 j=1

which is of the form (6.23) with µ2 − µ1 , θ = −1 2 (n1 + n−1 2 )σ ¯1, ¯2 − X Y =X

ϕ=

1 n1 µ1 + n2 µ2 , − 2 (n1 + n2 )σ 2 2σ

¯ 1 + n2 X ¯2, U = n1 X

ni 2 X X i=1 j=1

,

2 Xij .

For testing H0 : θ ≤ 0 (i.e., µ1 ≥ µ2 ) versus H1 : θ > 0, we consider V in Lemma 6.7 to be q −1 ¯2 − X ¯1) (X n1 + n−1 2 t(X) = p . (6.37) 2 2 [(n1 − 1)S1 + (n2 − 1)S2 ]/(n1 + n2 − 2) When θ = 0, t(X) is independent of U (Basu’s theorem) and satisfies the conditions in Lemma 6.7(i); the numerator and the denominator of t(X) (after division by σ) are independently distributed as N (0, 1) and

415

6.2. UMP Unbiased Tests

the chi-square distribution χ2n1 +n2 −2 , respectively. Hence t(X) has the tdistribution tn1 +n2 −2 and a UMPU test rejects H0 when t(X) > tn1 +n2 −2,α , where tn1 +n2 −2,α is the (1 − α)th quantile of the t-distribution tn1 +n2 −2 . This is the so-called (one-sided) two-sample t-test. For testing H0 : θ = 0 (i.e., µ1 = µ2 ) versus H1 : θ 6= 0, it follows from a similar argument used in the derivation of the (two-sided) one-sample t-test that a UMPU test rejects H0 when |t(X)| > tn1 +n2 −2,α/2 (exercise). This is the (two-sided) two-sample t-test. The power function of a two-sample t-test is related to a noncentral t-distribution. Normal linear models Consider linear model (3.25) with assumption A1, i.e., X = (X1 , ..., Xn ) is Nn (Zβ, σ 2 In ),

(6.38)

where β is a p-vector of unknown parameters, Z is the n × p matrix whose ith row is the vector Zi , Zi ’s are the values of a p-vector of deterministic covariates, and σ 2 > 0 is an unknown parameter. Assume that n > p and the rank of Z is r ≤ p. Let l ∈ R(Z) (the linear space generated by the rows of Z) and θ0 be a fixed constant. We consider the hypotheses H0 : l τ β ≤ θ 0

versus

H1 : lτ β > θ0

(6.39)

H0 : l τ β = θ 0

versus

H1 : lτ β 6= θ0 .

(6.40)

HΓ = ( Γ1 0 ),

(6.41)

or τ

−

τ

Since H = Z(Z Z) Z is a projection matrix of rank r, there exists an n × n orthogonal matrix Γ such that Γ = ( Γ 1 Γ2 )

and

where Γ1 is n×r and Γ2 is n×(n−r). Let Yj = Γτj X, j = 1, 2. Consider the transformation (Y1 , Y2 ) = Γτ X. Since Γτ Γ = In and X is Nn (Zβ, σ 2 In ), (Y1 , Y2 ) is Nn (Γτ Zβ, σ2 In ). It follows from (6.41) that E(Y2 ) = E(Γτ2 X) = Γτ2 Zβ = Γτ2 HZβ = 0. Let η = Γτ1 Zβ = E(Y1 ). Then the p.d.f. of (Y1 , Y2 ) is τ 1 η Y1 kY1 k2 + kY2 k2 kηk2 . exp − − σ2 2σ 2 2σ 2 (2πσ 2 )n/2

(6.42)

Since l in (6.39) or (6.40) is in R(Z), there exists λ ∈ Rn such that l = Z τ λ. Then lτ βˆ = λτ HX = λτ ΓΓτ HX = λτ Γ1 Γτ1 X = λτ Γ1 Y1 , (6.43)

416

6. Hypothesis Tests

where βˆ is the LSE defined by (3.27). By (6.43) and Theorem 3.6(ii), ˆ = lτ β = λτ Γ1 E(Y1 ) = aτ η, E(lτ β) where a = Γτ1 λ. Let η = (η1 , ..., ηr ) and a = (a1 , ..., ar ). Without loss of generality, we assume that a1 6= 0. Then the p.d.f. in (6.42) is of the form (6.23) with 1 η2 ηr a τ η − θ0 , Y = Y11 , , ϕ = − , , ..., θ= a1 σ 2 2σ 2 σ 2 σ2 2θ0 Y11 a2 Y11 ar Y11 , U = kY1 k2 + kY2 k2 − , Y12 − , ..., Y1r − a1 a1 a1

where Y1j is the jth component of Y1 . By Basu’s theorem, √ n − r(aτ Y1 − θ0 ) t(X) = kY2 k kak

is independent of U when aτ η = lτ β = θ0 . Note that kY2 k2 = SSR in (3.35) and kak2 = λτ Γ1 Γτ1 λ = λτ Hλ = lτ (Z τ Z)− l. Hence, by (6.43), lτ βˆ − θ0 p t(X) = p , lτ (Z τ Z)− l SSR/(n − r)

which has the t-distribution tn−r (Theorem 3.8). Using the same arguments in deriving the one-sample or two-sample t-test, we obtain that a UMPU test for the hypotheses in (6.39) rejects H0 when t(X) > tn−r,α , and that a UMPU test for the hypotheses in (6.40) rejects H0 when |t(X)| > tn−r,α/2 . Testing for independence in the bivariate normal family Suppose that X1 , ..., Xn are i.i.d. from a bivariate normal distribution, i.e., the p.d.f. of X = (X1 , ..., Xn ) is o n kY1 −µ1 k2 ρ(Y1 −µ1 )τ (Y2 −µ2 ) kY2 −µ2 k2 1 √ , (6.44) exp − + − 2 2 2 2 2 σ1 σ2 (1−ρ ) 2σ (1−ρ ) 2σ (1−ρ ) 2 n (2πσ1 σ2

1−ρ )

1

2

where Yj = (X1j , ..., Xnj ) and Xij is the jth component of Xi , j = 1, 2. Testing for independence of the two components of X1 (or Y1 and Y2 ) is equivalent to testing H0 : ρ = 0 versus H1 : ρ 6= 0. In some cases, one may also be interested in the one-sided hypotheses H0 : ρ ≤ 0 versus H1 : ρ > 0. It can be shown (exercise) that the p.d.f. in (6.44) is of the form (6.23) with ρ θ = σ1 σ2 (1−ρ 2 ) and ! n n n n n X X X X X 2 2 Xi1 Xi2 , U= Xi1 , Xi2 , Xi1 , Xi2 . Y = i=1

i=1

i=1

i=1

i=1

417

6.3. UMP Invariant Tests

The hypothesis ρ ≤ 0 is equivalent to θ ≤ 0. The sample correlation coefficient is #1/2 "X n n n X X 2 2 ¯ 1 )(Xi2 − X ¯2) ¯1) ¯2 ) R= (Xi1 − X (Xi1 − X (Xi2 − X , i=1

i=1

i=1

¯ j is the sample mean of X1j , ..., Xnj , and is independent of U when where X ρ = 0 (Basu’s theorem), j = 1, 2. To apply Lemma 6.7, we consider p √ V = n − 2R/ 1 − R2 . (6.45) It can be shown (exercise) that R is linear in Y and that V has the tdistribution tn−2 when ρ = 0. Hence, a UMPU test for H0 : ρ ≤ 0 versus H1 : ρ > 0 rejects H0 when V > tn−2,α and a UMPU test for H0 : ρ = 0 versus H1 : ρ 6= 0 rejects H0 when |V | > tn−2,α/2 , where tn−2,α is the (1 − α)th quantile of the t-distribution tn−2 .

6.3 UMP Invariant Tests In the previous section the unbiasedness principle is considered to derive an optimal test within the class of unbiased tests when a UMP test does not exist. In this section, we study the same problem with unbiasedness replaced by invariance under a given group of transformations. The principles of unbiasedness and invariance often complement each other in that each is successful in cases where the other is not.

6.3.1 Invariance and UMPI tests The invariance principle considered here is similar to that introduced in §2.3.2 (Definition 2.9) and in §4.2. Although a hypothesis testing problem can be treated as a particular statistical decision problem (see, e.g., Example 2.20), in the following definition we define invariant tests without using any loss function which is a basic element in statistical decision theory. However, the reader is encouraged to compare Definition 2.9 with the following definition. Definition 6.5. Let X be a sample from P ∈ P and G be a group (Definition 2.9(i)) of one-to-one transformations of X. (i) We say that the problem of testing H0 : P ∈ P0 versus H1 : P ∈ P1 is invariant under G if and only if both P0 and P1 are invariant under G in the sense of Definition 2.9(ii). (ii) In an invariant testing problem, a test T (X) is said to be invariant

418

6. Hypothesis Tests

under G if and only if T (g(x)) = T (x)

for all x and g.

(6.46)

(iii) A test of size α is said to be a uniformly most powerful invariant (UMPI) test if and only if it is UMP within the class of level α tests that are invariant under G. (iv) A statistic M (X) is said to be maximal invariant under G if and only if (6.46) holds with T replaced by M and M (x1 ) = M (x2 )

implies x1 = g(x2 ) for some g ∈ G.

(6.47)

The following result indicates that invariance reduces the data X to a maximal invariant statistic M (X) whose distribution may depend only on a functional of P that shrinks P. Proposition 6.2. Let M (X) be maximal invariant under G. (i) A test T (X) is invariant under G if and only if there is a function h such that T (x) = h(M (x)) for all x. (ii) Suppose that there is a functional θ(P ) on P satisfying θ(¯ g (P )) = θ(P ) for all g ∈ G and P ∈ P and θ(P1 ) = θ(P2 )

implies P1 = g¯(P2 ) for some g ∈ G

(i.e., θ(P ) is “maximal invariant”), where g¯(PX ) = Pg(X) is given in Definition 2.9(ii). Then the distribution of M (X) depends only on θ(P ). Proof. (i) If T (x) = h(M (x)) for all x, then T (g(x)) = h(M (g(x))) = h(M (x)) = T (x) so that T is invariant. If T is invariant and if M (x1 ) = M (x2 ), then x1 = g(x2 ) for some g and T (x1 ) = T (g(x2 )) = T (x2 ). Hence T is a function of M . (ii) Suppose that θ(P1 ) = θ(P2 ). Then P2 = g¯(P1 ) for some g ∈ G and for any event B in the range of M (X), P2 M (X) ∈ B = g¯(P1 ) M (X) ∈ B = P1 M (g(X)) ∈ B = P1 M (X) ∈ B . Hence the distribution of M (X) depends only on θ(P ).

In applications, maximal invariants M (X) and θ = θ(P ) are frequently real-valued. If the hypotheses of interest can be expressed in terms of θ, then there may exist a test UMP among those depending only on M (X) (e.g., when the distribution of M (X) is in a parametric family having monotone likelihood ratio). Such a test is then a UMPI test.

419

6.3. UMP Invariant Tests

Example 6.13 (Location-scale families). Suppose that X has the Lebesgue p.d.f. fi,µ (x) = fi (x1 − µ, ..., xn − µ), where n ≥ 2, µ ∈ R is unknown, and fi , i = 0, 1, are known Lebesgue p.d.f.’s. We consider the problem of testing H0 : X is from f0,µ

versus

H1 : X is from f1,µ .

(6.48)

Consider G = {gc : c ∈ R} with gc (x) = (x1 + c, ..., xn + c). For any gc ∈ G, it induces a transformation g¯c (fi,µ ) = fi,µ+c and the problem of testing H0 versus H1 in (6.48) is invariant under G. We now show that a maximal invariant under G is D(X) = (D1 , ..., Dn−1 ) = (X1 − Xn , ..., Xn−1 − Xn ). First, it is easy to see that D(X) is invariant under G. Let x = (x1 , ..., xn ) and y = (y1 , ..., yn ) be two points in the range of X. Suppose that xi − xn = yi − yn for i = 1, ..., n − 1. Putting c = yn − xn , we have yi = xi + c for all i. Hence, D(X) is maximal invariant under G. R By Proposition 1.8, D has the p.d.f. fi (d1 + t, ..., dn−1 + t, t)dt under Hi , i = 0, 1, which does not depend on µ. In fact, in this case Proposition 6.2 applies with M (X) = D(X) and θ(fi,µ ) = i. If we consider tests that are functions of D(X), then the problem of testing the hypotheses in (6.48) becomes one of testing a simple hypothesis versus a simple hypothesis. By Theorem 6.1, the test UMP among functions of D(X), which is then the UMPI test, rejects H0 in (6.48) when R R f1 (x1 + t, ..., xn + t)dt f (d + t, ..., dn−1 + t, t)dt R 1 1 R = > c, f0 (d1 + t, ..., dn−1 + t, t)dt f0 (x1 + t, ..., xn + t)dt where c is determined by the size of the UMPI test. The previous result can be extended to the case of a location-scale family where the p.d.f. of X is one of fi,µ,σ = σ1n fi x1σ−µ , ..., xnσ−µ , i = 0, 1, fi,µ,σ is symmetric about µ, the hypotheses of interest are given by (6.48) with fi,µ replaced by fi,µ,σ , and G = {gc,r : c ∈ R, r 6= 0} with gc,r (x) = (rx1 +c, ..., rxn +c). When n ≥ 3, it can be shown that a maximal invariant under G is W (X) = (W1 , ..., Wn−2 ), where Wi = (Xi − Xn )/(Xn−1 − Xn ), and that the p.d.f. of W does not depend on (µ, σ). A UMPI test can then be derived (exercise). The next example considers finding a maximal invariant in a problem that is not a location-scale family problem. Example 6.14. Let G be the set of n! permutations of the components of x ∈ Rn . Then a maximal invariant is the vector of order statistics. This is because a permutation of the components of x does not change the values of these components and two x’s with the same set of ordered components can be obtained from each other through a permutation of coordinates.

420

6. Hypothesis Tests

Suppose that P contains continuous c.d.f.’s on Rn . Let G be the class of all transformations of the form g(x) = (ψ(x1 ), ..., ψ(xn )), where ψ is continuous and strictly increasing. For x = (x1 , ..., xn ), let R(x) = (R1 , ..., Rn ) be the vector of ranks (§5.2.2), i.e., xi = x(Ri ) , where x(j) is the jth smallest value of xi ’s. Clearly, R(g(x)) = R(x) for any g ∈ G. For any x and y in Rn with R(x) = R(y), define ψ(t) to be linear between x(j) and x(j+1) , j = 1, ..., n−1, ψ(t) = t+(y(1) −x(1) ) for t ≤ x(1) , and ψ(t) = t+(y(n) −x(n) ) for t ≥ x(n) . Then ψ(xi ) = ψ(yi ), i = 1, ..., n. This shows that the vector of rank statistics is maximal invariant. When there is a sufficient statistic U (X), it is convenient first to reduce the data to U (X) before applying invariance. If there is a test T (U ) UMP among all invariant tests depending only on U , one would like to conclude that T (U ) is a UMPI test. Unfortunately, this may not be true in general, since it is not clear that for any invariant test based on X there is an equivalent invariant test based only on U (X). The following result provides a sufficient condition under which it is enough to consider invariant tests depending only on U (X). Its proof is omitted and can be found in Lehmann (1986, pp. 297-302). Proposition 6.3. Let G be a group of transformations on X (the range of X) and (G, BG , λ) be a measure space with a σ-finite λ. Suppose that the testing problem under consideration is invariant under G, that for any set A ∈ BX , the set of points (x, g) for which g(x) ∈ A is in σ(BX × BG ), and that λ(B) = 0 implies λ({h ◦ g : h ∈ B}) = 0 for all g ∈ G. Suppose further that there is a statistic U (X) sufficient for P ∈ P and that U (x1 ) = U (x2 ) implies U (g(x1 )) = U (g(x2 )) for all g ∈ G so that G induces a group GU of transformations on the range of U through gU (U (x)) = U (g(x)). Then, for any test T (X) invariant under G, there exists a test based on U (X) that is invariant under G (and GU ) and has the same power function as T (X). In many problems g(x) = ψ(x, g), where g ranges over a set G in Rm and ψ is a Borel function on Rn+m . Then the measurability condition in Proposition 6.3 is satisfied by choosing BG to be the Borel σ-field on G. In such cases it is usually not difficult to find a measure λ satisfying the condition in Proposition 6.3. Example 6.15. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ ∈ R and σ 2 > 0. The problem of testing H0 : σ 2 ≥ σ02 versus H1 : σ 2 < σ02 is invariant under G = {gc : c ∈ R} with gc (x) = (x1 + c, ..., xn + c). It ¯ S2) can be shown (exercise) that G and the sufficient statistic U = (X, satisfy the conditions in Proposition 6.3 with GU = {hc : c ∈ R} and hc (u1 , u2 ) = (u1 + c, u2 ), and that S 2 is maximal invariant under GU . It follows from Proposition 6.3, Corollary 6.1, and the fact that (n − 1)S 2 /σ02

6.3. UMP Invariant Tests

421

has the chi-square distribution χ2n−1 when σ 2 = σ02 that a UMPI test of size α rejects H0 when (n − 1)S 2 /σ02 ≤ χ2n−1,1−α , where χ2n−1,α is the (1 − α)th quantile of the chi-square distribution χ2n−1 . This test coincides with the UMPU test given in §6.2.3. Example 6.16. Let Xi1 , ..., Xini , i = 1, 2, be two independent samples i.i.d. from N (µi , σi2 ), i = 1, 2, respectively. The problem of testing H0 : σ22 /σ12 ≤ ∆0 versus H1 : σ22 /σ12 > ∆0 is invariant under G = {gc1 ,c2 ,r : ci ∈ R, i = 1, 2, r > 0} with gc1 ,c2 ,r (x1 , x2 ) = (rx11 + c1 , ..., rx1n1 + c1 , rx21 + c2 , ..., rx2n2 + c2 ). ¯1, X ¯ 2 , S12 , S22 ) It can be shown (exercise) that the sufficient statistic U = (X and G satisfy the conditions in Proposition 6.3 with GU = {hc1 ,c2 ,r : ci ∈ R, i = 1, 2, r > 0} and hc1 ,c2 ,r (u1 , u2 , u3 , u4 ) = (ru1 + c1 , ru2 + c2 , ru3 , ru4 ). A maximal invariant under GU is S2 /S1 . Let ∆ = σ22 /σ12 . Then (S22 /S12 )/∆ has an F-distribution and, therefore, V = S22 /S12 has a Lebesgue p.d.f. of the form f∆ (v) = C(∆)v (n2 −3)/2 [∆ + (n2 − 1)v/(n1 − 1)]−(n1 +n2 −2)/2 I(0,∞) (v), where C(∆) is a known function of ∆. It can be shown (exercise) that the family {f∆ : ∆ > 0} has monotone likelihood ratio in V so that a UMPI test of size α rejects H0 when V > Fn2 −1,n1 −1,α , where Fa,b,α is the (1 − α)th quantile of the F-distribution Fa,b . Again, this UMPI test coincides with the UMPU test given in §6.2.3. The following result shows that, in Examples 6.15 and 6.16, the fact that UMPI tests are the same as the UMPU tests is not a simple coincidence. Proposition 6.4. Consider a testing problem invariant under G. If there exists a UMPI test of size α, then it is unbiased. If there also exists a UMPU test of size α that is invariant under G, then the two tests have the same power function on P ∈ P1 . If either the UMPI test or the UMPU test is unique a.s. P, then the two tests are equal a.s. P. Proof. We only need to prove that a UMPI test of size α is unbiased. This follows from the fact that the test T ≡ α is invariant under G.

422

6. Hypothesis Tests

The next example shows an application of invariance in a situation where a UMPU test may not exist. Example 6.17. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ and σ 2 . Let θ = (µ − u)/σ, where u is a known constant. Consider the problem of testing H0 : θ ≤ θ0 versus H1 : θ > θ0 . Note that H0 is the same as P (X1 ≤ u) ≥ p0 for a known constant p0 = Φ(−θ0 ). Without loss of generality, we consider the case of u = 0. The problem is invariant under G = {gr : r > 0} with gr (x) = rx. By Proposition 6.3, we can consider tests that are functions of the sufficient ¯ S 2 ) only. A maximal invariant under G is t(X) = √nX/S. ¯ statistic (X, To find a UMPI test, it remains to find a test UMP among all tests that are functions of t(X). From √ the discussion in §1.3.1, t(X) has the noncentral t-distribution tn−1 ( nθ). Let fθ (t) be the Lebesgue p.d.f. √ of t(X), i.e., fθ is given by (1.32) with n replaced by n − 1 and δ = nθ. It can be shown (exercise) that the family of p.d.f.’s, {fθ (t) : θ ∈ R}, has monotone likelihood ratio in t. Hence, by Theorem 6.2, a UMPI test of√size α rejects H0 when t(X) > c, where c is the (1 − α)th quantile of tn−1 ( nθ0 ). In some problems, we may have to apply both unbiasedness and invariance principles. For instance, suppose that in the current problem we would like to test H0 : θ = θ0 versus H1 : θ 6= θ0 . The problem is still invariant under G. Following the previous discussion, we only need to consider tests that are functions of t(X). But a test UMP among functions of t(X) does not exist in this case. A test UMP among all unbiased tests of level α that are functions of t(X) rejects H0 when t(X) < c1 or t(X) > c2 , where c1 and c2 are determined by Z c 2 Z c2 d fθ0 (t)dt = 1 − α and fθ (t)dt =0 dθ c1 c1 θ=θ0

(see Exercise 26). This test is then UMP among all tests that are invariant and unbiased of level α. Whether it is also UMPU without the restriction to invariant tests is an open problem.

6.3.2 UMPI tests in normal linear models Consider normal linear model (6.38): X = Nn (Zβ, σ 2 In ), where β is a p-vector of unknown parameters, σ 2 > 0 is unknown, and Z is a fixed n × p matrix of rank r ≤ p < n. In §6.2.3, UMPU tests for testing

423

6.3. UMP Invariant Tests

(6.39) or (6.40) are derived. A frequently encountered problem in practice is to test H0 : Lβ = 0 versus H1 : Lβ 6= 0, (6.49) where L is an s × p matrix of rank s ≤ r and all rows of L are in R(Z). However, a UMPU test for (6.49) does not exist if s > 1. We now derive a UMPI test for testing (6.49). We use without proof the following result from linear algebra: there exists an orthogonal matrix Γ such that (6.49) is equivalent to H0 : η1 = 0

versus

H1 : η1 6= 0,

(6.50)

where η1 is the s-vector containing the first s components of η, η is the r-vector containing the first r components of ΓZβ, and the last n − r components of ΓZβ are 0’s. Let Y = ΓX. Then Y = Nn ((η, 0), σ 2 In ) with the p.d.f. given by (6.42). Let Y = (Y1 , Y2 ), where Y1 is an r-vector, and let Y1 = (Y11 , Y12 ), where Y11 is an s-vector. Define G = {gΛ,c,γ : c ∈ Rr−s , γ > 0, Λ is an s × s orthogonal matrix} with gΛ,c,γ (Y ) = γ(ΛY11 , Y12 + c, Y2 ). Testing (6.50) is invariant under G. By Proposition 6.3, we can restrict our attention to the sufficient statistic U = (Y1 , kY2 k2 ). The statistic M (U ) = kY11 k2 /kY2 k2

(6.51)

is invariant under GU , the group of transformations on the range of U defined by g˜Λ,c,γ (U (Y )) = U (gΛ,c,γ (Y )). We now show that M (U ) is maximal invariant under GU . Let li ∈ Rs , li 6= 0, and ti ∈ (0, ∞), i = 1, 2. If kl1 k2 /t21 = kl2 k2 /t22 , then t1 = γt2 with γ = kl1 k/kl2 k. Since l1 /kl1 k and l2 /kl2 k are two points having the same distance from the origin, there exists an orthogonal matrix Λ such that l1 /kl1 k = Λl2 /kl2 k, i.e., l1 = γΛl2 . (j) (j) This proves that if M (u(1) ) = M (u(2) ) with u(j) = (y11 , y12 , t2j ), then (1)

(2)

y11 = γΛy11 and t1 = γt2 for some γ > 0 and orthogonal matrix Λ and, (1) (2) therefore, u(1) = g˜Λ,c,γ (u(2) ) with c = γ −1 y12 − y12 . Thus, M (U ) is maximal invariant under GU . It can be shown (exercise) that W = M (U )(n − r)/s has the noncentral F-distribution Fs,n−r (θ) with θ = kη1 k2 /σ 2 (see §1.3.1). Let fθ (w) be the Lebesgue p.d.f. of W , i.e., fθ is given by (1.33) with n1 = s, n2 = n − r, and δ = θ. Note that under H0 , θ = 0 and fθ reduces to the p.d.f. of the central F-distribution Fs,n−r (Table 1.2, page 20). Also, it can be shown (exercise) that the ratio fθ1 (w)/f0 (w) is an increasing function of w for any given θ1 > 0. By Theorem 6.1, a UMPI test of size α for testing H0 : θ = 0

424

6. Hypothesis Tests

versus H1 : θ = θ1 rejects H0 when W > Fs,n−r,α , where Fs,n−r,α is the (1 − α)th quantile of the F-distribution Fs,n−r . Since this test does not depend on θ1 , by Lemma 6.1, it is also a UMPI test of size α for testing H0 : θ = 0 versus H1 : θ > 0, which is equivalent to testing (6.50). In applications it is not convenient to carry out the test by finding explicitly the orthogonal matrix Γ. Hence, we now express the statistic W in terms of X. Since Y = ΓX and E(Y ) = ΓE(X) = ΓZβ, kY1 − ηk2 + kY2 k2 = kX − Zβk2 and, therefore, min kY1 − ηk2 + kY2 k2 = min kX − Zβk2 , η

β

which is the same as ˆ 2 = SSR, kY2 k2 = kX − Z βk where βˆ is the LSE defined by (3.27). Similarly, kY11 k2 + kY2 k2 = min kX − Zβk2 . β:Lβ=0

If we define βˆH0 to be a solution of kX − Z βˆH0 k2 = min kX − Zβk2 , β:Lβ=0

which is called the LSE of β under H0 or the LSE of β subject to Lβ = 0, then ˆ 2 )/s (kX − Z βˆH0 k2 − kX − Z βk W = . (6.52) ˆ 2 /(n − r) kX − Z βk Thus, the UMPI test for (6.49) can be used without finding Γ. When s = 1, the UMPI test derived here is the same as the UMPU test for (6.40) given in §6.2.3. Example 6.18. Consider the one-way ANOVA model in Example 3.13: Xij = N (µi , σ 2 ),

j = 1, ..., ni , i = 1, ..., m,

and Xij ’s are independent. A common testing problem in applications is the test for homogeneity of means, i.e., H0 : µ1 = · · · = µm

versus

H1 : µi 6= µk for some i 6= k.

(6.53)

One can easily find a matrix L for which (6.53) is equivalent to (6.49). But it is not necessary to find such a matrix in order to compute the

425

6.3. UMP Invariant Tests

statistic W that defines the UMPI test. Note that the LSE of (µ1 , ..., µm ) ¯ 1· , ..., X ¯ m· ), where X ¯ i· is the sample mean based on Xi1 , ..., Xini , and is (X ¯ the sample mean based on all Xij ’s. Thus, the LSE under H0 is simply X, ˆ 2= SSR = kX − Z βk

ni m X X i=1 j=1

SST = kX − Z βˆH0 k2 = and SSA = SST − SSR = Then W =

¯ i· )2 , (Xij − X

ni m X X i=1 j=1

m X i=1

¯ 2, (Xij − X)

¯ i· − X) ¯ 2. ni (X

SSA/(m − 1) , SSR/(n − m)

Pm where n = i=1 ni . The name ANOVA comes from the fact that the UMPI test is carried out by comparing two sources of variation: the variation within each group of observations (measured by SSR) and the variation among m groups (measured by SSA), and that SSA + SSR = SST is the total variation in the data set. In this case, the distribution of W can also be derived using Cochran’s theorem (Theorem 1.5). See Exercise 75. Example 6.19. Consider the two-way balanced ANOVA model in Example 3.14: Xijk = N (µij , σ 2 ),

i = 1, ..., a, j = 1, ..., b, k = 1, ..., c,

Pa Pb Pa Pb where µij = µ+αi +βj +γij , i=1 αi = j=1 βj = i=1 γij = j=1 γij = 0, and Xijk ’s are independent. Typically the following hypotheses are of interest: H0 : αi = 0 for all i

versus

H1 : αi 6= 0 for some i,

(6.54)

H0 : βj = 0 for all j

versus

H1 : βj 6= 0 for some j,

(6.55)

H0 : γij = 0 for all i, j

versus

H1 : γij 6= 0 for some i, j.

(6.56)

and

In applications, αi ’s are effects of a factor A (a variable taking finitely many values), βj ’s are effects of a factor B, and γij ’s are effects of the interaction of factors A and B. Hence, testing hypotheses in (6.54), (6.55), and (6.56)

426

6. Hypothesis Tests

are the same as testing effects of factor A, of factor B, and of the interaction between A and B, respectively. ¯ ··· , The LSE’s of µ, αi , βj , and γij are given by (Example 3.14) µ ˆ=X ˆ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ α ˆ i = Xi·· − X··· , βj = X·j· − X··· , γˆij = Xij· − Xi·· − X·j· + X··· , and a dot is used to denote averaging over the indicated subscript. Let SSR =

b X c a X X

¯ ij· )2 , (Xijk − X

i=1 j=1 k=1

SSA = bc

a X i=1

SSB = ac

b X j=1

and SSC = c

b a X X i=1 j=1

¯ i·· − X ¯ ··· )2 , (X ¯ ·j· − X ¯ ··· )2 , (X

¯ ij· − X ¯ i·· − X ¯ ·j· + X ¯ ··· )2 . (X

Then, one can show (exercise) that for testing (6.54), (6.55), and (6.56), the statistics W in (6.52) (for the UMPI tests) are, respectively, SSA/(a − 1) , SSR/[(c − 1)ab]

SSB/(b − 1) , SSR/[(c − 1)ab]

and

SSC/[(a − 1)(b − 1)] . SSR/[(c − 1)ab]

We end this section with a discussion of testing for random effects in the following balanced one-way random effects model (Example 3.17): Xij = µ + Ai + eij ,

i = 1, ..., a, j = 1, ..., b,

(6.57)

where µ is an unknown parameter, Ai ’s are i.i.d. random effects from N (0, σa2 ), eij ’s are i.i.d. measurement errors from N (0, σ 2 ), and Ai ’s and eij ’s are independent. Consider the problem of testing H0 : σa2 /σ 2 ≤ ∆0

versus

H1 : σa2 /σ 2 > ∆0

(6.58)

for a given ∆0 . When ∆0 is small, hypothesis H0 in (6.58) means that the random effects are negligible relative to the measurement variation. Let (Yi1 , ..., Yib ) = Γ(Xi1 , ..., Xib ), where Γ is a√b × b orthogonal matrix whose elements in the first row are all equal to 1/ b. Then √ √ ¯ i· = b(µ + Ai + e¯i· ), i = 1, ..., a, Yi1 = bX

427

6.3. UMP Invariant Tests

√ are i.i.d. from N ( bµ, σ 2 + bσa2 ), Yij , i = 1, ..., a, j = 2, ..., b, are i.i.d. from N (0, σ 2 ), and Yij ’s are independent. The reason why E(Yij ) = 0 when j > 1 is because row j of Γ is orthogonal to the first row of Γ. Let Λ be an√a × a orthogonal matrix whose elements in the first√row are all √ equal to 1/ a and (U11 , ..., Ua1 ) = Λ(Y11 , ..., Ya1 ). Then U11 = aY¯·1 is N ( abµ, σ 2 + bσa2 ), Ui1 , i = 2, ..., a, are from N (0, σ 2 + bσa2 ), and Ui1 ’s are independent. Let Uij = Yij for j = 2, ..., b, i = 1, ..., a. The problem of testing (6.58) is invariant under the group of transformations that transform U11 to rU11 +c and Uij to rUij , (i, j) 6= (1, 1), where r > 0 and c ∈ R. It can be shown (exercise) that the maximal invariant under this group of transformations is SSA/SSR, where SSA =

a X

2 Ui1

and

i=2

SSR =

b a X X

Uij2 .

i=1 j=2

Note that H0 in (6.58) is equivalent to (σ 2 + bσa2 )/σ 2 ≤ 1 + b∆0 . Also, SSA/(σ 2 + bσa2 ) has the chi-square distribution χ2a−1 and SSR/σ 2 has the chi-square distribution χ2a(b−1) . Hence, the p.d.f. of the statistic W =

SSA/(a − 1) 1 1 + b∆0 SSR/[a(b − 1)]

is in a parametric family (indexed by the parameter (σ 2 + bσa2 )/σ 2 ) with monotone likelihood ratio in W . Thus, a UMPI test of size α for testing (6.58) rejects H0 when W > Fa−1,a(b−1),α , where Fa−1,a(b−1),α is the (1 − α)th quantile of the F-distribution Fa−1,a(b−1) . It remains to express W in terms of Xij ’s. Note that b b a X a b a X X X X X ¯ i· )2 SSR = Yij2 = e2ij − b¯ e2i· = (Xij − X i=1 j=2

i=1

j=1

i=1 j=1

and

SSA =

a X i=1

2 2 Ui1 − U11 =

a X i=1

Yi12 − aY¯·12 = b

a X i=1

¯ i· − X ¯ ·· )2 . (X

The SSR and SSA derived here are the same as those in Example 6.18 when ni = b for all i and m = a. It can also be seen that if ∆0 = 0, then testing (6.58) is equivalent to testing H0 : σa2 = 0 versus H1 : σa2 > 0 and the derived UMPI test is exactly the same as that in Example 6.18, although the testing problems are different in these two cases. Extensions to balanced two-way random effects models can be found in Lehmann (1986, §7.12).

428

6. Hypothesis Tests

6.4 Tests in Parametric Models A UMP, UMPU, or UMPI test often does not exist in a particular problem. In the rest of this chapter, we study some methods for constructing tests that have intuitive appeal and frequently coincide with optimal tests (UMP or UMPU tests) when optimal tests do exist. We consider tests in parametric models in this section, whereas tests in nonparametric models are studied in §6.5. When the hypothesis H0 is not simple, it is often difficult or even impossible to obtain a test that has exactly a given size α, since it is hard to find a population P that maximizes the power function of the test over all P ∈ P0 . In such cases a common approach is to find tests having asymptotic significance level α (Definition 2.13). This involves finding the limit of the power of a test at P ∈ P0 , which is studied in this section and §6.5. Throughout this section, we assume that a sample X is from P ∈ P = θ {Pθ : θ ∈ Θ}, Θ ⊂ Rk , fθ = dP dν exists w.r.t. a σ-finite measure ν for all θ, and the testing problem is H0 : θ ∈ Θ0

versus

H1 : θ ∈ Θ 1 ,

(6.59)

where Θ0 ∪ Θ1 = Θ and Θ0 ∩ Θ1 = ∅.

6.4.1 Likelihood ratio tests When both H0 and H1 are simple (i.e., both Θ0 = {θ0 } and Θ1 = {θ1 } are single-point sets), Theorem 6.1 applies and a UMP test rejects H0 when fθ1 (X) > c0 fθ0 (X)

(6.60)

for some c0 > 0. When c0 ≥ 1, (6.60) is equivalent to (exercise) fθ0 (X) θ0 , there is an LR test whose rejection region is the same as that of the UMP test T∗ given by (6.11). (ii) For testing the hypotheses in (6.12), there is an LR test whose rejection region is the same as that of the UMP test T∗ given by (6.15). (iii) For testing the hypotheses in (6.13) or (6.14), there is an LR test whose rejection region is equivalent to Y (X) < c1 or Y (X) > c2 for some constants c1 and c2 . Proof. (i) Let θˆ be the MLE of θ. Note that ℓ(θ) is increasing when θ ≤ θˆ ˆ Thus, and decreasing when θ > θ. ( 1 θˆ ≤ θ0 λ(X) = ℓ(θ0 ) θˆ > θ0 . ˆ ℓ(θ) ˆ < c. From the Then λ(X) < c is the same as θˆ > θ0 and ℓ(θ0 )/ℓ(θ) discussion in §4.4.2, θˆ is a strictly increasing function of Y . It can be ˆ − log ℓ(θ0 ) is strictly increasing in Y when θˆ > θ0 and shown that log ℓ(θ) strictly decreasing in Y when θˆ < θ0 (exercise). Hence, for any d ∈ R, ˆ < c is equivalent to Y > d for some c ∈ (0, 1). θˆ > θ0 and ℓ(θ0 )/ℓ(θ) (ii) The proof is similar to that in (i). Note that ( 1 θˆ < θ1 or θˆ > θ2 λ(X) = max{ℓ(θ1 ),ℓ(θ2 )} θ1 ≤ θˆ ≤ θ2 . ˆ ℓ(θ)

430

6. Hypothesis Tests

Hence λ(X) < c is equivalent to c1 < Y < c2 . (iii) The proof for (iii) is left as an exercise. Proposition 6.5 can be applied to problems concerning one-parameter exponential families such as the binomial, Poisson, negative binomial, and normal (with one parameter known) families. The following example shows that the same result holds in a situation where Proposition 6.5 is not applicable. Example 6.20. Consider the testing problem H0 : θ = θ0 versus H1 : θ 6= θ0 based on i.i.d. X1 , ..., Xn from the uniform distribution U (0, θ). We now show that the UMP test with rejection region X(n) > θ0 or X(n) ≤ θ0 α1/n given in Exercise 19(c) is an LR test. Note that ℓ(θ) = θ−n I(X(n) ,∞) (θ). Hence (X(n) /θ0 )n X(n) ≤ θ0 λ(X) = 0 X(n) > θ0 and λ(X) < c is equivalent to X(n) > θ0 or X(n) /θ0 < c1/n . Taking c = α ensures that the LR test has size α.

More examples of this kind can be found in §6.6. The next example considers multivariate θ. Example 6.21. Consider normal linear model (6.38) and the hypotheses in (6.49). The likelihood function in this problem is n/2 1 ℓ(θ) = 2πσ exp − 2σ1 2 kX − Zβk2 , 2

where θ = (β, σ2 ). Let βˆ be the LSE defined by (3.27). Since kX − Zβk2 ≥ ˆ 2 for any β, kX − Z βk o n n/2 1 ˆ 2 . exp − 2σ1 2 kX − Z βk ℓ(θ) ≤ 2πσ 2

Treating the right-hand side of the previous expression as a function of σ 2 , ˆ 2 /n and, it is easy to show that it has a maximum at σ 2 = σ ˆ 2 = kX − Z βk therefore, sup ℓ(θ) = (2πˆ σ 2 )−n/2 e−n/2 . θ∈Θ 2 ˆH = kX − Z βˆH0 k2 /n. Then Similarly, let βˆH0 be the LSE under H0 and σ 0 2 −n/2 −n/2 sup ℓ(θ) = (2πˆ σH ) e . 0

θ∈Θ0

Thus, λ(X) = (ˆ σ

2

2 n/2 /ˆ σH ) 0

=

ˆ 2 kX − Z βk kX − Z βˆH k2 0

!n/2

=

sW +1 n−r

−n/2

,

431

6.4. Tests in Parametric Models

where W is given in (6.52). This shows that LR tests are the same as the UMPI tests derived in §6.3.2. The one-sample or two-sample two-sided t-tests derived in §6.2.3 are special cases of LR tests. For a one-sample problem, we define β = µ and 2 ¯ σ Z = Jn , the n-vector of ones. Note that βˆ = X, ˆ 2 = (n − 1)S 2 /n, βˆH =0 0 2 2 2 2 ¯ . Hence (H0 : β = 0), and σ ˆH0 = kXk /n = (n − 1)S /n + X λ(X) = 1 +

¯2 nX (n − 1)S 2

−n/2 [t(X)]2 1+ , n−1

−n/2

=

Jn1 0

0 Jn2

√ ¯ has the t-distribution tn−1 under H0 . Thus, λ(X) < where t(X) = nX/S c is equivalent to |t(X)| > c0 , which is the rejection region of a one-sample two-sided t-test. For a two-sample problem, we let n = n1 + n2 , β = (µ1 , µ2 ), and Z=

.

Testing H0 : µ1 = µ2 versus H1 : µ1 6= µ2 is the same as testing (6.49) with ¯ and βˆ = (X ¯1, X ¯ 2 ), where X ¯ 1 and X ¯ 2 are L = ( 1 −1 ). Since βˆH0 = X the sample means based on X1 , ..., Xn1 and Xn1 +1 , ..., Xn , respectively, we have nˆ σ2 =

n1 X i=1

¯ 1 )2 + (Xi − X

n X

¯ 2 )2 = (n1 − 1)S12 + (n2 − 1)S22 (Xi − X

i=n1 +1

and 2 ¯1 − X ¯ 2 )2 + (n1 − 1)S 2 + (n2 − 1)S 2 . nˆ σH = (n − 1)S 2 = n−1 n1 n2 (X 1 2 0

Therefore, λ(X) < c is equivalent to |t(X)| > c0 , where t(X) is given by (6.37), and LR tests are the same as the two-sample two-sided t-tests in §6.2.3.

6.4.2 Asymptotic tests based on likelihoods As we can see from Proposition 6.5 and the previous examples, an LR test is often equivalent to a test based on a statistic Y (X) whose distribution under H0 can be used to determine the rejection region of the LR test with size α. When this technique fails, it is difficult or even impossible to find an LR test with size α, even if the c.d.f. of λ(X) is continuous. The following result shows that in the i.i.d. case we can obtain the asymptotic distribution (under H0 ) of the likelihood ratio λ(X) so that an LR test

432

6. Hypothesis Tests

having asymptotic significance level α can be obtained. Assume that Θ0 is determined by H0 : θ = g(ϑ), (6.63) where ϑ is a (k − r)-vector of unknown parameters and g is a continuously differentiable function from Rk−r to Rk with a full rank ∂g(ϑ)/∂ϑ. For example, if Θ = R2 and Θ0 = {(θ1 , θ2 ) ∈ Θ : θ1 = 0}, then ϑ = θ2 , g1 (ϑ) = 0, and g2 (ϑ) = ϑ. Theorem 6.5. Assume the conditions in Theorem 4.16. Suppose that H0 is determined by (6.63). Under H0 , −2 log λn →d χ2r , where λn = λ(X) and χ2r is a random variable having the chi-square distribution χ2r . Con2 sequently, the LR test with rejection region λn < e−χr,α /2 has asymptotic significance level α, where χ2r,α is the (1 − α)th quantile of the chi-square distribution χ2r . Proof. Without loss of generality, we assume that there exist an MLE θˆ and an MLE ϑˆ under H0 such that λn =

ˆ supθ∈Θ0 ℓ(θ) ℓ(g(ϑ)) = . ˆ supθ∈Θ ℓ(θ) ℓ(θ)

Following the proof of Theorem 4.17 in §4.5.2, we can obtain that √ nI1 (θ)(θˆ − θ) = n−1/2 sn (θ) + op (1), where sn (θ) = ∂ log ℓ(θ)/∂θ and I1 (θ) is the Fisher information about θ contained in X1 , and that ˆ − log ℓ(θ)] = n(θˆ − θ)τ I1 (θ)(θˆ − θ) + op (1). 2[log ℓ(θ) Then ˆ − log ℓ(θ)] = n−1 [sn (θ)]τ [I1 (θ)]−1 sn (θ) + op (1). 2[log ℓ(θ) Similarly, under H0 , ˆ − log ℓ(g(ϑ))] = n−1 [˜ 2[log ℓ(g(ϑ)) sn (ϑ)]τ [I˜1 (ϑ)]−1 s˜n (ϑ) + op (1), where s˜n (ϑ) = ∂ log ℓ(g(ϑ))/∂ϑ = D(ϑ)sn (g(ϑ)), D(ϑ) = ∂g(ϑ)/∂ϑ, and I˜1 (ϑ) is the Fisher information about ϑ (under H0 ) contained in X1 . Combining these results, we obtain that ˆ − log ℓ(g(ϑ))] ˆ −2 log λn = 2[log ℓ(θ)

= n−1 [sn (g(ϑ))]τ B(ϑ)sn (g(ϑ)) + op (1)

under H0 , where B(ϑ) = [I1 (g(ϑ))]−1 − [D(ϑ)]τ [I˜1 (ϑ)]−1 D(ϑ).

433

6.4. Tests in Parametric Models

By the CLT, n−1/2 [I1 (θ)]−1/2 sn (θ) →d Z, where Z = Nk (0, Ik ). Then, it follows from Theorem 1.10(iii) that, under H0 , −2 log λn →d Z τ [I1 (g(ϑ))]1/2 B(ϑ)[I1 (g(ϑ))]1/2 Z. Let D = D(ϑ), B = B(ϑ), A = I1 (g(ϑ)), and C = I˜1 (ϑ). Then (A1/2 BA1/2 )2 = A1/2 BABA1/2 = A1/2 (A−1 − Dτ C −1 D)A(A−1 − Dτ C −1 D)A1/2

= (Ik − A1/2 Dτ C −1 DA1/2 )(Ik − A1/2 Dτ C −1 DA1/2 ) = Ik − 2A1/2 Dτ C −1 DA1/2 + A1/2 Dτ C −1 DADτ C −1 DA1/2 = Ik − A1/2 Dτ C −1 DA1/2 = A1/2 BA1/2 ,

where the fourth equality follows from the fact that C = DADτ . This shows that A1/2 BA1/2 is a projection matrix. The rank of A1/2 BA1/2 is tr(A1/2 BA1/2 ) = tr(Ik − Dτ C −1 DA) = k − tr(C −1 DADτ ) = k − tr(C −1 C) = k − (k − r) = r.

Thus, by Exercise 51 in §1.6, Z τ [I1 (g(ϑ))]1/2 B(ϑ)[I1 (g(ϑ))]1/2 Z = χ2r . As an example, Theorem 6.5 can be applied to testing problems in Example 4.33 where the exact rejection region of the LR test of size α is difficult to obtain but the likelihood ratio λn can be calculated numerically. Tests whose rejection regions are constructed using asymptotic theory (so that these tests have asymptotic significance level α) are called asymptotic tests, which are useful when a test of exact size α is difficult to find. There are two popular asymptotic tests based on likelihoods that are asymptotically equivalent to LR tests. Note that the hypothesis in (6.63) is equivalent to a set of r ≤ k equations: H0 : R(θ) = 0,

(6.64)

where R(θ) is a continuously differentiable function from Rk to Rr . Wald (1943) introduced a test that rejects H0 when the value of ˆ τ {[C(θ)] ˆ τ [In (θ)] ˆ −1 C(θ)} ˆ −1 R(θ) ˆ Wn = [R(θ)] is large, where C(θ) = ∂R(θ)/∂θ, In (θ) is the Fisher information matrix based on X1 , ..., Xn , and θˆ is an MLE or RLE of θ. For testing H0 : θ = θ0

434

6. Hypothesis Tests

with a known θ0 , R(θ) = θ − θ0 and Wn simplifies to ˆ θˆ − θ0 ). Wn = (θˆ − θ0 )τ In (θ)( Rao (1947) introduced a score test that rejects H0 when the value of ˜ τ [In (θ)] ˜ −1 sn (θ) ˜ Rn = [sn (θ)] is large, where sn (θ) = ∂ log ℓ(θ)/∂θ is the score function and θ˜ is an MLE or RLE of θ under H0 in (6.64). Theorem 6.6. Assume the conditions in Theorem 4.16. (i) Under H0 given by (6.64), Wn →d χ2r and, therefore, the test rejects H0 if and only if Wn > χ2r,α has asymptotic significance level α, where χ2r,α is the (1 − α)th quantile of the chi-square distribution χ2r . (ii) The result in (i) still holds if Wn is replaced by Rn . Proof. (i) Using Theorems 1.12 and 4.17, √ ˆ − R(θ)] →d Nr 0, [C(θ)]τ [I1 (θ)]−1 C(θ) , n[R(θ)

where I1 (θ) is the Fisher information about θ contained in X1 . Under H0 , R(θ) = 0 and, therefore, ˆ τ {[C(θ)]τ [I1 (θ)]−1 C(θ)}−1 R(θ) ˆ →d χ2 n[R(θ)] r (Theorem 1.10). Then the result follows from Slutsky’s theorem (Theorem 1.11) and the fact that θˆ →p θ and I1 (θ) and C(θ) are continuous at θ. (ii) From the Lagrange multiplier, θ˜ satisfies ˜ + C(θ)λ ˜ n=0 sn (θ)

and

˜ = 0. R(θ)

Using Taylor’s expansion, one can show (exercise) that under H0 ,

and

[C(θ)]τ (θ˜ − θ) = op (n−1/2 )

(6.65)

sn (θ) − In (θ)(θ˜ − θ) + C(θ)λn = op (n1/2 ),

(6.66)

where In (θ) = nI1 (θ). Multiplying [C(θ)]τ [In (θ)]−1 to the left-hand side of (6.66) and using (6.65), we obtain that [C(θ)]τ [In (θ)]−1 C(θ)λn = −[C(θ)]τ [In (θ)]−1 sn (θ) + op (n−1/2 ),

(6.67)

which implies λτn [C(θ)]τ [In (θ)]−1 C(θ)λn →d χ2r

(6.68) ˜ n= (exercise). Then the result follows from (6.68) and the fact that C(θ)λ ˜ −sn (θ), In (θ) = nI1 (θ), and I1 (θ) is continuous at θ.

435

6.4. Tests in Parametric Models

Thus, Wald’s tests, Rao’s score tests, and LR tests are asymptotically ˆ not θ˜ = g(ϑ), ˆ equivalent. Note that Wald’s test requires computing θ, ˜ ˆ whereas Rao’s score test requires computing θ, not θ. On the other hand, an LR test requires computing both θˆ and θ˜ (or solving two maximization problems). Hence, one may choose one of these tests that is easy to compute in a particular application. The results in Theorems 6.5 and 6.6 can be extended to non-i.i.d. situations (e.g., the GLM in §4.4.2). We state without proof the following result. Theorem 6.7. Assume the conditions in Theorem 4.18. Consider the problem of testing H0 in (6.64) (or equivalently, (6.63)) with θ = (β, φ). Then the results in Theorems 6.5 and 6.6 still hold. Example 6.22. Consider the GLM (4.55)-(4.58) with ti ’s in a fixed interval (t0 , t∞ ), 0 < t0 ≤ t∞ < ∞. Then the Fisher information matrix In (θ) =

φ−1 Mn (β) 0

0 I˜n (β, φ)

,

where Mn (β) is given by (4.60) and I˜n (β, φ) is the Fisher information about φ. Consider the problem of testing H0 : β = β0 versus H1 : β 6= β0 , where ˆ φ) ˆ be the MLE (or β0 is a fixed vector. Then R(β, φ) = β − β0 . Let (β, RLE) of (β, φ). Then, Wald’s test is based on ˆ βˆ − β0 ) Wn = φˆ−1 (βˆ − β0 )τ Mn (β)( and Rao’s score test is based on ˜ sn (β0 )]τ [Mn (β0 )]−1 s˜n (β0 ), Rn = φ[˜ where s˜n (β) is given by (4.65) and φ˜ is a solution of ∂ log ℓ(β0 , φ)/∂φ = 0. It follows from Theorem 4.18 that both Wn and Rn are asymptotically distributed as χ2p under H0 . By Slutsky’s theorem, we may replace φˆ or φ˜ by any consistent estimator of φ. Wald’s tests, Rao’s score tests, and LR tests are typically consistent according to Definition 2.13(iii). They are also Chernoff-consistent (Definition 2.13(iv)) if α is chosen to be αn → 0 and χ2r,αn = o(n) as n → ∞ (exercise). Other asymptotic optimality properties of these tests are discussed in Wald (1943); see also Serfling (1980, Chapter 10).

436

6. Hypothesis Tests

6.4.3 χ2 -tests A test that is related to the asymptotic tests described in §6.4.2 is the so-called χ2 -test for testing cell probabilities in a multinomial distribution. Consider a sequence of n independent trials with k possible outcomes for each trial. Let pj > 0 be the cell probability of occurrence of the jth outcome in any given trial and Xj be the number of occurrences of the jth outcome in n trials. Then X = (X1 , ..., Xk ) has the multinomial distribution (Example 2.7) with the parameter p = (p1 , ..., pk ). Let ξi = (0, ..., 0, 1, 0, ..., 0), where the single nonzero component 1 is located in the jth position if thePith trial yields the jth outcome. Then ξ1 , ..., ξn are n i.i.d. and X/n = ξ¯ = i=1 ξi /n. By the CLT, √ √ n(ξ¯ − p) →d Nk (0, Σ), (6.69) Zn (p) = n X n −p =

√ where Σ = Var(X/ n) is a symmetric k × k matrix whose ith diagonal element is pi (1 − pi ) and (i, j)th off-diagonal element is −pi pj . Consider the problem of testing H0 : p = p0

versus

H1 : p 6= p0 ,

(6.70)

where p0 = (p01 , ..., p0k ) is a known vector of cell probabilities. A popular test for (6.70) is based on the following χ2 -statistic: χ2 =

k X (Xj − np0j )2 = kD(p0 )Zn (p0 )k2 , np 0j j=1

(6.71)

where Zn (p) is given by (6.69) and D(c) with c = (c1 , ..., ck ) is the k × k −1/2 diagonal matrix whose jth diagonal element is cj . Another popular test is based on the following modified χ2 -statistic: χ ˜2 =

k X (Xj − np0j )2 = kD(X/n)Zn (p0 )k2 . X j j=1

(6.72)

Note that X/n is an unbiased estimator of p. √ √ Theorem 6.8. Let φ = ( p1 , ..., pk ) and Λ be a k × k projection matrix. (i) If Λφ = aφ, then [Zn (p)]τ D(p)ΛD(p)Zn (p) →d χ2r , where χ2r has the chi-square distribution χ2r with r = tr(Λ) − a. (ii) The same result holds if D(p) in (i) is replaced by D(X/n).

437

6.4. Tests in Parametric Models

Proof. (i) Let D = D(p), Zn = Zn (p), and Z = Nk (0, Ik ). From (6.69) and Theorem 1.10, Znτ DΛDZn →d Z τ AZ

with A = Σ1/2 DΛDΣ1/2 .

From Exercise 51 in §1.6, the result in (i) follows if we can show that A2 = A (i.e., A is a projection matrix) and r = tr(A). Since Λ is a projection matrix and Λφ = aφ, a must be either 0 or 1. Note that DΣD = Ik − φφτ . Then A3 = Σ1/2 DΛDΣDΛDΣDΛDΣ1/2 = Σ1/2 D(Λ − aφφτ )(Λ − aφφτ )ΛDΣ1/2 = Σ1/2 D(Λ − 2aφφτ + a2 φφτ )ΛDΣ1/2 = Σ1/2 D(Λ − aφφτ )ΛDΣ1/2

= Σ1/2 DΛDΣDΛDΣ1/2 = A2 ,

which implies that the eigenvalues of A must be 0 or 1. Therefore, A2 = A. Also, tr(A) = tr[Λ(DΣD)] = tr(Λ − aφφτ ) = tr(Λ) − a. (ii) The result in (ii) follows from the result in (i) and X/n →p p.

Note that the χ2 -statistic in (6.71) and the modified χ2 -statistic in (6.72) are special cases of the statistics in Theorem 6.8(i) and (ii), respectively, with Λ = Ik satisfying Λφ = φ. Hence, a test of asymptotic significance level α for testing (6.70) rejects H0 when χ2 > χ2k−1,α (or χ ˜2 > χ2k−1,α ), where χ2k−1,α is the (1 − α)th quantile of χ2k−1 . These tests are called (asymptotic) χ2 -tests. Example 6.23 (Goodness of fit tests). Let Y1 , ..., Yn be i.i.d. from F . Consider the problem of testing H0 : F = F0

versus

H1 : F 6= F0 ,

(6.73)

where F0 is a known c.d.f. For instance, F0 = N (0, 1). One way to test (6.73) is to partition the range of Y1 into k disjoint events A1 , ..., Ak and test (6.70) with pj = PF (Aj ) and p0j = PF0 (Aj ), j = 1, ..., k. Let Xj be the number of Yi ’s in Aj , j = 1, ..., k. Based on Xj ’s, the χ2 -tests discussed previously can be applied to this problem and they are called goodness of fit tests. In the goodness of fit tests discussed in Example 6.23, F0 in H0 is known so that p0j ’s can be computed. In some cases, we need to test the following hypotheses that are slightly different from those in (6.73): H0 : F = Fθ

versus

H1 : F 6= Fθ ,

(6.74)

438

6. Hypothesis Tests

where θ is an unknown parameter in Θ ⊂ Rs . For example, Fθ = N (µ, σ 2 ), θ = (µ, σ 2 ). If we still try to test (6.70) with pj = PFθ (Aj ), j = 1, ..., k, the result in Example 6.23 is not applicable since p is unknown under H0 . A generalized χ2 -test for (6.74) can be obtained using the following result. Let p(θ) = (p1 (θ), ..., pk (θ)) be a k-vector of known functions of θ ∈ Θ ⊂ Rs , where s < k. Consider the testing problem H0 : p = p(θ)

H1 : p 6= p(θ).

versus

(6.75)

Note that (6.70) is the special case of (6.75) with s = 0, i.e., θ is known. Let θˆ be an MLE of θ under H0 . Then, by Theorem 6.5, the LR test that rejects H0 when −2 log λn > χ2k−s−1,α has asymptotic significance level α, where χ2k−s−1,α is the (1 − α)th quantile of χ2k−s−1 and λn =

k Y

j=1

ˆ Xj [pj (θ)]

(Xj /n)Xj .

ˆ Using the fact that pj (θ)/(X j /n) →p 1 under H0 and log(1 + x) = x − x2 /2 + o(|x|2 )

as |x| → 0,

we obtain that ! ˆ pj (θ) −2 log λn = −2 −1 Xj log 1 + Xj /n j=1 ! k k ˆ X X pj (θ) −1 + = −2 Xj Xj Xj /n j=1 j=1 k X

ˆ pj (θ) −1 Xj /n

!2

+ op (1)

k X ˆ 2 [Xj − npj (θ)] + op (1) = Xj j=1

k X ˆ 2 [Xj − npj (θ)] + op (1), = ˆ npj (θ) j=1

Pk ˆ = Pk Xj /n = 1. Dewhere the third equality follows from j=1 pj (θ) j=1 ˜2 to be the χ2 and χ ˜2 in (6.71) fine the generalized χ2 -statistics χ2 and χ ˆ We then have the and (6.72), respectively, with p0j ’s replaced by pj (θ)’s. following result. Theorem 6.9. Under H0 given by (6.75), the generalized χ2 -statistics converge in distribution to χ2k−s−1 . The χ2 -test with rejection region χ2 > χ2k−s−1,α (or χ ˜2 > χ2k−s−1,α ) has asymptotic significance level α, where χ2k−s−1,α is the (1 − α)th quantile of χ2k−s−1 .

439

6.4. Tests in Parametric Models

Theorem 6.9 can be applied to derive a goodness of fit test for hypotheses (6.74). However, one has to formulate (6.75) and compute an MLE of θ under H0 : p = p(θ), which is different from an MLE under H0 : F = Fθ unless (6.74) and (6.75) are the same; see Moore and Spruill (1975). The next example is the main application of Theorem 6.9. Example 6.24 (r × c contingency tables). The following r × c contingency table is a natural extension of the 2 × 2 contingency table considered in Example 6.12: B1 B2 ··· Br Total

A1 X11 X21 ··· Xr1 m1

A2 X12 X22 ··· Xr2 m2

··· ··· ··· ··· ··· ···

Ac X1c X2c ··· Xrc mc

Total n1 n2 ··· nr n

where Ai ’s are disjoint events with A1 ∪ · · · ∪ Ac = Ω (the sample space of a random experiment), Bi ’s are disjoint events with B1 ∪ · · · ∪ Br = Ω, and Xij is the observed frequency of the outcomes in Aj ∩ Bi . Similar to the case of the 2 × 2 contingency table discussed in Example 6.12, there are two important applications in this problem. We first consider testing independence of {Aj : j = 1, ..., c} and {Bi : i = 1, ..., r} with hypotheses H0 : pij = pi· p·j for all i, j

versus

H1 : pij 6= pi· p·j for some i, j,

where pij = P (Aj ∩ Bi ) = E(Xij )/n, pi· = P (Bi ), and p·j = P (Aj ), i = 1, ..., r, j = 1, ..., c. In this case, X = (Xij , i = 1, ..., r, j = 1, ..., c) has the multinomial distribution with parameters pij , i = 1, ..., r, j = ¯ i· = ni /n and X ¯ ·j = mj /n, 1, ..., c. Under H0 , MLE’s of pi· and p·j are X respectively, i = 1, ..., r, j = 1, ..., c (exercise). By Theorem 6.9, the χ2 -test rejects H0 when χ2 > χ2(r−1)(c−1),α , where χ2 =

c r X X ¯ i· X ¯ ·j )2 (Xij − nX ¯ i· X ¯ ·j nX i=1 j=1

(6.76)

and χ2(r−1)(c−1),α is the (1 − α)th quantile of the chi-square distribution χ2(r−1)(c−1) (exercise). One can also obtain the modified χ2 -test by replacing ¯ i· X ¯ ·j by Xij in the denominator of each term of the sum in (6.76). nX Next, suppose that (X1j , ..., Xrj ), j = 1, ..., c, are c independent random vectors having the multinomial distributions with parameters (p1j , ..., prj ), j = 1, ..., c, respectively. Consider the problem of testing whether c multinomial distributions are the same, i.e., H0 : pij = pi1 for all i, j

versus

H1 : pij 6= pi1 for some i, j.

440

6. Hypothesis Tests

It turns out that the rejection region of the χ2 -test given in Theorem 6.9 is still χ2 > χ2(r−1)(c−1),α with χ2 given by (6.76) (exercise). One can also obtain the LR test in this problem. When r = c = 2, the LR test is equivalent to Fisher’s exact test given in Example 6.12, which is a UMPU test. When r > 2 or c > 2, however, a UMPU test does not exist in this problem.

6.4.4 Bayes tests An LR test actually compares supθ∈Θ0 ℓ(θ) with supθ∈Θ1 ℓ(θ) for testing (6.59). Instead of comparing two maximum values, one may compare two R R averages such as π ˆj = Θj ℓ(θ)dΠ(θ)/ Θ ℓ(θ)dΠ(θ), j = 0, 1, where Π(θ) is ˆ1 > π ˆ0 . If Π is treated as a prior c.d.f., a c.d.f. on Θ, and reject H0 when π then π ˆj is the posterior probability of Θj , and this test is a particular Bayes action (see Exercise 18 in §4.6) and is called a Bayes test. In Bayesian analysis, one often considers the Bayes factor defined to be β=

posterior odds ratio π ˆ0 /ˆ π1 = , prior odds ratio π0 /π1

where πj = Π(Θj ) is the prior probability of Θj . Clearly, if there is a statistic sufficient for θ, then the Bayes test and Bayes factor depend only on the sufficient statistic. Consider the special case where Θ0 = {θ0 } and Θ1 = {θ1 } are simple hypotheses. For given X = x, π ˆj =

πj fθj (x) . π0 fθ0 (x) + π1 fθ1 (x)

Rejecting H0 when π ˆ1 > π ˆ0 is the same as rejecting H0 when π0 fθ1 (x) > . fθ0 (x) π1

(6.77)

This is equivalent to the UMP test T∗ in (6.3) (Theorem 6.1) with c = π0 /π1 and γ = 0. The Bayes factor in this case is β=

π ˆ0 π1 fθ (x) . = 0 π ˆ1 π0 fθ1 (x)

Thus, the UMP test T∗ in (6.3) is equivalent to the test that rejects H0 when the Bayes factor is small. Note that the rejection region given by (6.77) depends on prior probabilities, whereas the Bayes factor does not. When either Θ0 or Θ1 is not simple, however, Bayes factors also depend on the prior Π.

441

6.4. Tests in Parametric Models

If Π is an improper prior, the Bayes test is still defined as long as the posterior probabilities π ˆj are finite. However, the Bayes factor may not be well defined when Π is improper. Example 6.25. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with an unknown µ ∈ R and a known σ 2 > 0. Let the prior of µ be N (ξ, τ 2 ). Then the posterior of µ is N (µ∗ (x), c2 ), where µ∗ (x) =

σ2 nτ 2 ξ+ 2 x ¯ 2 +σ nτ + σ 2

nτ 2

c2 =

and

τ 2 σ2 nτ 2 + σ 2

(see Example 2.25). Consider first the problem of testing H0 : µ ≤ µ0 versus H1 : µ > µ0 . Let Φ be the c.d.f. of the standard normal. Then the posterior probability of Θ0 and the Bayes factor are, respectively, ξ−µ0 µ −µ (x) Φ 0 c∗ Φ µ0 −µ∗ (x) µ0τ−ξ . and β= π ˆ0 = Φ µ∗ (x)−µ0 c Φ

c

Φ

τ

It is interesting to see that if we let τ → ∞, which is the same as considering the improper prior Π = the Lebesgue measure on R, then µ0√ −¯ x , π ˆ0 → Φ σ/ n

which is exactly the p-value α ˆ (x) derived in Example 2.29. Consider next the problem of testing H0 : µ = µ0 versus H1 : µ 6= µ0 . In this case the prior c.d.f. cannot be continuous at µ0 . We consider Π(µ) = π0 I[µ0 ,∞) (µ) + (1 − π0 )Φ µ−ξ . Let ℓ(µ) be the likelihood function based τ on x¯. Then Z x ¯−ξ 1 ′ √ √ m1 (x) = , = ℓ(µ)dΦ µ−ξ Φ τ 2 2 2 2 µ6=µ0

τ +σ /n

τ +σ /n

where Φ′ (t) is the p.d.f. of the standard normal distribution, and −1 1 − π0 π0 ℓ(µ0 ) = 1+ , π ˆ0 = π0 ℓ(µ0 ) + (1 − π0 )m1 (x) π0 β where

√ ¯−µ 2 + σ 2 Φ′ x √0 nτ ℓ(µ0 ) σ/ n = β= m1 (x) σΦ′ √ 2x¯−ξ2 τ +σ /n

is the Bayes factor.

More discussions about Bayesian hypothesis tests can be found in Berger (1985, §4.3.3).

442

6. Hypothesis Tests

6.5 Tests in Nonparametric Models In a nonparametric problem, a UMP, UMPU, or UMPI test usually does not exist. In this section we study some nonparametric tests that have size α, limiting size α, or asymptotic significance level α. Consistency (Definition 2.13) of these nonparametric tests is also discussed. Nonparametric tests are derived using some intuitively appealing ideas. They are commonly referred to as distribution-free tests, since almost no assumption is imposed on the population under consideration. But a nonparametric test may not be as good as a parametric test (in terms of its power) when the parametric model is correct. This is very similar to the case where we consider parametric estimation methods versus nonparametric estimation methods.

6.5.1 Sign, permutation, and rank tests Three popular classes of nonparametric tests are introduced here. The first one is the class of sign tests. Let X1 , ..., Xn be i.i.d. random variables from F , u be a fixed constant, and p = F (u). Consider the problem of testing H0 : p ≤ p0 versus H1 : p > p0 , or testing H0 : p = p0 versus H1 : p 6= p0 , where p0 is a fixed constant in (0, 1). Let ∆i =

1 0

Xi − u ≤ 0 Xi − u > 0,

i = 1, ..., n.

Then ∆1 , ..., ∆n are i.i.d. binary random variables with p = P (∆i = 1). For testing H0 : p ≤ p0 versus H1 : p > p0 , it follows from Corollary 6.1 that the test Y >m 1 T∗ (Y ) = (6.78) γ Y =m 0 Y <m Pn is of size α and UMP among tests based on ∆i ’s, where Y = i=1 ∆i and m and γ satisfy (6.7). Although T∗ is of size α, we cannot conclude immediately that T∗ is a UMP test, since ∆1 , ..., ∆n may not be sufficient for F . However, it can be shown that T∗ is in fact a UMP test (Lehmann, 1986, pp. 106-107) in this particular case. Note that no assumption is imposed on F . For testing H0 : p = p0 versus H1 : p 6= p0 , it follows from Theorem 6.4 that the test Y < c1 or Y > c2 1 T∗ (Y ) = (6.79) γ Y = ci , i = 1, 2, i 0 c1 < Y < c2

443

6.5. Tests in Nonparametric Models

is of size α and UMP among unbiased tests based on ∆i ’s, where γ and ci ’s are chosen so that E(T∗ ) = α and E(T∗ Y ) = αnp0 when p = p0 . This test is in fact a UMPU test (Lehmann, 1986, p. 166). Since Y is equal to the number of nonnegative signs of (u − Xi )’s, tests based on T∗ in (6.78) or (6.79) are called sign tests. One can easily extend the sign tests to the case where p = P (X1 ∈ B) with any fixed event B. Another extension is to the case where we observe i.i.d. (X1 , Y1 ), ..., (Xn , Yn ) (matched pairs). By using ∆i = Xi − Yi − u, one can obtain sign tests for hypotheses concerning P (X1 − Y1 ≤ u). Next, we introduce the class of permutation tests. Let Xi1 , ..., Xini , i = 1, 2, be two independent samples i.i.d. from Fi , i = 1, 2, respectively, where Fi ’s are c.d.f.’s on R. In §6.2.3, we showed that the two-sample t-tests are UMPU tests for testing hypotheses concerning the means of Fi ’s, under the assumption that Fi ’s are normal with the same variance. Such types of testing problems arise from the comparison of two treatments. Suppose now we remove the normality assumption and replace it by a much weaker assumption that Fi ’s are in the nonparametric family F containing all continuous c.d.f.’s on R. Consider the problem of testing H0 : F1 = F2

versus

H1 : F1 6= F2 ,

(6.80)

which is the same as testing the equality of the means of Fi ’s when Fi ’s are normal with the same variance. Let X = (Xij , j = 1, ..., ni , i = 1, 2), n = n1 + n2 , and α be a given significance level. A test T (X) satisfying 1 X T (z) = α n!

(6.81)

z∈π(x)

is called a permutation test, where π(x) is the set of n! points obtained from x ∈ Rn by permuting the components of x. Permutation tests are of size α (exercise). Under the assumption that F1 (x) = F2 (x − θ) and F1 ∈ F containing all c.d.f.’s having Lebesgue p.d.f.’s that are continuous a.e., which is still much weaker than the assumption that Fi ’s are normal with the same variance, the class of permutation tests of size α is exactly the same as the class of unbiased tests of size α; see, for example, Lehmann (1986, p. 231). Unfortunately, a test UMP among all permutation tests of size α does not exist. In applications, we usually choose a Lebesgue p.d.f. h and define a permutation test h(X) > hm 1 T (X) = (6.82) γ h(X) = hm 0 h(X) < hm ,

444

6. Hypothesis Tests

where hm is the (m + 1)th largest value of the set {h(z) : z ∈ π(x)}, m is the integer part of αn!, and γ = αn! − m. This permutation test is optimal in some sense (Lehmann, 1986, §5.11). While the class of permutation tests is motivated by the unbiasedness principle, the third class of tests introduced here is motivated by the invariance principle. Consider first the one-sample problem in which X1 , ..., Xn are i.i.d. random variables from a continuous c.d.f. F and we would like to test H0 : F is symmetric about 0 versus H1 : F is not symmetric about 0. Let G be the class of transformations g(x) = (ψ(x1 ), ..., ψ(xn )), where ψ is ˜ continuous, odd, and strictly increasing. Let R(X) be the vector of ranks of ˜ |Xi |’s and R+ (X) (or R− (X)) be the subvector of R(X) containing ranks corresponding to positive (or negative) Xi ’s. It can be shown (exercise) that (R+ , R− ) is maximal invariant under G. Furthermore, sufficiency permits o a reduction from R+ and R− to R+ , the vector of ordered components of o R+ . A test based on R+ is called a (one-sample) signed rank test. Similar to the case of permutation tests, there is no UMP test within the class of signed rank tests. A common choice is the signed rank test that o rejects H0 when W (R+ ) is too large or too small, where o o o W (R+ ) = J(R+1 /n) + · · · + J(R+n /n), ∗

(6.83)

o J is a continuous and strictly increasing function on [0, 1], R+i is the ith o component of R+ , and n∗ is the number of positive Xi ’s. This is motivated by the fact that H0 is unlikely to be true if W in (6.83) is too large or too small. Note that W/n is equal to T(Fn ) with T given by (5.53) and J(t) = t, and the test based on W in (6.83) is the well-known one-sample Wilcoxon signed rank test. o Under H0 , P (R+ = y) = 2−n for each y ∈ Y containing 2n n∗ -tuples y = (y1 , ..., yn∗ ) satisfying 1 ≤ y1 < · · · < yn∗ ≤ n. Then, the following signed rank test is of size α:

1 T (X) = γ 0

o o W (R+ ) < c1 or W (R+ ) > c2 o W (R+ ) = ci , i = 1, 2 o ) < c2 , c1 < W (R+

(6.84)

where c1 and c2 are the (m + 1)th smallest and largest values of the set {W (y) : y ∈ Y}, m is the integer part of α2n /2, and γ = α2n /2 − m. Consider next the two-sample problem of testing (6.80) based on two independent samples, Xi1 , ..., Xini , i = 1, 2, i.i.d. from Fi , i = 1, 2, respectively. Let G be the class of transformations g(x) = (ψ(xij ), j = 1, ..., ni , i =

6.5. Tests in Nonparametric Models

445

1, 2), where ψ is continuous and strictly increasing. Let R(X) be the vector of ranks of all Xij ’s. In Example 6.14, we showed that R is maximal invariant under G. Again, sufficiency permits a reduction from R to R1o , the vector of ordered values of the ranks of X11 , ..., X1n1 . A test for (6.80) based on R1o is called a two-sample rank test. Under H0 , P (R1o = y) = n −1 for each y ∈ Y containing nn1 n1 -tuples y = (y1 , ..., yn1 ) satisfying n1 o o 1 ≤ y1 < · · · < yn1 ≤ n. Let R1o = (R11 , ..., R1n ). Then a commonly 1 o used two-sample rank test is given by (6.83)-(6.84) with R+i , n∗ , and 2n n o replaced by R1i , n1 , and n1 , respectively. When n1 = n2 , the statistic W/n is equal to T(Fn ) with T given by (5.55). When J(t) = t − 21 , this reduces to the well-known two-sample Wilcoxon rank test. A common feature of the permutation and rank tests previously introduced is that tests of size α can be obtained for each fixed sample size n, but the computation involved in determining the rejection regions {T (X) = 1} may be cumbersome if n is large. Thus, one may consider approximations to permutation and rank tests when n is large. Permutation tests can often be approximated by the two-sample t-tests derived in §6.2.3 (Lehmann, 1986, §5.13). Using the results in §5.2.2, we now derive one-sample signed rank tests having limiting size α (Definition 2.13(ii)), which can be viewed as signed rank tests of size approximately α when n is large. From the discussion in §5.2.2, W/n = T(Fn ) with a ̺∞ -Hadamard differentiable functional T given by (5.53) and, by Theorem 5.5, √ n[W/n − T(F )] →d N (0, σF2 ), where σF2 = E[φF (X1 )]2 , Z ∞ J ′ (F˜ (y))(˜διx − F˜ )(y)dF (y) + J(F˜ (x)) − T(F ) φF (x) = 0

(see (5.54)), and διx denotes the c.d.f. degenerated at x. Since F is continuous, F˜ (x) = F (x) − F (−x). Under H0 , F (x) = 1 − F (−x). Hence, σF2 under H0 is equal to v1 + v2 + 2v12 , where Z 1 ∞ v1 = Var J(F˜ (X1 )) = [J(F˜ (x))]2 dF˜ (x), 2 0 Z

∞

˜ ˜ ˜ J (F (y))(διX1 − F )(y)dF (y) ′

v2 = Var 0 Z ∞Z ∞ =E J ′ (F˜ (y))J ′ (F˜ (z))(˜διX1 − F˜ )(y)(˜διX1 − F˜ )(z)dF (y)dF (z) 0 0 Z Z 1 ∞ ∞ ′ ˜ = J (F (y))J ′ (F˜ (z))[F˜ (min{y, z}) − F˜ (y)F˜ (z)]dF˜ (y)dF˜ (z) 4 0 0 Z 1 J ′ (F˜ (y))J ′ (F˜ (z))F˜ (z)[1 − F˜ (y)]dF˜ (y)dF˜ (z), = 2 0 σ0 z1−α/2 → α,

i.e., T has limiting size α. Two-sample rank tests having limiting size α can be similarly derived (exercise).

6.5.2 Kolmogorov-Smirnov and Cram´ er-von Mises tests In this section we introduce two types of tests for hypotheses concerning continuous c.d.f.’s on R. Let X1 , ..., Xn be i.i.d. random variables from a continuous c.d.f. F . Suppose that we would like to test hypotheses (6.73), i.e., H0 : F = F0 versus H1 : F 6= F0 with a fixed F0 . Let Fn be the empirical c.d.f. and Dn (F ) = sup |Fn (x) − F (x)|,

(6.86)

x∈R

which is in fact the distance ̺∞ (Fn , F ). Intuitively, Dn (F0 ) should be small if H0 is true. From the results in §5.1.1, we know that Dn (F0 ) →a.s. 0 if and

6.5. Tests in Nonparametric Models

447

only if H0 is true. The statistic Dn (F0 ) is called the Kolmogorov-Smirnov statistic. Tests with rejection region Dn (F0 ) > c are called KolmogorovSmirnov tests. In some cases we would like to test “one-sided” hypotheses H0 : F = F0 versus H1 : F ≥ F0 , F 6= F0 , or H0 : F = F0 versus H1 : F ≤ F0 , F 6= F0 . The corresponding Kolmogorov-Smirnov statistic is Dn+ (F0 ) or Dn− (F0 ), where Dn+ (F ) = sup [Fn (x) − F (x)] (6.87) x∈R

and Dn− (F ) = sup [F (x) − Fn (x)]. x∈R

The rejection regions of one-sided Kolmogorov-Smirnov tests are, respectively, Dn+ (F0 ) > c and Dn− (F0 ) > c. Let X(1) < · · · < X(n) be the order statistics and define X(0) = −∞ and X(n+1) = ∞. Since Fn (x) = i/n when X(i) ≤ x < X(i+1) , i = 0, 1, ..., n, i + − F (x) sup Dn (F ) = max 0≤i≤n X(i) ≤x<X(i+1) n i = max − inf F (x) 0≤i≤n n X(i) ≤x<X(i+1) i − F (X(i) ) . = max 0≤i≤n n When F is continuous, F (X(i) ) is the ith order statistic of a sample of size n from the uniform distribution U (0, 1) irrespective of what F is. Therefore, the distribution of Dn+ (F ) does not depend on F , if we restrict our attention to continuous c.d.f.’s on R. The distribution of Dn− (F ) is the same as that of Dn+ (F ) because of symmetry (exercise). Since Dn (F ) = max{Dn+ (F ), Dn− (F )}, the distribution of Dn (F ) does not depend on F . This means that the distributions of Kolmogorov-Smirnov statistics are known under H0 . Theorem 6.10. Let Dn (F ) and Dn+ (F ) be defined by (6.86) and (6.87), respectively, for a continuous c.d.f. F on R. (i) For any fixed n, 0 t≤0 n Z un−i+2 Y n! du1 · · · dun 0 0.

j=−∞

6.5.3 Empirical likelihood ratio tests The method of likelihood ratio is useful in deriving tests under parametric models. In nonparametric problems, we now introduce a similar method based on the empirical likelihoods introduced in §5.1.2 and §5.1.4. Suppose that a sample X is from a population determined by a c.d.f. F ∈ F, where F is a class of c.d.f.’s on Rd . Consider the problem of testing H0 : T(F ) = t0

versus

H1 : T(F ) 6= t0 ,

(6.89)

where T is a functional from F to Rk and t0 is a fixed vector in Rk . Let ℓ(G), G ∈ F, be a given empirical likelihood, Fˆ be an MELE of F , and FˆH0 be an MELE of F under H0 , i.e., FˆH0 is an MELE of F subject to T(F ) = t0 . Then the empirical likelihood ratio is defined as λn (X) = ℓ(FˆH0 )/ℓ(Fˆ ). A test with rejection region λn (X) < c is called an empirical likelihood ratio test.

450

6. Hypothesis Tests

As a specific example, consider the following empirical likelihood (or nonparametric likelihood) when X = (X1 , ..., Xn ) with i.i.d. Xi ’s: ℓ(G) =

n Y

subject to pi ≥ 0,

pi

i=1

n X

pi = 1,

i=1

R where pi = PG ({xi }), i = 1, ..., n. Suppose that T(G) = u(x)dG(x) with a known function u(x) from Rd to Rr . Then Fˆ = Fn ; H0 in (6.89) with t0 = 0 is the same as the case where assumption (5.9) holds; FˆH0 is the MELE given by (5.11); and the empirical likelihood ratio is λn (X) = nn

n Y

pˆi ,

(6.90)

i=1

where pˆi is given by (5.12). An empirical likelihood ratio test with asymptotic significance level α can be obtained using the following result. Theorem 6.11. Assume the conditions in Theorem 5.4. Under the hypothesis H0 in (6.89) with t0 = 0 (i.e., (5.9) holds), −2 log λn →d χ2r , where λn = λn (X) is given by (6.90) and χ2r has the chi-square distribution χ2r . The proof of this result can be found in Owen (1988, 1990). In fact, the result in Theorem 6.11 holds for some other functionals T such as the median functional. We can also derive tests based on the profile empirical likelihoods discussed in §5.4.1. Consider an empirical likelihood ℓ(G) =

n Y

pi

subject to

i=1

pi ≥ 0,

n X i=1

pi = 1,

n X

pi ψ(xi , θ) = 0,

i=1

where θ is a k-vector of unknown parameters and ψ is a known function. Let θ = (ϑ, ϕ), where ϑ is an r-vector and ϕ is a (k − r)-vector. Suppose that we would like to test H0 : ϑ = ϑ0

versus

H1 : ϑ 6= ϑ0 ,

where ϑ0 is a fixed r-vector. Let θˆ be a maximum of the profile empirical likelihood ℓP (θ) given by (5.36) and let ϕˆ be a maximum of ℓP (ϕ) = ℓP (ϑ0 , ϕ). Then a profile empirical likelihood ratio test rejects H0 when λn (X) < c, where λn (X) =

n Y

ˆ τ ψ(xi , θ) ˆ 1 + [ξn (θ)] , 1 + [ζn (ϑ0 , ϕ)] ˆ τ ψ(xi , ϑ0 , ϕ) ˆ i=1

(6.91)

6.5. Tests in Nonparametric Models

451

ˆ satisfies θˆ and ϕˆ are maximum profile empirical likelihood estimators, ξn (θ) n X

ˆ ψ(xi , θ) = 0, τ ˆ ˆ 1 + [ξn (θ)] ψ(xi , θ)

i=1

ˆ satisfies and ζn (ϑ0 , ϕ) n X i=1

ψ(xi , ϑ0 , ϕ) ˆ = 0. τ 1 + [ζn (ϑ0 , ϕ)] ˆ ψ(xi , ϑ0 , ϕ) ˆ

Pn From the discussion in §5.4.1, θˆ is a solution of the GEE i=1 ψ(Xi , θ) = 0 when the dimension of ψ is k. Under some regularity conditions (e.g., the conditions in Proposition 5.3), Qin and Lawless (1994) showed that the result in Theorem 6.11 holds with λn (X) given by (6.91). Thus, a profile empirical likelihood ratio test with asymptotic significance level α can be obtained. Example 6.26. Let Y1 , ..., Yn be i.i.d. random 2-vectors from F . Consider the problem of testing H0 : µ1 = µ2 versus H1 : µ1 6= µ2 , where (µ1 , µ2 ) = E(Y1 ). Let Yi = (Yi1 , Yi2 ), Xi1 = Yi1 − Yi2 , Xi2 = Yi1 + Yi2 , and Xi = (Xi1 , Xi2 ), i = 1, ..., n. Then X1 , ..., Xn are i.i.d. with E(X1 ) = θ = (ϑ, ϕ), where ϑ = µ1 − µ2 and ϕ = µ1 + µ2 . The hypotheses of interest becomes H0 : ϑ = 0 versus H1 : ϑ 6= 0. To apply the profile empirical likelihood method, Pn we define ψ(x, θ) = x − θ, x ∈ R2 . Note that a solution of the GEE i=1 (Xi − θ) = 0 is the ¯ The profile empirical likelihood ratio is then given by sample mean θˆ = X. λn (X) =

n Y

¯ τ (Xi − X) ¯ 1 + [ξn (X)] , τ 1 + [ζn (0, ϕ)] ˆ [Xi − (0, ϕ)] ˆ i=1

¯ ζn (0, ϕ), where ξn (X), ˆ and ϕˆ satisfy n X i=1

n X i=1

¯ Xi − X τ ¯ ¯ = 0, 1 + [ξn (X)] (Xi − X)

Xi − (0, ϕ) ˆ = 0, τ 1 + [ζn (0, ϕ)] ˆ [Xi − (0, ϕ)] ˆ

and ℓP (0, ϕ) ˆ = maxϕ ℓP (0, ϕ) with ℓP (0, ϕ) =

n Y

1 . τ [X − (0, ϕ)]} n{1 + [ξ (0, ϕ)] n i i=1

Empirical likelihood ratio tests or profile empirical likelihood ratio tests in various other problems can be found, for example, in Owen (1988, 1990, 2001), Chen and Qin (1993), Qin (1993), and Qin and Lawless (1994).

452

6. Hypothesis Tests

6.5.4 Asymptotic tests We now introduce a simple method of constructing asymptotic tests (i.e., tests with asymptotic significance level α). This method works for almost all problems (parametric or nonparametric) in which the hypotheses being tested are H0 : θ = θ0 versus H1 : θ 6= θ0 , where θ is a vector of parameters, and an asymptotically normally distributed estimator of θ can be found. However, this simple method may not provide the best or even nearly best solution to the problem, especially when there are different asymptotically normally distributed estimators of θ. Let X be a sample of size n from a population P and θˆn be an estimator of θ, a k-vector of parameters related to P . Suppose that under H0 , Vn−1/2 (θˆn − θ) →d Nk (0, Ik ),

(6.92)

where Vn is the asymptotic covariance matrix of θˆn . If Vn is known when θ = θ0 , then a test with rejection region (θˆn − θ0 )τ Vn−1 (θˆn − θ0 ) > χ2k,α

(6.93)

has asymptotic significance level α, where χ2k,α is the (1 − α)th quantile of the chi-squared distribution χ2k . If the distribution of θˆn does not depend on the unknown population P under H0 and (6.92) holds, then a test with rejection region (6.93) has limiting size α. If Vn in (6.93) depends on the unknown population P even if H0 is true (θ = θ0 ), then we have to replace Vn in (6.93) by an estimator Vˆn . If, under H0 , Vˆn is consistent according to Definition 5.4, then the test having rejection region (6.93) with Vn replaced by Vˆn has asymptotic significance level α. Variance estimation methods introduced in §5.5 can be used to construct a consistent estimator Vˆn . In some cases result (6.92) holds for any P . Then, the following result shows that the test having rejection region (6.93) is asymptotically correct (§2.5.3), i.e., it is a consistent asymptotic test (Definition 2.13). Theorem 6.12. Assume that (6.92) holds for any P and that λ+ [Vn ] → 0, where λ+ [Vn ] is the largest eigenvalue of Vn . (i) The test having rejection region (6.93) (with a known Vn or Vn replaced by an estimator Vˆn that is consistent for any P ) is consistent. (ii) If we choose α = αn → 0 as n → ∞ and χ2k,1−αn λ+ [Vn ] = o(1), then the test in (i) is Chernoff-consistent. Proof. The proof of (ii) is left as an exercise. We only prove (i) for the −1/2 ˆ −1/2 case where Vn is known. Let Zn = Vn (θn − θ) and ln = Vn (θ − θ0 ). −1/2 Then kZn k = Op (1) and kln k = kVn (θ − θ0 )k → ∞ when θ 6= θ0 . The

6.5. Tests in Nonparametric Models

453

result follows from the fact that when θ 6= θ0 ,

(θˆn − θ0 )τ Vn−1 (θˆn − θ0 ) = kZn k2 + kln k2 + 2lnτ Zn ≥ kZn k2 + kln k2 − 2kln kkZn k = Op (1) + kln k2 [1 − op (1)]

and, therefore,

P (θˆn − θ0 )τ Vn−1 (θˆn − θ0 ) > χ2k,α → 1.

Example 6.27. Let X1 , ..., Xn be i.i.d. random variables from a symmetric c.d.f. F having finite variance and positive F ′ . Consider the problem of testing H0 : F is symmetric about 0 versus H1 : F is not symmetric about 0. Under H0 , there are many estimators satisfying (6.92). We consider the following five estimators: ¯ and θ = E(X1 ); (1) θˆn = X ˆ ˆ (2) θn = θ0.5 (the sample median) and θ = F −1 ( 12 ) (the median of F ); ¯ a (the a-trimmed sample mean defined by (5.77)) and θ = T(F ), (3) θˆn = X where T is given by (5.46) with J(t) = (1 − 2a)−1 I(a,1−a) (t), a ∈ (0, 12 ); (4) θˆn = the Hodges-Lehmann estimator (Example 5.8) and θ = F −1 ( 12 ); (5) θˆn = W/n − 12 , where W is given by (6.83) with J(t) = t, and θ = T(F ) − 12 with T given by (5.53). Although the θ’s in (1)-(5) are different in general, in all cases θ = 0 is equivalent to that H0 holds. ¯ it follows from the CLT that (6.92) holds with Vn = σ 2 /n for any For X, F , where σ 2 = Var(X1 ). From the SLLN, S 2 /n is a consistent estimator of ¯ and Vn for any F . Thus, the test having rejection region (6.93) with θˆn = X 2 Vn replaced by S /n is asymptotically correct. This test is asymptotically equivalent to the one-sample t-test derived in §6.2.3. From Theorem 5.10, θˆ0.5 satisfies (6.92) with Vn = 4−1 [F ′ (θ)]−2 n−1 for any F . A consistent estimator of Vn can be obtained using the bootstrap method considered in §5.5.3. Another consistent estimator of Vn can be obtained using Woodruff’s interval introduced in §7.4 (see Exercise 86 in §7.6). The test having rejection region (6.93) with θˆn = θˆ0.5 and Vn replaced by a consistent estimator is asymptotically correct. ¯ a satisfies (6.92) for any It follows from the discussion in §5.3.2 that X F . A consistent estimator of Vn can be obtained using formula (5.110) or the jackknife method in §5.5.2. The test having rejection region (6.93) ¯ a and Vn replaced by a consistent estimator is asymptotically with θˆn = X correct. From Example 5.8, the Hodges-Lehmann estimator satisfies (6.92) for R any F and Vn = 12−1 γ −2 n−1 under H0 , where γ = F ′ (x)dF (x). A

454

6. Hypothesis Tests

consistent estimator of Vn under H0 can be obtained using the result in Exercise 102 in §5.6. The test having rejection region (6.93) with θˆn = the Hodges-Lehmann estimator and Vn replaced by a consistent estimator is asymptotically correct. Note that all tests discussed so far are not of limiting size α, since the distributions of θˆn are still unknown under H0 . The test having rejection region (6.93) with θˆn = W/n − 21 and Vn = (12n)−1 is equivalent to the one-sample Wilcoxon signed rank test and is shown to have limiting size α (§6.5.1). Also, (6.92) is satisfied for any F (§5.2.2). Although Theorem 6.12 is not applicable, a modified proof of Theorem 6.12 can be used to show the consistency of this test (exercise). It is not clear which one of the five tests discussed here is to be preferred in general. The results for θˆn in (1)-(3) and (5) still hold for testing H0 : θ = 0 versus H1 : θ 6= 0 without the assumption that F is symmetric. An example of asymptotic tests for one-sided hypotheses is given in Exercise 123. Most tests in §6.1-§6.4 derived under parametric models are asymptotically correct even when the parametric model assumptions are removed. Some examples are given in Exercises 121-123. Finally, a study of asymptotic efficiencies of various tests can be found, for example, in Serfling (1980, Chapter 10).

6.6 Exercises 1. Prove Theorem 6.1 for the case of α = 0 or 1. 2. Assume the conditions in Theorem 6.1. Let β(P ) be the power function of a UMP test of size α ∈ (0, 1). Show that α < β(P1 ) unless P0 = P1 . 3. Let T∗ be given by (6.3) with c = c(α) for an α > 0. (a) Show that if α1 < α2 , then c(α1 ) ≥ c(α2 ). (b) Show that if α1 < α2 , then the type II error probability of T∗ of size α1 is larger than that of T∗ of size α2 . 4. Let H0 and H1 be simple and let α ∈ (0, 1). Suppose that T∗ is a UMP test of size α for testing H0 versus H1 and that β < 1, where β is the power of T∗ when H1 is true. Show that 1 − T∗ is a UMP test of size 1 − β for testing H1 versus H0 . 5. Let X be a sample of size 1 from a Lebesgue p.d.f. fθ . Find a UMP test of size α ∈ (0, 21 ) for H0 : θ = θ0 versus H1 : θ = θ1 when

6.6. Exercises

455

(a) fθ (x) = 2θ−2 (θ − x)I(0,θ) (x), θ0 < θ1 ; (b) fθ (x) = 2[θx + (1 − θ)(1 − x)]I(0,1) (x), 0 ≤ θ1 < θ0 ≤ 1; (c) fθ0 is the p.d.f. of N (0, 1) and fθ1 is the p.d.f. of the Cauchy distribution C(0, 1); (d) fθ0 (x) = 4xI(0, 21 ) (x) + 4(1 − x)I( 12 ,1) (x) and fθ1 (x) = I(0,1) (x); (e) fθ is the p.d.f. of the Cauchy distribution C(θ, 1) and θi = i; (f) fθ0 (x) = e−x I(0,∞) (x) and fθ1 (x) = 2−1 x2 e−x I(0,∞) (x). 6. Let X1 , ..., Xn be i.i.d. from a Lebesgue p.d.f. fθ . Find a UMP test of size α for H0 : θ = θ0 versus H1 : θ = θ1 in the following cases: (a) fθ (x) = e−(x−θ) I(θ,∞) (x), θ0 < θ1 ; (b) fθ (x) = θx−2 I(θ,∞) (x), θ0 6= θ1 . 7. Prove Proposition 6.1. 8. Let X ∈ Rn be a sample with a p.d.f. f w.r.t. a σ-finite measure ν. Consider the problem of testing H0 : f = fθ versus H1 : f = g, where θ ∈ Θ, fθ (x) is Borel on (Rn × Θ, σ(B n × F )), and (Θ, F , Λ) is a probability space. Let c > 0 be a constant and R 1 g(x) ≥ c Θ fθ (x)dΛ R φ∗ (x) = 0 g(x) < c Θ fθ (x)dΛ. R R Suppose that φ∗ (x)fθ (x)dν = supθ∈Θ φ∗ (x)fθ (x)dν = α for any θ ∈ Θ′ with Λ(Θ′ ) = 1. Show that φ∗ is a UMP test of size α. 9. Let f0 and f1 be Lebesgue integrable functions on R and φ∗ be the indicator function of the set R {x : f0 (x) < 0} ∪ {x : f0 (x) = 0, f1 (x) ≥ 0}. Show that φ∗ maximizes φ(x)f R 1 (x)dx over all RBorel functions φ on R satisfying 0 ≤ φ(x) ≤ 1 and φ(x)f0 (x)dx = φ∗ (x)f0 (x)dx.

10. Let F1 and F2 beR two c.d.f.’s on RR. Show that F1 (x) ≤ F2 (x) for all x if and only if g(x)dF2 (x) ≤ g(x)dF1 (x) for any nondecreasing function g. 11. Prove the claims in Example 6.5. 12. Show that the family {fθ : θ ∈ R} has monotone likelihood ratio, where fθ (x) = c(θ)h(x)I(a(θ),b(θ)) (x), h is a positive Lebesgue integrable function, and a and b are nondecreasing functions of θ. 13. Prove part (iv) and part (v) of Theorem 6.2. 14. Let X1 , ..., Xn be i.i.d. from a Lebesgue p.d.f. fθ , θ ∈ Θ ⊂ R. Find a UMP test of size α for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 when (a) fθ (x) = θ−1 e−x/θ I(0,∞) (x), θ > 0; (b) fθ (x) = θ−1 xθ−1 I(0,1) (x), θ > 0; (c) fθ (x) is the p.d.f. of N (1, θ); c (d) fθ (x) = θ−c cxc−1 e−(x/θ) I(0,∞) (x), θ > 0, where c > 0 is known.

456

6. Hypothesis Tests

15. Suppose that the distribution of X is in a family with monotone likelihood ratio in Y (X), where Y (X) has a continuous distribution. Consider the hypotheses H0 : θ ≤ θ0 versus H1 : θ > θ0 . Show that the p-value (§2.4.2) of the UMP test is given by Pθ0 (Y ≥ y), where y is the observed value of Y . 16. Let X1 , ..., Xm be i.i.d. from N (µx , σx2 ) and Y1 , ..., Yn be i.i.d. from N (µy , σy2 ). Suppose that Xi ’s and Yj ’s are independent. (a) When σx = σy = 1, find a UMP test of size α for testing H0 : µx ≤ µy versus H1 : µx > µy . (Hint: see Lehmann (1986, §3.9).) (b) When µx and µy are known, find a UMP test of size α for testing H0 : σx ≤ σy versus H1 : σx > σy . (Hint: see Lehmann (1986, §3.9).) 17. Let F and G be two known c.d.f.’s on R and X be a single observation from the c.d.f. θF (x) + (1 − θ)G(x), where θ ∈ [0, 1] is unknown. (a) Find a UMP test of size α for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 . (b) Show that the test T∗ (X) ≡ α is a UMP test of size α for testing H0 : θ ≤ θ1 or θ ≥ θ2 versus H1 : θ1 < θ < θ2 . 18. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (θ, θ + 1), θ ∈ R. Suppose that n ≥ 2. (a) Find the joint distribution of X(1) and X(n) . (b) Show that a UMP test of size α for testing H0 : θ ≤ 0 versus H1 : θ > 0 is of the form 0 X(1) < 1 − α1/n , X(n) < 1 T∗ (X(1) , X(n) ) = 1 otherwise. (c) Does the family of all possible distributions of (X(1) , X(n) ) have monotone likelihood ratio? (Hint: see Lehmann (1986, p. 115).) 19. Suppose that X1 , ..., Xn are i.i.d. from the discrete uniform distribution DU (1, ..., θ) (Table 1.1, page 18) with an unknown θ = 1, 2, .... (a) Consider H0 : θ ≤ θ0 versus H1 : θ > θ0 . Show that 1 X(n) > θ0 T∗ (X) = α X(n) ≤ θ0 is a UMP test of size α. (b) Consider H0 : θ = θ0 versus H1 : θ 6= θ0 . Show that 1 X(n) > θ0 or X(n) ≤ θ0 α1/n T∗ (X) = 0 otherwise is a UMP test of size α. (c) Show that the results in (a) and (b) still hold if the discrete uniform distribution is replaced by the uniform distribution U (0, θ), θ > 0.

6.6. Exercises

457

20. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(a, θ), a ∈ R, θ > 0. (a) Derive a UMP test of size α for testing H0 : a = a0 versus H1 : a 6= a0 , when θ is known. (b) For testing H0 : a = a0 versus H1 : a = a1 < a0 , show that any UMP test T∗ of size α satisfies βT∗ (a1 ) = 1 − (1 − α)e−n(a0 −a1 )/θ . (c) For testing H0 : a = a0 versus H1 : a = a1 < a0 , show that the power of any size α test that rejects H0 when Y ≤P c1 or Y ≥ c2 is the same as that in part (b), where Y = (X(1) − a0 )/ ni=1 (Xi − X(1) ). (d) Derive a UMP test of size α for testing H0 : a = a0 versus H1 : a 6= a0 . (e) Derive a UMP test of size α for testing H0 : θ = θ0 , a = a0 versus H1 : θ < θ0 , a < a0 . 21. Let X1 , ..., Xn be i.i.d. from the Pareto distribution P a(a, θ), θ > 0, a > 0. (a) Derive a UMP test of size α for testing H0 : a = a0 versus H1 : a 6= a0 when θ is known. (b) Derive a UMP test of size α for testing H0 : a = a0 , θ = θ0 versus H1 : θ > θ0 , a < a0 . 22. In Exercise 19(a) of §3.6, derive a UMP test of size α ∈ (0, 1) for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 , where θ0 is known and θ0 > (1 − α)−1/n . 23. In Exercise 55 of §2.6, derive a UMP test of size α for testing H0 : θ ≥ θ0 versus H1 : θ < θ0 based on data X1 , ..., Xn , where θ0 > 0 is a fixed value. 24. Prove part (ii) of Theorem 6.3. 25. Consider Example 6.10. Suppose that θ2 = −θ1 . Show that c2 = −c1 and discuss how to find the value of c2 . 26. Suppose that the distribution of X is in a family of p.d.f.’s indexed by a real-valued parameter θ; there is a real-valued sufficient statistic U (X) such that fθ2 (u)/fθ1 (u) is strictly increasing in u for θ1 < θ2 , where fθ (u) is the Lebesgue p.d.f. of U (X) and is continuous in u for each θ; and that for all θ1 < θ2 < θ3 and u1 < u2 < u3 , fθ1 (u1 ) fθ1 (u2 ) fθ1 (u3 ) fθ (u1 ) fθ (u2 ) fθ (u3 ) > 0. 2 2 2 f (u ) f (u ) f (u ) θ3 1 θ3 2 θ3 3 Show that the conclusions of Theorem 6.3 remain valid.

458

6. Hypothesis Tests

27. (p-values). Suppose that X has a distribution Pθ , where θ ∈ R is unknown. Consider a family of nonrandomized level α tests for H0 : θ = θ0 (or θ ≤ θ0 ) with rejection region Cα such that Pθ0 (X ∈ Cα ) = α for all 0 < α < 1 and Cα1 = ∩α>α1 Cα for all 0 < α1 < 1. (a) Show that the p-value is α ˆ (x) = inf{α : x ∈ Cα }. (b) Show that when θ = θ0 , α(X) ˆ has the uniform distribution U (0, 1). (c) If the tests with rejection regions Cα are unbiased of level α, show that under H1 , Pθ (ˆ α(X) ≤ α) ≥ α. 28. Suppose that X has the p.d.f. (6.10). Consider hypotheses (6.13) or (6.14). Show that a UMP test does not exist. (Hint: this follows from a consideration of the UMP tests for the one-sided hypotheses H0 : θ ≥ θ1 and H0 : θ ≤ θ2 .) 29. Consider Exercise 17 with H0 : θ ∈ [θ1 , θ2 ] versus H1 : θ 6∈ [θ1 , θ2 ], where 0 < θ1 ≤ θ2 < 1. (a) Show that a UMP test does not exist. (b) Obtain a UMPU test of size α. 30. In the proof of Theorem 6.4, show that (a) (6.30) is equivalent to (6.31); (b) (6.31) is equivalent to (6.29) with T∗ replaced by T ; (c) when 0 < α < 1, (α, αEθ0 (Y )) is an interior point of the set of points (Eθ0 (T ), Eθ0 (T Y )) as T ranges over all tests of the form T = T (Y ); (d) the UMPU tests are unique a.s. P if attention is restricted to tests depending on (Y, U ) and (Y, U ) has a continuous c.d.f. 31. Consider the decision problem in Example 2.20 with the 0-1 loss. Show that if a UMPU test of size α exists and is unique (in the sense that decision rules that are equivalent in terms of the risk are treated the same), then it is admissible. 32. Let X1 , ..., Xn be i.i.d. binary random variables with p = P (X1 = 1). (a) Determine the ci ’s and γi ’s in (6.15) and (6.16) for testing H0 : p ≤ 0.2 or p ≥ 0.7 when α = 0.1 and n = 15. Find the power of the UMP test (6.15) when p = 0.4. (b) Derive a UMPU test of size α for H0 : p = p0 versus H1 : p 6= p0 when n = 10, α = 0.05, and p0 = 0.4. 33. Suppose that X has the Poisson distribution P (θ) with an unknown θ > 0. Show that (6.29) reduces to cX 2 −1

x=c1

2

θ0x−1 e−θ0 X θci −1 e−θ0 + = 1 − α, (1 − γi ) 0 (x − 1)! (ci − 1)! +1 i=1

provided that c1 > 1.

6.6. Exercises

459

34. Let X be a random variable from the geometric distribution G(p). Find a UMPU test of size α for H0 : p = p0 versus H1 : p 6= p0 . 35. In Exercise 33 of §2.6, derive a UMPU test of size α ∈ (0, 1) for testing H0 : p ≤ 12 versus H1 : p > 12 . 36. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ and σ 2 . (a) Show how the power of the one-sample t-test depends on a noncentral t-distribution. (b) Show that the power of the one-sample t-test is an increasing function of (µ − µ0 )/σ for testing H0 : µ ≤ µ0 versus H1 : µ > µ0 , and of |µ − µ0 |/σ for testing H0 : µ = µ0 versus H1 : µ 6= µ0 . 37. Let X1 , ..., Xn be i.i.d. from the gamma distribution Γ(θ, γ) with unknown θ and γ. (a) For testing H0 : θ ≤ θ0 versus H1 : θ > θ0 and H0 : θ = θ0 versus H1 : θ 6= θ0 , show that Qnthere exist UMPU tests whose rejection ¯ regions are based on V = i=1 (Xi /X). (b) For testing H0 : γP≤ γ0 versus H Q1n : γ > γ0 , show that a UMPU n test rejects H0 when i=1 Xi > C( i=1 Xi ) for some function C.

38. Let X1 and X2 be independently distributed as the Poisson distributions P (λ1 ) and P (λ2 ), respectively. (a) Find a UMPU test of size α for testing H0 : λ1 ≥ λ2 versus H1 : λ1 < λ2 . (b) Calculate the power of the UMPU test in (a) when α = 0.1, (λ1 , λ2 ) = (0.1, 0.2), (1,2), (10,20), and (0.1,0.4). 39. Consider the binomial problem in Example 6.11. (a) Prove the claim about P (Y = y|U = u). (b) Find a UMPU test of size α for testing H0 : p1 ≥ p2 versus H1 : p1 < p2 . (c) Repeat (b) for H0 : p1 = p2 versus H1 : p1 6= p2 . 40. Let X1 and X2 be independently distributed as the negative binomial distributions N B(p1 , n1 ) and N B(p2 , n2 ), respectively, where ni ’s are known and pi ’s are unknown. (a) Show that there exists a UMPU test of size α for testing H0 : p1 ≤ p2 versus H1 : p1 > p2 . (b) Determine the conditional distribution PY |U =u in Theorem 6.4 when n1 = n2 = 1. 41. Let (X0 , X1 , X2 ) be a random vector having a multinomial distribution (Example 2.7) with k = 2, p0 = 1 − p1 − p2 , and unknown p1 ∈ (0, 1) and p2 ∈ (0, 1). Derive a UMPU test of size α for testing H0 : p0 = p2 , p1 = 2p(1 − p), p2 = (1 − p)2 versus H1 : H0 is not true, where p ∈ (0, 1) is unknown.

460

6. Hypothesis Tests

42. Consider Example 6.12. (a) Show that A and B are independent if and only if log pp11 = 22 p21 + log . log pp12 p22 22 (b) Derive a UMPU test of size α for testing H0 : P (A) = P (B) versus H1 : P (A) 6= P (B). 43. Let X1 and X2 be independently distributed according to p.d.f.’s given by (6.10) with ξ, η, θ, Y , and h replaced by ξi , ηi , θi , Yi , and hi , i = 1, 2, respectively. Show that there exists a UMPU test of size α for testing (a) H0 : η2 (θ2 ) − η1 (θ1 ) ≤ η0 versus H1 : η2 (θ2 ) − η1 (θ1 ) > η0 ; (b) H0 : η2 (θ2 ) + η1 (θ1 ) ≤ η0 versus H1 : η2 (θ2 ) + η1 (θ1 ) > η0 . 44. Let Xj , j = 1, 2, 3, be independent from the Poisson distributions P (λj ), j = 1, 2, 3, respectively. Show that there exists a UMPU test of size α for testing H0 : λ1 λ2 ≤ λ23 versus H1 : λ1 λ2 > λ23 . 45. Let Xij , i = 1, 2, j = 1, 2, be independent from the Poisson distributions P (λi pij ), where λi > 0, 0 < pij < 1, and pi1 + pi2 = 0. Derive a UMPU test of size α for testing H0 : p11 ≤ p21 versus H1 : p11 > p21 . 46. Let Xij be independent random variables satisfying P (Xij = 0) = θi , P (Xij = k) = (1 − θi )(1 − pi )j−1 pi , k = 1, 2, ..., where 0 < θi < 1 and 0 < pi < 1, j = 1, ..., ni and i = 1, 2. Derive a UMPU test of size α for testing H0 : p1 ≤ p2 versus H1 : p1 > p2 . 47. Let X11 , ..., X1n1 and X21 , ..., X2n2 be two independent samples i.i.d. from the gamma distributions Γ(θ1 , γ1 ) and Γ(θ2 , γ2 ), respectively. (a) Assume that θ1 and θ2 are known. For testing H0 : γ1 ≤ γ2 versus H1 : γ1 > γ2 and H0 : γ1 = γ2 versus H1 : γ1 6= γ2 , show that there exist UMPU tests and that the rejection regions can be determined by using beta distributions. (b) If θi ’s are unknown in (a), show that there exist UMPU tests and describe their general forms. (c) Assume that γ1 = γ2 (unknown). For testing H0 : θ1 ≤ θ2 versus H1 : θ1 > θ2 and H0 : θ1 = θ2 versus H1 : θ1 6= θ2 , show that there exist UMPU tests and describe their general forms. 48. Let N be a random variable with the following discrete p.d.f.: P (N = n) = C(λ)a(n)λn I{0,1,2,...} (n), where λ > 0 is unknown and a and C are known functions. Suppose that given N = n, X1 , ..., Xn are i.i.d. from the p.d.f. given in (6.10). Show that, based on (N, X1 , ..., XN ), there exists a UMPU test of size α for H0 : η(θ) ≤ η0 versus H1 : η(θ) > η0 .

6.6. Exercises

461

49. Let Xi1 , ..., Xini , i = 1, 2, be two independent samples i.i.d. from N (µi , σ 2 ), respectively, ni ≥ 2. Show that a UMPU test of size α for H0 : µ1 = µ2 versus H1 : µ1 6= µ2 rejects H0 when |t(X)| > tn1 +n2 −1,α/2 , where t(X) is given by (6.37) and tn1 +n2 −1,α is the (1 − α)th quantile of the t-distribution tn1 +n2 −1 . Derive the power function of this test. 50. In the two-sample problem discussed in §6.2.3, show that when n1 = n2 , a UMPU test of size α for testing H0 : σ22 = ∆0 σ12 versus H1 : σ22 6= ∆0 σ12 rejects H0 when 2 1−c S2 ∆0 S12 , > , max ∆0 S12 S22 c Rc where 0 f(n1 −1)/2,(n1 −1)/2 (v)dv = α/2 and fa,b is the p.d.f. of the beta distribution B(a, b). 51. Suppose that Xi = β0 + β1 ti + εi , where ti ’s are fixed constants that are not all the same, εi ’s are i.i.d. from N (0, σ 2 ), and β0 , β1 , and σ 2 are unknown parameters. Derive a UMPU test of size α for testing (a) H0 : β0 ≤ θ0 versus H1 : β0 > θ0 ; (b) H0 : β0 = θ0 versus H1 : β0 6= θ0 ; (c) H0 : β1 ≤ θ0 versus H1 : β1 > θ0 ; (d) H0 : β1 = θ0 versus H1 : β1 6= θ0 . 52. In the previous exercise, derive the power function in each of (a)-(d) in terms of a noncentral t-distribution. 53. Consider the normal linear model in §6.2.3 (i.e., model (3.25) with ε = Nn (0, σ 2 In )). For testing H0 : σ 2 ≤ σ02 versus H1 : σ 2 > σ02 and H0 : σ 2 = σ02 versus H1 : σ 2 6= σ02 , show that UMPU tests of size α are functions of SSR and their rejection regions can be determined using chi-square distributions. 54. In the problem of testing for independence in the bivariate normal family, show that (a) the p.d.f. in (6.44) is of the form (6.23) and identify ϕ; (b) the sample correlation coefficient R is independent of U when ρ = 0; (c) R is linear in Y , and V in (6.45) has the t-distribution tn−2 when ρ = 0. 55. Let X1 , ..., with the p.d.f. in (6.44) and n be i.i.d. bivariate normal PX n ¯ j )2 and S12 = Pn (Xi1 − X ¯ 1 )(Xi2 − X ¯ 2 ). let Sj2 = i=1 (Xij − X i=1 (a) Show that a UMPU test for testing H0 : σ2 /σ1 = ∆0 versus H1 : σ2 /σ1 6= ∆0 rejects H0 when

462

6. Hypothesis Tests q 2 2 2 > c. R = |∆20 S12 − S22 | (∆0 S1 + S22 )2 − 4∆20 S12

(b) Find the p.d.f. of R in (a) when σ2 /σ1 = ∆0 . (c) Assume that σ1 = σ2 . Show that a UMPU test for H0 : µ1 = µ2 versus H1 : µ1 6= µ2 rejects H0 when q 2 ¯2 − X ¯1| V = |X S1 + S22 − 2S12 > c. (d) Find the p.d.f. of V in (c) when µ1 = µ2 .

56. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors having the bivariate normal distribution with EX1 = EY1 = 0, Var(X1 ) = σx2 , Var(Y1 ) = σy2 , and Cov(X1 , Y1 ) = ρσx σy , where σx > 0, σy > 0, and ρ ∈ [0, 1) are unknown. Derive the form and exact distribution of a UMPU test of size α for testing H0 : ρ = 0 versus H1 : ρ > 0. 57. Let X1 , ..., Xn be i.i.d. from the P exponential distribution E(a, θ) with n unknown a and θ. Let V = 2 i=1 (Xi − X(1) ), where X(1) is the smallest order statistic. (a) For testing H0 : θ = 1 versus H1 : θ 6= 1, show that a UMPU test of size α rejects H0 when V < c1 or V > c2 , where ci ’s are determined by Z c2 Z c2 f2n−2 (v)dv = f2n (v)dv = 1 − α, c1

c1

and fm (v) is the p.d.f. of the chi-square distribution χ2m . (b) For testing H0 : a = 0 versus H1 : a 6= 0, show that a UMPU test of size α rejects H0 when X(1) < 0 or 2nX(1) /V > c, where c is determined by Z c (n − 1) (1 + v)−n dv = 1 − α. 0

58. Let X1 , ..., Xn be i.i.d. random variables from the uniform distribution U (θ, ϑ), −∞ < θ < ϑ < ∞. (a) Show that the conditional distribution of X(1) given X(n) = x is the distribution of the minimum of a sample of size n − 1 from the uniform distribution U (θ, x). (b) Find a UMPU test of size α for testing H0 : θ ≤ 0 versus H1 : θ > 0. 59. Let X1 , ..., Xn be independent random variables having the binomial distributions Bi(pi , ki ), i = 1, ..., n, respectively, where pi = ea+bti /(1 + ea+bti ), (a, b) ∈ R2 is unknown, and ti ’s are known covariate values that are not all the same. Derive the UMPU test of size α for testing (a) H0 : a ≥ 0 versus H1 : a < 0; (b) H0 : b ≥ 0 versus H1 : b < 0.

6.6. Exercises

463

60. In the previous exercise, derive approximations to the UMPU tests by considering the limiting distributions of the test statistics. 61. Let X = {x ∈ Rn : all components of x are nonzero} and G be the group of transformations g(x) = (cx1 , ..., cxn ), c > 0. Show that a maximal invariant under G is (sgn(xn ), x1 /xn , ..., xn−1 /xn ), where sgn(x) is 1 or −1 as x is positive or negative. 62. Let X1 , ..., Xn be i.i.d. with a Lebesgue p.d.f. σ −1 f (x/σ) and fi , i = 0, 1, be two known Lebesgue p.d.f.’s on R that are either 0 for x < 0 or symmetric about 0. Consider H0 : f = f0 versus H1 : f = f1 and G = {gr : r > 0} with gr (x) = rx. (a) Show that a UMPI test rejects H0 when R ∞ n−1 v f1 (vX1 ) · · · f1 (vXn )dv R0∞ > c. n−1 f0 (vX1 ) · · · f0 (vXn )dv 0 v

(b) Show that if f0 = N (0, 1) =Pe−|x|/2, then the UMPI Pnand f21 (x) n 1/2 test in (a) rejects H0 when ( i=1 Xi ) / i=1 |Xi | > c. (c) Show that if f0 (x) = I(0,1) (x) and f1 (x) = 2xI(0,1) (x), then the Q UMPI test in (a) rejects H0 when X(n) /( ni=1 Xi )1/n < c. (d) Find the value of c in part (c) when the UMPI test is of size α.

63. Consider the location-scale family problem (with unknown parameters µ and σ) in Example 6.13. (a) Show that W is maximal invariant under the given G. (b) Show that Proposition 6.2 applies and find the form of the functional θ(fi,µ,σ ). (c) Derive the p.d.f. of W (X) under Hi , i = 0, 1. (d) Obtain a UMPI test. 64. In Example 6.13, find the rejection region of the UMPI test when X1 , ..., Xn are i.i.d. and (a) f0,µ,σ is N (µ, σ 2 ) and f1,µ,σ is the p.d.f. of the uniform distribution U (µ − 21 σ, µ + 12 σ); (b) f0,µ,σ is N (µ, σ 2 ) and f1,µ,σ is the p.d.f. of the exponential distribution E(µ, σ); (c) f0,µ,σ is the p.d.f. of U (µ − 21 σ, µ + 12 σ) and f1,µ,σ is the p.d.f. of E(µ, σ); (d) f0,µ is N (µ, 1) and f1,µ (x) = exp{−ex−µ + x − µ}. 65. Prove the claims in Example 6.15. 66. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ and σ 2 . Consider the problem of testing H0 : µ = 0 versus H1 : µ 6= 0 and the group of transformations gc (Xi ) = cXi , c 6= 0.

464

6. Hypothesis Tests

(a) Show that the testing problem is invariant under G. (b) Show that the one-sample two-sided t-test in §6.2.3 is a UMPI test. 67. Prove the claims in Example 6.16. 68. Consider Example 6.16 with H0 and H1 replaced by H0 : µ1 = µ2 and H1 : µ1 6= µ2 , and with G changed to {gc1 ,c2 ,r : c1 = c2 ∈ R, r 6= 0}. (a) Show that the testing problem is invariant under G. (b) Show that the two-sample two-sided t-test in §6.2.3 is a UMPI test. 69. Show that the UMPU tests in Exercise 37(a) and Exercise 47(a) are also UMPI tests under G = {gr : r > 0} with gr (x) = rx. 70. In Example 6.17, show that t(X) has the noncentral t-distribution √ tn−1 ( nθ); the family {fθ (t) : θ ∈ R} has monotone likelihood ratio in t; and that for testing H0 : θ = θ0 versus H1 : θ 6= θ0 , a test that is UMP among all level α unbiased tests based on t(X) rejects H0 when t(X) < c1 or t(X) > c2 . (Hint: consider Exercise 26.) 71. Let X1 and X2 be independently distributed as the exponential distributions E(0, θi ), i = 1, 2, respectively. Define θ = θ1 /θ2 . (a) For testing H0 : θ ≤ 1 versus θ > 1, show that the problem is invariant under the group of transformations gc (x1 , x2 ) = (cx1 , cx2 ), c > 0, and that a UMPI test of size α rejects H0 when X2 /X1 > (1 − α)/α. (b) For testing H0 : θ = 1 versus θ 6= 1, show that the problem is invariant under the group of transformations in (a) and g(x1 , x2 ) = (x2 , x1 ), and that a UMPI test of size α rejects H0 when X1 /X2 > (2 − α)/α and X2 /X1 > (2 − α)/α. 72. Let X1 , ..., Xm and Y1 , ..., Yn be two independent samples i.i.d. from the exponential distributions E(a1 , θ1 ) and E(a2 , θ2 ), respectively. Let gr,c,d(x, y) = (rx1 + c, ..., rxm + c, ry1 + d, ..., ryn + d) and let G = {gr,c,d : r > 0, c ∈ R, d ∈ R}. (a) Show that a UMPI test of size P α for testing H0 : θ1P /θ2 ≥ ∆0 versus H1 : θ1 /θ2 < ∆0 rejects H0 when ni=1 (Yi −Y(1) ) > c m i=1 (Xi −X(1) ) for some constant c. (b) Find the value of c in (a). (c) Show that the UMPI test in (a) is also a UMPU test. 73. Let M (U ) be given by (6.51) and W = M (U )(n − r)/s. (a) Show that W has the noncentral F-distribution Fs,n−r (θ). (b) Show that fθ1 (w)/f0 (w) is an increasing function of w for any given θ1 > 0.

465

6.6. Exercises

74. Consider normal linear model (6.38). Show that (a) the UMPI test derived in §6.3.2 for testing (6.49) is the same as the UMPU test for (6.40) given in §6.2.3 when s = 1 and θ0 = 0; (b) the test with the rejection region W > Fs,n−r,α is a UMPI test of size α for testing H0 : Lβ = θ0 versus H1 : Lβ 6= θ0 , where W is given by (6.52), θ0 is a fixed constant, L is the same as that in (6.49), and Fs,n−r,α is the (1 − α)th quantile of the F-distribution Fs,n−r . 75. In Examples 6.18-6.19, (a) prove the claim in Example 6.19; (b) derive the distribution of W by applying Cochran’s theorem. 76. (Two-way additive model). Assume that Xij ’s are independent and Xij = N (µij , σ 2 ),

i = 1, ..., a, j = 1, ..., b, Pa Pb where µij = µ + αi + βj and i=1 αi = j=1 βj = 0. Derive the forms of the UMPI tests in §6.3.2 for testing (6.54) and (6.55). 77. (Three-way additive model). Assume that Xijk ’s are independent and Xijk = N (µijk , σ 2 ),

i = 1, ..., a, j = 1, ..., b, k = 1, ..., c, Pa Pb Pc where µijk = µ+αi +βj +γk and i=1 αi = j=1 βj = k=1 γk = 0. Derive the UMPI test based on the W in (6.52) for testing H0 : αi = 0 for all i versus H1 : αi 6= 0 for some i. 78. Let X1 , ..., Xm and Y1 , ..., Yn be independently normally distributed with a common unknown variance σ 2 and means E(Xi ) = µx + βx (ui − u ¯),

E(Yj ) = µy + βy (vj − v¯), Pm where ui ’s and vj ’s are known constants, u ¯ = m−1 i=1 ui , v¯ = P n n−1 i=1 vi , and µx , µy , βx , and βy are unknown. Derive the UMPI test based on the W in (6.52) for testing (a) H0 : βx = βy versus H1 : βx 6= βy ; (b) H0 : βx = βy and µx = µy versus H1 : βx 6= βy or µx 6= µy . 79. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. from a bivariate normal distribution with unknown means, variances, and correlation coefficient ρ. (a) Show that the problem of testing H0 : ρ ≤ ρ0 versus H1 : ρ > ρ0 is invariant under G containing transformations rXi + c, sYi + d, i = 1, ..., n, where r > 0, s > 0, c ∈ R, and d ∈ R. Show that a UMPI test rejects H0 when R > c, where R is the sample

Stephen Fienberg

Ingram Olkin

Springer Texts in Statistics Alfred: Elements of Statistics for the Life and Social Sciences Berger: An Introduction to Probability and Stochastic Processes Bilodeau and Brenner: Theory of Multivariate Statistics Blom: Probability and Statistics: Theory and Applications Brockwell and Davis: Introduction to Times Series and Forecasting, Second Edition Chow and Teicher: Probability Theory: Independence, Interchangeability, Martingales, Third Edition Christensen: Advanced Linear Modeling: Multivariate, Time Series, and Spatial Data: Nonparametric Regression and Response Surface Maximization, Second Edition Christensen: Log-Linear Models and Logistic Regression, Second Edition Christensen: Plane Answers to Complex Questions: The Theory of Linear Models, Third Edition Creighton: A First Course in Probability Models and Statistical Inference Davis: Statistical Methods for the Analysis of Repeated Measurements Dean and Voss: Design and Analysis of Experiments du Toit, Steyn, and Stumpf: Graphical Exploratory Data Analysis Durrett: Essentials of Stochastic Processes Edwards: Introduction to Graphical Modelling, Second Edition Finkelstein and Levin: Statistics for Lawyers Flury: A First Course in Multivariate Statistics Jobson: Applied Multivariate Data Analysis, Volume I: Regression and Experimental Design Jobson: Applied Multivariate Data Analysis, Volume II: Categorical and Multivariate Methods Kalbfleisch: Probability and Statistical Inference, Volume I: Probability, Second Edition Kalbfleisch: Probability and Statistical Inference, Volume II: Statistical Inference, Second Edition Karr: Probability Keyfitz: Applied Mathematical Demography, Second Edition Kiefer: Introduction to Statistical Inference Kokoska and Nevison: Statistical Tables and Formulae Kulkarni: Modeling, Analysis, Design, and Control of Stochastic Systems Lange: Applied Probability Lehmann: Elements of Large-Sample Theory Lehmann: Testing Statistical Hypotheses, Second Edition Lehmann and Casella: Theory of Point Estimation, Second Edition Lindman: Analysis of Variance in Experimental Design Lindsey: Applying Generalized Linear Models (continued after index)

Jun Shao

Mathematical Statistics Second Edition

Jun Shao Department of Statistics University of Wisconsin, Madison Madison, WI 53706-1685 USA [email protected]

Editorial Board George Casella Department of Statistics University of Florida Gainesville, FL 32611-8545 USA

Stephen Fienberg Department of Statistics Carnegie Mellon University Pittsburgh, PA 15213-3890 USA

Ingram Olkin Department of Statistics Stanford University Stanford, CA 94305 USA

With 7 figures.

Library of Congress Cataloging-in-Publication Data Shao, Jun. Mathematical statistics / Jun Shao.—2nd ed. p. cm.— (Springer texts in statistics) Includes bibliographical references and index. ISBN 0-387-95382-5 (alk. paper) 1. Mathematical statistics. I. Title. II. Series. QA276.S458 2003 519.5—dc21 2003045446 ISBN 0-387-95382-5

Printed on acid-free paper.

ISBN-13 978-0-387-95382-3 © 2003 Springer Science+Business Media, LLC. All rights reserved. This work may not be translated or copied in whole or in part without the written permission of the publisher (Springer Science+Business Media, LLC., 233 Spring St., New York, N.Y., 10013, USA), except for brief excerpts in connection with reviews or scholarly analysis. Use in connection with any form of information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed is forbidden. The use in this publication of trade names, trademarks, service marks, and similar terms, even if they are not identified as such, is not to be taken as an expression of opinion as to whether or not they are subject to proprietary rights. Printed in the United States of America. th

9 8 7 6 5 4 (corrected printing as of 4 printing, 2007) springer.com

To Guang, Jason, and Annie

Preface to the First Edition This book is intended for a course entitled Mathematical Statistics offered at the Department of Statistics, University of Wisconsin-Madison. This course, taught in a mathematically rigorous fashion, covers essential materials in statistical theory that a first or second year graduate student typically needs to learn as preparation for work on a Ph.D. degree in statistics. The course is designed for two 15-week semesters, with three lecture hours and two discussion hours in each week. Students in this course are assumed to have a good knowledge of advanced calculus. A course in real analysis or measure theory prior to this course is often recommended. Chapter 1 provides a quick overview of important concepts and results in measure-theoretic probability theory that are used as tools in mathematical statistics. Chapter 2 introduces some fundamental concepts in statistics, including statistical models, the principle of sufficiency in data reduction, and two statistical approaches adopted throughout the book: statistical decision theory and statistical inference. Each of Chapters 3 through 7 provides a detailed study of an important topic in statistical decision theory and inference: Chapter 3 introduces the theory of unbiased estimation; Chapter 4 studies theory and methods in point estimation under parametric models; Chapter 5 covers point estimation in nonparametric settings; Chapter 6 focuses on hypothesis testing; and Chapter 7 discusses interval estimation and confidence sets. The classical frequentist approach is adopted in this book, although the Bayesian approach is also introduced (§2.3.2, §4.1, §6.4.4, and §7.1.3). Asymptotic (large sample) theory, a crucial part of statistical inference, is studied throughout the book, rather than in a separate chapter. About 85% of the book covers classical results in statistical theory that are typically found in textbooks of a similar level. These materials are in the Statistics Department’s Ph.D. qualifying examination syllabus. This part of the book is influenced by several standard textbooks, such as Casella and vii

viii

Preface to the First Edition

Berger (1990), Ferguson (1967), Lehmann (1983, 1986), and Rohatgi (1976). The other 15% of the book covers some topics in modern statistical theory that have been developed in recent years, including robustness of the least squares estimators, Markov chain Monte Carlo, generalized linear models, quasi-likelihoods, empirical likelihoods, statistical functionals, generalized estimation equations, the jackknife, and the bootstrap. In addition to the presentation of fruitful ideas and results, this book emphasizes the use of important tools in establishing theoretical results. Thus, most proofs of theorems, propositions, and lemmas are provided or left as exercises. Some proofs of theorems are omitted (especially in Chapter 1), because the proofs are lengthy or beyond the scope of the book (references are always provided). Each chapter contains a number of examples. Some of them are designed as materials covered in the discussion section of this course, which is typically taught by a teaching assistant (a senior graduate student). The exercises in each chapter form an important part of the book. They provide not only practice problems for students, but also many additional results as complementary materials to the main text. The book is essentially based on (1) my class notes taken in 1983-84 when I was a student in this course, (2) the notes I used when I was a teaching assistant for this course in 1984-85, and (3) the lecture notes I prepared during 1997-98 as the instructor of this course. I would like to express my thanks to Dennis Cox, who taught this course when I was a student and a teaching assistant, and undoubtedly has influenced my teaching style and textbook for this course. I am also very grateful to students in my class who provided helpful comments; to Mr. Yonghee Lee, who helped me to prepare all the figures in this book; to the Springer-Verlag production and copy editors, who helped to improve the presentation; and to my family members, who provided support during the writing of this book. Madison, Wisconsin January 1999

Jun Shao

Preface to the Second Edition In addition to correcting typos and errors and making a better presentation, the main effort in preparing this new edition is adding some new material to Chapter 1 (Probability Theory) and a number of new exercises to each chapter. Furthermore, two new sections are created to introduce semiparametric models and methods (§5.1.4) and to study the asymptotic accuracy of confidence sets (§7.3.4). The structure of the book remains the same. In Chapter 1 of the new edition, moment generating and characteristic functions are treated in more detail and a proof of the uniqueness theorem is provided; some useful moment inequalities are introduced; discussions on conditional independence, Markov chains, and martingales are added, as a continuation of the discussion of conditional expectations; the concepts of weak convergence and tightness are introduced; proofs to some key results in asymptotic theory, such as the dominated convergence theorem and monotone convergence theorem, the L´evy-Cram´er continuity theorem, the strong and weak laws of large numbers, and Lindeberg’s central limit theorem, are included; and a new section (§1.5.6) is created to introduce Edgeworth and Cornish-Fisher expansions. As a result, Chapter 1 of the new edition is self-contained for important concepts, results, and proofs in probability theory with emphasis in statistical applications. Since the original book was published in 1999, I have been using it as a textbook for a two-semester course in mathematical statistics. Exercise problems accumulated during my teaching are added to this new edition. Some exercises that are too trivial have been removed. In the original book, indices on definitions, examples, theorems, propositions, corollaries, and lemmas are included in the subject index. In the new edition, they are in a separate index given in the end of the book (prior to the author index). A list of notation and a list of abbreviations, which are appendices of the original book, are given after the references.

ix

x

Preface to the Second Edition

The most significant change in notation is the notation for a vector. In the text of the new edition, a k-dimensional vector is denoted by c = (c1 , ..., ck ), whether it is treated as a column or a row vector (which is not important if matrix algebra is not considered). When matrix algebra is involved, any vector c is treated as a k × 1 matrix (a column vector) and its transpose cτ is treated as a 1 × k matrix (a row vector). Thus, for c = (c1 , ..., ck ), cτ c = c21 + · · · + c2k and ccτ is the k × k matrix whose (i, j)th element is ci cj . I would like to thank reviewers of this book for their constructive comments, the Springer-Verlag production and copy editors, students in my classes, and two teaching assistants, Mr. Bin Cheng and Dr. Hansheng Wang, who provided help in preparing the new edition. Any remaining errors are of course my own responsibility, and a correction of them may be found on my web page http://www.stat.wisc.edu/˜ shao. Madison, Wisconsin April, 2003

Jun Shao

Contents Preface to the First Edition

vii

Preface to the Second Edition

ix

Chapter 1. Probability Theory

1

1.1 Probability Spaces and Random Elements . . . . . . . . . . .

1

1.1.1 σ-fields and measures . . . . . . . . . . . . . . . . . .

1

1.1.2 Measurable functions and distributions . . . . . . . .

6

1.2 Integration and Differentiation . . . . . . . . . . . . . . . . .

10

1.2.1 Integration . . . . . . . . . . . . . . . . . . . . . . . .

10

1.2.2 Radon-Nikodym derivative . . . . . . . . . . . . . . .

15

1.3 Distributions and Their Characteristics . . . . . . . . . . . .

17

1.3.1 Distributions and probability densities

. . . . . . . .

17

1.3.2 Moments and moment inequalities . . . . . . . . . . .

28

1.3.3 Moment generating and characteristic functions . . .

32

1.4 Conditional Expectations . . . . . . . . . . . . . . . . . . . .

37

1.4.1 Conditional expectations . . . . . . . . . . . . . . . .

37

1.4.2 Independence

. . . . . . . . . . . . . . . . . . . . . .

41

1.4.3 Conditional distributions . . . . . . . . . . . . . . . .

43

1.4.4 Markov chains and martingales . . . . . . . . . . . . .

45

1.5 Asymptotic Theory . . . . . . . . . . . . . . . . . . . . . . .

49

1.5.1 Convergence modes and stochastic orders . . . . . . .

50

1.5.2 Weak convergence . . . . . . . . . . . . . . . . . . . .

56

1.5.3 Convergence of transformations . . . . . . . . . . . .

59

1.5.4 The law of large numbers . . . . . . . . . . . . . . . .

62

1.5.5 The central limit theorem . . . . . . . . . . . . . . . .

67

xi

xii

Contents

1.5.6 Edgeworth and Cornish-Fisher expansions . . . . . .

70

1.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

74

Chapter 2. Fundamentals of Statistics

91

2.1 Populations, Samples, and Models . . . . . . . . . . . . . . .

91

2.1.1 Populations and samples . . . . . . . . . . . . . . . .

91

2.1.2 Parametric and nonparametric models . . . . . . . . .

94

2.1.3 Exponential and location-scale families . . . . . . . .

96

2.2 Statistics, Sufficiency, and Completeness . . . . . . . . . . . . 100 2.2.1 Statistics and their distributions . . . . . . . . . . . . 100 2.2.2 Sufficiency and minimal sufficiency

. . . . . . . . . . 103

2.2.3 Complete statistics . . . . . . . . . . . . . . . . . . . 109 2.3 Statistical Decision Theory . . . . . . . . . . . . . . . . . . . 113 2.3.1 Decision rules, loss functions, and risks . . . . . . . . 113 2.3.2 Admissibility and optimality . . . . . . . . . . . . . . 116 2.4 Statistical Inference . . . . . . . . . . . . . . . . . . . . . . . 122 2.4.1 Point estimators . . . . . . . . . . . . . . . . . . . . . 122 2.4.2 Hypothesis tests . . . . . . . . . . . . . . . . . . . . . 125 2.4.3 Confidence sets

. . . . . . . . . . . . . . . . . . . . . 129

2.5 Asymptotic Criteria and Inference . . . . . . . . . . . . . . . 131 2.5.1 Consistency . . . . . . . . . . . . . . . . . . . . . . . 132 2.5.2 Asymptotic bias, variance, and mse . . . . . . . . . . 135 2.5.3 Asymptotic inference . . . . . . . . . . . . . . . . . . 139 2.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142 Chapter 3. Unbiased Estimation

161

3.1 The UMVUE . . . . . . . . . . . . . . . . . . . . . . . . . . . 161 3.1.1 Sufficient and complete statistics . . . . . . . . . . . . 162 3.1.2 A necessary and sufficient condition . . . . . . . . . . 166 3.1.3 Information inequality . . . . . . . . . . . . . . . . . . 169 3.1.4 Asymptotic properties of UMVUE’s . . . . . . . . . . 172 3.2 U-Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 3.2.1 Some examples . . . . . . . . . . . . . . . . . . . . . . 174 3.2.2 Variances of U-statistics . . . . . . . . . . . . . . . . . 176 3.2.3 The projection method . . . . . . . . . . . . . . . . . 178

Contents

xiii

3.3 The LSE in Linear Models . . . . . . . . . . . . . . . . . . . 182 3.3.1 The LSE and estimability . . . . . . . . . . . . . . . . 182 3.3.2 The UMVUE and BLUE . . . . . . . . . . . . . . . . 186 3.3.3 Robustness of LSE’s . . . . . . . . . . . . . . . . . . . 189 3.3.4 Asymptotic properties of LSE’s . . . . . . . . . . . . 193 3.4 Unbiased Estimators in Survey Problems . . . . . . . . . . . 195 3.4.1 UMVUE’s of population totals . . . . . . . . . . . . . 195 3.4.2 Horvitz-Thompson estimators . . . . . . . . . . . . . 199 3.5 Asymptotically Unbiased Estimators . . . . . . . . . . . . . . 204 3.5.1 Functions of unbiased estimators . . . . . . . . . . . . 204 3.5.2 The method of moments . . . . . . . . . . . . . . . . 207 3.5.3 V-statistics . . . . . . . . . . . . . . . . . . . . . . . . 210 3.5.4 The weighted LSE . . . . . . . . . . . . . . . . . . . . 213 3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217 Chapter 4. Estimation in Parametric Models

231

4.1 Bayes Decisions and Estimators . . . . . . . . . . . . . . . . 231 4.1.1 Bayes actions . . . . . . . . . . . . . . . . . . . . . . . 231 4.1.2 Empirical and hierarchical Bayes methods . . . . . . 236 4.1.3 Bayes rules and estimators . . . . . . . . . . . . . . . 239 4.1.4 Markov chain Monte Carlo . . . . . . . . . . . . . . . 245 4.2 Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251 4.2.1 One-parameter location families . . . . . . . . . . . . 251 4.2.2 One-parameter scale families . . . . . . . . . . . . . . 255 4.2.3 General location-scale families . . . . . . . . . . . . . 257 4.3 Minimaxity and Admissibility . . . . . . . . . . . . . . . . . 261 4.3.1 Estimators with constant risks . . . . . . . . . . . . . 261 4.3.2 Results in one-parameter exponential families . . . . 265 4.3.3 Simultaneous estimation and shrinkage estimators . . 267 4.4 The Method of Maximum Likelihood . . . . . . . . . . . . . 273 4.4.1 The likelihood function and MLE’s . . . . . . . . . . 273 4.4.2 MLE’s in generalized linear models . . . . . . . . . . 279 4.4.3 Quasi-likelihoods and conditional likelihoods . . . . . 283 4.5 Asymptotically Efficient Estimation . . . . . . . . . . . . . . 286 4.5.1 Asymptotic optimality . . . . . . . . . . . . . . . . . 286

xiv

Contents

4.5.2 Asymptotic efficiency of MLE’s and RLE’s . . . . . . 290 4.5.3 Other asymptotically efficient estimators . . . . . . . 295 4.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 299 Chapter 5. Estimation in Nonparametric Models 5.1 Distribution Estimators . . . . . . . . . . . . . . . 5.1.1 Empirical c.d.f.’s in i.i.d. cases . . . . . . . 5.1.2 Empirical likelihoods . . . . . . . . . . . . 5.1.3 Density estimation . . . . . . . . . . . . . . 5.1.4 Semi-parametric methods . . . . . . . . . . 5.2 Statistical Functionals . . . . . . . . . . . . . . . . 5.2.1 Differentiability and asymptotic normality 5.3

5.4

5.5

5.6

. . . . . . .

. . . . . . .

. . . . . . .

319 319 320 323 330 333 338 338

5.2.2 L-, M-, and R-estimators and rank statistics . . . Linear Functions of Order Statistics . . . . . . . . . . . . 5.3.1 Sample quantiles . . . . . . . . . . . . . . . . . . . 5.3.2 Robustness and efficiency . . . . . . . . . . . . . . 5.3.3 L-estimators in linear models . . . . . . . . . . . . Generalized Estimating Equations . . . . . . . . . . . . . 5.4.1 The GEE method and its relationship with others 5.4.2 Consistency of GEE estimators . . . . . . . . . . . 5.4.3 Asymptotic normality of GEE estimators . . . . . Variance Estimation . . . . . . . . . . . . . . . . . . . . . 5.5.1 The substitution method . . . . . . . . . . . . . . 5.5.2 The jackknife . . . . . . . . . . . . . . . . . . . . . 5.5.3 The bootstrap . . . . . . . . . . . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . .

. . . . . . . . . . . . . .

343 351 351 355 358 359 360 363 367 371 372 376 380 383

. . . . . . . .

393 393 394 397 401 404 404 406 410

Chapter 6. Hypothesis Tests 6.1 UMP Tests . . . . . . . . . . . . . . . . . . . 6.1.1 The Neyman-Pearson lemma . . . . . 6.1.2 Monotone likelihood ratio . . . . . . . 6.1.3 UMP tests for two-sided hypotheses . 6.2 UMP Unbiased Tests . . . . . . . . . . . . . 6.2.1 Unbiasedness, similarity, and Neyman 6.2.2 UMPU tests in exponential families . 6.2.3 UMPU tests in normal families . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . structure . . . . . . . . . . . .

. . . . . . . .

. . . . . . . .

xv

Contents

6.3 UMP Invariant Tests . . . . . . . . . . . . . . . . . . . . . . 417 6.3.1 Invariance and UMPI tests . . . . . . . . . . . . . . . 417 6.3.2 UMPI tests in normal linear models . . . . . . . . . . 422 6.4 Tests in Parametric Models . . . . . . . . . . . . . . . . . . . 428 6.4.1 Likelihood ratio tests . . . . . . . . . . . . . . . . . . 428 6.4.2 Asymptotic tests based on likelihoods . . . . . . . . . 431 6.4.3 χ2 -tests . . . . . . . . . . . . . . . . . . . . . . . . . . 436 6.4.4 Bayes tests . . . . . . . . . . . . . . . . . . . . . . . . 440 6.5 Tests in Nonparametric Models . . . . . . . . . . . . . . . . . 442 6.5.1 Sign, permutation, and rank tests . . . . . . . . . . . 442 6.5.2 Kolmogorov-Smirnov and Cram´er-von Mises tests . . 446 6.5.3 Empirical likelihood ratio tests . . . . . . . . . . . . . 449 6.5.4 Asymptotic tests . . . . . . . . . . . . . . . . . . . . . 452 6.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 454 Chapter 7. Confidence Sets

471

7.1 Construction of Confidence Sets . . . . . . . . . . . . . . . . 471 7.1.1 Pivotal quantities . . . . . . . . . . . . . . . . . . . . 471 7.1.2 Inverting acceptance regions of tests . . . . . . . . . . 477 7.1.3 The Bayesian approach . . . . . . . . . . . . . . . . . 480 7.1.4 Prediction sets . . . . . . . . . . . . . . . . . . . . . . 482 7.2 Properties of Confidence Sets . . . . . . . . . . . . . . . . . . 484 7.2.1 Lengths of confidence intervals . . . . . . . . . . . . . 484 7.2.2 UMA and UMAU confidence sets . . . . . . . . . . . 488 7.2.3 Randomized confidence sets

. . . . . . . . . . . . . . 491

7.2.4 Invariant confidence sets . . . . . . . . . . . . . . . . 493 7.3 Asymptotic Confidence Sets

. . . . . . . . . . . . . . . . . . 495

7.3.1 Asymptotically pivotal quantities . . . . . . . . . . . 495 7.3.2 Confidence sets based on likelihoods . . . . . . . . . . 497 7.3.3 Confidence intervals for quantiles . . . . . . . . . . . 501 7.3.4 Accuracy of asymptotic confidence sets . . . . . . . . 503 7.4 Bootstrap Confidence Sets . . . . . . . . . . . . . . . . . . . 505 7.4.1 Construction of bootstrap confidence intervals . . . . 506 7.4.2 Asymptotic correctness and accuracy . . . . . . . . . 509 7.4.3 High-order accurate bootstrap confidence sets . . . . 515

xvi

Contents

7.5 Simultaneous Confidence Intervals . . . . . 7.5.1 Bonferroni’s method . . . . . . . . . 7.5.2 Scheff´e’s method in linear models . 7.5.3 Tukey’s method in one-way ANOVA 7.5.4 Confidence bands for c.d.f.’s . . . . 7.6 Exercises . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . models . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

519 519 520 523 525 527

References

543

List of Notation

555

List of Abbreviations

557

Index of Definitions, Main Results, and Examples

559

Author Index

571

Subject Index

575

Chapter 1

Probability Theory Mathematical statistics relies on probability theory, which in turn is based on measure theory. The present chapter provides some principal concepts and notational conventions of probability theory, and some important results that are useful tools in statistics. A more complete account of probability theory can be found in a standard textbook, for example, Billingsley (1986), Chung (1974), or Lo`eve (1977). The reader is assumed to be familiar with set operations and set functions (mappings) in advanced calculus.

1.1 Probability Spaces and Random Elements In an elementary probability course, one defines a random experiment to be an experiment whose outcome cannot be predicted with certainty, and the probability of A (a collection of possible outcomes) to be the fraction of times that the outcome of the random experiment results in A in a large number of trials of the random experiment. A rigorous and logically consistent definition of probability was given by A. N. Kolmogorov in his measure-theoretic fundamental development of probability theory in 1933 (Kolmogorov, 1933).

1.1.1 σ-fields and measures Let Ω be a set of elements of interest. For example, Ω can be a set of numbers, a subinterval of the real line, or all possible outcomes of a random experiment. In probability theory, Ω is often called the outcome space, whereas in statistical theory, Ω is called the sample space. This is because in probability and statistics, Ω is usually the set of all possible outcomes of a random experiment under study. 1

2

1. Probability Theory

A measure is a natural mathematical extension of the length, area, or volume of subsets in the one-, two-, or three-dimensional Euclidean space. In a given sample space Ω, a measure is a set function defined for certain subsets of Ω. It is necessary for this collection of subsets to satisfy certain properties, which are given in the following definition. Definition 1.1. Let F be a collection of subsets of a sample space Ω. F is called a σ-field (or σ-algebra) if and only if it has the following properties. (i) The empty set ∅ ∈ F. (ii) If A ∈ F , then the complement Ac ∈ F . (iii) If Ai ∈ F, i = 1, 2, ..., then their union ∪Ai ∈ F. A pair (Ω, F ) consisting of a set Ω and a σ-field F of subsets of Ω is called a measurable space. The elements of F are called measurable sets in measure theory or events in probability and statistics. Since ∅c = Ω, it follows from (i) and (ii) in Definition 1.1 that Ω ∈ F if F is a σ-field on Ω. Also, it follows from (ii) and (iii) that if Ai ∈ F , i = 1, 2, ..., and F is a σ-field, then the intersection ∩Ai ∈ F . This can be c shown using DeMorgan’s law: (∩Ai ) = ∪Aci . For any given Ω, there are two trivial σ-fields. The first one is the collection containing exactly two elements, ∅ and Ω. This is the smallest possible σ-field on Ω. The second one is the collection of all subsets of Ω, which is called the power set and is the largest σ-field on Ω. Let us now consider some nontrivial σ-fields. Let A be a nonempty proper subset of Ω (A ⊂ Ω, A 6= Ω). Then (verify) {∅, A, Ac , Ω}

(1.1)

is a σ-field. In fact, this is the smallest σ-field containing A in the sense that if F is any σ-field containing A, then the σ-field in (1.1) is a subcollection of F. In general, the smallest σ-field containing C, a collection of subsets of Ω, is denoted by σ(C) and is called the σ-field generated by C. Hence, the σ-field in (1.1) is σ({A}). Note that σ({A, Ac }), σ({A, Ω}), and σ({A, ∅}) are all the same as σ({A}). Of course, if C itself is a σ-field, then σ(C) = C. On the real line R, there is a special σ-field that will be used almost exclusively. Let C be the collection of all finite open intervals on R. Then B = σ(C) is called the Borel σ-field. The elements of B are called Borel sets. The Borel σ-field B k on the k-dimensional Euclidean space Rk can be similarly defined. It can be shown that all intervals (finite or infinite), open sets, and closed sets are Borel sets. To illustrate, we now show that, on the real line, B = σ(O), where O is the collection of all open sets. Typically, one needs to show that σ(C) ⊂ σ(O) and σ(O) ⊂ σ(C). Since an open interval is an open set, C ⊂ O and, hence, σ(C) ⊂ σ(O) (why?). Let U be an open set. Then U can be expressed as a union of a sequence of finite open

1.1. Probability Spaces and Random Elements

3

intervals (see Royden (1968, p.39)). Hence, U ∈ σ(C) (Definition 1.1(iii)) and O ⊂ σ(C). By the definition of σ(O), σ(O) ⊂ σ(C). This completes the proof. Let C ⊂ Rk be a Borel set and let BC = {C ∩ B : B ∈ B k }. Then (C, BC ) is a measurable space and BC is called the Borel σ-field on C. Now we can introduce the notion of a measure. Definition 1.2. Let (Ω, F ) be a measurable space. A set function ν defined on F is called a measure if and only if it has the following properties. (i) 0 ≤ ν(A) ≤ ∞ for any A ∈ F. (ii) ν(∅) = 0. (iii) If Ai ∈ F , i = 1, 2, ..., and Ai ’s are disjoint, i.e., Ai ∩ Aj = ∅ for any i 6= j, then ! ∞ ∞ [ X ν Ai = ν(Ai ). i=1

i=1

The triple (Ω, F, ν) is called a measure space. If ν(Ω) = 1, then ν is called a probability measure and we usually denote it by P instead of ν, in which case (Ω, F, P ) is called a probability space. Although measure is an extension of length, area, or volume, sometimes it can be quite abstract. For example, the following set function is a measure: ∞ A ∈ F, A 6= ∅ ν(A) = (1.2) 0 A = ∅.

Since a measure can take ∞ as its value, we must know how to do arithmetic with ∞. In this book, it suffices to know that (1) for any x ∈ R, ∞+x = ∞, x ∞ = ∞ if x > 0, x ∞ = −∞ if x < 0, and 0 ∞ = 0; (2) ∞ + ∞ = ∞; and (3) ∞a = ∞ for any a > 0. However, ∞ − ∞ or ∞/∞ is not defined. The following examples provide two very important measures in probability and statistics. Example 1.1 (Counting measure). Let Ω be a sample space, F the collection of all subsets, and ν(A) the number of elements in A ∈ F (ν(A) = ∞ if A contains infinitely many elements). Then ν is a measure on F and is called the counting measure. Example 1.2 (Lebesgue measure). There is a unique measure m on (R, B) that satisfies m([a, b]) = b − a (1.3)

for every finite interval [a, b], −∞ < a ≤ b < ∞. This is called the Lebesgue measure. If we restrict m to the measurable space ([0, 1], B[0,1]), then m is a probability measure.

4

1. Probability Theory

If Ω is countable in the sense that there is a one-to-one correspondence between Ω and the set of all integers, then one can usually consider the trivial σ-field that contains all subsets of Ω and a measure that assigns a value to every subset of Ω. When Ω is uncountable (e.g., Ω = R or [0, 1]), it is not possible to define a reasonable measure for every subset of Ω; for example, it is not possible to find a measure on all subsets of R and still satisfy property (1.3). This is why it is necessary to introduce σ-fields that are smaller than the power set. The following result provides some basic properties of measures. Whenever we consider ν(A), it is implicitly assumed that A ∈ F. Proposition 1.1. Let (Ω, F, ν) be a measure space. (i) (Monotonicity). If A ⊂ B, then ν(A) ≤ ν(B). (ii) (Subadditivity). For any sequence A1 , A2 , ..., ! ∞ ∞ [ X ν Ai ≤ ν(Ai ). i=1

i=1

(iii) (Continuity). If A1 ⊂ A2 ⊂ A3 ⊂ · · · (or A1 ⊃ A2 ⊃ A3 ⊃ · · · and ν(A1 ) < ∞), then ν lim An = lim ν (An ) , n→∞

where

lim An =

n→∞

∞ [

n→∞

Ai

or =

i=1

∞ \

Ai

i=1

!

.

Proof. We prove (i) only. The proofs of (ii) and (iii) are left as exercises. Since A ⊂ B, B = A ∪ (Ac ∩ B) and A and Ac ∩ B are disjoint. By Definition 1.2(iii), ν(B) = ν(A) + ν(Ac ∩ B), which is no smaller than ν(A) since ν(Ac ∩ B) ≥ 0 by Definition 1.2(i). There is a one-to-one correspondence between the set of all probability measures on (R, B) and a set of functions on R. Let P be a probability measure. The cumulative distribution function (c.d.f.) of P is defined to be F (x) = P ((−∞, x]) ,

x ∈ R.

(1.4)

Proposition 1.2. (i) Let F be a c.d.f. on R. Then (a) F (−∞) = limx→−∞ F (x) = 0; (b) F (∞) = limx→∞ F (x) = 1; (c) F is nondecreasing, i.e., F (x) ≤ F (y) if x ≤ y; (d) F is right continuous, i.e., limy→x,y>x F (y) = F (x). (ii) Suppose that a real-valued function F on R satisfies (a)-(d) in part (i). Then F is the c.d.f. of a unique probability measure on (R, B).

1.1. Probability Spaces and Random Elements

5

The Cartesian product of sets (or collections of sets) Γi , i ∈ I = {1, ..., k} (or {1, 2, ...}) is defined asQ the set of all (a1 , ..., ak ) (or (a1 , a2 , ...)), ai ∈ Γi , i ∈ I, and is denoted by i∈I Γi = Γ1 × · · Q · × Γk (or Γ1 × Γ2 × · · ·). Let (Ωi , Fi ), i ∈ I, be measurable spaces. Since a σi∈I Fi is not necessarily Q Q F is called the product σ-field on the product space field, σ i i∈I Ωi Q Q i∈I Q is denoted by and Ω , σ F (Ω , F ). As an example, i i i i i∈I i∈I i∈I consider (Ωi , Fi ) = (R, B), i = 1, ..., k. Then the product space is Rk and it can be shown that the product σ-field is the same as the Borel σ-field on Rk , which is the σ-field generated by the collection of all open sets in Rk . In Example 1.2, the usual length of an interval [a, b] ⊂ R is the same as the Lebesgue measure of [a, b]. Consider a rectangle [a1 , b1 ] × [a2 , b2 ] ⊂ R2 . The usual area of [a1 , b1 ] × [a2 , b2 ] is (b1 − a1 )(b2 − a2 ) = m([a1 , b1 ])m([a2 , b2 ]),

(1.5)

i.e., the product of the Lebesgue measures of two intervals [a1 , b1 ] and [a2 , b2 ]. Note that [a1 , b1 ] × [a2 , b2 ] is a measurable set by the definition of the product σ-field. Is m([a1 , b1 ])m([a2 , b2 ]) the same as the value of a measure defined on the product σ-field? The following result answers this question for any product space generated by a finite number of measurable spaces. (Its proof can be found in Billingsley (1986, pp. 235-236).) Before introducing this result, we need the following technical definition. A measure ν on (Ω, F ) is said to be σ-finite if and only if there exists a sequence {A1 , A2 , ...} such that ∪Ai = Ω and ν(Ai ) < ∞ for all i. Any finite measure (such as a probability measure) is clearly σ-finite. The Lebesgue measure in Example 1.2 is σ-finite, since R = ∪An with An = (−n, n), n = 1, 2, .... The counting measure in Example 1.1 is σ-finite if and only if Ω is countable. The measure defined by (1.2), however, is not σ-finite. Proposition 1.3 (Product measure theorem). Let (Ωi , Fi , νi ), i = 1, ..., k, be measure spaces with σ-finite measures, where k ≥ 2 is an integer. Then there exists a unique σ-finite measure on the product σ-field σ(F1 ×· · ·×Fk ), called the product measure and denoted by ν1 × · · · × νk , such that ν1 × · · · × νk (A1 × · · · × Ak ) = ν1 (A1 ) · · · νk (Ak ) for all Ai ∈ Fi , i = 1, ..., k. In R2 , there is a unique measure, the product measure m× m, for which m × m([a1 , b1 ] × [a2 , b2 ]) is equal to the value given by (1.5). This measure is called the Lebesgue measure on (R2 , B 2 ). The Lebesgue measure on (R3 , B 3 ) is m × m × m, which equals the usual volume for a subset of the form [a1 , b1 ] × [a2 , b2 ] × [a3 , b3 ]. The Lebesgue measure on (Rk , B k ) for any positive integer k is similarly defined. The concept of c.d.f. can be extended to Rk . Let P be a probability

6

1. Probability Theory

measure on (Rk , B k ). The c.d.f. (or joint c.d.f.) of P is defined by F (x1 , ..., xk ) = P ((−∞, x1 ] × · · · × (−∞, xk ]) ,

xi ∈ R.

(1.6)

Again, there is a one-to-one correspondence between probability measures and joint c.d.f.’s on Rk . Some properties of a joint c.d.f. are given in Exercise 10 in §1.6. If F (x1 , ..., xk ) is a joint c.d.f., then Fi (x) =

lim

xj →∞,j=1,...,i−1,i+1,...,k

F (x1 , ..., xi−1 , x, xi+1 , ..., xk )

is a c.d.f. and is called the ith marginal c.d.f. Apparently, marginal c.d.f.’s are determined by their joint c.d.f. But a joint c.d.f. cannot be determined by k marginal c.d.f.’s. There is one special but important case in which a joint c.d.f. F is determined by its k marginal c.d.f. Fi ’s through F (x1 , ..., xk ) = F1 (x1 ) · · · Fk (xk ),

(x1 , ..., xk ) ∈ Rk ,

(1.7)

in which case the probability measure corresponding to F is the product measure P1 × · · · × Pk with Pi being the probability measure corresponding to Fi . Proposition 1.3 can be extended to cases involving infinitely many measure spaces (Billingsley, 1986). In particular, if (Rk , B k , Pi ), i = 1, 2, ..., are probability spaces, then there is a product probability measure P on Q ∞ k k k i=1 (R , B ) such that for any positive integer l and Bi ∈ B , i = 1, ..., l, P (B1 × · · · × Bl × Rk × Rk × · · ·) = P1 (B1 ) · · · Pl (Bl ).

1.1.2 Measurable functions and distributions Since Ω can be quite arbitrary, it is often convenient to consider a function (mapping) f from Ω to a simpler space Λ (often Λ = Rk ). Let B ⊂ Λ. Then the inverse image of B under f is f −1 (B) = {f ∈ B} = {ω ∈ Ω : f (ω) ∈ B}. The inverse function f −1 need not exist for f −1 (B) to be defined. The reader is asked to verify the following properties: (a) f −1 (B c ) = (f −1 (B))c for any B ⊂ Λ; (b) f −1 (∪Bi ) = ∪f −1 (Bi ) for any Bi ⊂ Λ, i = 1, 2, .... Let C be a collection of subsets of Λ. We define f −1 (C) = {f −1 (C) : C ∈ C}. Definition 1.3. Let (Ω, F ) and (Λ, G) be measurable spaces and f a function from Ω to Λ. The function f is called a measurable function from (Ω, F) to (Λ, G) if and only if f −1 (G) ⊂ F.

1.1. Probability Spaces and Random Elements

7

If Λ = R and G = B (Borel σ-field), then f is said to be Borel measurable or is called a Borel function on (Ω, F) (or with respect to F). In probability theory, a measurable function is called a random element and denoted by one of X, Y , Z,.... If X is measurable from (Ω, F ) to (R, B), then it is called a random variable; if X is measurable from (Ω, F ) to (Rk , B k ), then it is called a random k-vector. If X1 , ..., Xk are random variables defined on a common probability space, then the vector (X1 , ..., Xk ) is a random k-vector. (As a notational convention, any vector c ∈ Rk is denoted by (c1 , ..., ck ), where ci is the ith component of c.) If f is measurable from (Ω, F) to (Λ, G), then f −1 (G) is a sub-σ-field of F (verify). It is called the σ-field generated by f and is denoted by σ(f ). Now we consider some examples of measurable functions. If F is the collection of all subsets of Ω, then any function f is measurable. Let A ⊂ Ω. The indicator function for A is defined as 1 ω∈A IA (ω) = 0 ω 6∈ A. For any B ⊂ R,

∅ A −1 IA (B) = Ac Ω

0 6∈ B, 1 6∈ B 0 6∈ B, 1 ∈ B 0 ∈ B, 1 6∈ B 0 ∈ B, 1 ∈ B.

Then σ(IA ) is the σ-field given in (1.1). If A is a measurable set, then IA is a Borel function. Note that σ(IA ) is a much smaller σ-field than the original σ-field F. This is another reason why we introduce the concept of measurable functions and random variables, in addition to the reason that it is easy to deal with numbers. Often the σ-field F (such as the power set) contains too many subsets and we are only interested in some of them. One can then define a random variable X with σ(X) containing subsets that are of interest. In general, σ(X) is between the trivial σ-field {∅, Ω} and F, and contains more subsets if X is more complicated. For the simplest function IA , we have shown that σ(IA ) contains only four elements. The class of simple functions is obtained by taking linear combinations of indicators of measurable sets, i.e., ϕ(ω) =

k X

ai IAi (ω),

(1.8)

i=1

where A1 , ..., Ak are measurable sets on Ω and a1 , ..., ak are real numbers. One can show directly that such a function is a Borel function, but it

8

1. Probability Theory

follows immediately from Proposition 1.4. Let A1 , ..., Ak be a partition of Ω, i.e., Ai ’s are disjoint and A1 ∪ · · · ∪ Ak = Ω. Then the simple function ϕ given by (1.8) with distinct ai ’s exactly characterizes this partition and σ(ϕ) = σ({A1 , ..., Ak }). Proposition 1.4. Let (Ω, F) be a measurable space. (i) f is Borel if and only if f −1 (a, ∞) ∈ F for all a ∈ R. (ii) If f and g are Borel, then so are f g and af + bg, where a and b are real numbers; also, f /g is Borel provided g(ω) 6= 0 for any ω ∈ Ω. (iii) If f1 , f2 , ... are Borel, then so are supn fn , inf n fn , lim supn fn , and lim inf n fn . Furthermore, the set o n A = ω ∈ Ω : lim fn (ω) exists n→∞

is an event and the function limn→∞ fn (ω) h(ω) = f1 (ω)

ω∈A ω 6∈ A

is Borel. (iv) Suppose that f is measurable from (Ω, F ) to (Λ, G) and g is measurable from (Λ, G) to (∆, H). Then the composite function g◦f is measurable from (Ω, F) to (∆, H). (v) Let Ω be a Borel set in Rp . If f is a continuous function from Ω to Rq , then f is measurable. Proposition 1.4 indicates that there are many Borel functions. In fact, it is hard to find a non-Borel function. The following result is very useful in technical proofs. Let f be a nonnegative Borel function on (Ω, F ). Then there exists a sequence of simple functions {ϕn } satisfying 0 ≤ ϕ1 ≤ ϕ2 ≤ · · · ≤ f and limn→∞ ϕn = f (Exercise 17 in §1.6). Let (Ω, F , ν) be a measure space and f be a measurable function from (Ω, F) to (Λ, G). The induced measure by f , denoted by ν◦f −1 , is a measure on G defined as ν ◦ f −1 (B) = ν(f ∈ B) = ν f −1 (B) , B ∈ G. (1.9)

It is usually easier to deal with ν ◦ f −1 than to deal with ν since (Λ, G) is usually simpler than (Ω, F ). Furthermore, subsets not in σ(f ) are not involved in the definition of ν ◦ f −1 . As we discussed earlier, in some cases we are only interested in subsets in σ(f ). If ν = P is a probability measure and X is a random variable or a random vector, then P ◦ X −1 is called the law or the distribution of X and

1.1. Probability Spaces and Random Elements

9

is denoted by PX . The c.d.f. of PX defined by (1.4) or (1.6) is also called the c.d.f. or joint c.d.f. of X and is denoted by FX . On the other hand, for any c.d.f. or joint c.d.f. F , there exists at least one random variable or vector (usually there are many) defined on some probability space for which FX = F . The following are some examples of random variables and their c.d.f.’s. More examples can be found in §1.3.1. Example 1.3 (Discrete c.d.f.’s). Let a1 < a2 < · · · be a sequence of real numbers P∞ and let pn , n = 1, 2, ..., be a sequence of positive numbers such that n=1 pn = 1. Define F (x) =

Pn

i=1

an ≤ x < an+1 , n = 1, 2, ... −∞ < x < a1 .

pi

0

(1.10)

Then F is a stepwise c.d.f. It has a jump of size pn at each an and is flat between an and an+1 , n = 1, 2, .... Such a c.d.f. is called a discrete c.d.f. and the corresponding random variable is called a discrete random variable. We can easily obtain a random variable having F in (1.10) as its c.d.f. For example, let Ω = {a1 , a2 , ...}, F be the collection of all subsets of Ω, X

P (A) =

i:ai ∈A

pi ,

A ∈ F,

(1.11)

and X(ω) = ω. One can show that P is a probability measure and the c.d.f. of X is F in (1.10). Example 1.4 (Continuous c.d.f.’s). Opposite to the class of discrete c.d.f.’s is the class of continuous c.d.f.’s. Without the concepts of integration and differentiation introduced in the next section, we can only provide a few examples of continuous c.d.f.’s. One such example is the uniform c.d.f. on the interval [a, b] defined as

F (x) =

0

x−a b−a

1

−∞ < x < a a≤x 0 and a∞ = R 0 if a = 0, the right-hand side of (1.12) is always well defined, although ϕdν = ∞ is possible. Note that different ai ’s and Ai ’s may produce the same function ϕ; for example, with Ω = R, 2I(0,1) (x) + I[1,2] (x) = I(0,2] (x) + I(0,1) (x). However, one can show Rthat different representations of ϕ in (1.8) produce the same value for ϕdν so that the integral of a nonnegative simple function is well defined. Next, we consider a nonnegative Borel function f . Definition 1.4(b). Let f be a nonnegative Borel function and let Sf be the collection of all nonnegative simple functions of the form (1.8) satisfying ϕ(ω) ≤ f (ω) for any ω ∈ Ω. The integral of f w.r.t. ν is defined as Z Z f dν = sup ϕdν : ϕ ∈ Sf . Hence, for any Borel function f ≥ 0, there exists a sequence Rof simple functions ϕ1 , ϕ2 , ... such that 0 ≤ ϕi ≤ f for all i and limn→∞ ϕn dν = R f dν.

11

1.2. Integration and Differentiation

Finally, for a Borel function f , we first define the positive part of f by f+ (ω) = max{f (ω), 0} and the negative part of f by f− (ω) = max{−f (ω), 0}. Note that f+ and f− are nonnegative Borel functions, f (ω) = f+ (ω) − f− (ω), and |f (ω)| = f+ (ω) + f− (ω).

R Definition 1.4(c). Let f Rbe a Borel function. We say that f dν exists if R and only if at least one of f+ dν and f− dν is finite, in which case Z Z Z (1.13) f dν = f+ dν − f− dν. R R When both f+ dν and f− dν are finite, we say that f is integrable. Let A be a measurable set and IA be its indicator function. The integral of f over A is defined as Z Z f dν = IA f dν. A

Note that a Borel function f is integrable if and only if |f | is integrable. It is convenient to define the integral of a measurable function f from ¯ B), ¯ where R ¯ = R ∪ {−∞, ∞}, B¯ = σ(B ∪ {{∞}, {−∞}}). (Ω, F , ν) to (R, R Let A+ = {f = ∞} and A f+ dν − R= {f = −∞}.R If ν(A+ ) = 0, we define R to be IAc+ f+ dν; otherwise f+ dν = ∞. f− dν is similarly defined. If at R R R least one of f+ dν and f− dν is finite, then f dν is defined by (1.13). The integral of f may be denoted differently whenever there is a need to indicateR the variable(s) to Rbe integrated andR the integration domain; for R example, Ω f dν, f (ω)dν,R f (ω)dν(ω), or f (ω)ν(dω), and so on. In probability and statistics, XdP is usually written as EX or E(X) and called the Rexpectation or expected value R of X. If F isR the c.d.f. of P on (Rk , B k ), f (x)dP is also denoted by f (x)dF (x) or f dF . Example 1.5. Let Ω be a countable set, F be all subsets of Ω, and ν be the counting measure given in Example 1.1. For any Borel function f , it can be shown (exercise) that Z X f dν = f (ω). (1.14) ω∈Ω

Example 1.6. If Ω = R and ν is the Lebesgue measure, then the Lebesgue R Rb integral of f over an interval [a, b] is written as [a,b] f (x)dx = a f (x)dx, which agrees with the Riemann integral in calculus when the latter is well

12

1. Probability Theory

defined. However, there are functions for which the Lebesgue integrals are defined but not the Riemann integrals. We now introduce some properties of integrals. The proof of the following result is left to the reader. Proposition 1.5 (Linearity of integrals). Let (Ω, F, ν) be a measure space and fRand g be Borel functions. R R (i) If f dν Rexists and aR ∈ R, then (afR)dν exists R and is equal to a f dν. R(ii) If both f dν and gdν existRand fRdν + gdν is well defined, then (f + g)dν exists and is equal to f dν + gdν.

If N is an event with ν(N ) = 0 and a statement holds for all ω in the complement N c , then the statement is said to hold a.e. (almost everywhere) ν (or simply a.e. if the measure ν is clear from the context). If ν is a probability measure, then a.e. may be replaced by a.s. (almost surely).

Proposition 1.6. LetR (Ω, F, ν) R be a measure space and f and g be Borel. (i) If f ≤ g a.e., thenR f dν ≤ gdν, provided that the integrals exist. (ii) If f ≥ 0 a.e. and f dν = 0, then f = 0 a.e. Proof. (i) The proof for part (i) is left to the reader. (ii) Let A = {f > 0} and An = {f ≥ n−1 }, n = 1, 2, .... Then An ⊂ A for any n and limn→∞ An = ∪An = A (why?). By Proposition 1.1(iii), limn→∞ ν(An ) = ν(A). Using part (i) and Proposition 1.5, we obtain that Z Z Z n−1 ν(An ) = n−1 IAn dν ≤ f IAn dν ≤ f dν = 0 for any n. Hence ν(A) = 0 and f = 0 a.e.

R R Some direct consequences of Proposition 1.6(i) are: R R | f dν| R ≤ |f |dν; if f ≥ 0 a.e., then f dν ≥ 0; and if f = g a.e., then f dν = gdν. It is sometimes required to know whether the following interchange of two operations is valid: Z Z lim fn dν = lim fn dν, (1.15) n→∞

n→∞

where {fn : n = 1, 2, ...} is a sequence of Borel functions. Note that we only require limn→∞ fn exists a.e. Also, limn→∞ fn is Borel (Proposition 1.4). The following example shows that (1.15) is not always true.

Example 1.7. Consider (R, B) and the Lebesgue measure. Define fn (x) = nI[0,n−1 ] (x), n = 1, 2, .... Then limn→∞ fn (x) = 0 for all x but x = 0. Since the Lebesgue measure Rof a single point set is 0 (see Example 1.2), lim R n→∞ fn (x) = 0 a.e. and limn→∞ fn (x)dx R = 0. On the other hand, fn (x)dx = 1 for any n and, hence, limn→∞ fn (x)dx = 1.

13

1.2. Integration and Differentiation

The following result gives sufficient conditions under which (1.15) holds. Theorem 1.1. Let f1 , f2 , ... be a sequence of Borel functions on (Ω, F , ν). (i) (Fatou’s lemma). If fn ≥ 0, then Z Z lim inf fn dν ≤ lim inf fn dν. n

n

(ii) (Dominated convergence theorem). If limn→∞ fn = f a.e. and there exists an integrable function g such that |fn | ≤ g a.e., then (1.15) holds. (iii) (Monotone convergence theorem). If 0 ≤ f1 ≤ f2 ≤ · · · and limn→∞ fn = f a.e., then (1.15) holds. Proof. The results in (i) and (iii) are equivalent (exercise). R Applying Fatou’s Rlemma to functions g + f and g − f , we obtain that (g + f )dν ≤ n n R R lim inf n R(g + fn )dν and (g − fR )dν ≤ lim inf n (g − fn )dν (which is the same as (f − g)dνR ≥ lim supn (fn − R g)dν). Since g is Rintegrable,R these results imply that f dν ≤ lim inf n fn dν ≤ lim supn fn dν ≤ f dν. Hence, the result in (i) implies the result in (ii). It remains to show part (iii). Let f, f1 , fR2 , ... be given R in part (iii). From Proposition 1.6(i), there exists limn→∞ fn dν ≤ f dν. Let ϕ be a simple function with 0 ≤ R ϕ ≤ f and let Aϕ = {ϕ > 0}. Suppose that ν(Aϕ ) = ∞. Then f dν = ∞. Let a = 2−1 minω∈Aϕ ϕ(ω) and An = {fn > a}. Then a > 0, A1 ⊂ A2 ⊂ · · ·, and Aϕ ⊂ ∪AnR (why?). By R Proposition 1.1, ν(An ) → ν(∪An ) ≥ ν(Aϕ ) = ∞ and, hence, fn dν ≥ An fn dν ≥ aν(An ) → ∞. Suppose now ν(Aϕ ) < ∞. By Egoroff’s theorem (Exercise 20 in §1.6), for any ǫ > 0, there is B ⊂ AϕR with ν(B) R < ǫ such that fn converges to f uniformly on Aϕ ∩ B c . Hence, fn dν ≥ Aϕ ∩B c fn dν → R R R R R = ϕdν − B ϕdν ≥ ϕdν − ǫ maxω ϕ(ω). Since Aϕ ∩B c f dν ≥ Aϕ ∩B c ϕdν R R ǫ is arbitrary, lim R n→∞ fRn dν ≥ ϕdν. Since ϕ is arbitrary, by Definition 1.4(b), limn→∞ fn dν ≥ f dν. This completes the proof.

Example 1.8 (Interchange of differentiation and integration). Let (Ω, F , ν) be a measure space and, for any fixed θ ∈ R, let f (ω, θ) be a Borel function on Ω. Suppose that ∂f (ω, θ)/∂θ exists a.e. for θ ∈ (a, b) ⊂ R and that |∂f (ω, θ)/∂θ| ≤ g(ω) a.e., where g is an integrable function on Ω. Then, for each θ ∈ (a, b), ∂f (ω, θ)/∂θ is integrable and, by Theorem 1.1(ii), Z Z d ∂f (ω, θ) f (ω, θ)dν = dν. dθ ∂θ

Theorem 1.2 (Change of variables). Let f be measurable from (Ω, F, ν) to (Λ, G) and g be Borel on (Λ, G). Then Z Z g ◦ f dν = gd(ν ◦ f −1 ), (1.16) Ω

Λ

i.e., if either integral exists, then so does the other, and the two are the same.

14

1. Probability Theory

The reader is encouraged to provide a proof. A complete proof is in Billingsley (1986, p. 219). This the change of variable formula R result extends R for Riemann integrals, i.e., g(y)dy = g(f (x))f ′ (x)dx, y = f (x). Result (1.16) is very important in probability and statistics. R Let X be a random variable on a probability space (Ω, F, P ). IfREX = Ω XdP exists, then usually it is much simpler to compute EX = R xdPX , where PX = P ◦ X −1 is the law of X. Let Y be a random vector from Ω to Rk and k Rg be Borel from RR to R. According to (1.16), Eg(Y ) can be computed as g(y)dP or Y Rk R xdPg(Y ) , depending on which of PY and Pg(Y ) is easier to handle. As a more specific example, consider k = 2, Y = (X1 , X2 ), and g(Y ) = X1 + X2 . Using Proposition 1.5(ii), E(X1 + X2 ) = EX1 + EX2 R R and, hence, E(X1 + X2 ) = R xdPX1 + R xdPX2 . Then we need to handle Rtwo integrals involving PX1 and PX2 . On the other hand, E(X1 + X2 ) = R xdPX1 +X2 , which involves one integral w.r.t. PX1 +X2 . Unless we have some knowledge about the joint c.d.f. of (X1 , X2 ), it is not easy to obtain PX1 +X2 . The following theorem states how to evaluate an integral w.r.t. a product measure via iterated integration. The reader is encouraged to prove this theorem. A complete proof can be found in Billingsley (1986, pp. 236-238). Theorem 1.3 (Fubini’s theorem). Let νi be a σ-finite measure on (Ωi , Fi ), Q2 i = 1, 2, and let f be a Borel function on i=1 (Ωi , Fi ). Suppose that either f ≥ 0 or f is integrable w.r.t. ν1 × ν2 . Then Z g(ω2 ) = f (ω1 , ω2 )dν1 Ω1

exists a.e. ν2 and defines a Borel function on Ω2 whose integral w.r.t. ν2 exists, and Z Z Z f (ω1 , ω2 )dν1 × ν2 = f (ω1 , ω2 )dν1 dν2 . Ω1 ×Ω2

Ω2

Ω1

This result Q can be naturally extended to the integral w.r.t. the product measure on ki=1 (Ωi , Fi ) for any finite positive integer k.

Example 1.9. Let Ω1 = Ω2 = {0, 1, 2, ...}, and ν1 = ν2 be the counting measure (Example 1.1). A function f on Ω1 ×Ω2 defines a double sequence. R If f ≥ 0 or |f |dν1 × ν2 < ∞, then Z ∞ ∞ ∞ X ∞ X X X f (i, j) = f (i, j) (1.17) f dν1 × ν2 = i=0 j=0

j=0 i=0

(by Theorem 1.3 and Example 1.5). Thus, a double series can be summed in either order, if it is summable or f ≥ 0.

15

1.2. Integration and Differentiation

1.2.2 Radon-Nikodym derivative Let (Ω, F, ν) be a measure space and f be a nonnegative Borel function. One can show that the set function Z λ(A) = f dν, A ∈ F , (1.18) A

is a measure on (Ω, F) (verify). Note that ν(A) = 0 implies

λ(A) = 0.

(1.19)

If (1.19) holds for two measures λ and ν defined on the same measurable space, then we say λ is absolutely continuous w.r.t. ν and write λ ≪ ν. Formula (1.18) gives us not only a way of constructing measures, but also a method of computing measures of measurable sets. Let ν be a wellknown measure (such as the Lebesgue measure or the counting measure) and λ a relatively unknown measure. If we can find a function f such that (1.18) holds, then computing λ(A) can be done through integration. A necessary condition for (1.18) is clearly λ ≪ ν. The following result shows that λ ≪ ν is also almost sufficient for (1.18). Theorem 1.4 (Radon-Nikodym theorem). Let ν and λ be two measures on (Ω, F) and ν be σ-finite. If λ ≪ ν, then there exists a nonnegative Borel function f on RΩ such that (1.18) holds. Furthermore, f is unique a.e. ν, i.e., if λ(A) = A gdν for any A ∈ F, then f = g a.e. ν.

The proof of this theorem can be found in Billingsley (1986, pp. 443444). If (1.18) holds, then the function f is called the Radon-Nikodym derivative or density of λ w.r.t. ν and is denoted by dλ/dν. R A useful consequence of Theorem 1.4 is that if f is Borel on (Ω, F ) and f dν = 0 for any A ∈ F, then f = 0 a.e. A R If f dν = 1 for an f ≥ 0 a.e. ν, then λ given by (1.18) is a probability measure and f is called its probability density function (p.d.f.) w.r.t. ν. For any probability measure P on (Rk , B k ) corresponding to a c.d.f. F or a random vector X, if P has a p.d.f. f w.r.t. a measure ν, then f is also called the p.d.f. of F or X w.r.t. ν.

Example 1.10 (p.d.f. of a discrete c.d.f.). Consider the discrete c.d.f. F in (1.10) of Example 1.3 with its probability measure given by (1.11). Let Ω = {a1 , a2 , ...} and ν be the counting measure on the power set of Ω. By Example 1.5, Z X P (A) = f dν = f (ai ), A ⊂ Ω, (1.20) A

ai ∈A

16

1. Probability Theory

where f (ai ) = pi , i = 1, 2, .... That is, f is the p.d.f. of P or F w.r.t. ν. Hence, any discrete c.d.f. has a p.d.f. w.r.t. counting measure. A p.d.f. w.r.t. counting measure is called a discrete p.d.f. Example 1.11. Let F be a c.d.f. Assume that F is differentiable in the usual sense in calculus. Let f be the derivative of F . From calculus, Z x F (x) = f (y)dy, x ∈ R. (1.21) −∞

Let P be theR probability measure corresponding to F . It can be shown that P (A) = A f dm for any A ∈ B, where m is the Lebesgue measure on R. Hence, f is the p.d.f. of P or F w.r.t. Lebesgue measure. In this case, the Radon-Nikodym derivative is the same as the usual derivative of F in calculus. A continuous c.d.f. may not have a p.d.f. w.r.t. Lebesgue measure. A necessary and sufficient condition for a c.d.f. F having a p.d.f. w.r.t. Lebesgue measure is that F is absolute continuous in the sense that for any ǫ > 0, there exists a δ > 0 such collection of disjoint P that for each finite P bounded open intervals (ai , bi ), (bi −ai ) < δ implies [F (bi )−F (ai )] < ǫ. Absolute continuity is weaker than differentiability, but is stronger than continuity. Thus, any discontinuous c.d.f. (such as a discrete c.d.f.) is not absolute continuous. Note that every c.d.f. is differentiable a.e. Lebesgue measure (Chung, 1974, Chapter 1). Hence, if f is the p.d.f. of F w.r.t. Lebesgue measure, then f is the usual derivative of F a.e. Lebesgue measure and (1.21) holds. In such a case probabilities can be computed through integration. It can be shown that the uniform and exponential c.d.f.’s in Example 1.4 are absolute continuous and their p.d.f.’s are, respectively, ( 1 a≤x0 (1 − p)x−1 p, x = 1, 2, ... pet /[1 − (1 − p)et ], t < − log(1 − p) 1/p (1 − p)/p2 p ∈ [0, 1] N m (nx ) r−x r x = 0, 1, ..., min{r, n}, r − x ≤ m No explicit form rn/N rnm(N − r)/[N 2 (N − 1)] r, n, m = 1, 2, ..., N = n + m r x−1 x−r , x = r, r + 1, ... r−1 p (1 − p) r rt p e /[1 − (1 − p)et ]r , t < − log(1 − p) r/p r(1 − p)/p2 p ∈ [0, 1], r = 1, 2, ... −(log p)−1 x−1 (1 − p)x , x = 1, 2, ... log[1 − (1 − p)et ]/ log p, t ∈ R −(1 − p)/(p log p) −(1 − p)[1 + (1 − p)/ log p]/(p2 log p) p ∈ (0, 1)

All p.d.f.’s are w.r.t. counting measure.

19

1.3. Distributions and Their Characteristics

forms, whereas many others do not and they have to be evaluated numerically or computed using tables or software. There are p.d.f.’s that are neither discrete nor Lebesgue. Example 1.12. Let X be a random variable on (Ω, F , P ) whose c.d.f. FX has a Lebesgue p.d.f. fX and FX (c) < 1, where c is a fixed constant. Let Y = min{X, c}, i.e., Y is the smaller of X and c. Note that Y −1 ((−∞, x]) = Ω if x ≥ c and Y −1 ((−∞, x]) = X −1 ((∞, x]) if x < c. Hence Y is a random variable and the c.d.f. of Y is 1 x≥c FY (x) = x < c. FX (x) This c.d.f. is discontinuous at c, since FX (c) < 1. Thus, it does not have a Lebesgue p.d.f. It is not discrete either. Does PY , the probability measure corresponding to FY , have a p.d.f. w.r.t. some measure? Define a probability measure on (R, B), called point mass at c, by 1 c∈A διc (A) = A∈B (1.22) 0 c 6∈ A, (which is a special case of the discrete uniform distribution in Table 1.1). Then PY ≪ m + διc , where m is the Lebesgue measure, and the p.d.f. of PY is x>c 0 dPY (x) = (1.23) 1 − FX (c) x=c d(m + διc ) x < c. fX (x)

A p.d.f. corresponding to a joint c.d.f. is called a joint p.d.f. The following is a joint Lebesgue p.d.f. on Rk that is important in statistics: f (x) = (2π)−k/2 [Det(Σ)]−1/2 e−(x−µ) k

τ

Σ−1 (x−µ)/2

,

x ∈ Rk ,

(1.24)

where µ ∈ R , Σ is a positive definite k × k matrix, Det(Σ) is the determinant of Σ and, when matrix algebra is involved, any k-vector c is treated as a k × 1 matrix (column vector) and cτ denotes its transpose (row vector). The p.d.f. in (1.24) and its c.d.f. are called the k-dimensional multivariate normal p.d.f. and c.d.f., and both are denoted by Nk (µ, Σ). Random vectors distributed as Nk (µ, Σ) are also denoted by Nk (µ, Σ) for convenience. The normal distribution N (µ, σ 2 ) in Table 1.2 is a special case of Nk (µ, Σ) with k = 1. In particular, N (0, 1) is called the standard normal distribution. When Σ is a nonnegative definite but singular matrix, we define X to be Nk (µ, Σ) if and only if cτ X is N (cτ µ, cτ Σc) for any c ∈ Rk (N (a, 0) is defined to be the c.d.f. of the point mass at a), which is an important property of Nk (µ, Σ) with a nonsingular Σ (Exercise 81). Another important joint p.d.f. will be introduced in Example 2.7.

20

1. Probability Theory

Table 1.2. Distributions on R with Lebesgue p.d.f.’s Uniform U (a, b)

Normal 2

N (µ, σ )

Exponential E(a, θ)

Chi-square χ2k

Gamma Γ(α, γ)

Beta B(α, β)

Cauchy C(µ, σ)

p.d.f. m.g.f. Expectation Variance Parameter p.d.f.

(b − a)−1 I(a,b) (x) (ebt − eat )/[(b − a)t], t ∈ R (a + b)/2 (b − a)2 /12 a, b ∈ R, a < b 2 2 √ 1 e−(x−µ) /2σ 2πσ

m.g.f.

eµt+σ

Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter

µ σ2 µ ∈ R, σ > 0 θ−1 e−(x−a)/θ I(a,∞) (x) eat (1 − θt)−1 , t < θ−1 θ+a θ2 θ > 0, a ∈ R 1 xk/2−1 e−x/2 I(0,∞) (x) Γ(k/2)2k/2 −k/2 , t < 1/2 (1 − 2t) k 2k k = 1, 2, ... 1 α−1 −x/γ e I(0,∞) (x) Γ(α)γ α x −α (1 − γt) , t < γ −1 αγ αγ 2 γ > 0, α > 0 Γ(α+β) α−1 (1 − x)β−1 I(0,1) (x) Γ(α)Γ(β) x No explicit form α/(α + β) αβ/[(α + β + 1)(α + β)2 ] α > 0, β > 0 h i−1 x−µ 2 1 πσ 1 + σ

p.d.f. ch.f. Expectation Variance Parameter

√

2 2

t /2

, t∈R

e −1µt−σ|t| Does not exist Does not exist µ ∈ R, σ > 0

21

1.3. Distributions and Their Characteristics

Table 1.2. (continued) t-distribution tn

F-distribution Fn,m

Log-normal LN (µ, σ 2 )

Weibull W (α, θ)

DE(µ, θ) Pareto P a(a, θ)

Logistic LG(µ, σ)

x2 n

−(n+1)/2

Γ[(n+1)/2] √ nπΓ(n/2)

ch.f. Expectation Variance Parameter

No explicit form 0, (n > 1) n/(n − 2), (n > 2) n = 1, 2, ...

p.d.f. ch.f. Expectation Variance Parameter p.d.f. ch.f. Expectation Variance Parameter p.d.f. ch.f. Expectation Variance

Double Exponential

p.d.f.

Parameter p.d.f. m.g.f. Expectation Variance Parameter p.d.f. ch.f. Expectation Variance Parameter p.d.f. m.g.f. Expectation Variance Parameter

1+

nn/2 mm/2 Γ[(n+m)/2]xn/2−1 I (x) Γ(n/2)Γ(m/2)(m+nx)(n+m)/2 (0,∞)

No explicit form m/(m − 2), (m > 2) 2m2 (n + m − 2)/[n(m − 2)2 (m − 4)], (m > 4) n = 1, 2, ..., m = 1, 2, ... 2 2 √ 1 x−1 e−(log x−µ) /2σ I(0,∞) (x) 2πσ No explicit form 2 eµ+σ /2 2 2 e2µ+σ (eσ − 1) µ ∈ R, σ > 0 α α−1 −xα /θ e I(0,∞) (x) θx No explicit form −1 + 1) θ1/α Γ(α n 2 o 2/α Γ(2α−1 + 1) − Γ(α−1 + 1) θ

θ > 0, α > 0 1 −|x−µ|/θ 2θ e eµt /(1 − θ2 t2 ), |t| < θ−1 µ 2θ2 µ ∈ R, θ > 0 θaθ x−(θ+1) I(a,∞) (x) No explicit form θa/(θ − 1), (θ > 1) θa2 /[(θ − 1)2 (θ − 2)], (θ > 2) θ > 0, a > 0 σ −1 e−(x−µ)/σ /[1 + e−(x−µ)/σ ]2 eµt Γ(1 + σt)Γ(1 − σt), |t| < σ −1 µ σ 2 π 2 /3 µ ∈ R, σ > 0

22

1. Probability Theory

If a random k-vector (X1 , ..., Xk ) has a joint p.d.f. f w.r.t. a product measure ν1 × · · · × νk defined on B k , then Xi has the following marginal p.d.f. w.r.t. νi : Z f (x1 , ..., xi−1 , x, xi+1 , ..., xk )dν1 · · · dνi−1 dνi+1 · · · dνk . fi (x) = Rk−1

Let F be the joint c.d.f. of a random k-vector (X1 , ..., Xk ) and Fi be the marginal c.d.f. of Xi , i = 1, ..., k. If (1.7) holds, then random variables X1 , ..., Xk are said to be independent. From the discussion in the end of §1.1.1, this independence means that the probability measure corresponding to F is the product measure of the k probability measures corresponding to Fi ’s. The meaning of independence is further discussed in §1.4.2. If (X1 , ..., Xk ) has a joint p.d.f. f w.r.t. a product measure ν1 × · · · × νk defined on B k , then X1 , ..., Xk are independent if and only if f (x1 , ..., xk ) = f1 (x1 ) · · · fk (xk ),

(x1 , ..., xk ) ∈ Rk ,

(1.25)

where fi is the p.d.f. of Xi w.r.t. νi , i = 1, ..., k. For example, using (1.24), one can show (exercise) that the components of Nk (µ, Σ) are independent if and only if Σ is a diagonal matrix. The following lemma is useful in considering the independence of functions of independent random variables. Lemma 1.1. Let X1 , ..., Xn be independent random variables. Then random variables g(X1 , ..., Xk ) and h(Xk+1 , ..., Xn ) are independent, where g and h are Borel functions and k is an integer between 1 and n. Lemma 1.1 can be proved directly (exercise). But it is a simple consequence of an equivalent definition of independence introduced in §1.4.2. Let X1 , ..., Xk be random variables. If Xi and Xj are independent for every pair i 6= j, then X1 , ..., Xk are said to be pairwise independent. If X1 , ..., Xk are independent, then clearly they are pairwise independent. However, the converse is not true. The following is an example. Example 1.13. Let X1 and X2 be independent random variables each assuming the values 1 and −1 with probability 0.5, and X3 = X1 X2 . Let Ai = {Xi = 1}, i = 1, 2, 3. Then P (Ai ) = 0.5 for any i and P (A1 )P (A2 )P (A3 ) = 0.125. However, P (A1 ∩A2 ∩A3 ) = P (A1 ∩A2 ) = P (A1 )P (A2 ) = 0.25. This implies that (1.7) does not hold and, hence, X1 , X2 , X3 are not independent. We now show that X1 , X2 , X3 are pairwise independent. It is enough to show that X1 and X3 are independent. Let Bi = {Xi = −1}, i = 1, 2, 3. Note that A1 ∩ A3 = A1 ∩ A2 , A1 ∩ B3 = A1 ∩ B2 , B1 ∩ A3 = B1 ∩ B2 , and B1 ∩ B3 = B1 ∩ A2 . Then the result follows from the fact that P (Ai ) = P (Bi ) = 0.5 for any i and X1 and X2 are independent.

23

1.3. Distributions and Their Characteristics

The random variable Y in Example 1.12 is a transformation of the random variable X. Transformations of random variables or vectors are frequently used in statistics. For a random variable or vector X, g(X) is a random variable or vector as long as g is measurable (Proposition 1.4). How do we find the c.d.f. (or p.d.f.) of g(X) when the c.d.f. (or p.d.f.) of X is known? In many cases, the most effective method is direct computation. Example 1.12 is one example. The following is another one. Example 1.14. Let X be a random variable with c.d.f. FX and Lebesgue −1 p.d.f. fX , and let Y = X 2 . √Since √ Y ((−∞, x]) is empty if x < 0 and −1 −1 equals Y ([0, x]) = X ([− x, x ]) if x ≥ 0, the c.d.f. of Y is FY (x) = P ◦ Y −1 ((−∞, x]) √ √ = P ◦ X −1 ([− x, x ]) √ √ = FX ( x) − FX (− x) if x ≥ 0 and FY (x) = 0 if x < 0. Clearly, the Lebesgue p.d.f. of FY is √ √ 1 fY (x) = √ [fX ( x) + fX (− x)]I(0,∞) (x). 2 x

(1.26)

In particular, if

2 1 fX (x) = √ e−x /2 , (1.27) 2π which is the Lebesgue p.d.f. of the standard normal distribution N (0, 1) (Table 1.2), then 1 fY (x) = √ e−x/2 I(0,∞) (x), 2πx

which is the Lebesgue p.d.f. for the chi-square distribution χ21 (Table 1.2). This is actually an important result in statistics. In some cases, one may apply the following general result whose proof is left to the reader. Proposition 1.8. Let X be a random k-vector with a Lebesgue p.d.f. fX and let Y = g(X), where g is a Borel function from (Rk , B k ) to (Rk , B k ). Let A1 , ..., Am be disjoint sets in B k such that Rk − (A1 ∪ · · · ∪ Am ) has Lebesgue measure 0 and g on Aj is one-to-one with a nonvanishing Jacobian, i.e., the determinant Det(∂g(x)/∂x) 6= 0 on Aj , j = 1, ..., m. Then Y has the following Lebesgue p.d.f.: fY (x) =

m X Det (∂hj (x)/∂x) fX (hj (x)) , j=1

where hj is the inverse function of g on Aj , j = 1, ..., m.

24

1. Probability Theory

One may apply Proposition 1.8 to obtain result (1.26) in Example 1.14, √ using A1 √ = (−∞, 0), A2 = (0, ∞), and√g(x) = x2 . Note that h1 (x) = − x, h2 (x) = x, and |dhj (x)/dx| = 1/(2 x). Another immediate application of Proposition 1.8 is to show that Y = AX is Nk (Aµ, AΣAτ ) when X is Nk (µ, Σ), where Σ is positive definite, A is a k × k matrix of rank k, and Aτ denotes the transpose of A. Example 1.15. Let X = (X1 , X2 ) be a random 2-vector having a joint Lebesgue p.d.f. fX . Consider first the transformation g(x) = (x1 , x1 + x2 ). Using Proposition 1.8, one can show that the joint p.d.f. of g(X) is fg(X) (x1 , y) = fX (x1 , y − x1 ), where y = x1 + x2 (note that the Jacobian equals 1). The marginal p.d.f. of Y = X1 + X2 is then Z fY (y) = fX (x1 , y − x1 )dx1 . In particular, if X1 and X2 are independent, then Z fY (y) = fX1 (x1 )fX2 (y − x1 )dx1 .

(1.28)

Next, consider the transformation h(x1 , x2 ) = (x1 /x2 , x2 ), assuming that X2 6= 0 a.s. Using Proposition 1.8, one can show that the joint p.d.f. of h(X) is fh(X) (z, x2 ) = |x2 |fX (zx2 , x2 ),

where z = x1 /x2 . The marginal p.d.f. of Z = X1 /X2 is Z fZ (z) = |x2 |fX (zx2 , x2 )dx2 . In particular, if X1 and X2 are independent, then Z fZ (z) = |x2 |fX1 (zx2 )fX2 (x2 )dx2 .

(1.29)

A number of results can be derived from (1.28) and (1.29). For example, if X1 and X2 are independent and both have the standard normal p.d.f. given by (1.27), then, by (1.29), the Lebesgue p.d.f. of Z = X1 /X2 is Z 2 2 1 |x2 |e−(1+z )x2 /2 dx2 fZ (z) = 2π Z 1 ∞ −(1+z2 )x = e dx π 0 1 , = π(1 + z 2 )

25

1.3. Distributions and Their Characteristics

which is the p.d.f. of the Cauchy distribution C(0, 1) in Table 1.2. Another application of formula (1.29) leads to the following important result in statistics. Example 1.16 (t-distribution and F-distribution). Let X1 and X2 be independent random variables having the chi-square distributions χ2n1 and χ2n2 (Table 1.2), respectively. By (1.29), the p.d.f. of Z = X1 /X2 is fZ (z) =

z n1 /2−1 I(0,∞) (z) 2(n1 +n2 )/2 Γ(n1 /2)Γ(n2 /2)

Z

0 n1 /2−1

=

∞

(n1 +n2 )/2−1 −(1+z)x2 /2

x2

e

dx2

z Γ[(n1 + n2 )/2] I(0,∞) (z), Γ(n1 /2)Γ(n2 /2) (1 + z)(n1 +n2 )/2

where the last equality follows from the fact that 1 (n +n )/2−1 −x2 /2 x 1 2 e I(0,∞) (x2 ) 2(n1 +n2 )/2 Γ[(n1 + n2 )/2] 2 is the p.d.f. of the chi-square distribution χ2n1 +n2 . Using Proposition 1.8, one can show that the p.d.f. of Y = (X1 /n1 )/(X2 /n2 ) = (n2 /n1 )Z is the p.d.f. of the F-distribution Fn1 ,n2 given in Table 1.2. Let U1 be a random variable having the standard normal distribution N (0, 1) and U2 a random variable having the chi-square distribution χ2n . Using the same argument, one canp show that if U1 and U2 are independent, then the distribution of T = U1 / U2 /n is the t-distribution tn given in Table 1.2. This result can also be derived using the result given in this example as follows. Let X1 = U12 and X2 = U2 . Then X1 and X2 are independent (which can be shown directly but follows from Lemma 1.1). By Example 1.14, the distribution of X1 is χ21 . Then Y = X1 /(X2 /n) has the F-distribution F1,n and its Lebesgue p.d.f. is nn/2 Γ[(n + 1)/2]x−1/2 √ I(0,∞) (x). πΓ(n/2)(n + x)(n+1)/2 Note that

√ Y √ T = − Y

U1 ≥ 0 U1 < 0.

The result follows from Proposition 1.8 and the fact that P ◦ T −1 ((−∞, −t]) = P ◦ T −1 ([t, ∞)) ,

t > 0.

(1.30)

If a random variable T satisfies (1.30), i.e., T and −T have the same distribution, then T and its c.d.f. and p.d.f. (if it exists) are said to be

26

1. Probability Theory

symmetric about 0. If T has a Lebesgue p.d.f. fT , then T is symmetric about 0 if and only if fT (x) = fT (−x) for any x > 0. T and its c.d.f. and p.d.f. are said to be symmetric about a (or symmetric for simplicity) if and only if T − a is symmetric about 0 for a fixed a ∈ R. The c.d.f.’s of t-distributions are symmetric about 0 and the normal, Cauchy, and double exponential c.d.f.’s are symmetric. The chi-square, t-, and F-distributions in the previous examples are special cases of the following noncentral chi-square, t-, and F-distributions, which are useful in some statistical problems. Let X1 , ..., Xn be independent random variables and Xi = N (µi , σ 2 ), i = 1, ..., n. The distribution of the random variable Y = (X12 +· · ·+Xn2 )/σ 2 is called the noncentral chi-square distribution and denoted by χ2n (δ), where δ = (µ21 + · · · + µ2n )/σ 2 is the noncentrality parameter. The chi-square distribution χ2k in Table 1.2 is a special case of the noncentral chi-square distribution χ2k (δ) with δ = 0 and, therefore, is called a central chi-square distribution. It can be shown (exercise) that Y has the following Lebesgue p.d.f.: ∞ X (δ/2)j e−δ/2 f2j+n (x), (1.31) j! j=0 where fk (x) is the Lebesgue p.d.f. of the chi-square distribution χ2k . It follows from the definition of noncentral chi-square distributions that if Y1 , ..., Yk are independent random variables and Yi has the noncentral chisquare distribution χ2ni (δi ), i = 1, ..., k, then Y = Y1 + · · · + Yk has the noncentral chi-square distribution χ2n1 +···+nk (δ1 + · · · + δk ). The result for the t-distribution in Example 1.16 can be extended to the case where U1 has a nonzero expectation µ (U2 still has thepχ2n distribution and is independent of U1 ). The distribution of T = U1 / U2 /n is called the noncentral t-distribution and denoted by tn (δ), where δ = µ is the noncentrality parameter. Using the same argument as that in Example 1.15, one can show (exercise) that T has the following Lebesgue p.d.f.: Z ∞ √ 1 (n−1)/2 −[(x y/n−δ)2 +y]/2 √ y e dy. (1.32) 2(n+1)/2 Γ(n/2) πn 0 The t-distribution tn in Example 1.16 is called a central t-distribution, since it is a special case of the noncentral t-distribution tn (δ) with δ = 0. Similarly, the result for the F-distribution in Example 1.16 can be extended to the case where X1 has the noncentral chi-square distribution χ2n1 (δ), X2 has the central chi-square distribution χ2n2 , and X1 and X2 are independent. The distribution of Y = (X1 /n1 )/(X2 /n2 ) is called the noncentral F-distribution and denoted by Fn1 ,n2 (δ), where δ is the noncentrality parameter. The F-distribution Fn1 ,n2 in Example 1.16 is called a central

27

1.3. Distributions and Their Characteristics

F-distribution, since it is a special case of the noncentral F-distribution Fn1 ,n2 (δ) with δ = 0. It can be shown (exercise) that the noncentral Fdistribution Fn1 ,n2 (δ) has the following Lebesgue p.d.f.: ∞ X n1 (δ/2)j n1 x f2j+n1 ,n2 , (1.33) e−δ/2 j!(2j + n1 ) 2j + n1 j=0 where fk1 ,k2 (x) is the Lebesgue p.d.f. of the central F-distribution Fk1 ,k2 given in Table 1.2. Using some results from linear algebra, we can prove the following result useful in analysis of variance (Scheff´e, 1959; Searle, 1971). Theorem 1.5. (Cochran’s theorem). Suppose that X = Nn (µ, In ) and X τ X = X τ A1 X + · · · + X τ Ak X,

(1.34)

where In is the n × n identity matrix and Ai is an n × n symmetric matrix with rank ni , i = 1, ..., k. A necessary and sufficient condition that X τ Ai X has the noncentral chi-square distribution χ2ni (δi ), i = 1, ..., k, and X τ Ai X’s are independent is n = n1 + · · · + nk , in which case δi = µτ Ai µ and δ1 + · · · + δk = µτ µ. Proof. Suppose that X τ Ai X, i = 1, ..., k, are independent and X τ Ai X has the χ2ni (δi ) distribution. Then X τ X has the χ2n1 +···+nk (δ1 + · · · + δk ) distribution. By definition, X τ X has the noncentral chi-square distribution χ2n (µτ µ). By (1.34), n = n1 + · · · + nk and δ1 + · · · + δk = µτ µ. Suppose now that n = n1 + · · · + nk . From linear algebra, for each i there exists cij ∈ Rn , j = 1, ..., ni , such that X τ Ai X = ±(cτi1 X)2 ± · · · ± (cτini X)2 .

(1.35)

Let Ci be the n × ni matrix whose jth column is cij , and C τ = (C1 , ..., Ck ). By (1.34) and (1.35), X τ X = X τ C τ ∆CX with an n × n diagonal matrix ∆ whose diagonal elements are either 1 or −1. This implies C τ ∆C = In . Thus, C is of full rank and, hence, ∆ = (C τ )−1 C −1 , which is positive definite. This shows ∆ = In , which implies C τ C = In and n1 +···+ni−1 +ni

X τ Ai X =

X

Yj2 ,

(1.36)

j=n1 +···+ni−1 +1

where Yj is the jth component of Y = CX. Note that Y = Nn (Cµ, In ) (Exercise 43). Hence Yj ’s are independent and Yj = N (λj , 1), where λj is the jth component of Cµ. This shows that X τ Ai X has the χ2ni (δi ) distribution with δi = λ2n1 +···+ni−1 +1 + · · · + λ2n1 +···+ni−1 +ni . Letting X = µ in (1.36) and (1.34), we obtain that δi = µτ Ai µ and δ1 + · · · + δk = µτ C τ Cµ = µτ µ. Finally, from (1.36) and Lemma 1.1, we conclude that X τ Ai X, i = 1, ..., k, are independent.

28

1. Probability Theory

1.3.2 Moments and moment inequalities We have defined the expectation of a random variable in §1.2.1. It is an important characteristic of a random variable. In this section, we introduce moments, which are some other important characteristics of a random variable or vector. Let X be a random variable. If EX k is finite, where k is a positive integer, then EX k is called the kth moment of X or PX (the distribution of X). If E|X|a < ∞ for some real number a, then E|X|a is called the ath absolute moment of X or PX . If µ = EX and E(X − µ)k are finite for a positive integer k, then E(X − µ)k is called the kth central moment of X or PX . If E|X|a < ∞ for an a > 0, then E|X|t < ∞ for any positive t < a and EX k is finite for any positive integer k ≤ a (Exercise 54). The expectation and the second central moment (if they exist) are two important characteristics of a random variable (or its distribution) in statistics. They are listed in Tables 1.1 and 1.2 for those useful distributions. The expectation, also called the mean in statistics, is a measure of the central location of the distribution of a random variable. The second central moment, also called the variance in statistics, is a measure of dispersion or spread of a random variable. The variance of a random variable X is denoted by Var(X). The variance is always nonnegative. If the variance of X is 0, then X is equal to its mean a.s. (Proposition 1.6). The squared root of the variance is called the standard deviation, another important characteristic of a random variable in statistics. The concept of mean and variance can be extended to random vectors. The expectation of a random matrix M with (i, j)th element Mij is defined to be the matrix whose (i, j)th element is EMij . Thus, for a random kvector X = (X1 , ..., Xk ), its mean is EX = (EX1 , ..., EXk ). The extension of variance is the variance-covariance matrix of X defined as Var(X) = E(X − EX)(X − EX)τ , which is a k × k symmetric matrix whose diagonal elements are variances of Xi ’s. The (i, j)th element of Var(X), i 6= j, is E(Xi − EXi )(Xj − EXj ), which is called the covariance of Xi and Xj and is denoted by Cov(Xi , Xj ). Let c ∈ Rk and X = (X1 , ..., Xk ) be a random k-vector. Then Y = τ c X is a random variable and, by Proposition 1.5 (linearity of integrals), EY = cτ EX if EX exists. Also, when Var(X) is finite (i.e., all elements of Var(X) are finite), Var(Y ) = E(cτ X − cτ EX)2 = E[cτ (X − EX)(X − EX)τ c] = cτ [E(X − EX)(X − EX)τ ]c = cτ Var(X)c.

29

1.3. Distributions and Their Characteristics

Since Var(Y ) ≥ 0 for any c ∈ Rk , the matrix Var(X) is nonnegative definite. Consequently, [Cov(Xi , Xj )]2 ≤ Var(Xi )Var(Xj ),

i 6= j.

(1.37)

An important quantity in p statistics is the correlation coefficient defined to be ρXi ,Xj = Cov(Xi , Xj )/ Var(Xi )Var(Xj ), which, by inequality (1.37), is always between −1 and 1. It is a measure of relationship between Xi and Xj ; if ρXi ,Xj is positive (or negative), then Xi and Xj tend to be positively (or negatively) related; if ρXi ,Xj = ±1, then P (Xi = c1 ± c2 Xj ) = 1 with some constants c1 and c2 > 0; if ρXi ,Xj = 0 (i.e., Cov(Xi , Xj ) = 0), then Xi and Xj are said to be uncorrelated. If Xi and Xj are independent, then they are uncorrelated. This follows from the following more general result. If X1 , ..., Xn are independent random variables and E|X1 · · · Xn | < ∞, then, by Fubini’s theorem and the fact that the joint c.d.f. of (X1 , ..., Xn ) corresponds to a product measure, we obtain that E(X1 · · · Xn ) = EX1 · · · EXn .

(1.38)

In fact, pairwise independence of X1 , ..., Xn implies that Xi ’s are uncorrelated, since Cov(Xi , Xj ) involves only a pair of random variables. However, the converse is not necessarily true: uncorrelated random variables may not be pairwise independent. Examples can be found in Exercises 60-61. Let RM = {y ∈ Rk : y = M x with some x ∈ Rk } for any k × k symmetric matrix M . If a random k-vector X has a finite Var(X), then P (X − EX ∈ RVar(X) ) = 1. This means that if the rank of Var(X) is r < k, then X is in a subspace of Rk with dimension r. Consequently, if PX ≪ Lebesgue measure on Rk , then the rank of Var(X) is k. Example 1.17. Let X be a random k-vector having the Nk (µ, Σ) distribution. It can be shown (exercise) that EX = µ and Var(X) = Σ. Thus, µ and Σ in (1.24) are the mean vector and the variance-covariance matrix of X. If Σ is a diagonal matrix (i.e., all components of X are uncorrelated), then by (1.25), the components of X are independent. This shows an important property of random variables having normal distributions: they are independent if and only if they are uncorrelated. There are many useful inequalities related to moments. The inequality in (1.37) is in fact the well-known Cauchy-Schwartz inequality whose general form is [E(XY )]2 ≤ EX 2 EY 2 , (1.39)

where X and Y are random variables with a well-defined E(XY ). Inequality (1.39) is a special case of the following H¨ older’s inequality: E|XY | ≤ (E|X|p )1/p (E|Y |q )1/q ,

(1.40)

30

1. Probability Theory

where p and q are constants satisfying p > 1 and p−1 + q −1 = 1. To show inequality (1.40), we use the following inequality (Exercise 62): xt y 1−t ≤ tx + (1 − t)y,

(1.41)

where x and y are nonnegative real numbers and t ∈ (0, 1). If either E|X|p or E|Y |q is ∞, then (1.40) holds. Hence we can assume that both E|X|p and E|Y |q are finite. Let a = (E|X|p )1/p and b = (E|Y |q )1/q . If either a = 0 or b = 0, then the equality in (1.40) holds because of Proposition 1.6(ii). Assume now a 6= 0 and b 6= 0. Letting x = |X/a|p , y = |Y /b|q , and t = p−1 in (1.41), we obtain that XY |X|p |Y |q ≤ + . ab pap qbq Taking expectations on both sides of this expression, we obtain that E|XY | E|X|p E|Y |q 1 1 ≤ + = + = 1, ab pap qaq p q which is (1.40). In fact, the equality in (1.40) holds if and only if α|X|p = β|Y |q a.s. for some nonzero constants α and β (Exercise 62). Using H¨older’s inequality, we can prove Liapounov’s inequality (E|X|r )1/r ≤ (E|X|s )1/s ,

(1.42)

where r and s are constants satisfying 1 ≤ r ≤ s, and Minkowski’s inequality (E|X + Y |p )1/p ≤ (E|X|p )1/p + (E|Y |p )1/p , (1.43) where X and Y are random variables and p is a constant larger than or equal to 1 (Exercise 63). Minkowski’s inequality can be extended to the case of more than two random variables (Exercise 63). The following inequality is a tightened form of Minkowski’s inequality due to Esseen and von Bahr (1965). Let X1 , ..., Xn be independent random variables with mean 0 and E|Xi |p < ∞, i = 1, ..., n, where p is a constant in [1, 2]. Then X p n X n E Xi ≤ Cp E|Xi |p , i=1

(1.44)

i=1

where Cp is a constant depending only on p. When 1 < p < 2, inequality (1.44) can be proved (Exercise 63) using inequality |a + b|p ≤ |a|p + psgn(a)|a|p−1 b + Cp |b|p ,

a ∈ R, b ∈ R,

31

1.3. Distributions and Their Characteristics

where sgn(x) is 1 or −1 as x is positive or negative and Cp =

sup (|1 + x|p − 1 − px)/|x|p .

x∈R,x6=0

For p ≥ 2, there is a similar inequality due to Marcinkiewicz and Zygmund: n n X p Cp X Xi ≤ 1−p/2 E|Xi |p , E n i=1 i=1

(1.45)

where Cp is a constant depending only on p. A proof of inequality (1.45) can be found in Lo`eve (1977, p. 276). Recall from calculus that a subset A of Rk is convex if and only if x ∈ A and y ∈ A imply tx + (1 − t)y ∈ A for any t ∈ [0, 1]; a function f from a convex A ⊂ Rk to R is convex if and only if f (tx + (1 − t)y) ≤ tf (x) + (1 − t)f (y),

x ∈ A, y ∈ A, t ∈ [0, 1];

(1.46)

and f is strictly convex if and only if (1.46) holds with ≤ replaced by the strict inequality . By Proposition 1.6(ii), Ef (X) > f (EX) unless P (f (X) = Ef (X)) = 1. Example 1.18. A direct application of Jensen’s inequality (1.47) is that if X is a nonconstant positive random variable with finite mean, then (EX)−1 < E(X −1 )

and

E(log X) < log(EX),

32

1. Probability Theory

since t−1 and − log t are convex functions on (0, ∞). Another application is to prove the following inequality related to entropy. Let f and g be positive integrable functions on a measure space with a σ-finite measure ν. R R If f dν ≥ gdν > 0, then one can show (exercise) that Z f f log dν ≥ 0. (1.49) g The next inequality, Chebyshev’s inequality, is almost trivial but very useful and famous. Let X be a random variable and ϕ a nonnegative and nondecreasing function on [0, ∞) satisfying ϕ(−t) = ϕ(t). Then, for each constant t ≥ 0, Z ϕ(t)P (|X| ≥ t) ≤ ϕ(X)dP ≤ Eϕ(X), (1.50) {|X|≥t}

where both inequalities in (1.50) follow from Proposition 1.6(i) and the first inequality also uses the fact that on the set {|X| ≥ t}, ϕ(X) ≥ ϕ(t). The most familiar application of (1.50) is when ϕ(t) = |t|p for p ∈ (0, ∞), in which case inequality (1.50) is also called Markov’s inequality. Chebyshev’s inequality, sometimes together with one of the moment inequalities introduced in this section, can be used to yield a desired upper bound for the “tail” probability P (|X| ≥ t). For example, let Y be a random variable with mean µ and variance σ 2 . Then X = (Y − µ)/σ has mean 0 and variance 1 and, by (1.50) with ϕ(t) = t2 , P (|X| ≥ 2) ≤ 41 . This means that the probability that the random variable |Y − µ| exceeds twice its standard deviation is bounded by 14 . Similarly, we can also claim that the probability of |Y − µ| exceeding 3σ is bounded by 19 . These bounds are rough but they can be applied to any random variable with a finite variance. Other applications of Chebyshev’s inequality can be found in §1.5. In some cases, we need an improvement over inequality (1.50) when X is of some special form. Let Y1 , ..., Yn be independent random variables having finite variances. The following inequality is due to H´ajek and R`enyi: ! l n X 1 X 2 P max cl (Yi − EYi ) > t ≤ 2 c Var(Yi ), t > 0, (1.51) 1≤l≤n t i=1 i i=1 where ci ’s are positive constants satisfying c1 ≥ c2 ≥ · · · ≥ cn . If ci = 1 for all i, then inequality (1.51) reduces to the famous Kolmogorov’s inequality. A proof for (1.51) is given in Sen and Singer (1993, pp. 65-66).

1.3.3 Moment generating and characteristic functions Moments are important characteristics of a distribution, but they do not determine a distribution in the sense that two different distributions may

33

1.3. Distributions and Their Characteristics

have the same moments of all orders. Functions that determine a distribution are introduced in the following definition. Definition 1.5. Let X be a random k-vector. (i) The moment generating function (m.g.f.) of X or PX is defined as τ

ψX (t) = Eet

X

,

t ∈ Rk .

(ii) The characteristic function (ch.f.) of X or PX is defined as √ √ τ φX (t) = Ee −1t X = E[cos(tτ X)] + −1 E[sin(tτ X)], t ∈ Rk . Obviously ψX (0) = φX (0) = 1 for any random vector X. The ch.f. is complex-valued and always well defined. In fact, any ch.f. is bounded by 1 and is a uniformly continuous function on Rk (exercise). The m.g.f. is nonnegative but may be ∞ everywhere except at t = 0 (Example 1.19). If the m.g.f. is finite in a neighborhood of 0 ∈ Rk , then φX (t) can be obtained √ by replacing t in ψX (t) by −1t. Tables 1.1 and 1.2 contain the m.g.f. (or ch.f. when the m.g.f. is ∞ everywhere except at 0) for distributions useful in statistics. For a linear transformation Y = Aτ X + c, where A is a k × m matrix and c ∈ Rm , it follows from Definition 1.5 that τ

ψY (u) = ec u ψX (Au) and φY (u) = e

√ −1cτ u

u ∈ Rm . (1.52) For a random variable X, if its m.g.f. is finite at t and −t for a t 6= 0, then X has finite moments and absolute moments of any order. To compute moments of X using its m.g.f., a condition stronger than the finiteness of the m.g.f. at some t 6= 0 is needed. Consider a random k-vector X. If ψX is finite in a neighborhood of 0, then µr1 ,...,rk = E(X1r1 · · · Xkrk ) is finite for any nonnegative integers r1 , ..., rk , where Xj is the jth component of X, and ψX has the power series expansion ψX (t) =

X

(r1 ,...,rk )∈Z

φX (Au),

µr1 ,...,rk tr11 · · · trkk r1 ! · · · rk !

(1.53)

for t in the neighborhood of 0, where tj is the jth component of t and Z ⊂ Rk containing vectors whose components are nonnegative integers. Consequently, the components of X have finite moments of all orders and ∂ r1 +···+rk ψX (t) rk r1 E(X1 · · · Xk ) = , ∂tr11 · · · ∂trkk t=0 which are also called moments of X. In particular, ∂ψX (t) ∂ 2 ψX (t) = EX, = E(XX τ ), ∂t t=0 ∂t∂tτ t=0

(1.54)

34

1. Probability Theory (p)

and, when k = 1 and p is a positive integer, ψX (0) = EX p , where g (p) (t) denotes the pth order derivative of a function g(t). If 0 < ψX (t) < ∞, then κX (t) = log ψX (t) is called the cumulant generating function of X or PX . If 0 < ψX (t) < ∞ for t in a neighborhood of 0, then κX has a power series expansion similar to that in (1.53): X

κX (t) =

(r1 ,...,rk )∈Z

κr1 ,...,rk tr11 · · · trkk , r1 ! · · · rk !

(1.55)

where κr1 ,...,rk ’s are called cumulants of X. There is a one-to-one correspondence between the set of moments and the set of cumulants. An example for the case of k = 1 is given in Exercise 68. When ψX is not finite, finite moments of X can be obtained by differentiating its ch.f. φX . Suppose that E|X1r1 · · · Xkrk | < ∞ for some nonnegative integers r1 , ..., rk . Let r = r1 + · · · + rk and √

τ

√ τ ∂ r e −1t X r/2 r1 X1 · · · Xkrk e −1t X . g(t) = r1 rk = (−1) ∂t1 · · · ∂tk

Then |g(t)| ≤ |X1r1 · · · Xkrk |, which is integrable. Hence, from Example 1.8, √ τ ∂ r φX (t) r/2 E X1r1 · · · Xkrk e −1t X r1 rk = (−1) ∂t1 · · · ∂tk

and

In particular,

(1.56)

∂ r φX (t) = (−1)r/2 E(X1r1 · · · Xkrk ). ∂tr11 · · · ∂trkk t=0

√ ∂φX (t) = −1EX, ∂t t=0

∂ 2 φX (t) = −E(XX τ ), ∂t∂tτ t=0 (p)

and, if k = 1 and p is a positive integer, then φX (0) = (−1)p/2 EX p , provided that all moments involved are finite. In fact, when k = 1, if φX has a finite derivative of even order p at t = 0, then EX p < ∞ (see, e.g., Chung, 1974, pp. 166-168). 2 2

Example 1.19. Let X = N (µ, σ 2 ). From Table 1.2, ψX (t) = eµt+σ t /2 . A ′ ′′ direct calculation shows that EX = ψX (0) = µ, EX 2 = ψX (0) = σ 2 + µ2 , (3) (4) EX 3 = ψX (0) = 3σ 2 µ+µ3 , and EX 4 = ψX (0) = 3σ 4 +6σ 2 µ2 +µ4 . If µ = 0, then EX p = 0 when p is an odd integer and EX p = (p−1)(p−3) · · · 3·1σ p when p is an even integer (exercise). The cumulant generating function of X is κX (t) = log ψX (t) = µt + σ 2 t2 /2. Hence, κ1 = µ, κ2 = σ 2 , and κr = 0 for r = 3, 4, ....

35

1.3. Distributions and Their Characteristics

We now find a random variable having finite moments of all order but having an m.g.f. = ∞ except for t = 0. Let Pn be the probability meaP∞ sure for the N (0, σn2 ) distribution, n = 1, 2, .... Then P = n=1 2−n Pn is a probability measure (Exercise 35). Let X be a random variable hav2 2 ing distribution P . Since the m.g.f. of N (0, σn2 ) is eσn t /2 , it follows from P∞ −n σ2 t2 /2 . When Fubini’s theorem that the m.g.f. of X is ψX (t) = n=1 2 e n σn2 = n2 ,PψX (t) = ∞ for any t 6= 0 but EX k = 0 for any odd integer k and ∞ EX k = n=1 2−n (k − 1)(k − 3) · · · 1nk < ∞ for any even integer k. When √ 2 σn2 = n, ψX (t) = (2e−t /2 − 1)−1 for |t| < log 4 and, hence, the moments of X can be obtained by differentiating ψX . For example, EX = φ′X (0) = 0 and EX 2 = φ′′X (0) = 2. A fundamental fact about ch.f.’s is that there is a one-to-one correspondence between the set of all distributions on Rk and the set of all ch.f.’s defined on Rk . The same fact is true for m.g.f.’s, but we have to focus on distributions having m.g.f.’s finite in neighborhoods of 0. Theorem 1.6. (Uniqueness). Let X and Y be random k-vectors. (i) If φX (t) = φY (t) for all t ∈ Rk , then PX = PY . (ii) If ψX (t) = ψY (t) < ∞ for all t in a neighborhood of 0, then PX = PY . Proof. (i) The result follows from the following inversion formula whose proof can be found, for example, in Billingsley (1986, p. 395): for any a = (a1 , ..., ak ) ∈ Rk , b = (b1 , ..., bk ) ∈ Rk , and (a, b] = (a1 , b1 ] × · · · × (ak , bk ] satisfying PX (the boundary of (a, b]) = 0, √ √ Z c Z c k φX (t1 , ..., tk ) Y e− −1ti ai − e− −1ti bi ··· dti . PX (a, b] = lim k/2 (2π)k c→∞ −c ti −c (−1) i=1

(ii) First consider the case of k = 1. From es|x| ≤ esx + e−sx , we conclude that |X| has an m.g.f. that is finite in the neighborhood (−c, c) for some c >√0 and |X| has finite moments of all order. Using the inequality √ Pn √ |e −1tx [e −1ax − j=0 ( −1ax)j /j!]| ≤ |ax|n+1 /(n + 1)!, we obtain that n X √ |a|n+1 E|X|n+1 √ aj j −1tX φX (t + a) − E[( , −1X) e ] ≤ j! (n + 1)! j=0 which together with (1.53) and (1.56) imply that, for any t ∈ R, φX (t + a) =

∞ (j) X φ (t) X

j=0

j!

aj ,

|a| < c.

(1.57)

Similarly, (1.57) holds with φX replaced by φY . Under the assumption that ψX = ψY < ∞ in a neighborhood of 0, X and Y have the same moments of (j) (j) all order. By (1.56), φX (0) = φY (0) for all j = 1, 2, ..., which and (1.57)

36

1. Probability Theory

with t = 0 imply that φX and φY are the same on the interval (−c, c) and hence have identical derivatives there. Considering t = c − ǫ and −c + ǫ for an arbitrarily small ǫ > 0 in (1.57) shows that φX and φY also agree on (−2c + ǫ, 2c − ǫ) and hence on (−2c, 2c). By the same argument φX and φY are the same on (−3c, 3c) and so on. Hence, φX (t) = φY (t) for all t and, by part (i), PX = PY . Consider now the general case of k ≥ 2. If PX 6= PY , then by part (i) there exists t ∈ Rk such that φX (t) 6= φY (t). Then φtτ X (1) 6= φtτ Y (1), which implies that Ptτ X 6= Ptτ Y . But ψX = ψY < ∞ in a neighborhood of 0 ∈ Rk implies that ψtτ X = ψtτ Y < ∞ in a neighborhood of 0 ∈ R and, by the proved result for k = 1, Ptτ X = Ptτ Y . This contradiction shows that PX = PY . Applying result (1.38) and Lemma 1.1, we obtain that ψX+Y (t) = ψX (t)ψY (t)

and φX+Y (t) = φX (t)ψY (t),

t ∈ Rk , (1.58)

for independent random k-vectors X and Y . This result, together with Theorem 1.6, provides a useful tool to obtain distributions of sums of independent random vectors with known distributions. The following example is an illustration. Example 1.20. Let Xi , i = 1, ..., k, be independent random variables and Xi have the gamma distribution Γ(αi , γ) (Table 1.2), i = 1, ..., k. From Table 1.2, Xi has the m.g.f. ψXi (t) = (1 − γt)−αi , t < γ −1 , i = 1, ..., k. By result (1.58), the m.g.f. of Y = X1 + · · · + Xk is equal to ψY (t) = (1 − γt)−(α1 +···+αk ) , t < γ −1 . From Table 1.2, the gamma distribution Γ(α1 + · · · + αk , γ) has the m.g.f. ψY (t) and, hence, is the distribution of Y (by Theorem 1.6). Similarly, result (1.52) and Theorem 1.6 can be used to determine distributions of linear transformations of random vectors with known distributions. The following is another interesting application of Theorem 1.6. Note that a random variable X is symmetric about 0 (defined according to (1.30)) if and only if X and −X have the same distribution, which can then be used as the definition of a random vector X symmetric about 0. We now show that X is symmetric about 0 if and only if its ch.f. φX is realvalued. If X and −X have the same distribution, then by Theorem 1.6, φX (t) = φ−X (t). From (1.52), φ−X (t) = φX (−t). Then φX (t) = φX (−t). Since sin(−tτ X) = − sin(tτ X) and cos(tτ X) = cos(−tτ X), this proves E[sin(tτ X)] = 0 and, thus, φX is real-valued. Conversely, if φX is realvalued, then φX (t) = E[cos(tτ X)] and φ−X (t) = φX (−t) = φX (t). By Theorem 1.6, X and −X must have the same distribution. Other applications of ch.f.’s can be found in §1.5.

1.4. Conditional Expectations

37

1.4 Conditional Expectations In elementary probability the conditional probability of an event B given an event A is defined as P (B|A) = P (A ∩ B)/P (A), provided that P (A) > 0. In probability and statistics, however, we sometimes need a notion of “conditional probability” even for A’s with P (A) = 0; for example, A = {Y = c}, where c ∈ R and Y is a random variable having a continuous c.d.f. General definitions of conditional probability, expectation, and distribution are introduced in this section, and they are shown to agree with those defined in elementary probability in special cases.

1.4.1 Conditional expectations Definition 1.6. Let X be an integrable random variable on (Ω, F , P ). (i) Let A be a sub-σ-field of F. The conditional expectation of X given A, denoted by E(X|A), is the a.s.-unique random variable satisfying the following two conditions: (a) E(X|A) is measurable from (Ω, A) to (R, B); R R (b) A E(X|A)dP = A XdP for any A ∈ A. (Note that the existence of E(X|A) follows from Theorem 1.4.) (ii) Let B ∈ F . The conditional probability of B given A is defined to be P (B|A) = E(IB |A). (iii) Let Y be measurable from (Ω, F , P ) to (Λ, G). The conditional expectation of X given Y is defined to be E(X|Y ) = E[X|σ(Y )]. Essentially, the σ-field σ(Y ) contains “the information in Y ”. Hence, E(X|Y ) is the “expectation” of X given the information provided by σ(Y ). The following useful result shows that there is a Borel function h defined on the range of Y such that E(X|Y ) = h ◦ Y . Lemma 1.2. Let Y be measurable from (Ω, F ) to (Λ, G) and Z a function from (Ω, F ) to Rk . Then Z is measurable from (Ω, σ(Y )) to (Rk , B k ) if and only if there is a measurable function h from (Λ, G) to (Rk , B k ) such that Z = h ◦ Y . The function h in E(X|Y ) = h ◦ Y is a Borel function on (Λ, G). Let y ∈ Λ. We define E(X|Y = y) = h(y) to be the conditional expectation of X given Y = y. Note that h(y) is a function on Λ, whereas h ◦ Y = E(X|Y ) is a function on Ω. For a random vector X, E(X|A) is defined as the vector of conditional expectations of components of X.

38

1. Probability Theory

Example 1.21. Let X be an integrable random variable on (Ω, F , P ), A1 , A2 , ... be disjoint events on (Ω, F , P ) such that ∪Ai = Ω and P (Ai ) > 0 for all i, and let a1 , a2 , ... be distinct real numbers. Define Y = a1 IA1 + a2 IA2 + · · ·. We now show that R ∞ X Ai XdP IAi . E(X|Y ) = (1.59) P (Ai ) i=1 We need to verify (a) and (b) in Definition 1.6 with A = σ(Y ). Since σ(Y ) = σ({A1 , A2 , ...}), it is clear that the function on the right-hand side of (1.59) is measurable on (Ω, σ(Y )). For any B ∈ B, Y −1 (B) = ∪i:ai ∈B Ai . Using properties of integrals, we obtain that Z X Z XdP = XdP Y −1 (B)

=

i:ai ∈B Ai R ∞ X Ai XdP i=1

=

Z

P Ai ∩ Y −1 (B)

P (Ai ) "∞ R X A XdP i

Y −1 (B)

i=1

P (Ai )

#

IAi dP.

This verifies (b) and thus (1.59) holds. R Let h be a Borel function on R satisfying h(ai ) = Ai XdP/P (Ai ). Then, by (1.59), E(X|Y ) = h ◦ Y and E(X|Y = y) = h(y). Let A ∈ F and X = IA . Then P (A|Y ) = E(X|Y ) =

∞ X P (A ∩ Ai ) i=1

P (Ai )

IAi ,

which equals P (A ∩ Ai )/P (Ai ) = P (A|Ai ) if ω ∈ Ai . Hence, the definition of conditional probability in Definition 1.6 agrees with that in elementary probability. The next result generalizes the result in Example 1.21 to conditional expectations of random variables having p.d.f.’s. Proposition 1.9. Let X be a random n-vector and Y a random m-vector. Suppose that (X, Y ) has a joint p.d.f. f (x, y) w.r.t. ν × λ, where ν and λ are σ-finite measures on (Rn , B n ) and (Rm , B m ), respectively. Let g(x, y) be a Borel function on Rn+m for which E|g(X, Y )| < ∞. Then R g(x, Y )f (x, Y )dν(x) R E[g(X, Y )|Y ] = a.s. (1.60) f (x, Y )dν(x)

39

1.4. Conditional Expectations

Proof. Denote the right-hand side of (1.60) by h(Y ). By Fubini’s theorem, h is Borel. Then, by Lemma R 1.2, h(Y ) is Borel on (Ω, σ(Y )). Also, by Fubini’s theorem, fY (y) = f (x, y)dν(x) is the p.d.f. of Y w.r.t. λ. For B ∈ Bm, Z Z h(Y )dP = h(y)dPY Y −1 (B) B Z R g(x, y)f (x, y)dν(x) R fY (y)dλ(y) = f (x, y)dν(x) B Z = g(x, y)f (x, y)dν × λ Rn ×B Z = g(x, y)dP(X,Y ) Rn ×B Z = g(X, Y )dP, Y −1 (B)

where the first and the last equalities follow from Theorem 1.2, the second and the next to last equalities follow from the definition of h and p.d.f.’s, and the third equality follows from Theorem 1.3 (Fubini’s theorem). For a random vector (X, Y ) with a joint p.d.f. f (x, y) w.r.t. ν × λ, define the conditional p.d.f. of X given Y = y to be fX|Y (x|y) =

f (x, y) , fY (y)

(1.61)

R where fY (y) = f (x, y)dν(x) is the marginal p.d.f. of Y w.r.t. λ. One can easily check that for each fixed y with fY (y) > 0, fX|Y (x|y) in (1.61) is a p.d.f. w.r.t. ν. Then equation (1.60) can be rewritten as Z E[g(X, Y )|Y ] = g(x, Y )fX|Y (x|Y )dν(x). Again, this agrees with the conditional expectation defined in elementary probability (i.e., the conditional expectation of g(X, Y ) given Y is equal to the expectation of g(X, Y ) w.r.t. the conditional p.d.f. of X given Y ). Now we list some useful properties of conditional expectations. The proof is left to the reader. Proposition 1.10. Let X, Y , X1 , X2 , ... be integrable random variables on (Ω, F , P ) and A be a sub-σ-field of F. (i) If X = c a.s., c ∈ R, then E(X|A) = c a.s. (ii) If X ≤ Y a.s., then E(X|A) ≤ E(Y |A) a.s. (iii) If a ∈ R and b ∈ R, then E(aX + bY |A) = aE(X|A) + bE(Y |A) a.s.

40

1. Probability Theory

(iv) E[E(X|A)] = EX. (v) E[E(X|A)|A0 ] = E(X|A0 ) = E[E(X|A0 )|A] a.s., where A0 is a sub-σfield of A. (vi) If σ(Y ) ⊂ A and E|XY | < ∞, then E(XY |A) = Y E(X|A) a.s. (vii) If X and Y are independent and E|g(X, Y )| < ∞ for a Borel function g, then E[g(X, Y )|Y = y] = E[g(X, y)] a.s. PY . (viii) If EX 2 < ∞, then [E(X|A)]2 ≤ E(X 2 |A) a.s. (ix) (Fatou’s lemma). If Xn ≥ 0 for any n, then E lim inf n Xn A ≤ lim inf n E(Xn |A) a.s. (x) (Dominated convergence theorem). Suppose that |Xn | ≤ Y for any n and Xn →a.s. X. Then E(Xn |A) →a.s. E(X|A). Although part (vii) of Proposition 1.10 can be proved directly, it is a consequence of a more general result given in Theorem 1.7(i). Since E(X|A) is defined only for integrable X, a version of monotone convergence theorem (i.e., 0 ≤ X1 ≤ X2 ≤ · · · and Xn →a.s. X imply E(Xn |A) →a.s. E(X|A)) becomes a special case of Proposition 1.10(x). It can also be shown (exercise) that H¨ older’s inequality (1.40), Liapounov’s inequality (1.42), Minkowski’s inequality (1.43), and Jensen’s inequality (1.47) hold a.s. with the expectation E replaced by the conditional expectation E(·|A). As an application, we consider the following example. Example 1.22. Let X be a random variable on (Ω, F , P ) with EX 2 < ∞ and let Y be a measurable function from (Ω, F , P ) to (Λ, G). One may wish to predict the value of X based on an observed value of Y . Let g(Y ) be a predictor, i.e., g ∈ ℵ = {all Borel functions g with E[g(Y )]2 < ∞}. Each predictor is assessed by the “mean squared prediction error” E[X − g(Y )]2 . We now show that E(X|Y ) is the best predictor of X in the sense that E[X − E(X|Y )]2 = min E[X − g(Y )]2 . g∈ℵ

(1.62)

First, Proposition 1.10(viii) implies E(X|Y ) ∈ ℵ. Next, for any g ∈ ℵ, E[X − g(Y )]2 = E[X − E(X|Y ) + E(X|Y ) − g(Y )]2 = E[X − E(X|Y )]2 + E[E(X|Y ) − g(Y )]2

+ 2E{[X − E(X|Y )][E(X|Y ) − g(Y )]} = E[X − E(X|Y )]2 + E[E(X|Y ) − g(Y )]2 + 2E E{[X − E(X|Y )][E(X|Y ) − g(Y )]|Y } = E[X − E(X|Y )]2 + E[E(X|Y ) − g(Y )]2 + 2E{[E(X|Y ) − g(Y )]E[X − E(X|Y )|Y ]} = E[X − E(X|Y )]2 + E[E(X|Y ) − g(Y )]2 ≥ E[X − E(X|Y )]2 ,

1.4. Conditional Expectations

41

where the third equality follows from Proposition 1.10(iv), the fourth equality follows from Proposition 1.10(vi), and the last equality follows from Proposition 1.10(i), (iii), and (vi).

1.4.2 Independence Definition 1.7. Let (Ω, F, P ) be a probability space. (i) Let C be a collection of subsets in F . Events in C are said to be independent if and only if for any positive integer n and distinct events A1 ,...,An in C, P (A1 ∩ A2 ∩ · · · ∩ An ) = P (A1 )P (A2 ) · · · P (An ). (ii) Collections Ci ⊂ F, i ∈ I (an index set that can be uncountable), are said to be independent if and only if events in any collection of the form {Ai ∈ Ci : i ∈ I} are independent. (iii) Random elements Xi , i ∈ I, are said to be independent if and only if σ(Xi ), i ∈ I, are independent. The following result is useful for checking the independence of σ-fields. Lemma 1.3. Let Ci , i ∈ I, be independent collections of events. Suppose that each Ci has the property that if A ∈ Ci and B ∈ Ci , then A ∩ B ∈ Ci . Then σ(Ci ), i ∈ I, are independent. An immediate application of Lemma 1.3 is to show (exercise) that random variables Xi , i = 1, ..., k, are independent according to Definition 1.7 if and only if (1.7) holds with F being the joint c.d.f. of (X1 , ..., Xk ) and Fi being the marginal c.d.f. of Xi . Hence, Definition 1.7(iii) agrees with the concept of independence of random variables discussed in §1.3.1. It is easy to see from Definition 1.7 that if X and Y are independent random vectors, then so are g(X) and h(Y ) for Borel functions g and h. Since the independence in Definition 1.7 is equivalent to the independence discussed in §1.3.1, this provides a simple proof of Lemma 1.1. For two events A and B with P (A) > 0, A and B are independent if and only if P (B|A) = P (B). This means that A provides no information about the probability of the occurrence of B. The following result is a useful extension. Proposition 1.11. Let X be a random variable with E|X| < ∞ and let Yi be random ki -vectors, i = 1, 2. Suppose that (X, Y1 ) and Y2 are independent. Then E[X|(Y1 , Y2 )] = E(X|Y1 ) a.s.

42

1. Probability Theory

Proof. First, E(X|Y1 ) is Borel on (Ω, σ(Y1 , Y2 )), since σ(Y1 ) ⊂ σ(Y1 , Y2 ). Next, we need to show that for any Borel set B ∈ B k1 +k2 , Z Z XdP = E(X|Y1 )dP. (1.63) (Y1 ,Y2 )−1 (B)

(Y1 ,Y2 )−1 (B)

If B = B1 × B2 , where Bi ∈ B ki , then (Y1 , Y2 )−1 (B) = Y1−1 (B1 ) ∩ Y2−1 (B2 ) and Z Z E(X|Y1 )dP = IY −1 (B1 ) IY −1 (B2 ) E(X|Y1 )dP 1

Y1−1 (B1 )∩Y2−1 (B2 )

= = = =

Z

Z

Z

Z

2

Z

IY −1 (B1 ) E(X|Y1 )dP IY −1 (B2 ) dP 1 2 Z IY −1 (B1 ) XdP IY −1 (B2 ) dP 1

2

IY −1 (B1 ) IY −1 (B2 ) XdP 1

2

Y1−1 (B1 )∩Y2−1 (B2 )

XdP,

where the second and the next to last equalities follow from result (1.38) and the independence of (X, Y1 ) and Y2 , and the third equality follows from the fact that E(X|Y1 ) is the conditional expectation of X given Y1 . This shows that (1.63) holds for B = B1 × B2 . We can show that the collection H = {B ⊂ Rk1 +k2 : B satisfies (1.63)} is a σ-field. Since we have already shown that B k1 × B k2 ⊂ H, B k1 +k2 = σ(B k1 × B k2 ) ⊂ H and thus the result follows. Clearly, the result in Proposition 1.11 still holds if X is replaced by h(X) for any Borel h and, hence, P (A|Y1 , Y2 ) = P (A|Y1 ) a.s. for any A ∈ σ(X),

(1.64)

if (X, Y1 ) and Y2 are independent. If Y1 is a constant and Y = Y2 , (1.64) reduces to P (A|Y ) = P (A) a.s. for any A ∈ σ(X), if X and Y are independent, i.e., σ(Y ) does not provide any additional information about the stochastic behavior of X. This actually provides another equivalent but more intuitive definition of the independence of X and Y (or two σ-fields). With a nonconstant Y1 , we say that given Y1 , X and Y2 are conditionally independent if and only if (1.64) holds. Then the result in Proposition 1.11 can be stated as: if Y2 and (X, Y1 ) are independent, then given Y1 , X and Y2 are conditionally independent. It is important to know that the result in Proposition 1.11 may not be true if Y2 is independent of X but not (X, Y1 ) (Exercise 96).

43

1.4. Conditional Expectations

1.4.3 Conditional distributions The conditional p.d.f. was introduced in §1.4.1 for random variables having p.d.f.’s w.r.t. some measures. We now consider conditional distributions in general cases where we may not have any p.d.f. Let X and Y be two random vectors defined on a common probability space. It is reasonable to consider P [X −1 (B)|Y = y] as a candidate for the conditional distribution of X, given Y = y, where B is any Borel set. However, since conditional probability is defined almost surely, for any fixed y, P [X −1 (B)|Y = y] may not be a probability measure. The first part of the following theorem (whose proof can be found in Billingsley (1986, pp. 460-461)) shows that there exists a version of conditional probability such that P [X −1 (B)|Y = y] is a probability measure for any fixed y. Theorem 1.7. (i) (Existence of conditional distributions). Let X be a random n-vector on a probability space (Ω, F, P ) and A be a sub-σ-field of F . Then there exists a function P (B, ω) on B n × Ω such that (a) P (B, ω) = P [X −1 (B)|A] a.s. for any fixed B ∈ B n , and (b) P (·, ω) is a probability measure on (Rn , B n ) for any fixed ω ∈ Ω. Let Y be measurable from (Ω, F , P ) to (Λ, G). Then there exists PX|Y (B|y) such that (a) PX|Y (B|y) = P [X −1 (B)|Y = y] a.s. PY for any fixed B ∈ B n , and (b) PX|Y (·|y) is a probability measure on (Rn , B n ) for any fixed y ∈ Λ. Furthermore, if E|g(X, Y )| < ∞ with a Borel function g, then Z E[g(X, Y )|Y = y] = E[g(X, y)|Y = y] = g(x, y)dPX|Y (x|y) a.s. PY . Rn

(ii) Let (Λ, G, P1 ) be a probability space. Suppose that P2 is a function from B n × Λ to R and satisfies (a) P2 (·, y) is a probability measure on (Rn , B n ) for any y ∈ Λ, and (b) P2 (B, ·) is Borel for any B ∈ B n . Then there is a unique probability measure P on (Rn × Λ, σ(B n × G)) such that, for B ∈ B n and C ∈ G, Z P (B × C) = P2 (B, y)dP1 (y). (1.65) C

m

m

Furthermore, if (Λ, G) = (R , B ), and X(x, y) = x and Y (x, y) = y define the coordinate random vectors, then PY = P1 , PX|Y (·|y) = P2 (·, y), and the probability measure in (1.65) is the joint distribution of (X, Y ), which has the following joint c.d.f.: Z F (x, y) = PX|Y (−∞, x]|z dPY (z), x ∈ Rn , y ∈ Rm , (1.66) (−∞,y]

where (−∞, a] denotes (−∞, a1 ] × · · · × (−∞, ak ] for a = (a1 , ..., ak ).

44

1. Probability Theory

For a fixed y, PX|Y =y = PX|Y (·|y) is called the conditional distribution of X given Y = y. Under the conditions in Theorem 1.7(i), if Y is a random m-vector and (X, Y ) has a p.d.f. w.r.t. ν × λ (ν and λ are σ-finite measures on (Rn , B n ) and (Rm , B m ), respectively), then fX|Y (x|y) defined in (1.61) is the p.d.f. of PX|Y =y w.r.t. ν for any fixed y. The second part of Theorem 1.7 states that given a distribution on one space and a collection of conditional distributions (which are conditioned on values of the first space) on another space, we can construct a joint distribution in the product space. It is sometimes called the “two-stage experiment theorem” for the following reason. If Y ∈ Rm is selected in stage 1 of an experiment according to its marginal distribution PY = P1 , and X is chosen afterward according to a distribution P2 (·, y), then the combined two-stage experiment produces a jointly distributed pair (X, Y ) with distribution P(X,Y ) given by (1.65) and PX|Y =y = P2 (·, y). This provides a way of generating dependent random variables. The following is an example. Example 1.23. A market survey is conducted to study whether a new product is preferred over the product currently available in the market (old product). The survey is conducted by mail. Questionnaires are sent along with the sample products (both new and old) to N customers randomly selected from a population, where N is a positive integer. Each customer is asked to fill out the questionnaire and return it. Responses from customers are either 1 (new is better than old) or 0 (otherwise). Some customers, however, do not return the questionnaires. Let X be the number of ones in the returned questionnaires. What is the distribution of X? If every customer returns the questionnaire, then (from elementary probability) X has the binomial distribution Bi(p, N ) in Table 1.1 (assuming that the population is large enough so that customers respond independently), where p ∈ (0, 1) is the overall rate of customers who prefer the new product. Now, let Y be the number of customers who respond. Then Y is random. Suppose that customers respond independently with the same probability π ∈ (0, 1). Then PY is the binomial distribution Bi(π, N ). Given Y = y (an integer between 0 and N ), PX|Y =y is the binomial distribution Bi(p, y) if y ≥ 1 and the point mass at 0 (see (1.22)) if y = 0. Using (1.66) and the fact that binomial distributions have p.d.f.’s w.r.t. counting measure, we obtain that the joint c.d.f. of (X, Y ) is F (x, y) =

y X

k=0

=

N k PX|Y =k (−∞, x] π (1 − π)N −k k

y min{x,k} X X k N k pj (1 − p)k−j π (1 − π)N −k j k j=0

k=0

45

1.4. Conditional Expectations

for x = 0, 1, ..., y, y = 0, 1, ..., N . The marginal c.d.f. FX (x) = F (x, ∞) = F (x, N ). The p.d.f. of X w.r.t. counting measure is N X k

N k fX (x) = p (1 − p) π (1 − π)N −k x k k=x k−x N −k N X N − x π − πp 1−π N x N −x = (πp) (1 − πp) 1 − πp 1 − πp k−x x k=x N = (πp)x (1 − πp)N −x x x

k−x

for x = 0, 1, ..., N . It turns out that the marginal distribution of X is the binomial distribution Bi(πp, N ).

1.4.4 Markov chains and martingales As applications of conditional expectations, we introduce here two important types of dependent sequences of random variables. Markov chains A sequence of random vectors {Xn : n = 1, 2, ...} is said to be a Markov chain or Markov process if and only if P (B|X1 , ..., Xn ) = P (B|Xn ) a.s., B ∈ σ(Xn+1 ), n = 2, 3, ....

(1.67)

Comparing (1.67) with (1.64), we conclude that (1.67) implies that Xn+1 (tomorrow) is conditionally independent of (X1 , ..., Xn−1 ) (the past), given Xn (today). But (X1 , ..., Xn−1 ) is not necessarily independent of (Xn , Xn+1 ). Clearly, a sequence of independent random vectors forms a Markov chain since, by Proposition 1.11, both quantities on two sides of (1.67) are equal to P (B) for independent Xi ’s. The following example describes some Markov processes of dependent random variables. Example 1.24 (First-order autoregressive processes). Let ε1 , ε2 , ... be independent random variables defined on a probability space, X1 = ε1 , and Xn+1 = ρXn + εn+1 , n = 1, 2, ..., where ρ is a constant in R. Then {Xn } is called a first-order autoregressive process. We now show that for any B ∈ B and n = 1, 2, ..., P (Xn+1 ∈ B|X1 , ..., Xn ) = Pεn+1 (B − ρXn ) = P (Xn+1 ∈ B|Xn ) a.s.,

46

1. Probability Theory

where B − y = {x ∈ R : x + y ∈ B}, which implies that {Xn } is a Markov chain. For any y ∈ R, Z Pεn+1 (B − y) = P (εn+1 + y ∈ B) = IB (x + y)dPεn+1 (x) and, by Fubini’s theorem, Pεn+1 (B − y) is Borel. Hence, Pεn+1 (B − ρXn ) is Borel w.r.t. σ(Xn ) and, thus, is Borel w.r.t. σ(X1 , ..., Xn ). Let Bj ∈ B, j = 1, ..., n, and A = ∩nj=1 Xj−1 (Bj ). Since εn+1 + ρXn = Xn+1 and εn+1 is independent of (X1 , ..., Xn ), it follows from Theorem 1.2 and Fubini’s theorem that Z Z Z Pεn+1 (B − ρXn )dP = dPεn+1 (t)dPX (x) A

xj ∈Bj ,j=1,...,n

= =

Z

t∈B−ρxn

xj ∈Bj ,j=1,...,n,xn+1 ∈B −1 (B) , P A ∩ Xn+1

dP(X,εn+1 ) (x, t)

where X and x denote (X1 , ..., Xn ) and (x1 , ..., xn ), respectively, and xn+1 denotes ρxn + t. Using this and the argument in the end of the proof for Proposition 1.11, we obtain P (Xn+1 ∈ B|X1 , ..., Xn ) = Pεn+1 (B − ρXn ) a.s. The proof for Pεn+1 (B − ρXn ) = P (Xn+1 ∈ B|Xn ) a.s. is similar and simpler. The following result provides some characterizations of Markov chains. Proposition 1.12. A sequence of random vectors {Xn } is a Markov chain if and only if one of the following three conditions holds. (a) For any n = 2, 3, ... and any integrable h(Xn+1 ) with a Borel function h, E[h(Xn+1 )|X1 , ..., Xn ] = E[h(Xn+1 )|Xn ] a.s. (b) For any n = 1, 2, ... and B ∈ σ(Xn+1 , Xn+2 , ...), P (B|X1 , ..., Xn ) = P (B|Xn ) a.s. (c) For any n = 2, 3, ..., A ∈ σ(X1 , ..., Xn ), and B ∈ σ(Xn+1 , Xn+2 , ...), P (A ∩ B|Xn ) = P (A|Xn )P (B|Xn ) a.s. Proof. (i) It is clear that (a) implies (1.67). If h is a simple function, then (1.67) and Proposition 1.10(iii) imply (a). If h is nonnegative, then by Exercise 17 there are nonnegative simple functions h1 ≤ h2 ≤ · · · ≤ h such that hj → h. Then (1.67) together with Proposition 1.10(iii) and (x) imply (a). Since h = h+ − h− , we conclude that (1.67) implies (a). (ii) It is also clear that (b) implies (1.67). We now show that (1.67) implies (b). Note that σ(Xn+1 , Xn+2 , ...) = σ ∪∞ j=1 σ(Xn+1 , ..., Xn+j ) (Exercise 19). Hence, it suffices to show that P (B|X1 , ..., Xn ) = P (B|Xn ) a.s. for B ∈ σ(Xn+1 , ..., Xn+j ) for any j = 1, 2, .... We use induction. The result for j = 1 follows from (1.67). Suppose that the result holds for any B ∈

47

1.4. Conditional Expectations

σ(Xn+1 , ..., Xn+j ). To show the result for any B ∈ σ(Xn+1 , ..., Xn+j+1 ), it is enough (why?) to show that for any B1 ∈ σ(Xn+j+1 ) and any B2 ∈ σ(Xn+1 , ..., Xn+j ), P (B1 ∩ B2 |X1 , ..., Xn ) = P (B1 ∩ B2 |Xn ) a.s. From the proof in (i), the induction assumption implies E[h(Xn+1 , ..., Xn+j )|X1 , ..., Xn ] = E[h(Xn+1 , ..., Xn+j )|Xn ]

(1.68)

for any Borel function h. The result follows from E(IB1 IB2 |X1 , ..., Xn ) = E[E(IB1 IB2 |X1 , ..., Xn+j )|X1 , ..., Xn ] = E[IB2 E(IB1 |X1 , ..., Xn+j )|X1 , ..., Xn ] = E[IB2 E(IB1 |Xn+j )|X1 , ..., Xn ]

= E[IB2 E(IB1 |Xn+j )|Xn ] = E[IB2 E(IB1 |Xn , ..., Xn+j )|Xn ] = E[E(IB1 IB2 |Xn , ..., Xn+j )|Xn ] = E(IB1 IB2 |Xn ) a.s.,

where the first and last equalities follow from Proposition 1.10(v), the second and sixth equalities follow from Proposition 1.10(vi), the third and fifth equalities follow from (1.67), and the fourth equality follows from (1.68). (iii) Let A ∈ σ(X1 , ..., Xn ) and B ∈ σ(Xn+1 , Xn+2 , ...). If (b) holds, then E(IA IB |Xn ) = E[E(IA IB |X1 , ..., Xn )|Xn ] = E[IA E(IB |X1 , ..., Xn )|Xn ] = E[IA E(IB |Xn )|Xn ] = E(IA |Xn )E(IB |Xn ), which is (c). Assume that (c) holds. Let A1 ∈ σ(Xn ), A2 ∈ σ(X1 , ..., Xn−1 ), and B ∈ σ(Xn+1 , Xn+2 , ...). Then Z Z E(IB |Xn )dP = IA2 E(IB |Xn )dP A1 ∩A2 A1 Z = E[IA2 E(IB |Xn )|Xn ]dP A1 Z = E(IA2 |Xn )E(IB |Xn )dP A1 Z = E(IA2 IB |Xn )dP A1

= P (A1 ∩ A2 ∩ B).

Since disjoint unions of events of the form A1 ∩ A2 as specified above generate σ(X1 , ..., Xn ), this shows that E(IB |Xn ) = E(IB |X1 , ..., Xn ) a.s., which is (b). Note that condition (b) in Proposition 1.12 can be stated as “the past and the future are conditionally independent given the present”, which is a property of any Markov chain. More discussions and applications of Markov chains can be found in §4.1.4.

48

1. Probability Theory

Martingales Let {Xn } be a sequence of integrable random variables on a probability space (Ω, F , P ) and F1 ⊂ F2 ⊂ · · · ⊂ F be a sequence of σ-fields such that σ(Xn ) ⊂ Fn , n = 1, 2, .... The sequence {Xn , Fn : n = 1, 2, ...} is said to be a martingale if and only if E(Xn+1 |Fn ) = Xn a.s., n = 1, 2, ...,

(1.69)

a submartingale if and only if (1.69) holds with = replaced by ≥, and a supermartingale if and only if (1.69) holds with = replaced by ≤. {Xn } is said to be a martingale (submartingale or supermartingale) if and only if {Xn , σ(X1 , ..., Xn )} is a martingale (submartingale or supermartingale). From Proposition 1.10(v), if {Xn , Fn } is a martingale (submartingale or supermartingale), then so is {Xn }. A simple property of a martingale (or a submartingale) {Xn , Fn } is that E(Xn+j |Fn ) = Xn a.s. (or E(Xn+j |Fn ) ≥ Xn a.s.) and EX1 = EXj (or EX1 ≤ EX2 ≤ · · ·) for any j = 1, 2, ... (exercise). For any probability space (Ω, F, P ) and σ-fields F1 ⊂ F2 ⊂ · · · ⊂ F, we can always construct a martingale {E(Y |Fn )} by using an integrable random variable Y . Another way to construct a martingale is to use a sequence of independent integrable random variables {εn } by letting Xn = ε1 + · · · + εn , n = 1, 2, .... Since E(Xn+1 |X1 , ..., Xn ) = E(Xn + εn+1 |X1 , ..., Xn ) = Xn + Eεn+1 a.s., {Xn } is a martingale if Eεn = 0 for all n, a submartingale if Eεn ≥ 0 for all n, and a supermartingale if Eεn ≤ 0 for all n. Note that in Example 1.24 with ρ = 1, {Xn } is shown to be a Markov chain. The next example provides another example of martingales. Example 1.25 (Likelihood ratio). Let (Ω, F , P ) be a probability space, Q be a probability measure on F, and F1 ⊂ F2 ⊂ · · · ⊂ F be a sequence of σ-fields. Let Pn and Qn be P and Q restricted to Fn , respectively, n = 1, 2, .... Suppose that Qn ≪ Pn for each n. Then {Xn , Fn } is a martingale, where Xn = dQn /dPn (the Radon-Nikodym derivative of Qn w.r.t. Pn ), n = 1, 2, ... (exercise). Suppose now that {Yn } is a sequence of random variables on (Ω, F , P ), Fn = σ(Y1 , ..., Yn ) and that there exists a σfinite measure νn on Fn such that Pn ≪ νn and νn ≪ Pn , n = 1, 2, .... Let pn (Y1 , ..., Yn ) = dPn /dνn and qn (Y1 , ..., Yn ) = dQn /dνn . By Proposition 1.7(iii), Xn = qn (Y1 , ..., Yn )/pn (Y1 , ..., Yn ), which is called a likelihood ratio in statistical terms. The following results contain some useful properties of martingales and submartingales.

1.5. Asymptotic Theory

49

Proposition 1.13. Let ϕ be a convex function on R. (i) If {Xn , Fn } is a martingale and ϕ(Xn ) is integrable for all n, then {ϕ(Xn ), Fn } is a submartingale. (ii) If {Xn , Fn } is a submartingale, ϕ(Xn ) is integrable for all n, and ϕ is nondecreasing, then {ϕ(Xn ), Fn } is a submartingale. Proof. (i) Note that ϕ(Xn ) = ϕ(E(Xn+1 |Fn )) ≤ E[ϕ(Xn+1 |Fn )] a.s. by Jensen’s inequality for conditional expectations (Exercise 89(c)). (ii) Since ϕ is nondecreasing and {Xn , Fn } is a submartingale, ϕ(Xn ) ≤ ϕ(E(Xn+1 |Fn )) ≤ E[ϕ(Xn+1 |Fn )] a.s. An application of Proposition 1.13 shows that if {Xn , Fn } is a submartingale, then so is {(Xn )+ , Fn }; if {Xn , Fn } is a martingale, then {|Xn |, Fn } is a submartingale and so are {|Xn |p , Fn }, where p > 1 is a constant, and {|Xn |(log |Xn |)+ , Fn }, provided that |Xn |p and |Xn |(log |Xn |)+ are integrable for all n. Proposition 1.14 (Doob’s decomposition). Let {Xn , Fn } be a submartingale. Then Xn = Yn + Zn , n = 1, 2, ..., where {Yn , Fn } is a martingale, 0 = Z1 ≤ Z2 ≤ · · ·, and EZn < ∞ for all n. Furthermore, if supn E|Xn | < ∞, then supn E|Yn | < ∞ and supn EZn < ∞. Proof. Define η1 = ξ1 , ζ1 = 0, ηn = Xn −Xn−1 −E(X P n −Xn−1 |Fn−1 P), and ζn = E(Xn − Xn−1 |Fn−1 ) for n ≥ 2. Then Yn = ni=1 ηi and Zn = ni=1 ζi satisfy Xn = Yn + Zn and the required conditions (exercise). Assume now that supn E|Xn | < ∞. Since EY1 = EYn for any n and Zn ≤ |Xn | − Yn , EZn ≤ E|Xn | − EY1 . Hence supn EZn < ∞. Also, |Yn | ≤ |Xn | + Zn . Hence supn E|Yn | < ∞. The following martingale convergence theorem, due to Doob, has many applications (see, e.g., Example 1.27 in §1.5.1). Its proof can be found, for example, in Billingsley (1986, pp. 490-491). Proposition 1.15. Let {Xn , Fn } be a submartingale. If c = supn E|Xn | < ∞, then limn→∞ Xn = X a.s., where X is a random variable satisfying E|X| ≤ c.

1.5 Asymptotic Theory Asymptotic theory studies limiting behavior of random variables (vectors) and their distributions. It is an important tool for statistical analysis. A more complete coverage of asymptotic theory in statistical analysis can be found in Serfling (1980), Shorack and Wellner (1986), Sen and Singer (1993), Barndorff-Nielsen and Cox (1994), and van der Vaart (1998).

50

1. Probability Theory

1.5.1 Convergence modes and stochastic orders There are several convergence modes for random variables/vectors. Let r > 0 be a constant. For any c = (c1 , ..., ck ) ∈ Rk , we define kckr = Pk ( j=1 |cj |r )1/r . If r ≥ 1, then kckr is the Lr -distance between 0 and c. √ When r = 2, the subscript r is omitted and kck = kck2 = cτ c. Definition 1.8. Let X, X1 , X2 , . . . be random k-vectors defined on a probability space. (i) We say that the sequence {Xn } converges to X almost surely (a.s.) and write Xn →a.s. X if and only if limn→∞ Xn = X a.s. (ii) We say that {Xn } converges to X in probability and write Xn →p X if and only if, for every fixed ǫ > 0, lim P (kXn − Xk > ǫ) = 0.

n→∞

(1.70)

(iii) We say that {Xn } converges to X in Lr (or in rth moment) and write Xn →Lr X if and only if lim EkXn − Xkrr = 0,

n→∞

where r > 0 is a fixed constant. (iv) Let F , Fn , n = 1, 2, ..., be c.d.f.’s on Rk and P , Pn , n = 1, ..., be their corresponding probability measures. We say that {Fn } converges to F weakly (or {Pn } converges to P weakly) and write Fn →w F (or Pn →w P ) if and only if, for each continuity point x of F , lim Fn (x) = F (x).

n→∞

We say that {Xn } converges to X in distribution (or in law) and write Xn →d X if and only if FXn →w FX . The a.s. convergence has already been considered in previous sections. The concept of convergence in probability, convergence in Lr , or a.s. convergence represents a sense in which, for n sufficiently large, Xn and X approximate each other as functions on the original probability space. The concept of convergence in distribution in Definition 1.8(iv), however, depends only on the distributions FXn and FX (or probability measures PXn and PX ) and does not necessitate that Xn and X are close in any sense; in fact, Definition 1.8(iv) still makes sense even if X and Xn ’s are not defined on the same probability space. In Definition 1.8(iv), it is not required that limn→∞ Fn (x) = F (x) for every x. However, if F is a continuous function, then we have the following stronger result.

51

1.5. Asymptotic Theory

Proposition 1.16 (P´ olya’s theorem). If Fn →w F and F is continuous on Rk , then lim sup |Fn (x) − F (x)| = 0. n→∞ x∈Rk

A useful characterization of a.s. convergence is given in the following lemma. Lemma 1.4. For random k-vectors X, X1 , X2 , . . . on a probability space, Xn →a.s. X if and only if for every ǫ > 0, ! ∞ [ {kXm − Xk > ǫ} = 0. (1.71) lim P n→∞

∪∞ n=1

m=n

∩∞ m=n

Proof. Let Aj = {kXm − Xk ≤ j −1 }, j = 1, 2, .... By Proposition 1.1(iii) and DeMorgan’s law, (1.71) holds for every ǫ > 0 if and only if P (Aj ) = 1 for every j, which is equivalent to P (∩∞ j=1 Aj ) = 1. The result follows from ∩∞ A = {ω : lim X (ω) = X(ω)} (exercise). n→∞ n j=1 j The following result describes the relationship among the four convergence modes in Definition 1.8. Theorem 1.8. Let X, X1 , X2 , . . . be random k-vectors. (i) If Xn →a.s. X, then Xn →p X. (ii) If Xn →Lr X for an r > 0, then Xn →p X. (iii) If Xn →p X, then Xn →d X. (iv) (Skorohod’s theorem). If Xn →d X, then there are random vectors Y, Y1 , Y2 , ... defined on a common probability space such that PY = PX , PYn = PXn , n = 1, 2,...,Pand Yn →a.s. Y . ∞ (v) If, for every ǫ > 0, n=1 P (kXn − Xk ≥ ǫ) < ∞, then Xn →a.s. X. (vi) If Xn →p X, then there is a subsequence {Xnj , j = 1, 2, ...} such that Xnj →a.s. X as j → ∞. (vii) If Xn →d X and P (X = c) = 1, where c ∈ Rk is a constant vector, then Xn →p c. (viii) Suppose that Xn →d X. Then, for any r > 0, lim EkXn krr = EkXkrr < ∞

n→∞

if and only if {kXn krr } is uniformly integrable in the sense that lim sup E kXn krr I{kXn kr >t} = 0. t→∞ n

(1.72)

(1.73)

The proof of Theorem 1.8 is given after the following discussion and example.

52

1. Probability Theory

The converse of Theorem 1.8(i), (ii), or (iii) is generally not true (see Example 1.26 and Exercise 116). Note that part (iv) of Theorem 1.8 (Skorohod’s theorem) is not a converse of part (i), but it is an important result in probability theory. It is useful when we study convergence of quantities related to FXn and FX when Xn →d X (see, e.g., the proofs of Theorems 1.8 and 1.9). Part (v) of Theorem 1.8 indicates that the converse of part (i) is true under the additional condition that P (kXn − Xk ≥ ǫ) tends to 0 fast enough. Part (vi) provides a partial converse of part (i) whereas part (vii) is a partial converse of part (iii). A consequence of Theorem 1.8(viii) is that if Xn →p X and {kXn − Xkrr } is uniformly integrable, then Xn →Lr X; i.e., the converse of Theorem 1.8(ii) is true under the additional condition of uniform integrability. A useful sufficient condition for uniform integrability of {kXn krr } is that sup EkXn kr+δ 0. Some other sufficient conditions are given in Exercises 117-120. Example 1.26. Let θn = 1 + n−1 and Xn be a random variable having the exponential distribution E(0, θn ) (Table 1.2), n = 1, 2, .... Let X be a random variable having the exponential distribution E(0, 1). For any x > 0, FXn (x) = 1 − e−x/θn → 1 − e−x = FX (x)

as n → ∞. Since FXn (x) ≡ 0 ≡ FX (x) for x ≤ 0, we have shown that Xn →d X. Is it true that Xn →p X? This question cannot be answered without any further information about the random variables X and Xn . We consider two cases in which different answers can be obtained. First, suppose that Xn ≡ θn X (then Xn has the given c.d.f.). Note that Xn − X = (θn − 1)X = n−1 X, which has the c.d.f. (1 − e−nx )I[0,∞) (x). Hence P (|Xn − X| ≥ ǫ) = e−nǫ → 0 for any ǫ > 0. In fact, by Theorem 1.8(v), Xn →a.s. X; since E|Xn − X|p = n−p EX p < ∞ for any p > 0, Xn →Lp X for any p > 0. Next, suppose that Xn and X are independent random variables. Using result (1.28) and the fact that the p.d.f.’s for Xn and −X are θn−1 e−x/θn I(0,∞) (x) and ex I(−∞,0) (x), respectively, we obtain that Z ǫZ θn−1 e−x/θn ey−x I(0,∞) (x)I(−∞,x) (y)dxdy, P (|Xn − X| ≤ ǫ) = −ǫ

which converges to (by the dominated convergence theorem) Z ǫ Z e−x ey−x I(0,∞) (x)I(−∞,x) (y)dxdy = 1 − e−ǫ . −ǫ

53

1.5. Asymptotic Theory

Thus, P (|Xn − X| ≥ ǫ) → e−ǫ > 0 for any ǫ > 0 and, therefore, {Xn } does not converge to X in probability. The previous discussion, however, indicates how to construct the random variables Yn and Y in Theorem 1.8(iv) for this example. The following famous result is used in the proof of Theorem 1.8(v). Its proof is left to the reader. Lemma 1.5. (Borel-Cantelli lemma). Let An be a sequence of events in a ∞ ∞ probability P∞ space and lim supn An = ∩n=1 ∪m=n Am . (i) If n=1 P (An ) < ∞, then P (lim supn An ) = 0. P∞ (ii) If A1 , A2 , ... are pairwise independent and n=1 P (An ) = ∞, then P (lim supn An ) = 1. Proof of Theorem 1.8. (i) The result follows from Lemma 1.4, since (1.71) implies (1.70). (ii) The result follows from Chebyshev’s inequality with ϕ(t) = |t|r . (iii) For any c = (c1 , ..., ck ) ∈ Rk , define (−∞, c] = (−∞, c1 ]×· · ·×(−∞, ck ]. Let x be a continuity point of FX , ǫ > 0 be given, and Jk be the k-vector of ones. Then {X ∈ (−∞, x − ǫJk ], Xn 6∈ (−∞, x]} ⊂ {kXn − Xk > ǫ} and FX (x − ǫJk ) = P X ∈ (−∞, x − ǫJk ] ≤ P Xn ∈ (−∞, x] + P X ∈ (−∞, x − ǫJk ], Xn 6∈ (−∞, x] ≤ FXn (x) + P (kXn − Xk > ǫ) . Letting n → ∞, we obtain that FX (x − ǫJk ) ≤ lim inf n FXn (x). Similarly, we can show that FX (x + ǫJk ) ≥ lim supn FXn (x). Since ǫ is arbitrary and FX is continuous at x, FX (x) = limn→∞ FXn (x). (iv) The proof of this part can be found in Billingsley (1986, pp. 399-402). (v) Let An = {kXn − Xk ≥ ǫ}. The result follows from Lemma 1.4, Lemma 1.5(i), and Proposition 1.1(iii). (vi) From (1.70), for every j = 1, 2, ..., there is a positive integer nj such that P (kXnj − Xk > 2−j ) < 2−j . For any ǫ > 0, there is aP kǫ such that for −j j ≥ kǫ , P (kXnj − Xk > ǫ) < P (kXnj − Xk > 2−j ). Since ∞ = 1, it j=1 2 follows from the result in (v) that Xnj →a.s. X as j → ∞. (vii) The proof for this part is left as an exercise. (viii) First, by part (iv), we may assume that Xn →a.s. X (why?). Assume that {kXn krr } is uniformly integrable. Then supn EkXn krr < ∞ (why?) and by Fatou’s lemma (Theorem 1.1(i)), EkXkrr ≤ lim inf n EkXn krr < ∞. Hence, (1.72) follows if we can show that lim sup EkXn krr ≤ EkXkrr .

(1.75)

n

For any ǫ > 0 and t > 0, let An = {kXn −Xkr ≤ ǫ} and Bn = {kXn kr > t}.

54

1. Probability Theory

Then EkXn krr = E(kXn krr IAcn ∩Bn ) + E(kXn krr IAcn ∩Bnc ) + E(kXn krr IAn ) ≤ E(kXn krr IBn ) + tr P (Acn ) + EkXn IAn krr .

For r ≤ 1, kXn IAn krr ≤ (kXn − Xkrr + kXkrr )IAn and EkXn IAn krr ≤ E[(kXn − Xkrr + kXkrr )IAn ] ≤ ǫr + EkXkrr . For r > 1, an application of Minkowski’s inequality leads to EkXn IAn krr = Ek(Xn − X)IAn + XIAn krr r ≤ E [k(Xn − X)IAn kr + kXIAn kr ] or n 1/r 1/r ≤ [Ek(Xn − X)IAn krr ] + [EkXIAn krr ] or n ≤ ǫ + [EkXkrr ]1/r .

In any case, since ǫ is arbitrary, lim supn EkXn IAn krr ≤ EkXkrr . This result and the previously established inequality imply that lim sup EkXn krr ≤ lim sup E(kXn krr IBn ) + tr lim P (Acn ) n

n→∞

n

≤

+ lim sup EkXn IAn krr n sup E(kXn krr I{kXn kr >t} ) n

+ EkXkrr ,

since P (Acn ) → 0. Since {kXn krr } is uniformly integrable, letting t → ∞ we obtain (1.75). Suppose now that (1.72) holds. Let ξn = kXn krr IBnc − kXkrr IBnc . Then ξn →a.s. 0 and |ξn | ≤ tr + kXkrr , which is integrable. By the dominated convergence theorem, Eξn → 0; this and (1.72) imply that E(kXn krr IBn ) − E(kXkrr IBn ) → 0. From the definition of Bn , Bn ⊂ {kXn − Xkr > t/2} ∪ {kXkr > t/2}. Since EkXkrr < ∞, it follows from the dominated convergence theorem that E(kXkrr I{kXn −Xkr >t/2} ) → 0 as n → ∞. Hence, lim sup E(kXn krr IBn ) ≤ lim sup E(kXkrr IBn ) ≤ E(kXkrr I{kXkr >t/2} ). n

n

Letting t → ∞, it follows from the dominated convergence theorem that lim lim sup E(kXn krr IBn ) ≤ lim E(kXkrr I{kXkr >t/2} ) = 0.

t→∞

n

This proves (1.73).

t→∞

1.5. Asymptotic Theory

55

Example 1.27. As an application of Theorem 1.8(viii) and Proposition 1.15, we consider again the prediction problem in Example 1.22. Suppose that we predict a random variable X by a random n-vector Y = (Y1 , ..., Yn ). It is shown in Example 1.22 that Xn = E(X|Y1 , ..., Yn ) is the best predictor in terms of the mean squared prediction error, when EX 2 < ∞. We now show that Xn →a.s. X when n → ∞ under the assumption that σ(X) ⊂ F∞ = σ(Y1 , Y2 , ...) (i.e., X provides no more information than Y1 , Y2 , ...). From the discussion in §1.4.4, {Xn } is a martingale. Also, supn E|Xn | ≤ supn E[E(|X||Y1 , ..., Yn )] = E|X| < ∞. Hence, by Proposition 1.15, Xn →a.s. Z for some random variable Z. We now need to show Z = X a.s. Since σ(X) ⊂ F∞ , X = E(X|F∞ ) a.s. Hence, it suffices to show that Z = E(X|F∞ ) a.s. Since EXn2 ≤ EX 2 < ∞ (why?), condition (1.74) holds for sequence {|Xn |} and, hence, {|Xn |} is uniformly integrable. By Theorem R R 1.8(viii), E|Xn − Z| → 0, which implies A Xn dP → A ZdP for any event A. for n ≥ m and R Note thatR if A ∈ σ(Y1 , ..., Ym ), then A ∈ σ(Y1 , ..., Yn ) ∞ X dP = XdP . This implies that for any A ∈ ∪ n j=1 σ(Y1 , ..., Yj ), R A RA ∞ XdP = ZdP . Since ∪ σ(Y , ..., Y ) generates F 1 j ∞ , we conclude j=1 A A R R that A XdP = A ZdP for any A ∈ F∞ and thus Z = E(X|F∞ ) a.s. In the proof above, the condition EX 2 < ∞ is used only for showing the uniform integrability of {|Xn |}. But by Exercise 120, {|Xn |} is uniformly integrable as long as E|X| < ∞. Hence Xn →a.s. X is still true if the condition EX 2 < ∞ is replaced by E|X| < ∞. We now introduce the notion of O( · ), o( · ), and stochastic O( · ) and o( · ). In calculus, two sequences of real numbers, {an } and {bn }, satisfy an = O(bn ) if and only if |an | ≤ c|bn | for all n and a constant c; and an = o(bn ) if and only if an /bn → 0 as n → ∞. Definition 1.9. Let X1 , X2 , ... be random vectors and Y1 , Y2 , ... be random variables defined on a common probability space. (i) Xn = O(Yn ) a.s. if and only if P (kXn k = O(|Yn |)) = 1. (ii) Xn = o(Yn ) a.s. if and only if Xn /Yn →a.s. 0. (iii) Xn = Op (Yn ) if and only if, for any ǫ > 0, there is a constant Cǫ > 0 such that supn P (kXn k ≥ Cǫ |Yn |) < ǫ. (iv) Xn = op (Yn ) if and only if Xn /Yn →p 0. Note that Xn = op (Yn ) implies Xn = Op (Yn ); Xn = Op (Yn ) and Yn = Op (Zn ) implies Xn = Op (Zn ); but Xn = Op (Yn ) does not imply Yn = Op (Xn ). The same conclusion can be obtained if Op ( · ) and op ( · ) are replaced by O( · ) a.s. and o( · ) a.s., respectively. Some results related to Op are given in Exercise 127. For example, if Xn →d X for a random variable X, then Xn = Op (1). Since an = O(1) means that {an } is bounded, {Xn } is said to be bounded in probability if Xn = Op (1).

56

1. Probability Theory

1.5.2 Weak convergence We now discuss more about convergence in distribution or weak convergence of probability measures. A sequence {Pn } of probability measures on (Rk , B k ) is tight if for every ǫ > 0, there is a compact set C ⊂ Rk such that inf n Pn (C) > 1 − ǫ. If {Xn } is a sequence of random k-vectors, then the tightness of {PXn } is the same as the boundedness of {kXn k} in probability. The proof of the following result can be found in Billingsley (1986, pp. 392-395). Proposition 1.17. Let {Pn } be a sequence of probability measures on (Rk , B k ). (i) Tightness of {Pn } is a necessary and sufficient condition that for every subsequence {Pni } there exists a further subsequence {Pnj } ⊂ {Pni } and a probability measure P on (Rk , B k ) such that Pnj →w P as j → ∞. (ii) If {Pn } is tight and if each subsequence that converges weakly at all converges to the same probability measure P , then Pn →w P . The following result gives some useful sufficient and necessary conditions for convergence in distribution. Theorem 1.9. Let X, X1 , X2 , . . . be random k-vectors. (i) Xn →d X is equivalent to any one of the following conditions: (a) E[h(Xn )] → E[h(X)] for every bounded continuous function h; (b) lim supn PXn (C) ≤ PX (C) for any closed set C ⊂ Rk ; (c) lim inf n PXn (O) ≥ PX (O) for any open set O ⊂ Rk . (ii) (L´evy-Cram´er continuity theorem). Let φX , φX1 , φX2 , ... be the ch.f.’s of X, X1 , X2 , ..., respectively. Xn →d X if and only if limn→∞ φXn (t) = φX (t) for all t ∈ Rk . (iii) (Cram´er-Wold device). Xn →d X if and only if cτ Xn →d cτ X for every c ∈ Rk . Proof. (i) First, we show Xn →d X implies (a). By Theorem 1.8(iv) (Skorohod’s theorem), there exists a sequence of random vectors {Yn } and a random vector Y such that PYn = PXn for all n, PY = PX and Yn →a.s. Y . For bounded continuous h, h(Yn ) →a.s. h(Y ) and, by the dominated convergence theorem, E[h(Yn )] → E[h(Y )]. Then (a) follows from E[h(Xn )] = E[h(Yn )] for all n and E[h(X)] = E[h(Y )]. Next, we show (a) implies (b). Let C be a closed set and fC (x) = inf{kx − yk : y ∈ C}. Then fC is continuous. For j = 1, 2, ..., define ϕj (t) = I(−∞,0] + (1 − jt)I(0,j −1 ] . Then hj (x) = ϕj (fC (x)) is continuous and bounded, hj ≥ hj+1 , j = 1, 2, ..., and hj (x) → IC (x) as j → ∞. Hence lim supn PXn (C) ≤ limn→∞ E[hj (Xn )] = E[hj (X)] for each j (by (a)). By the dominated convergence theorem, E[hj (X)] → E[IC (X)] = PX (C).

1.5. Asymptotic Theory

57

This proves (b). For any open set O, Oc is closed. Hence, (b) is equivalent to (c). Now, we show (b) and (c) imply Xn →d X. For x = (x1 , ..., xk ) ∈ Rk , let (−∞, x] = (−∞, x1 ] × · · · × (−∞, xk ] and (−∞, x) = (−∞, x1 ) × · · · × (−∞, xk ). From (b) and (c), PX (−∞, x) ≤ lim inf n PXn (−∞, x) ≤ liminf n FXn (x) ≤ lim supn FXn (x) = lim supn PXn (−∞, x] ≤ PX (−∞, x] = FX (x). If x is a continuity point of FX , then PX (−∞, x) = FX (x). This proves Xn →d X and completes the proof of (i). √ τ (ii) From (a) of part (i), Xn →d X implies φXn (t) → φX (t), since e −1t x = √ cos(tτ x) + −1 sin(tτ x) and cos(tτ x) and sin(tτ x) are bounded continuous functions for any fixed t. Suppose now that k = 1 and that φXn (t) → φX (t) for every t ∈ R. By Fubini’s theorem, Z Z ∞ Z u √ 1 u 1 −1tx [1 − φXn (t)]dt = (1 − e )dt dPXn (x) u −u −∞ u −u Z ∞ sin ux dPXn (x) 1− =2 ux −∞ Z 1 dPXn (x) 1− ≥2 |ux| {|x|>2u−1 } ≥ PXn (−∞, −2u−1) ∪ (2u−1 , ∞) for any u > 0. Since φX is continuous at 0 and φX (0) = 1, for any ǫ > 0 Ru there is a u > 0 such that u−1 −u [1 − φX (t)]dt < ǫ/2. Since φXn → φX , Ru by the dominated convergence theorem, supn {u−1 −u [1 − φXn (t)]dt} < ǫ. Hence, Z u 1 inf PXn [−2u−1 , 2u−1 ] ≥ 1 − sup [1 − φXn (t)]dt ≥ 1 − ǫ, n u −u n

i.e., {PXn } is tight. Let {PXnj } be any subsequence that converges to a probability measure P . By the first part of the proof, φXnj → φ, which is the ch.f. of P . By the convergence of φXn , φ = φX . By Theorem 1.6(i), P = PX . By Proposition 1.17(ii), Xn →d X. Consider now the case where k ≥ 2 and φXn → φX . Let Ynj be the jth component of Xn and Yj be the jth component of X. Then φYnj → φYj for each j. By the proof for the case of k = 1, Ynj →d Yj . By Proposition 1.17(i), {PYnj } is tight, j = 1, ..., k. This implies that {PXn } is tight (why?). Then the proof for Xn →d X is the same as that for the case of k = 1. (iii) From (1.52), φcτ Xn (u) = φXn (uc) and φcτ X (u) = φX (uc) for any u ∈ R and any c ∈ Rk . Hence, convergence of φXn to φX is equivalent to convergence of φcτ Xn to φcτ X for every c ∈ Rk . Then the result follows from part (ii).

58

1. Probability Theory

Example 1.28. Let X1 , ..., Xn be independent random variables having a common c.d.f. and Tn = X1 + · · · + Xn , n = 1, 2, .... Suppose that E|X1 | < ∞. It follows from (1.56) and a result in calculus that the ch.f. of X1 satisfies √ φX1 (t) = φX1 (0) + −1µt + o(|t|) as |t| → 0, where µ = EX1 . From (1.52) and (1.58), the ch.f. of Tn /n is √ n n −1µt t t +o φTn /n (t) = φX1 = 1+ n n n for any t ∈ R, as n → ∞. Since (1 + cn /n)n → ec for any complex sequence √ {cn } satisfying cn → c, we obtain that φTn /n (t) → e −1µt , which is the ch.f. of the distribution degenerated at µ (i.e., the point mass probability measure at µ; see (1.22)). By Theorem 1.9(ii), Tn /n →d µ. From Theorem 1.8(vii), this also shows that Tn /n →p µ. Similarly, µ = 0 and σ 2 = Var(X1 ) < ∞ imply 2 n t σ 2 t2 +o φTn /√n (t) = 1 − 2n n 2 2

for any t ∈ R, which implies that φTn /√n (t) → e−σ t /2 , the ch.f. of √ N (0, σ 2 ). Hence Tn / n →d N (0, σ 2 ). (Recall that N (µ, σ 2 ) denotes a random variable having the N (µ, σ 2 ) distribution.) If µ 6= 0, a transforma√ tion of Yi = Xi − µ leads to (Tn − nµ)/ n →d N (0, σ 2 ). Suppose now that X1 , ..., Xn are random k-vectors and µ = EX1 and Σ = Var(X1 ) are finite. For any√fixed c ∈ Rk , it follows from the previous discussion that (cτ Tn − ncτ µ)/ n →d N (0, cτ Σc). From Theorem 1.9(iii) and a property √ of the normal distribution (Exercise 81), we conclude that (Tn − nµ)/ n →d Nk (0, Σ). Example 1.28 shows that Theorem 1.9(ii) together with some properties of ch.f.’s can be applied to show convergence in distribution for sums of independent random variables (vectors). The following is another example. Example 1.29. Let X1 , ..., Xn be independent random variables having a common Lebesgue p.d.f. f (x) = (1 − cos x)/(πx2 ). Then the ch.f. of X1 is max{1 − |t|, 0} (Exercise 73) and the ch.f. of Tn /n = (X1 + · · · + Xn )/n is n |t| max 1 − , 0 → e−|t| , t ∈ R. n Since e−|t| is the ch.f. of the Cauchy distribution C(0, 1) (Table 1.2), we conclude that Tn /n →d X, where X has the Cauchy distribution C(0, 1). Does this result contradict the first result in Example 1.28?

1.5. Asymptotic Theory

59

Other examples of applications of Theorem 1.9 are given in Exercises 135-140 in §1.6. The following result can be used to check whether Xn →d X when X has a p.d.f. f and Xn has a p.d.f. fn . Proposition 1.18 (Scheff´e’s theorem). Let {fn } be a sequence of p.d.f.’s on Rk w.r.t. a measure ν. Suppose that R limn→∞ fn (x) = f (x) a.e. ν and f (x) is a p.d.f. w.r.t. ν. Then limn→∞ |fn (x) − f (x)|dν = 0. Proof. Let gn (x) = [f (x) − fn (x)]I{f ≥fn } (x), n = 1, 2,.... Then Z Z |fn (x) − f (x)|dν = 2 gn (x)dν. Since 0 ≤ gn (x) ≤ f (x) for all x and gn → 0 a.e. ν, the result follows from the dominated convergence theorem. As an example, consider the Lebesgue p.d.f. fn of the t-distribution tn (Table 1.2), n = 1, 2,.... One can show (exercise) that fn → f , where f is the standard normal p.d.f. This is an important result in statistics.

1.5.3 Convergence of transformations Transformation is an important tool in statistics. For random vectors Xn converging to X in some sense, we often want to know whether g(Xn ) converges to g(X) in the same sense. The following result provides an answer to this question in many problems. Its proof is left to the reader. Theorem 1.10. Let X, X1 , X2 , ... be random k-vectors defined on a probability space and g be a measurable function from (Rk , B k ) to (Rl , B l ). Suppose that g is continuous a.s. PX . Then (i) Xn →a.s. X implies g(Xn ) →a.s. g(X); (ii) Xn →p X implies g(Xn ) →p g(X); (iii) Xn →d X implies g(Xn ) →d g(X). Example 1.30. (i) Let X1 , X2 , ... be random variables. If Xn →d X, where X has the N (0, 1) distribution, then Xn2 →d Y , where Y has the chi-square distribution χ21 (Example 1.14). (ii) Let (Xn , Yn ) be random 2-vectors satisfying (Xn , Yn ) →d (X, Y ), where X and Y are independent random variables having the N (0, 1) distribution, then Xn /Yn →d X/Y , which has the Cauchy distribution C(0, 1) (§1.3.1). (iii) Under the conditions in part (ii), max{Xn , Yn } →d max{X, Y }, which has the c.d.f. [Φ(x)]2 (Φ(x) is the c.d.f. of N (0, 1)). In Example 1.30(ii) and (iii), the condition that (Xn , Yn ) →d (X, Y ) cannot be relaxed to Xn →d X and Yn →d Y (exercise); i.e., we need the

60

1. Probability Theory

convergence of the joint c.d.f. of (Xn , Yn ). This is different when →d is replaced by →p or →a.s. . The following result, which plays an important role in probability and statistics, establishes the convergence in distribution of Xn + Yn or Xn Yn when no information regarding the joint c.d.f. of (Xn , Yn ) is provided. Theorem 1.11 (Slutsky’s theorem). Let X, X1 , X2 , ..., Y1 , Y2 , ... be random variables on a probability space. Suppose that Xn →d X and Yn →p c, where c is a fixed real number. Then (i) Xn + Yn →d X + c; (ii) Yn Xn →d cX; (iii) Xn /Yn →d X/c if c 6= 0. Proof. We prove (i) only. The proofs of (ii) and (iii) are left as exercises. Let t ∈ R and ǫ > 0 be fixed constants. Then FXn +Yn (t) = P (Xn + Yn ≤ t)

≤ P ({Xn + Yn ≤ t} ∩ {|Yn − c| < ǫ}) + P (|Yn − c| ≥ ǫ) ≤ P (Xn ≤ t − c + ǫ) + P (|Yn − c| ≥ ǫ)

and, similarly, FXn +Yn (t) ≥ P (Xn ≤ t − c − ǫ) − P (|Yn − c| ≥ ǫ). If t − c, t − c + ǫ, and t − c − ǫ are continuity points of FX , then it follows from the previous two inequalities and the hypotheses of the theorem that FX (t − c − ǫ) ≤ lim inf FXn +Yn (t) ≤ lim sup FXn +Yn (t) ≤ FX (t − c + ǫ). n

n

Since ǫ can be arbitrary (why?), lim FXn +Yn (t) = FX (t − c).

n→∞

The result follows from FX+c (t) = FX (t − c). An application of Theorem 1.11 is given in the proof of the following important result. Theorem 1.12. Let X1 , X2 , ... and Y be random k-vectors satisfying an (Xn − c) →d Y,

(1.76)

where c ∈ Rk and {an } is a sequence of positive numbers with limn→∞ an = ∞. Let g be a function from Rk to R. (i) If g is differentiable at c, then an [g(Xn ) − g(c)] →d [∇g(c)]τ Y,

(1.77)

61

1.5. Asymptotic Theory

where ∇g(x) denotes the k-vector of partial derivatives of g at x. (ii) Suppose that g has continuous partial derivatives of order m > 1 in a neighborhood of c, with all the partial derivatives of order j, 1 ≤ j ≤ m− 1, vanishing at c, but with the mth-order partial derivatives not all vanishing at c. Then k k X 1 X ∂mg m Yi1 · · · Yim , (1.78) an [g(Xn ) − g(c)] →d ··· m! i =1 ∂x i 1 · · · ∂xim x=c i =1 1

m

where Yj is the jth component of Y . Proof. We prove (i) only. The proof of (ii) is similar. Let

Zn = an [g(Xn ) − g(c)] − an [∇g(c)]τ (Xn − c). If we can show that Zn = op (1), then by (1.76), Theorem 1.9(iii), and Theorem 1.11(i), result (1.77) holds. The differentiability of g at c implies that for any ǫ > 0, there is a δǫ > 0 such that |g(x) − g(c) − [∇g(c)]τ (x − c)| ≤ ǫkx − ck (1.79) whenever kx − ck < δǫ . Let η > 0 be fixed. By (1.79),

P (|Zn | ≥ η) ≤ P (kXn − ck ≥ δǫ ) + P (an kXn − ck ≥ η/ǫ). Since an → ∞, (1.76) and Theorem 1.11(ii) imply Xn →p c. By Theorem 1.10(iii), (1.76) implies an kXn − ck →d kY k. Without loss of generality, we can assume that η/ǫ is a continuity point of FkY k . Then lim sup P (|Zn | ≥ η) ≤ lim P (kXn − ck ≥ δǫ ) n

n→∞

+ lim P (an kXn − ck ≥ η/ǫ) n→∞

= P (kY k ≥ η/ǫ). The proof is complete since ǫ can be arbitrary. In statistics, we often need a nondegenerated limiting distribution of an [g(Xn ) − g(c)] so that probabilities involving an [g(Xn ) − g(c)] can be approximated by the c.d.f. of [∇g(c)]τ Y , if (1.77) holds. Hence, result (1.77) is not useful for this purpose if ∇g(c) = 0, and in such cases result (1.78) may be applied. A useful method in statistics, called the delta-method, is based on the following corollary of Theorem 1.12. Corollary 1.1. Assume the conditions of Theorem 1.12. If Y has the Nk (0, Σ) distribution, then an [g(Xn ) − g(c)] →d N (0, [∇g(c)]τ Σ∇g(c)) .

62

1. Probability Theory

Example 1.31. Let {Xn } be a sequence of random variables satisfying √ n(Xn − c) →d N (0, 1). Consider the function g(x) = x2 . If c 6= 0, then an √ 2 application of Corollary 1.1 gives that n(Xn − c2 ) →d N (0, 4c2 ). If c = 0, the first-order derivative of g at 0 is 0 but the second-order derivative of g ≡ 2. Hence, an application of result (1.78) gives that nXn2 →d [N (0, 1)]2 , which has the chi-square distribution χ21 (Example 1.14). The last result can also be obtained by applying Theorem 1.10(iii).

1.5.4 The law of large numbers The law of large numbers concerns the limiting behavior of sums of independent random variables. The weak law of large numbers (WLLN) refers to convergence in probability, whereas the strong law of large numbers (SLLN) refers to a.s. convergence. The following lemma is useful in establishing the SLLN. Its proof is left as an exercise. Lemma 1.6. (Kronecker’s lemma). Let xnP∈ R, an ∈ R, 0 < an ≤ ∞ an+1P , n = 1, 2, ..., and an → ∞. If the series n=1 xn /an converges, then n a−1 n i=1 xi → 0.

Our first result gives the WLLN and SLLN for a sequence of independent and identically distributed (i.i.d.) random variables.

Theorem 1.13. Let X1 , X2 , ... be i.i.d. random variables. (i) (The WLLN). A necessary and sufficient condition for the existence of a sequence of real numbers {an } for which n

1X Xi − an →p 0 n i=1

(1.80)

is that nP (|X1 | > n) → 0, in which case we may take an = E(X1 I{|X1 |≤n} ). (ii) (The SLLN). A necessary and sufficient condition for the existence of a constant c for which n 1X Xi →a.s. c (1.81) n i=1 is that E|X1 | < ∞, in which case c = EX1 and n

1X ci (Xi − EX1 ) →a.s. 0 n i=1 for any bounded sequence of real numbers {ci }.

(1.82)

63

1.5. Asymptotic Theory

Proof. (i) We prove the sufficiency. The proof of necessity can be found in Petrov (1975). Consider a sequence of random variables obtained by truncating Xj ’s at n: Ynj = Xj I{|Xj |≤n} . Let Tn = X1 + · · · + Xn and Zn = Yn1 + · · · + Ynn . Then P (Tn 6= Zn ) ≤

n X j=1

P (Ynj 6= Xj ) = nP (|X1 | > n) → 0.

(1.83)

For any ǫ > 0, it follows from Chebyshev’s inequality that 2 Zn − EZn Var(Zn ) EYn1 Var(Yn1 ) >ǫ ≤ ≤ , = P n ǫ 2 n2 ǫ2 n ǫ2 n

where the last equality follows from the fact that Ynj , j = 1, ..., n, are i.i.d. From integration by parts, we obtain that Z Z 2 EYn1 1 2 n = x2 dF|X1 | (x) = xP (|X1 | > x)dx − nP (|X1 | > n), n n [0,n] n 0

which converges to 0 since nP (|X1 | > n) → 0 (why?). This proves that (Zn − EZn )/n →p 0, which together with (1.83) and the fact that EYnj = E(X1 I{|X1 |≤n} ) imply the result. (ii) For the sufficiency, let Yn = Xn I{|Xn |≤n} , n = 1, 2, .... Let m > 0 be an integer smaller than n. If we define ci = i−1 for i ≥ m, Z1 = · · · = Zm−1 = 0, Zm = Y1 + · · · + Ym , Zi = Yi , i = m + 1, ..., n, and apply the H´ajek-R`enyi inequality (1.51) to Zi ’s, then we obtain that for any ǫ > 0, m n 1 X 1 X Var(Yi ) P max |ξl | > ǫ ≤ 2 2 Var(Yi ) + 2 , (1.84) m≤l≤n ǫ m i=1 ǫ i=m+1 i2 where ξn = n−1 that

Pn

i=1 (Zi

− EZi ) (= n−1

Pn

i=1 (Yi

− EYi ) if l ≥ m). Note

n ∞ ∞ X X X E(X12 I{j−1 n) =

∞ X

n=1

P (|X1 | > n) < ∞

∞ (Exercise 54) and Lemma 1.5(i) that P (∩∞ n=1 ∪m=n {Xm 6= Ym }) = 0, i.e., there is an event A with P (A) = 1 such that if ω ∈ A, then Xn (ω) = Yn (ω) for sufficiently large n. This implies n

n

1X 1X Xi − Yi →a.s. 0, n i=1 n i=1

(1.85)

which proves the sufficiency. The proof of (1.82) is left as an exercise. We now prove the necessity. Suppose that (1.81) holds for some c ∈ R. Then Xn Tn n − 1 Tn−1 c = −c− − c + →a.s. 0. n n n n−1 n

From Exercise 114, Xn /n →a.s. 0 and the i.i.d. assumption on Xn ’s imply ∞ X

n=1

P (|Xn | ≥ n) =

∞ X

n=1

P (|X1 | ≥ n) < ∞,

which implies E|X1 | < ∞ (Exercise 54). From the proved sufficiency, c = EX1 . If E|X1 | < ∞, then an in (1.80) converges to EX1 and result (1.80) is actually established in Example 1.28 in a much simpler way. On the other hand, if E|X1 | < ∞, then the stronger result (1.81) can be obtained. Some results for the case of E|X1 | = ∞ can be found in Exercise 148 in §1.6 and Theorem 5.4.3 in Chung (1974). The next result is for sequences of independent but not necessarily identically distributed random variables.

65

1.5. Asymptotic Theory

Theorem 1.14. Let X1 , X2 , ... be independent random variables with finite expectations. (i) (The SLLN). If there is a constant p ∈ [1, 2] such that ∞ X E|Xi |p i=1

then

ip

< ∞,

(1.86)

n

1X (Xi − EXi ) →a.s. 0. n i=1

(1.87)

(ii) (The WLLN). If there is a constant p ∈ [1, 2] such that n 1 X E|Xi |p = 0, n→∞ np i=1

(1.88)

lim

then

n

1X (Xi − EXi ) →p 0. n i=1

(1.89)

Proof. (i) Consider again the truncated Xn : Yn = Xn I{|Xn |≤n} , n = 1, 2, .... Since Xn2 I{|Xn |≤n} ≤ n2−p |Xn |p , ∞ ∞ ∞ X X E(Xn2 I{|Xn |≤n} ) X E|Xn |p EYn2 = ≤ < ∞. n2 n2 np n=1 n=1 n=1

It follows from the proof of Theorem 1.13(ii) that n−1 0. Also, ∞ X

n=1

P (Xn 6= Yn ) =

∞ X

n=1

P (|Xn | > n) ≤

Pn

i=1 (Yi −EYi )

→a.s.

∞ X E|Xn |p < ∞. np n=1

Hence, it follows from the proof of Theorem 1.13(ii) that (1.85) holds. Finally, ∞ ∞ ∞ X |E(Xn − Yn )| X E(|Xn |I{|Xn |>n} ) X E|Xn |p = ≤ < ∞, n n np n=1 n=1 n=1

P which together with Lemma 1.6 imply that n−1 ni=1 |E(Xi − Yi )| → 0 and thus (1.87) holds. (ii) For any ǫ > 0, an application of Chebyshev’s inequality and inequality (1.44) leads to ! n n X 1 X > ǫ ≤ Cp P (X − EX ) E|Xi − EXi |p , i i n i=1 ǫp np i=1 which converges to 0 under (1.88). This proves (1.89).

66

1. Probability Theory

Note that (1.86) implies (1.88) (Lemma 1.6). The result in Theorem 1.14(i) is called Kolmogorov’s SLLN when p = 2 and is due to Marcinkiewicz and Zygmund when 1 ≤ p < 2. An obvious sufficient condition for (1.86) with p ∈ (1, 2] is supn E|Xn |p < ∞. For dependent random variables, a result for Markov chains introduced in §1.4.4 is discussed in §4.1.4. We now consider martingales studied in §1.4.4. First, consider the WLLN. Inequality (1.44) still holds if the independence assumption of Xi ’s is replaced by the martingale assumption on Pn the sequence { i=1 (Xi −EXi )} (why?). Hence, from the proof of Theorem 1.14(ii) we conclude that (1.89) still holds if P the independence assumption of Xi ’s in Theorem 1.14 is replaced by that { ni=1 (Xi − EXi )} is a martingale. A result similar to the SLLN in Theorem 1.14(i) can be established ifP the independence assumption of Xi ’s is replaced by that the sequence n { i=1 (Xi − EXi )} is a martingale and if condition (1.86) is replaced by ∞ X E(|Xn |p |X1 , ..., Xn−1 ) < ∞ a.s., np n=2

which is the same as (1.86) if Xi ’s are independent. The proof of this martingale SLLN and many other versions of WLLN and SLLN can be found in standard probability textbooks, for example, Chung (1974) and Lo`eve (1977). The WLLN and SLLN have many applications in probability and statistics. The following is an example. Other examples can be found in later chapters. Example 1.32. Let f and g be continuous functions on [0, 1] satisfying 0 ≤ f (x) ≤ Cg(x) for all x, where C > 0 is a constant. We now show that lim

n→∞

Z

0

1

Z

0

1

···

Z

0

1

R1 Pn f (x)dx f (xi ) i=1 Pn dx1 dx2 · · · dxn = R01 g(x ) i i=1 0 g(x)dx

(1.90)

R1 (assuming that 0 g(x)dx 6= 0). Let X1 , X2 , ... be i.i.d. random variables having the uniform distribution on [0, 1]. By Theorem 1.2, E[f (X1 )] = R1 R1 0 f (x)dx < ∞ and E[g(X1 )] = 0 g(x)dx < ∞. By the SLLN (Theorem 1.13(ii)), n 1X f (Xi ) →a.s. E[f (X1 )], n i=1

and the same result holds when f is replaced by g. By Theorem 1.10(i), Pn f (Xi ) E[f (X1 )] Pi=1 . (1.91) →a.s. n E[g(X1 )] g(X ) i i=1

67

1.5. Asymptotic Theory

Since the random variable on the left-hand side of (1.91) is bounded by C, result (1.90) follows from the dominated convergence theorem and the fact that the left-hand side of (1.90) is the expectation of the random variable on the left-hand side of (1.91). Moment inequalities introduced in §1.3.2 play important roles in proving convergence theorems. They can also be P used to obtain convergence n rates of tail probabilities of the form P |n−1 i=1 (Xi − EXi )| > t . For example, an application of the Esseen-von Bahr, Marcinkiewicz-Zygmund, and Chebyshev inequalities produces ! n 1 X if 1 < p < 2 O(n1−p ) P (Xi − EXi ) > t ≤ if p ≥ 2 O(n−p/2 ) n i=1

for independent random variables X1 , ..., Xn with supn E|Xn |p < ∞.

1.5.5 The central limit theorem The WLLN and SLLN may not be useful in approximating the distributions of (normalized) sums of independent random variables. We need to use the central limit theorem (CLT), which plays a fundamental role in statistical asymptotic theory. Theorem 1.15 (Lindeberg’s CLT). Let {Xnj , j = 1, ..., kn } be independent P n random variables with 0 < σn2 = Var( kj=1 Xnj ) < ∞, n = 1, 2,..., and kn → ∞ as n → ∞. If kn X j=1

then

E (Xnj − EXnj )2 I{|Xnj −EXnj |>ǫσn } = o(σn2 ) for any ǫ > 0, (1.92) kn 1 X (Xnj − EXnj ) →d N (0, 1). σn j=1

(1.93)

Proof. Considering (Xnj − EXnj )/σn , without loss of generality we may assume EXnj = 0 and σn2 = 1 in this proof. Let t ∈ R be given. From the √ √ −1tx inequality |e − (1 + −1tx − t2 x2 /2)| ≤ min{|tx|2 , |tx|3 }, the ch.f. of Xnj satisfies 2 φXnj (t) − 1 − t2 σnj /2 ≤ E min{|tXnj |2 , |tXnj |3 } , (1.94) 2 = Var(Xnj ). For any ǫ > 0, the right-hand side of (1.94) is where σnj bounded by E(|tXnj |3 I{|Xnj |ǫ} ) → ǫ2 for arbitrary ǫ > 0. Hence

lim max

n→∞ j≤kn

2 σnj = 0. σn2

(1.96)

2 (Note that σn2 = 1 is assumed for convenience.) This implies that 1 − t2 σnj are all between 0 and 1 for large enough n. Using the inequality

|a1 · · · am − b1 · · · bm | ≤

m X j=1

|aj − bj |

for any complex numbers aj ’s and bj ’s with |aj | ≤ 1 and |bj | ≤ 1, j = 1, ..., m, we obtain that kn kn kn Y −t2 σ2 /2 Y −t2 σ2 /2 X 2 2 2 2 nj nj 1 − t σnj /2 ≤ e − − 1 − t σnj /2 , e j=1

j=1

j=1

Pkn 4

4 4 2 x which is bounded by t j=1 σnj ≤ t maxj≤kn σnj → 0, since |e −1−x| ≤ P n 2 σnj = σn2 = 1. Also, x2 /2 if |x| ≤ 12 and kj=1

kn kn Y Y 2 2 1 − t φ (t) − σ /2 Xnj nj j=1

j=1

is bounded by the quantity on the left-hand side of (1.95) and, hence, converges to 0 by (1.95). Thus, kn Y

j=1

φXnj (t) =

kn Y

2

e−t

2 σnj /2

2

+ o(1) = e−t

/2

+ o(1).

j=1

P n Xnj converges to the ch.f. of N (0, 1) for This shows that the ch.f. of kj=1 every t. By Theorem 1.9(ii), the result follows. Condition (1.92) is called Lindeberg’s condition. From the proof of Theorem 1.15, Lindeberg’s condition implies (1.96), which is called Feller’s condition. Feller’s condition (1.96) means that all terms in the sum σn2 =

69

1.5. Asymptotic Theory Pkn

2 σnj are uniformly negligible as n → ∞. If Feller’s condition is assumed, then Lindeberg’s condition is not only sufficient but also necessary for result (1.93), which is the well-known Lindeberg-Feller CLT. A proof can be found in Billingsley (1986, pp. 373-375). Note that neither Lindeberg’s condition nor Feller’s condition is necessary for result (1.93) (Exercise 158). A sufficient condition for Lindeberg’s condition is the following Liapounov’s condition, which is somewhat easier to verify: j=1

kn X j=1

E|Xnj − EXnj |2+δ = o(σn2+δ ) for some δ > 0.

(1.97)

Example 1.33. Let X1 , X2 , ... be independent random variables. Suppose that distribution Bi(pi , 1), i = 1, 2,..., and that σn2 = Pn Pn Xi has the binomial i=1 Var(Xi ) = i=1 pi (1 − pi ) → ∞ as n → ∞. For each i, EXi = p and E|X − EX |3 = (1 − pi )3 pi + p3i (1 − pi ) ≤ 2pi (1 − pi ). Hence i i i Pn 3 2 i=1 E|Xi − EXi | ≤ 2σn , i.e., Liapounov’s condition (1.97) holds with δ = 1. Thus, by Theorem 1.15, n 1 X (Xi − pi ) →d N (0, 1). σn i=1

(1.98)

It can be shown (exercise) that the condition σn → ∞ is also necessary for result (1.98). The following are useful corollaries of Theorem 1.15 (and Theorem 1.9(iii)). Corollary 1.2 is in fact proved in Example 1.28. The proof of Corollary 1.3 is left as an exercise. Corollary 1.2 (Multivariate CLT). Let X1 , ..., Xn be i.i.d. random kvectors with a finite Σ = Var(X1 ). Then n

1 X √ (Xi − EX1 ) →d Nk (0, Σ). n i=1 Corollary 1.3. Let Xni ∈ Rmi , i = 1, ..., kn , be independent random vectors with mi ≤ m (a fixed integer), n = 1, 2,..., kn → ∞ as n → ∞, and inf i,n λ− [Var(Xni )] > 0, where λ− [A] is the smallest eigenvalue of A. Let cni ∈ Rmi be vectors such that ! X kn max kcni k2 kcni k2 = 0. lim n→∞

1≤i≤kn

i=1

70

1. Probability Theory

(i) Suppose that supi,n EkXni k2+δ < ∞ for some δ > 0. Then kn X i=1

cτni (Xni

#1/2 "X kn τ − EXni ) Var(cni Xni ) →d N (0, 1).

(1.99)

i=1

(ii) Suppose that whenever mi = mj , 1 ≤ i < j ≤ kn , n = 1, 2, ..., Xni and Xnj have the same distribution with EkXni k2 < ∞. Then (1.99) holds. Applications of these corollaries can be found in later chapters. An extension of Lindeberg’s CLT is the so-called martingale CLT. In Theorem 1.15, if the independence assumption of Xnj , j = 1, ..., kn , is replaced by that {Yn } is a martingale and kn 1 X E[(Xnj − EXnj )2 |Xn1 , ..., Xn(j−1) ] →p 1, σn2 j=1

P n where Yn = kj=1 (Xnj − EXnj ) when n ≤ kn , Yn = Ykn when n > kn , and Xn0 is defined to be 0, then result (1.93) still holds (see, e.g., Billingsley, 1986, p. 498 and Sen and Singer 1993, p. 120). More results on the CLT can be found, for example, in Serfling (1980) and Shorack and Wellner (1986). Let Yn be a sequence of random variables, {µn } and {σn } be sequences of real numbers such that σn > 0 for all n, and (Yn − µn )/σn →d N (0, 1). Then, by Proposition 1.16, lim sup |F(Yn −µn )/σn (x) − Φ(x)| = 0,

n→∞ x

(1.100)

where Φ is the c.d.f. of N (0, 1). This implies that for any sequence of real n numbers {cn }, limn→∞ |P (Yn ≤ cn ) − Φ cnσ−µ | = 0, i.e., P (Yn ≤ cn ) can n n be approximated by Φ cnσ−µ , regardless of whether {cn } has a limit. Since n t−µn 2 Φ σn is the c.d.f. of N (µn , σn ), Yn is said to be asymptotically distributed Pkn τ cni Xni as N (µn , σn2 ) or simply asymptotically normal. For example, i=1 in Corollary 1.3 is asymptotically normal. This can be extended to ranPn dom vectors. For example, i=1 Xi in Corollary 1.2 is asymptotically distributed as Nk (nEX1 , nΣ).

1.5.6 Edgeworth and Cornish-Fisher expansions Let {Yn } be a sequence of random variables satisfying (1.100) and Wn = (Yn − µn )/σn . The convergence speed of (1.100) can be used to assess whether Φ provides a good approximation to the c.d.f. FWn . Also, sometimes we would like to find an approximation to FWn that is better than

71

1.5. Asymptotic Theory

Φ in terms of convergence speed. The Edgeworth expansion is a useful tool for these purposes. P To illustrate the idea, let Wn = n−1/2 ni=1 (Xi −µ)/σ, where X1 , X2 , ... are i.i.d. random variables with EX1 = µ and Var(X1 ) = σ 2 . Assume that the m.g.f. of Z = (X1 − µ)/σ is finite and positive in a neighborhood of 0. From (1.55), the cumulant generating function of Z has the expansion κ(t) =

∞ X κj j=1

j!

tj ,

where κj , j = 1, 2, ..., are cumulants of Z (e.g., κ1 = 0, κ2 = 1, κ3 = EZ 3 , and κ4 = EZ 4 − 3), and the m.g.f. of Wn is equal to 2 X ∞ √ n κ j tj t + , ψn (t) = exp{κ(t/ n)} = exp 2 j!n(j−2)/2 j=3 where exp{x} denotes the exponential function ex . Using the series expan2 sion for et /2 , we obtain that 2

ψn (t) = et

/2

2

+ n−1/2 r1 (t)et

/2

2

+ · · · + n−j/2 rj (t)et

/2

+ ···,

(1.101)

where rj is a polynomial of degree 3j depending on κ3 , ..., κj+2 but not on n, j = 1, 2, .... For example, it can be shown (exercise) that r1 (t) = 16 κ3 t3

and r2 (t) =

1 4 24 κ4 t

+

1 2 6 72 κ3 t .

(1.102)

R R 2 Since ψn (t) = etx dFWn (x) and et /2 = etx dΦ(x), expansion (1.101) suggests the inverse expansion FWn (x) = Φ(x) + n−1/2 R1 (x) + · · · + n−j/2 Rj (x) + · · · , R 2 where Rj (x) is a function satisfying etx dRj (x) = rj (t)et /2 , j = 1, 2, .... j d 1 Let ∇j = dx j be the differential operator and ∇ = ∇ . Then Rj (x) = rj (−∇)Φ(x), j = 1, 2, ..., where rj (−∇) is interpreted as a differential operator. Thus, Rj ’s can be obtained once rj ’s are derived. It follows from (1.102) (exercise) that R1 (x) = − 16 κ3 (x2 − 1)Φ′ (x)

(1.103)

and 1 κ4 x(x2 − 3) + R2 (x) = −[ 24

1 2 4 72 κ3 x(x

− 10x2 + 15)]Φ′ (x).

(1.104)

A rigorous statement of the Edgeworth expansion for a more general Wn is given in the following theorem whose proof can be found in Hall (1992).

72

1. Probability Theory

Theorem 1.16 (Edgeworth expansions). Let m be a positive integer and X1 , X2 , √ ... be i.i.d. random k-vectors having Pn finite m+2 moments. Consider −1 ¯ ¯ Wn = nh(X)/σ h , where X = n i=1 Xi , h is a Borel function on Rk that is m + 2 times continuously differentiable in a neighborhood of µ = EX1 , h(µ) = 0, and σh2 = [∇h(µ)]τ Var(X1 )∇h(µ) > 0. Assume that lim sup |φX1 (t)| < 1,

(1.105)

ktk→∞

where φX1 is the ch.f. of X1 . Then, FWn admits the Edgeworth expansion m X pj (x)Φ′ (x) 1 sup FWn (x) − Φ(x) − = o nm/2 , nj/2 x

(1.106)

j=1

where pj (x) is a polynomial of degree at most 3j − 1, odd for even j and even for odd j, with coefficients depending on the first m + 2 moments of X1 , j = 1, ..., m. In particular, (1.107) p1 (x) = −c1 σh−1 + 6−1 c2 σh−3 (x2 − 1) P P P P P k k k with c1 = 2−1 ki=1 kj=1 aij µij and c2 = i=1 j=1 l=1 ai aj al µijl + Pk Pk Pk Pk 3 i=1 j=1 l=1 h=1 ai aj alh µil µjh , where ai is the ith component of ∇h(µ), aij is the (i, j)th element of the Hessian matrix ∇2 h(µ), µij = E(Yi Yj ), µijl = E(Yi Yj Yl ), and Yi is the ith component of X1 − µ. Condition (1.105) is Cram´er’s continuity condition. It is satisfied if one component of X1 has a Lebesgue p.d.f. The polynomial pj with j ≥ 2 may be derived using the method in deriving (1.103) and (1.104), but the derivation is usually complicated (see Hall (1992)). Under the conditions of Theorem 1.16, the convergence speed of (1.100) Pm is O(n−1/2 ) and, as an approximation to FWn , Φ+ j=1 n−j/2 pj Φ′ is better than Φ, since its convergence speed is o(n−m/2 ). The results in Theorem 1.16 can be applied to many cases, as the following example indicates. ¯ = n−1 Pn Xi with i.i.d. random variables X1 , X2 , Example 1.34. Let X i=1 ... satisfying condition (1.105). First, consider the normalized random √ ¯ variable Wn = n(X − µ)/σ, where µ = EX1 and σ 2 = Var(X1 ). Then, Theorem 1.16 can be applied with h(x) = x − µ and σh2 = σ 2 , and the Edgeworth expansion in (1.106) holds if E|X1 |m+2 < ∞. In this case, results (1.103) and (1.104) imply that pj (x) = Rj (x)/Φ′ (x), j = 1, 2. √ ¯ Next, consider the studentized random variable Wn = n(X − µ)/ˆ σ, P n ¯ 2 . Assuming that EX 2m+4 < ∞ and applywhere σ ˆ 2 = n−1 i=1 (Xi − X) 1 ing Theorem 1.16 to random vectors (Xi , Xi2 ), i = 1, 2, ..., and h(x, y) =

73

1.5. Asymptotic Theory

(x−µ)/

p

(y − x2 ), we obtain the Edgeworth expansion (1.106) with σh = 1, p1 (x) = 16 κ3 (2x2 + 1)

(exercise). Furthermore, it can be found in Hall (1992, p. 73) that p2 (x) =

1 2 12 κ4 x(x

− 3) −

1 2 4 18 κ3 x(x

+ 2x2 − 3) − 14 x(x2 + 3).

√ σ 2 − σ 2 ). Theorem 1.16 can be Consider now the random variable n(ˆ 2 applied to random vectors (Xi , Xi ), i = 1, 2, ..., and h(x, y) = (y − x2 − σ 2 ). Assume that EX12m+4 < ∞. It can be √ shown (exercise) that the Edgeworth expansion in (1.106) holds with Wn = n(ˆ σ 2 − σ 2 )/σh , σh2 = E(X1 − µ)4 − 4 σ , and p1 (x) = (ν4 − 1)−1/2 [1 − 16 (ν4 − 1)−1 (ν6 − 3ν4 − 6ν32 + 2)(x2 − 1)], where νj = σ −j E(X1 − µ)j , j = 3, ..., 6. √ σ 2 − σ 2 )/ˆ τ, Finally, consider random variable Wn = n(ˆ Pnthe studentized 2 −1 4 4 ¯ where τˆ = n (X − X) − σ ˆ . Theorem 1.16 can be applied to i i=1 random vectors (Xi , Xi2 , Xi3 , Xi4 ), i = 1, 2, ..., and h(x, y, z, w) = (y − x2 − σ 2 )[w − y 2 − 4xz + 8x2 y − 4x4 ]−1/2 . Assume that EX14m+8 < ∞. It can be shown (exercise) that the Edgeworth expansion in (1.106) holds with σh2 = 1 and p1 (x) = −(ν4 −1)−3/2 [ 21 (4ν32 +ν4 −ν6 ) + 13 (3ν32 +3ν4 −ν6 −2)(x2 −1)]. An inverse Edgeworth expansion is referred to as a Cornish-Fisher expansion, which is useful in statistics (see §7.4). For α ∈ (0, 1), let zα = Φ−1 (α). Since the c.d.f. FWn may not be strictly increasing and continuous, we define wnα = inf{x : FWn (x) ≥ α}. The following result can be found in Hall (1992). Theorem 1.17 (Cornish-Fisher expansions). Under the conditions of Theorem 1.16, wnα admits the Cornish-Fisher expansion m X qj (zα ) 1 =o , (1.108) sup wnα − zα − nj/2 nm/2 ǫ 0, there is a δǫ such that ν(A) < δǫ and A ∈ F imply A |f |dν < ǫ.

25. Prove that part (i) and part (iii) of Theorem 1.1 are equivalent. 26. Prove Theorem 1.2.

27. Prove Theorem 1.3. (Hint: first consider simple nonnegative f .) 28. Consider Example 1.9. Show that (1.17) does not hold for i=j 1 f (i, j) = −1 i=j−1 0 otherwise. Does this contradict Fubini’s theorem?

29. Let f be a nonnegative Borel function on (Ω, F , ν) with a σ-finite ν, A = {(ω, x) ∈ Ω × R : 0 ≤ x ≤ f (ω)}, and Rm be the Lebesgue measure on (R, B). Show that A ∈ σ(F × B) and Ω f dν = ν × m(A).

1.6. Exercises

77

R 30. For any c.d.f. F and any a ≥ 0, show that [F (x + a) − F (x)]dx = a.

31. (Integration by parts). Let F and G be two c.d.f.’s on R. Show that if F andR G have no common points of discontinuity Rin the interval (a, b], then (a,b] G(x)dF (x) = F (b)G(b) − F (a)G(a) − (a,b] F (x)dG(x).

32. Let f be a Borel function on R2 such that f (x, y) = 0 for each x ∈ R and y 6∈ Cx , where m(Cx ) = 0 for each x and m is the Lebesgue measure. Show that f (x, y) = 0 for each y 6∈ C and x 6∈ By , where m(C) = 0 and m(By ) = 0 for each y 6∈ C. 33. RConsider Example 1.11. Show that if (1.21) holds, then R P (A) = f (x)dx for any Borel set A. (Hint: A = {A : P (A) = A A f (x)dx} is a σ-field containing all sets of the form (−∞, x].) 34. Prove Proposition 1.7. P∞ 35. Let {an } be a sequence of positive numbers satisfying n=1 an = 1 and let {Pn } be a sequence ofPprobability measures on a common ∞ measurable space. Define P = n=1 an Pn . (a) Show that P is a probability measure. (b) Show that Pn ≪ ν for all n and a measure only if P ≪ ν P∞ ν if and dPn and, when P ≪ ν and ν is σ-finite, dP = a . n n=1 dν dν (c) Derive the Lebesgue p.d.f. of P when Pn is the gamma distribution Γ(α, n−1 ) (Table 1.2) with α > 1 and an is proportional to n−α . 36. Let Fi be a c.d.f. having a Lebesgue p.d.f. fi , i = 1, 2. Assume that there is a c ∈ R such that F1 (c) < F2 (c). Define −∞ < x < c F1 (x) F (x) = c ≤ x < ∞. F2 (x) Show that the probability measure P corresponding to F satisfies P ≪ m + διc and find dP/d(m + διc ), where m + διc is given in (1.23). 37. Let (X, Y ) be a random 2-vector with the following Lebesgue p.d.f.: 8xy 0≤x≤y≤1 f (x, y) = 0 otherwise. Find the marginal p.d.f.’s of X and Y . Are X and Y independent? 38. Let (X, Y, Z) be a random 3-vector with the following Lebesgue p.d.f.: 1−sin x sin y sin z 0 ≤ x ≤ 2π, 0 ≤ y ≤ 2π, 0 ≤ z ≤ 2π 8π 3 f (x, y, z) = 0 otherwise. Show that X, Y , and Z are not independent, but are pairwise independent.

78

1. Probability Theory

39. Prove Lemma 1.1 without using Definition 1.7 for independence. 40. Let X be a random variable having a continuous c.d.f. F . Show that Y = F (X) has the uniform distribution U (0, 1) (Table 1.2). 41. Let U be a random variable having the uniform distribution U (0, 1) and let F be a c.d.f. Show that the c.d.f. of Y = F −1 (U ) is F , where F −1 (t) = inf{x ∈ R : F (x) ≥ t}. 42. Prove Proposition 1.8. 43. Let X = Nk (µ, Σ) with a positive definite Σ. (a) Let Y = AX + c, where A is an l × k matrix of rank l ≤ k and c ∈ Rl . Show that Y has the Nl (Aµ + c, AΣAτ ) distribution. (b) Show that the components of X are independent if and only if Σ is a diagonal matrix. (c) Let Λ be positive definite and Y = Nm (η, Λ) be independent of X. Show that (X, Y ) has the Nk+m ((µ, η), D) distribution, where D is a block diagonal matrix whose two diagonal blocks are Σ and Λ. 44. Let X be a random variable having the Lebesgue p.d.f. Derive the p.d.f. of Y = sin X.

2x π 2 I(0,π) (x).

45. Let Xi , i = 1, 2, 3, be independent random variables having the same Lebesgue p.d.f. f (x) = e−x I(0,∞) (x). Obtain the joint Lebesgue p.d.f. of (Y1 , Y2 , Y3 ), where Y1 = X1 + X2 + X3 , Y2 = X1 /(X1 + X2 ), and Y3 = (X1 + X2 )/(X1 + X2 + X3 ). Are Yi ’s independent? 46. Let X1 and X2 be independent random variables having the standard normalp distribution. Obtain the joint Lebesgue p.d.f. of (Y1 , Y2 ), where Y1 = X12 + X22 and Y2 = X1 /X2 . Are Yi ’s independent?

random variables and Y = X1 + X2 . 47. Let X1 and X2 be independent R Show that FY (y) = FX2 (y − x)dFX1 (x).

48. Show that the Lebesgue p.d.f.’s given by (1.31) and (1.33) are the p.d.f.’s of the χ2n (δ) and Fn1 ,n2 (δ) distributions, respectively. 49. Show that the Lebesgue p.d.f. given by (1.32) is the p.d.f. of the tn (δ) distribution.

50. Let X = Nn (µ, In ) and A be an n × n symmetric matrix. Show that if X τ AX has the χ2r (δ) distribution, then A2 = A, r is the rank of A, and δ = µτ Aµ. 51. Let X = Nn (µ, In ). Apply Cochran’s theorem (Theorem 1.5) to show that if A2 = A, then X τ AX has the noncentral chi-square distribution χ2r (δ), where A is an n × n symmetric matrix, r is the rank of A, and δ = µτ Aµ.

79

1.6. Exercises

and Xi =PN (0, σi2 ), i = 1, ..., n. Let 52. Let XP 1 , ..., Xn be independent Pn n n −2 −2 ˜ 2 . Apply ˜ X = i=1 σi Xi / i=1 σi and S˜2 = i=1 σi−2 (Xi − X) 2 2 ˜ ˜ Cochran’s theorem to show that X and S are independent and that S˜2 has the chi-square distribution χ2n−1 . 53. Let X = Nn (µ, In ) and Ai be an n × n symmetric matrix satisfying A2i = Ai , i = 1, 2. Show that a necessary and sufficient condition that X τ A1 X and X τ A2 X are independent is A1 A2 = 0. 54. Let X be a random variable and a > 0. Show that E|X|a < ∞ if and P∞ only if n=1 na−1 P (|X| ≥ n) < ∞.

55. Let X be a random variable. RShow that R0 ∞ (a) if EX exists, then EX = 0 P (X > x)dx − −∞ P (X ≤ x)dx; P∞ (b) if X has range {0, 1, 2, ...}, then EX = n=1 P (X ≥ n).

56. Let T be a random variable having the noncentral t-distribution tn (δ). Show that p (a) E(T ) = δΓ((n − 1)/2) n/2/Γ(n/2) when n > 1; h i2 2 ) δ 2 n Γ((n−1)/2) − when n > 2. (b) Var(T ) = n(1+δ n−2 2 Γ(n/2)

57. Let F be a random variable having the noncentral F-distribution Fn1 ,n2 (δ). Show that (n1 +δ) (a) E(F) = nn12 (n when n2 > 2; 2 −2) (b) Var(F) =

2n22 [(n1 +δ)2 +(n2 −2)(n1 +2δ)] n21 (n2 −2)2 (n2 −4)

when n2 > 4.

58. Let X = Nk (µ, Σ) with a positive definite Σ. (a) Show that EX = µ and Var(X) = Σ. (b) Let A be an l × k matrix and B be an m × k matrix. Show that AX and BX are independent if and only if AΣB τ = 0. (c) Suppose that k = 2, X = (X1 , X2 ), µ = 0, Var(X1 ) =pVar(X2 ) = 1, and Cov(X1 , X2 ) = ρ. Show that E(max{X1 , X2 }) = (1 − ρ)/π.

59. Let X be a random variable and g and h be nondecreasing functions on R. Show that Cov(g(X), h(X)) ≥ 0 when E|g(X)h(X)| < ∞.

60. Let X be a random variable with EX 2 < ∞ and let Y = |X|. Suppose that X has a Lebesgue p.d.f. symmetric about 0. Show that X and Y are uncorrelated, but they are not independent. 61. Let (X, Y ) be a random 2-vector with the following Lebesgue p.d.f.: −1 π x2 + y 2 ≤ 1 f (x, y) = 0 x2 + y 2 > 1. Show that X and Y are uncorrelated, but are not independent.

80

1. Probability Theory

62. Show that inequality (1.41) holds and that when 0 < E|X|p < ∞ and 0 < E|Y |q < ∞, the equality in (1.40) holds if and only if α|X|p = β|Y |q a.s. for some nonzero constants α and β. 63. Prove the following inequalities. (a) Liapounov’s inequality (1.42). (b) Minkowski’s inequality (1.43). (Hint: apply H¨older’s inequality to random variables |X + Y |p−1 and |X|.) (c) (Cr -inequality). E|X + Y |r ≤ Cr (E|X|r + E|Y |r ), where X and Y are random variables, r is a positive constant, and Cr = 1 if 0 < r ≤ 1 and Cr = 2r−1 if r > 1. (d) Let Xi be a random variable with E|Xi |p < ∞, i = 1, ..., n, where p is a constant larger than 1. Show that #p ) ( n " n n 1 X p 1X 1X p p 1/p E . Xi ≤ min E|Xi | , (E|Xi | ) n i=1 n i=1 n i=1 (e) Inequality (1.44). (Hint: prove the case of n = 2 first and then use induction.) (f) Inequality (1.49).

64. Show that the following functions of x are convex and discuss whether they are strictly convex. (a) |x − a|p , where p ≥ 1 and a ∈ R. (b) x−p , x ∈ (0, ∞), where p > 0. (c) ecx , where c ∈ R. (d) x log x, x ∈ (0, ∞). (e) g(ϕ(x)), x ∈ (a, b), where −∞ ≤ a < b ≤ ∞, ϕ is convex on (a, b), and g is convex Pk and nondecreasing on the range Qkof ϕ. (f) ϕ(x) = i=1 ci ϕi (xi ), x = (x1 , ..., xk ) ∈ i=1 Xi , where ci is a positive constant and ϕi is convex on Xi , i = 1, ..., k. 65. Let X = Nk (µ, Σ) with a positive definite Σ. τ τ (a) Show that the m.g.f. of X is et µ+t Σt/2 . (b) Show that EX = µ and Var(X) = Σ by applying (1.54). ′ (c) When k = 1 (Σ = σ 2 ), show that EX = ψX (0) = µ, EX 2 = (3) (4) ′′ ψX (0) = σ 2 + µ2 , EX 3 = ψX (0) = 3σ 2 µ + µ3 , and EX 4 = ψX (0) = 3σ 4 + 6σ 2 µ2 + µ4 . (d) In part (c), show that if µ = 0, then EX p = 0 when p is an odd integer and EX p = (p − 1)(p − 3) · · · 3 · 1σ p when p is an even integer. 66. Let X be a random variable having the gamma distribution Γ(α, γ). Find moments EX p , p = 1, 2, ..., by differentiating the m.g.f. of X. 67. Let X be a random variable with finite EetX and Ee−tX for a t 6= 0. Show that E|X|a < ∞ for any a > 0.

1.6. Exercises

81

68. Let X be a random variable having ψX (t) < ∞ for t in a neighborhood of 0. Show that the moments and cumulants of X satisfy the following equations: µ1 = κ1 , µ2 = κ2 + κ21 , µ3 = κ3 + 3κ1 κ2 + κ31 , and µ4 = κ4 + 3κ22 + 4κ1 κ3 + 6κ21 κ2 + κ41 , where µi and κi are the ith moment and cumulant of X, respectively. 69. Let X be a discrete random variable taking values 0,1,2.... The probability generating function of X is defined to be ρX (t) = E(tX ). Show that (a) ρX (t) = ψX (log t), where ψX is the m.g.f. of X; p (b) d ρdtXp(t) t=1 = E[X(X − 1) · · · (X − p + 1)] for any positive integer p, if ρX is finite in a neighborhood of 1. 70. Let Y be a random variable having the noncentral chi-square distribution χ2k (δ). Show that √ √ √ (a) the ch.f. of Y is (1 − 2 −1t)−k/2 e −1δt/(1−2 −1t) ; (b) E(Y ) = k + δ and Var(Y ) = 2k + 4δ.

71. Let φ be a ch.f. on Rk . Show that |φ| ≤ 1 and φ is uniformly continuous on Rk . √ and b are real numbers, 72. For a complex number√z = a+ −1b, where Pn aP n z¯ is defined to be a − −1b. Show that i=1 j=1 φ(ti − tj )zi z¯j ≥ 0, where φ is a ch.f. on Rk , t1 , ..., tn are k-vectors, and z1 , ..., zn are complex numbers. 73. Show that the following functions of t ∈ R are ch.f.’s, where a > 0 and b > 0 are constants: (a) a2 /(a2 + t2 ); √ (b) (1 + ab − abe −1t )−1/b ; (c) max{1 − |t|/a, 0}; (d) 2(1 − cos at)/(a2 t2 ); a (e) e−|t| , where 0 < a ≤ 2; 2 (f) |φ| R , where φ is a ch.f. on R; (g) φ(ut)dG(u), where φ is a ch.f. on R and G is a c.d.f. on R.

74. Let φn be the ch.f. of a probability measure P Pn , n = 1, 2,.... Let {an } ∞ be a sequence of nonnegative numbers with n=1 an = 1. Show that P∞ n=1 an φn is a ch.f. and find its corresponding probability measure. R 75. Let X be a random variable whose ch.f. φX satisfies |φX (t)|dt < ∞. √ R Show that (2π)−1 e− −1xt φX (t)dt is the Lebesgue p.d.f. of X. 76. A random variable PX or its distribution is of the lattice type if and only if FX (x) = ∞ j=−∞ pj I{a+jd} (x), x ∈ R, where a, d, pj ’s are

82

1. Probability Theory P constants, d > 0, pj ≥ 0, and ∞ j=−∞ pj = 1. Show that X is of the lattice type if and only if its ch.f. satisfies |φX (t)| = 1 for some t 6= 0.

77. Let φ be a ch.f. on R. Show that (a) if |φ(t )| = |φ(t2 )| = 1 and t1 /t2 is an irrational number, then √1 φ(t) = e −1at for some constant a; (b) if tn → 0, tn 6= 0, and |φ(tn )| = 1, then the result in (a) holds; (c) | cos t| is not a ch.f., although cos t is a ch.f. 78. Let X1 , ..., Xk be independent random variables and Y = X1 +· · ·+Xk . Prove the following statements, using Theorem 1.6 and result (1.58). (a) If Xi has the binomial distribution Bi(p, ni ), i = 1, ..., k, then Y has the binomial distribution Bi(p, n1 + · · · + nk ). (b) If Xi has the Poisson distribution P (θi ), i = 1, ..., k, then Y has the Poisson distribution P (θ1 + · · · + θk ). (c) If Xi has the negative binomial distribution N B(p, ri ), i = 1, ..., k, then Y has the negative binomial distribution N B(p, r1 + · · · + rk ). (d) If Xi has the exponential distribution E(0, θ), i = 1, ..., k, then Y has the gamma distribution Γ(k, θ). (e) If Xi has the Cauchy distribution C(0, 1), i = 1, ..., k, then Y /k has the same distribution as X1 . 79. Find an example of two random variables X and Y such that X and Y are not independent but their ch.f.’s satisfy φX (t)φY (t) = φX+Y (t) for all t ∈ R. 80. Let X1 , X2 , ... be independent random variables having the exponential distribution E(0, θ). For given t > 0, let Y be the maximum of n such that Tn ≤ t, where T0 = 0 and Tn = X1 + · · · + Xn , n = 1, 2, .... Show that Y has the Poisson distribution P (t/θ). 81. Let Σ be a k × k nonnegative definite matrix. (a) For a nonsingular Σ, show that X is Nk (µ, Σ) if and only if cτ X is N (cτ µ, cτ Σc) for any c ∈ Rk . (b) For a singular Σ, we define X to be Nk (µ, Σ) if and only if cτ X is N (cτ µ, cτ Σc) for any c ∈ Rk (N (a, 0) is the c.d.f. of the point mass at a). Show that the results in Exercise 43(a)-(c), Exercise 58(a)-(b), and Exercise 65(a) still hold for X = Nk (µ, Σ) with a singular Σ. 82. Let (X1 , X2 ) be Nk (µ, Σ) with a k × k positive definite Σ11 Σ12 , Σ= Σ21 Σ22 where X1 is a random l-vector and Σ11 is an l × l matrix. Show that the conditional Lebesgue p.d.f. of X2 given X1 = x1 is −1 Nk−l µ2 + Σ21 Σ−1 11 (x1 − µ1 ), Σ22 − Σ21 Σ11 Σ12 ,

1.6. Exercises

83

where µi = EXi , i = 1, 2. (Hint: consider X2 − µ2 − Σ21 Σ−1 11 (X1 − µ1 ) and X1 − µ1 .) 83. Let X be an integrable random variable with a Lebesgue p.d.f. fX and let Y = g(X), where g is a function with positive derivative on (0, ∞) and g(x) = g(−x). Find an expression for E(X|Y ) and verify that it is indeed the conditional expectation. 84. Prove Lemma 1.2. (Hint: first consider simple functions.) 85. Prove Proposition 1.10. (Hint for proving (ix): first show that 0 ≤ X1 ≤ X2 ≤ · · · and Xn →a.s. X imply E(Xn |A) →a.s. E(X|A).) 86. Let X and Y be integrable random variables on (Ω, F, P ) and A ⊂ F be a σ-field. Show that E[Y E(X|A)] = E[XE(Y |A)], assuming that both integrals exist. 87. Let X, X1 , X2 , ... be a sequence of random variables on (Ω, F, P ) and A ⊂ F be a σ-field. Suppose that E(Xn Y ) → E(XY ) for every integrable (or bounded) random variable Y . Show that E[E(Xn |A)Y ] → E[E(X|A)Y ] for every integrable (or bounded) random variable Y . 88. Let X be a nonnegative integrable randomRvariable on (Ω,F , P ) and ∞ A ⊂ F be a σ-field. Show that E(X|A) = 0 P X > t|A dt a.s.

89. Let X and Y be random variables on (Ω, F , P ) and A ⊂ F be a σfield. Prove the following inequalities for conditional expectations. (a) If E|X|p < ∞ and E|Y |q < ∞ for constants p and q with p > 1 and p−1 + q −1 = 1, then E(|XY ||A) ≤ [E(|X|p |A)]1/p [E(|Y |q |A)]1/q a.s. (b) If E|X|p < ∞ and E|Y |p < ∞ for a constant p ≥ 1, then [E(|X + Y |p |A)]1/p ≤ [E(|X|p |A)]1/p + [E(|Y |p |A)]1/p a.s. (c) If f is a convex function on R, then f (E(X|A)) ≤ E[f (X)|A] a.s.

90. Let X and Y be random variables on a probability space with Y = E(X|Y ) a.s. and let ϕ be a nondecreasing convex function on [0, ∞). (a) Show that if Eϕ(|X|) < ∞, then Eϕ(|Y |) < ∞. (b) Find an example in which Eϕ(|Y |) < ∞ but Eϕ(|X|) = ∞. (c) Suppose that Eϕ(|X|) = Eϕ(|Y |) < ∞ and ϕ is strictly convex and strictly increasing. Show that X = Y a.s. 91. Let X, Y , and Z be random variables on a probability space. Suppose that E|X| < ∞ and Y = h(Z) with a Borel h. Show that (a) if X and Z are independent and E|Z| < ∞, then E(XZ|Y ) = E(X)E(Z|Y ) a.s.; (b) if E[f (X)|Z] = f (Y ) for all bounded continuous functions f on R, then X = Y a.s.; (c) if E[f (X)|Z] ≥ f (Y ) for all bounded, continuous, nondecreasing functions f on R, then X ≥ Y a.s.

84

1. Probability Theory

92. Prove Lemma 1.3. 93. Show that random variables Xi , i = 1, ..., n, are independent according to Definition 1.7 if and only if (1.7) holds with F being the joint c.d.f. of Xi ’s and Fi being the marginal c.d.f. of Xi . 94. Show that a random variable X is independent of itself if and only if X is constant a.s. Can X and f (X) be independent for a Borel f ? 95. Let X, Y , and Z be independent random variables on a probability space and let U = X + Z and V = Y + Z. Show that given Z, U and V are conditionally independent. 96. Show that the result in Proposition 1.11 may not be true if Y2 is independent of X but not (X, Y1 ). 97. Let X and Y be independent random variables on a probability space. Show that if E|X|a < ∞ for some a ≥ 1 and E|Y | < ∞, then E|X + Y |a ≥ E|X + EY |a . 98. Let PY be a discrete distribution on {0, 1, 2, ...} and PX|Y =y be the binomial distribution Bi(p, y). Let (X, Y ) be the random vector having the joint c.d.f. given by (1.66). Show that (a) if Y has the Poisson distribution P (θ), then the marginal distribution of X is the Poisson distribution P (pθ); (b) if Y + r has the negative binomial distribution N B(π, r), then the marginal distribution of X + r is the negative binomial distribution N B(π/[1 − (1 − p)(1 − π)], r). 99. Let X1 , X2 , ... be i.i.d. random variables and Y be a discrete random variable taking positivePinteger values. Assume that Y and Xi ’s are Y independent. Let Z = i=1 Xi . (a) Obtain the ch.f. of Z. (b) Show that EZ = EY EX1 . (c) Show that Var(Z) = EY Var(X1 ) + Var(Y )(EX1 )2 . 100. Let X, Y , and Z be random variables having a positive joint Lebesgue p.d.f. Let fX|Y (x|y) and fX|Y,Z (x|y, z) be respectively the conditional p.d.f. of X given Y and the conditional p.d.f. of X given (Y, Z), as defined by (1.61). Show that Var(1/fX|Y (X|Y )|X) ≤ Var(1/fX|Y,Z (X|Y, Z)|X) a.s., where Var(ξ|ζ) = E{[ξ − E(ξ|ζ)]2 |ζ} for any random variables ξ and ζ with Eξ 2 < ∞. 101. Let {Xn } be a Markov chain. Show that if g is a one-to-one Borel function, then {g(Xn )} is also a Markov chain. Give an example to show that {g(Xn )} may not be a Markov chain in general.

1.6. Exercises

85

102. A sequence of random vectors {Xn } is said to be a Markov chain of order r for a positive integer r if P (B|X1 , ..., Xn ) = P (B|Xn−r+1 , ..., Xn ) a.s. for any B ∈ σ(Xn+1 ) and n = r, r + 1, .... (a) Let s > r be two positive integers. Show that if {Xn } is a Markov chain of order r, then it is a Markov chain of order s. (b) Let {Xn } be a sequence of random variables, r be a positive integer, and Yn = (Xn , Xn+1 , ..., Xn+r−1 ). Show that {Yn } is a Markov chain if and only if {Xn } is a Markov chain of order r. (c) (Autoregressive process of order r). Let {εn } be a sequence of independent random variables and r be a positive P integer. Show that r {Xn } is a Markov chain of order r, where Xn = j=1 ρj Xn−j + εn and ρj ’s are constants. 103. Show that if {Xn , Fn } is a martingale (or a submartingale), then E(Xn+j |Fn ) = Xn a.s. (or E(Xn+j |Fn ) ≥ Xn a.s.) and EX1 = EXj (or EX1 ≤ EX2 ≤ · · ·) for any j = 1, 2, .... 104. Show that {Xn } in Example 1.25 is a martingale. 105. Let {Xj } and {Zj } be sequences of random variables and let fn and gn denote the Lebesgue p.d.f.’s of Yn = (X1 , ..., Xn ) and (Z1 , ..., Zn ), respectively, n = 1, 2, .... Define λn = −gn (Yn )/fn (Yn )I{fn (Yn )>0} , n = 1, 2, .... Show that {λn } is a submartingale. 106. Let {Yn } be a sequence of independent random variables. (a) Suppose that EYn = 0 for all n. Let X1 = Y1 and Xn+1 = Xn + Yn+1 hn (X1 , ..., Xn ), n ≥ 2, where {hn } is a sequence of Borel functions. Show that {Xn } is a martingale. (b) that EYn = 0 and Var(Yn ) = σ 2 for all n. Let Xn = PnSuppose 2 ( j=1 Yj ) − nσ 2 . Show that {Xn } is a martingale. (c) Suppose that Yn > 0 and EYn = 1 for all n. Let Xn = Y1 · · · Yn . Show that {Xn } is a martingale. 107. Prove the claims in the proof of Proposition 1.14. 108. Show that every sequence of integrable random variables is the sum of a supermartingale and a submartingale. 109. Let {Xn } be a martingale. Show that if {Xn } is bounded either above or below, then supn E|Xn | < ∞. 2 110. Let {Xn } be a martingale satisfying PmEX1 = 0 and EXn 0, n=1 P (|Xn | ≥ ǫ) < ∞.

115. Let X1 , X2 , ... be a sequence of identically distributed random variables with a finite E|X1 | and let Yn = n−1 maxi≤n |Xi |. Show that (a) Yn →L1 0; (b) Yn →a.s. 0. 116. Let X, X1 , X2 , ... be random variables. Find an example for each of the following cases: (a) Xn →p X, but {Xn } does not converge to X a.s. (b) Xn →p X, but {Xn } does not converge to X in Lp for any p > 0. (c) Xn →d X, but {Xn } does not converge to X in probability (do not use Example 1.26). (d) Xn →p X, but {g(Xn )} does not converge to g(X) in probability for some function g.

117. Let X1 , X2 , ... be random variables. Show that (a) {|Xn |} is uniformly integrable if and only if supn E|Xn | < ∞ and, for any ǫ > 0, there is a δǫ > 0 such that supn E(|Xn |IA ) < ǫ for any event A with P (A) < δǫ ; (b) supn E|Xn |1+δ < ∞ for a δ > 0 implies that {|Xn |} is uniformly integrable. 118. Let X, X1 , X2 , ... be random variables satisfying P (|Xn | ≥ c) ≤ P (|X| ≥ c) for all n and c > 0. Show that if E|X| < ∞, then {|Xn |} is uniformly integrable. 119. Let X1 , X2 , ... and Y1 , Y2 , ... be random variables. Show that (a) if {|Xn |} and {|Yn |} are uniformly integrable, then {|Xn + Yn |} is uniformly integrable; Pn (b) if {|Xn |} is uniformly integrable, then {|n−1 i=1 Xi |} is uniformly integrable. 120. Let Y be an integrable random variable and {Fn } be a sequence of σ-fields. Show that {|E(Y |Fn )|} is uniformly integrable. 121. Let X, Y, X1 , X2 , ... be random variables satisfying Xn →p X and P (|Xn | ≤ |Y |) = 1 for all n. Show that if E|Y |r < ∞ for some r > 0, then Xn →Lr X. 122. Let X1 , X2 , ... be a sequence of random k-vectors. Show that Xn →p 0 if and only if E[kXn k/(1 + kXn k)] → 0.

1.6. Exercises

87

123. Let X, X1 , X2 , ... be random variables. Show that Xn →p X if and only if, for any subsequence {nk } of integers, there is a further subsequence {nj } ⊂ {nk } such that Xnj →a.s. X as j → ∞. 124. Let X1 , X2 , ... be a sequence of random variables satisfying |Xn | ≤ C1 and Var(Xn ) ≥ C2 for all n, where Ci ’s are positive constants. Show that Xn →p 0 does not hold. 125. Prove Lemma 1.5. P∞(Hint for part (ii): use Chebyshev’s inequality to show that P ( n=1 IAn = ∞) = 1, which can be shown to be equivalent to the result in (ii).) 126. Prove part (vii) of Theorem 1.8. 127. Let X, X1 , X2 , ..., Y1 , Y2 , ..., Z1 , Z2 , ... be random variables. Prove the following statements. (a) If Xn →d X, then Xn = Op (1). (b) If Xn = Op (Zn ) and P (Yn = 0) = 0, then Xn Yn = Op (Yn Zn ). (c) If Xn = Op (Zn ) and Yn = Op (Zn ), then Xn + Yn = Op (Zn ). (d) If E|Xn | = O(an ), then Xn = Op (an ), where an ∈ (0, ∞). (e) If Xn →a.s. X, then supn |Xn | = Op (1). 128. Let {Xn } and {Yn } be two sequences of random variables such that Xn = Op (1) and P (Xn ≤ t, Yn ≥ t + ǫ) + P (Xn ≥ t + ǫ, Yn ≤ t) = o(1) for any fixed t ∈ R and ǫ > 0. Show that Xn − Yn = op (1). 129. Let {Fn } be a sequence of c.d.f.’s on R, Gn (x) = Fn (an x + cn ), and Hn (x) = Fn (bn x+ dn ), where {an } and {bn } are sequences of positive numbers and {cn } and {dn } are sequences of real numbers. Suppose that Gn →w G and Hn →w H, where G and H are nondegenerate c.d.f.’s. Show that an /bn → a > 0, (cn − dn )/an → b ∈ R, and H(ax + b) = G(x) for all x ∈ R. 130. Let {Pn } be a sequence of probability measures on (R, B) and f be a R nonnegative Borel function such that supn f dPn < ∞ and f (x) → 0 as |x| → ∞. Show that {Pn } is tight. 131. Let P, P1 , P2 , ... be probability measures on (Rk , B k ). Show that if Pn (O) → P (O) for every open subset of R, then Pn (B) → P (B) for every B ∈ B k . 132. Let P, P1 , P2 , ... be probability measures on (R, B). Show that Pn →w P if and only if there exists a dense subset D of R such that limn→∞ Pn ((a, b]) = P ((a, b]) for any a < b, a ∈ D and b ∈ D. 133. Let Fn , n = 0, 1, 2, ..., be c.d.f.’s such that Fn →w F0 . Let Gn (U ) = sup{x : Fn (x) ≤ U }, n = 0, 1, 2, ..., where U is a random variable having the uniform U (0, 1) distribution. Show that Gn (U ) →p G0 (U ).

88

1. Probability Theory

134. Let P, P1 , P2 , ... be probability measures on (R, B). Suppose that Pn →w P and {gn } is a sequence of bounded R continuous R functions on R converging uniformly to g. Show that gn dPn → gdP .

135. Let X, X1 , X2 , ... be random k-vectors and Y, Y1 , Y2 , ... be random lvectors. Suppose that Xn →d X, Yn →d Y , and Xn and Yn are independent for each n. Show that (Xn , Yn ) converges in distribution to a random (k + l)-vector. 136. Let X1 , X2 , ... be independent variables with P (Xn = ±2−n ) Prandom n 1 = 2 , n = 1, 2, .... Show that i=1 Xi →d U , where U has the uniform distribution U (−1, 1). 137. Let {Xn } and {Yn } be two sequences of random variables. Suppose that Xn →d X and that PYn |Xn =xn →w PY almost surely for every sequence of numbers {xn }, where X and Y are independent random variables. Show that Xn + Yn →d X + Y . 138. Let X1 , X2 , ... be i.i.d. random variables having the ch.f. of the form 1 − c|t|a + o(|t|a )Pas t → 0, where 0 < a ≤ 2. Determine the constants b and u so that ni=1 Xi /(bnu ) converges in distribution to a random a variable having ch.f. e−|t| . 139. Let X, X1 , X2 , ... be random k-vectors and A1 , A2 , ... be events. Suppose that Xn →d X. Show that Xn IAn →d X if and only if P (An ) → 1. 140. Let Xn be a random variable having the N (µn , σn2 ) distribution, n = 1, 2,..., and X be a random variable having the N (µ, σ 2 ) distribution. Show that Xn →d X if and only if µn → µ and σn → σ. 141. Suppose that Xn is a random variable having the binomial distribution Bi(pn , n). Show that if npn → θ > 0, then Xn →d X, where X has the Poisson distribution P (θ). 142. Let fn be the Lebesgue p.d.f. of the t-distribution tn , n = 1, 2,.... Show that fn (x) → f (x) for any x ∈ R, where f is the Lebesgue p.d.f. of the standard normal distribution. 143. Prove Theorem 1.10. 144. Show by example that Xn →d X and Yn →d Y does not necessarily imply that g(Xn , Yn ) →d g(X, Y ), where g is a continuous function. 145. Prove Theorem 1.11(ii)-(iii) and Theorem 1.12(ii). Extend Theorem 1.12(i) to the case where g is a function from Rp to Rq with 2 ≤ q ≤ p. 146. Let U1 , U2 , ... be i.i.d. random variables having the uniform distribuQn √ −1/n tion on [0, 1] and Yn = ( i=1 Ui ) . Show that n(Yn − e) →d N (0, e2 ).

89

1.6. Exercises −1 147. Prove Lemma Pn1.6. (Hint: an where bn = i=1 xi /ai .)

Pn

i=1

xi = bn − a−1 n

Pn−1 i=0

bi (ai+1 − ai ),

148. In Theorem 1.13, (a) prove (1.82) for bounded ci ’s whenPE|X1 | < ∞; n (b) show that if EX1 = ∞, then n−1 i=1 Xi →P a.s. ∞; n (c) show that if E|X1 | = ∞, then P (lim supn {| i=1 Xi | > cn}) = P (lim supn {|XP n | > cn}) = 1 for any fixed positive constant c, and n lim supn |n−1 i=1 Xi | = ∞ a.s.

that for x = 3, 4, ..., 149. Let X1 , ..., Xn be i.i.d. random variables such P∞ −2 P (X1 = ±x) = (2cx2 logP x)−1 , where c = / log x. Show x=3 x n −1 that E|X1 | = ∞ but n X → 0, using Theorem 1.13(i). i p i=1

150. Let X1 , X2 , ... be i.i.d. random variables satisfying P (X1 = 2j ) = 2−j , j = 1, 2, .... Show that the WLLN does not hold for {Xn }, i.e., (1.80) does not hold for any {an }. Suppose that, as 151. Let X1 , XP 2 , ... be independent random variables. Pn n −2 2 n → ∞, P (|X | > n) → 0 and n E(X i i I{|Xi |≤n} ) → i=1 i=1P n 0. Show that (Tn − bn )/n →p 0, where Tn = i=1 Xi and bn = P n E(X I ). i {|Xi |≤n} i=1 Pn 152. Let Tn = i=1 Xi , where Xn ’s are independent random variables satisfying P (Xn = ±nθ ) = 0.5 and θ > 0 is a constant. Show that (a) when θ < 0.5, Tn /n →a.s. 0; (b) when θ ≥ 1, Tn /n →p 0 does not hold. 153. Let X2 , X3 , ... bepa sequence of independent random variables satisfying P (Xn = ± n/ log n) = 0.5. Show that (1.86) does not hold for p ∈ [1, 2] but (1.88) is satisfied for p = 2 and, thus, (1.89) holds. 154. Let X1 , ..., Xn beP i.i.d. random variables with Var(X1 ) < ∞. Show n that [n(n + 1)]−1 j=1 jXj →p EX1 . P ¯ = n Xi /n. 155. Let {Xn } be a sequence of random variables and let X i=1 ¯ →a.s. 0. (a) Show that if Xn →a.s. 0, then X ¯ →Lr 0, where r ≥ 1 is a constant. (b) Show that if Xn →Lr 0, then X (c) Show that the result in (b) may not be true for r ∈ (0, 1). ¯ →p 0. (d) Show that Xn →p 0 may not imply X 156. Let X1 , ..., Xn be random variables and {µn }, {σn }, {an }, and {bn } be sequences of real numbers with σn ≥ 0 and an ≥ 0. Suppose that Xn is asymptotically distributed as N (µn , σn2 ). Show that an Xn + bn is asymptotically distributed as N (µn , σn2 ) if and only if an → 1 and [µn (an − 1) + bn ]/σn → 0.

90

1. Probability Theory

157. Show that Liapounov’s condition (1.97) implies Lindeberg’s condition (1.92). 158. Let X1 , XP 2 , ... be a sequence of independent random variables and σn2 = Var( nj=1 Xj ). −n (a) Show that if Xn = N (0, Pn2 ), n = 1, 2, ..., then Feller’s condition (1.96) does not hold but j=1 (Xj − EXj )/σn →d N (0, 1). (b) Show that the result in (a) is still true if X1 has the uniform distribution U (−1, 1) and Xn = N (0, 2n−1 ), n = 2, 3, .... 159. In Example 1.33, show that (a) the condition σn2 → ∞ is also necessary for (1.98); Pn −1 (b) n P i=1 (Xi − pi ) →Lr 0 for any constant r > 0; (c) n−1 ni=1 (Xi − pi ) →a.s. 0.

160. Prove Corollary 1.3.

161. Suppose that Xn is a random variable having the binomial distribution Bi(θ, n), where 0 < θ < 1, n = 1, 2,.... Define Yn = log(Xn /n) when√Xn ≥ 1 and Yn = 1 when Xn = 0. Show that Yn →a.s. log θ and n(Yn − log θ) →d N 0, 1−θ . Establish similar results when θ Xn has the Poisson distribution P (nθ). 162. Let X1 , X2 , ... be independent random variables such that Xj has the uniform distribution on [−j, j], j = 1, 2,.... Show that Lindeberg’s condition is satisfied and state the resulting CLT. 163. Let X1 , X2 , ... be independent random variables such that for j = 1, 2,..., P (Xj = ±j a ) = 6−1 j −2(a−1) and P (Xj = 0) = 1 − 3−1 j −2(a−1) , where a > 1 is a constant. Show that Lindeberg’s condition is satisfied if and only if a < 1.5. 164. Let X1 , X2 , ... be independent random variables with P (Xj = ±j a ) = P (Xj = 0) = 1/3, where a > 0, j = 1, 2,.... Can we apply Theorem 1.15 to {Xj } by checking Liapounov’s condition (1.97)? 165. Let {X Pn } be a sequence of independent random variables.PSuppose that nj=1 (Xj − EXj )/σn →d N (0, 1), where σn2 = Var( nj=1 Xj ). Pn Show that n−1 j=1 (Xj − EXj ) →p 0 if and only if σn = o(n). p 166. Consider Exercise 152. Show that Tn / Var(Tn ) →d N (0, 1) and, when 0.5 ≤ θ < 1, Tn /n →p 0 does not hold. 167. Prove (1.102)-(1.104). √ ¯ √ 168. In Example 1.34, prove σh2 = 1 for n(X − µ)/ˆ σ and n(ˆ σ 2 − σ 2 )/ˆ τ and derive the expressions for p1 (x) in all four cases.

Chapter 2

Fundamentals of Statistics This chapter discusses some fundamental concepts of mathematical statistics. These concepts are essential for the material in later chapters.

2.1 Populations, Samples, and Models A typical statistical problem can be described as follows. One or a series of random experiments is performed; some data from the experiment(s) are collected; and our task is to extract information from the data, interpret the results, and draw some conclusions. In this book we do not consider the problem of planning experiments and collecting data, but concentrate on statistical analysis of the data, assuming that the data are given. A descriptive data analysis can be performed to obtain some summary measures of the data, such as the mean, median, range, standard deviation, etc., and some graphical displays, such as the histogram and boxand-whisker diagram, etc. (see, e.g., Hogg and Tanis (1993)). Although this kind of analysis is simple and requires almost no assumptions, it may not allow us to gain enough insight into the problem. We focus on more sophisticated methods of analyzing data: statistical inference and decision theory.

2.1.1 Populations and samples In statistical inference and decision theory, the data set is viewed as a realization or observation of a random element defined on a probability space (Ω, F, P ) related to the random experiment. The probability measure P is called the population. The data set or the random element that produces 91

92

2. Fundamentals of Statistics

the data is called a sample from P . The size of the data set is called the sample size. A population P is known if and only if P (A) is a known value for every event A ∈ F. In a statistical problem, the population P is at least partially unknown and we would like to deduce some properties of P based on the available sample. Example 2.1 (Measurement problems). To measure an unknown quantity θ (for example, a distance, weight, or temperature), n measurements, x1 , ..., xn , are taken in an experiment of measuring θ. If θ can be measured without errors, then xi = θ for all i; otherwise, each xi has a possible measurement error. In descriptive data analysis, a few summary measures may be calculated, for example, the sample mean n

x ¯= and the sample variance

1X xi n i=1 n

1 X 2 s = (xi − x ¯) . n − 1 i=1 2

However, what is the relationship between x ¯ and θ? Are they close (if not equal) in some sense? The sample variance s2 is clearly an average of squared deviations of xi ’s from their mean. But, what kind of information does s2 provide? Finally, is it enough to just look at x¯ and s2 for the purpose of measuring θ? These questions cannot be answered in descriptive data analysis. In statistical inference and decision theory, the data set, (x1 , ..., xn ), is viewed as an outcome of the experiment whose sample space is Ω = Rn . We usually assume that the n measurements are obtained in n independent trials of the experiment. Hence, we can define a random n-vector Qn X = (X1 , ..., Xn ) on i=1 (R, B, P ) whose realization is (x1 , ..., xn ). The population in this problem is P (note that the product probability measure is determined by P ) and is at least partially unknown. The random vector X is a sample and n is the sample size. Define n

X ¯= 1 X Xi n i=1

and

(2.1)

n

S2 =

1 X ¯ 2. Xi − X n − 1 i=1

(2.2)

¯ and S 2 are random variables that produce x Then X ¯ and s2 , respectively. Questions raised previously can be answered if some assumptions are imposed on the population P , which are discussed later.

93

2.1. Populations, Samples, and Models

When the sample (X1 , ..., Xn ) has i.i.d. components, which is often the case in applications, the population is determined by the marginal distribution of Xi . Example 2.2 (Life-time testing problems). Let x1 , ..., xn be observed lifetimes of some electronic components. Again, in statistical inference and decision theory, x1 , ..., xn are viewed as realizations of independent random variables X1 , ..., Xn . Suppose that the components are of the same type so that it is reasonable to assume that X1 , ..., Xn have a common marginal c.d.f. F . Then the population is F , which is often unknown. A quantity of interest in this problem is 1 − F (t) with a t > 0, which is the probability that a component does not fail at time t. It is possible that all xi ’s are smaller (or larger) than t. Conclusions about 1 − F (t) can be drawn based on data x1 , ..., xn when certain assumptions on F are imposed. Example 2.3 (Survey problems). A survey is often conducted when one is not able to evaluate all elements in a collection P = {y1 , ..., yN } containing N values in Rk , where k and N are finite positive integers but N may be very large. PN Suppose that the quantity of interest is the population total Y = i=1 yi . In a survey, a subset s of n elements are selected from {1, ..., N } and values yi , i ∈ s, are obtained. Can we draw some conclusion about Y based on data yi , i ∈ s? How do we define some random variables that produce the survey data? First, we need to specify how s is selected. A commonly used probability sampling plan can be described as follows. Assume that every element in {1, ..., N } can be selected at most once, i.e., we consider sampling without replacement. Let S be the collection of all subsets of n distinct elements from {1, ..., N }, Fs be the collection of all subsets of S, and p be a probability measure on (S, Fs ). Any s ∈ S is selected with probability p(s). Note that p(s) is a known value whenever s is given. Let X1 , ..., Xn be random variables such that P (X1 = yi1 , ..., Xn = yin ) =

p(s) , n!

s = {i1 , ..., in } ∈ S.

(2.3)

Then (yi , i ∈ s) can be viewed as a realization of the sample (X1 , ..., Xn ). If p(s) is constant, then the sampling plan is called the simple random sampling (without replacement) and (X1 , ..., Xn ) is called a simple random sample. Although X1 , ..., Xn are identically distributed, they are not necessarily independent. Thus, unlike in the previous two examples, the population in this problem may not be specified by the marginal distributions of Xi ’s. The population is determined by P and the known selection probability measure p. For this reason, P is often treated as the population. Conclusions about Y and other characteristics of P can be drawn based on data yi , i ∈ s, which are discussed later.

94

2. Fundamentals of Statistics

2.1.2 Parametric and nonparametric models A statistical model (a set of assumptions) on the population P in a given problem is often postulated to make the analysis possible or easy. Although testing the correctness of postulated models is part of statistical inference and decision theory, postulated models are often based on knowledge of the problem under consideration. Definition 2.1. A set of probability measures Pθ on (Ω, F ) indexed by a parameter θ ∈ Θ is said to be a parametric family if and only if Θ ⊂ Rd for some fixed positive integer d and each Pθ is a known probability measure when θ is known. The set Θ is called the parameter space and d is called its dimension. A parametric model refers to the assumption that the population P is in a given parametric family. A parametric family {Pθ : θ ∈ Θ} is said to be identifiable if and only if θ1 6= θ2 and θi ∈ Θ imply Pθ1 6= Pθ2 . In most cases an identifiable parametric family can be obtained through reparameterization. Hence, we assume in what follows that every parametric family is identifiable unless otherwise stated. Let P be a family of populations and ν be a σ-finite measure on (Ω, F ). If P ≪ ν for all P ∈ P, then P is said to be dominated by ν, in which case P dPθ can be identified by the family of densities { dP dν : P ∈ P} (or { dν : θ ∈ Θ} for a parametric family). Many examples of parametric families can be obtained from Tables 1.1 and 1.2 in §1.3.1. All parametric families from Tables 1.1 and 1.2 are dominated by the counting measure or the Lebesgue measure on R. Example 2.4 (The k-dimensional normal family). Consider the normal distribution Nk (µ, Σ) given by (1.24) for a fixed positive integer k. An important parametric family in statistics is the family of normal distributions P = {Nk (µ, Σ) : µ ∈ Rk , Σ ∈ Mk }, where Mk is a collection of k ×k symmetric positive definite matrices. This family is dominated by the Lebesgue measure on Rk . In the measurement problem described in Example 2.1, Xi ’s are often i.i.d. from the N (µ, σ 2 ) distribution. Hence, we can impose a parametric model on the population, i.e., P ∈ P = {N (µ, σ 2 ) : µ ∈ R, σ 2 > 0}. The normal parametric model is perhaps not a good model for the lifetime testing problem described in Example 2.2, since clearly Xi ≥ 0 for all i. In practice, the normal family {N (µ, σ 2 ) : µ ∈ R, σ 2 > 0} can be used for a life-time testing problem if one puts some restrictions on µ and σ so that P (Xi < 0) is negligible. Common parametric models for

2.1. Populations, Samples, and Models

95

life-time testing problems are the exponential model (containing the exponential distributions E(0, θ) with an unknown parameter θ; see Table 1.2 in §1.3.1), the gamma model (containing the gamma distributions Γ(α, γ) with unknown parameters α and γ), the log-normal model (containing the log-normal distributions LN (µ, σ2 ) with unknown parameters µ and σ), the Weibull model (containing the Weibull distributions W (α, θ) with unknown parameters α and θ), and any subfamilies of these parametric families (e.g., a family containing the gamma distributions with one known parameter and one unknown parameter). The normal family is often not a good choice for the survey problem discussed in Example 2.3. In a given problem, a parametric model is not useful if the dimension of Θ is very high. For example, the survey problem described in Example 2.3 has a natural parametric model, since the population P can be indexed by the parameter θ = (y1 , ..., yN ). If there is no restriction on the y-values, however, the dimension of the parameter space is kN , which is usually much larger than the sample size n. If there are some restrictions on the y-values (for example, yi ’s are nonnegative integers no larger than a fixed integer m), then the dimension of the parameter space is at most m + 1 and the parametric model becomes useful. A family of probability measures is said to be nonparametric if it is not parametric according to Definition 2.1. A nonparametric model refers to the assumption that the population P is in a given nonparametric family. There may be almost no assumption on a nonparametric family, for example, the family of all probability measures on (Rk , B k ). But in many applications, we may use one or a combination of the following assumptions to form a nonparametric family on (Rk , B k ): (1) The joint c.d.f.’s are continuous. (2) The joint c.d.f.’s have finite moments of order ≤ a fixed integer. (3) The joint c.d.f.’s have p.d.f.’s (e.g., Lebesgue p.d.f.’s). (4) k = 1 and the c.d.f.’s are symmetric. For instance, in Example 2.1, we may assume a nonparametric model with symmetric and continuous c.d.f.’s. The symmetry assumption may not be suitable for the population in Example 2.2, but the continuity assumption seems to be reasonable. In statistical inference and decision theory, methods designed for parametric models are called parametric methods, whereas methods designed for nonparametric models are called nonparametric methods. However, nonparametric methods are used in a parametric model when parametric methods are not effective, such as when the dimension of the parameter

96

2. Fundamentals of Statistics

space is too high (Example 2.3). On the other hand, parametric methods may be applied to a semi-parametric model, which is a nonparametric model having a parametric component. Some examples are provided in §5.1.4.

2.1.3 Exponential and location-scale families In this section, we discuss two types of parametric families that are of special importance in statistical inference and decision theory. Definition 2.2 (Exponential families). A parametric family {Pθ : θ ∈ Θ} dominated by a σ-finite measure ν on (Ω, F ) is called an exponential family if and only if dPθ (ω) = exp [η(θ)]τ T (ω) − ξ(θ) h(ω), dν

ω ∈ Ω,

(2.4)

where exp{x} = ex , T is a random p-vector with a fixed positive integer p, η is a function R from Θ to Rp , h is a nonnegative Borel function on (Ω, F), and ξ(θ) = log Ω exp{[η(θ)]τ T (ω)}h(ω)dν(ω) .

In Definition 2.2, T and h are functions of ω only, whereas η and ξ are functions of θ only. Ω is usually Rk . The representation (2.4) of an exponential family is not unique. In fact, any transformation η˜(θ) = Dη(θ) with a p × p nonsingular matrix D gives another representation (with T replaced by T˜ = (Dτ )−1 T ). A change of the measure that dominates the Rfamily also changes the representation. For example, if we define λ(A) = hdν for any A ∈ F, then we obtain an exponential family with densities A dPθ (ω) = exp [η(θ)]τ T (ω) − ξ(θ) . dλ

(2.5)

In an exponential family, consider the reparameterization η = η(θ) and fη (ω) = exp η τ T (ω) − ζ(η) h(ω), ω ∈ Ω, (2.6) R where ζ(η) = log Ω exp{η τ T (ω)}h(ω)dν(ω) . This is the canonical form for the family, which is not unique for the reasons discussed previously. The new parameter η is called the natural parameter. The new parameter space Ξ = {η(θ) : θ ∈ Θ}, a subset of Rp , is called the natural parameter space. An exponential family in canonical form is called a natural exponential family. If there is an open set contained in the natural parameter space of an exponential family, then the family is said to be of full rank. Example 2.5. Let Pθ be the binomial distribution Bi(θ, n) with parameter θ, where n is a fixed positive integer. Then {Pθ : θ ∈ (0, 1)} is an

2.1. Populations, Samples, and Models

97

exponential family, since the p.d.f. of Pθ w.r.t. the counting measure is n o n θ + n log(1 − θ) I{0,1,...,n} (x) fθ (x) = exp x log 1−θ x θ , ξ(θ) = −n log(1 − θ), and h(x) = nx I{0,1,...,n} (x)). (T (x) = x, η(θ) = log 1−θ θ If we let η = log 1−θ , then Ξ = R and the family with p.d.f.’s n I{0,1,...,n} (x) fη (x) = exp {xη − n log(1 + eη )} x is a natural exponential family of full rank. Example 2.6. The normal family {N (µ, σ 2 ) : µ ∈ R, σ > 0} is an exponential family, since the Lebesgue p.d.f. of N (µ, σ 2 ) can be written as 1 µ 1 2 µ2 √ exp x − x − − log σ . σ2 2σ 2 2σ 2 2π

µ2 Hence, T (x) = (x, −x2 ), η(θ) = σµ2 , 2σ1 2 , θ = (µ, σ 2 ), ξ(θ) = 2σ 2 + log σ, √ and h(x) = 1/ 2π. Let η = (η1 , η2 ) = σµ2 , 2σ1 2 . Then Ξ = R × (0, ∞) and we can obtain √ a natural exponential family of full rank with ζ(η) = η12 /(4η2 ) + log(1/ 2η2 ). A subfamily of the previous normal family, {N (µ, µ2 ) : µ ∈ R, µ 6= 0}, is also an exponential family with the natural parameter η = µ1 , 2µ1 2 and natural parameter space Ξ = {(x, y) : y = 2x2 , x ∈ R, y > 0}. This exponential family is not of full rank. For an exponential family, (2.5) implies that there is a nonzero measure λ such that dPθ (ω) > 0 for all ω and θ. (2.7) dλ We can use this fact to show that a family of distributions is not an exponential family. For example, consider the family of uniform distributions, i.e., Pθ is U (0, θ) with an unknown θ ∈ (0, ∞). If {Pθ : θ ∈ (0, ∞)} is an exponential family, then from the previous discussion we have a nonzero measure λ such that (2.7) holds. For any t > 0, there is a θ < t such that Pθ ([t, ∞)) = 0, which with (2.7) implies that λ([t, ∞)) = 0. Also, for any t ≤ 0, Pθ ((−∞, t]) = 0, which with (2.7) implies that λ((−∞, t]) = 0. Since t is arbitrary, λ ≡ 0. This contradiction implies that {Pθ : θ ∈ (0, ∞)} cannot be an exponential family. The reader may verify which of the parametric families from Tables 1.1 and 1.2 are exponential families. As another example, we consider an important exponential family containing multivariate discrete distributions.

98

2. Fundamentals of Statistics

Example 2.7 (The multinomial family). Consider an experiment having k + 1 possiblePoutcomes with pi as the probability for the ith outcome, k i = 0, 1, ..., k, i=0 pi = 1. In n independent trials of this experiment, let Xi be the number of trials resulting in the ith outcome, i = 0, 1, ..., k. Then the joint p.d.f. (w.r.t. counting measure) of (X0 , X1 , ..., Xk ) is n! px0 px1 · · · pxk k IB (x0 , x1 , ..., xk ), x0 !x1 ! · · · xk ! 0 1 P where B = {(x0 , x1 , ..., xk ) : xi ’s are integers ≥ 0, ki=0 xi = n} and θ = (p0 , p1 , ..., pk ). The distribution of (X0 , X1 , ..., Xk ) is called the multinomial distribution, which is an extension of the binomial distribution. In fact, the marginal c.d.f. of each Xi is the binomial distribution Bi(pi , n). Let P Θ = {θ ∈ Rk+1 : 0 < pi < 1, ki=0 pi = 1}. The parametric family {fθ : θ ∈ Θ} is called the multinomial family. Let x = (x0 , x1 , ..., xk ), η = (log p0 , log p1 , ..., log pk ), and h(x) = [n!/(x0 !x1 ! · · · xk !)]IB (x). Then fθ (x0 , x1 , ..., xk ) =

fθ (x0 , x1 , ..., xk ) = exp {η τ x} h(x),

x ∈ Rk+1 .

(2.8)

Hence, the multinomial family is a natural exponential family with natural parameter η. However, representation (2.8) does not provide an exponential family of full rank, since there is no open set of Rk+1 contained in the natural parameter space. A reparameterization leads to anPexponential P family with full rank. Using the fact that ki=0 Xi = n and ki=0 pi = 1, we obtain that fθ (x0 , x1 , ..., xk ) = exp {η∗τ x∗ − ζ(η∗ )} h(x),

x ∈ Rk+1 ,

(2.9)

where x∗ = (x1 , ..., xk ), η∗ = (log(p1 /p0 ), ..., log(pk /p0 )), and ζ(η∗ ) = −n log p0 . The η∗ -parameter space is Rk . Hence, the family of densities given by (2.9) is a natural exponential family of full rank. If X1 , ..., Xm are independent random vectors with p.d.f.’s in exponential families, then the p.d.f. of (X1 , ..., Xm ) is again in an exponential family. The following result summarizes some other useful properties of exponential families. Its proof can be found in Lehmann (1986). Theorem 2.1. Let P be a natural exponential family given by (2.6). (i) Let T = (Y, U ) and η = (ϑ, ϕ), where Y and ϑ have the same dimension. Then, Y has the p.d.f. fη (y) = exp{ϑτ y − ζ(η)} w.r.t. a σ-finite measure depending on ϕ. In particular, T has a p.d.f. in a natural exponential family. Furthermore, the conditional distribution of Y given U = u has the p.d.f. (w.r.t. a σ-finite measure depending on u) fϑ,u (y) = exp{ϑτ y − ζu (ϑ)},

99

2.1. Populations, Samples, and Models

which is in a natural exponential family indexed by ϑ. (ii) If η0 is an interior point of the natural parameter space, then the m.g.f. ψη0 of Pη0 ◦ T −1 is finite in a neighborhood of 0 and is given by ψη0 (t) = exp{ζ(η0 + t) − ζ(η0 )}. R Furthermore, if f is a Borel function satisfying |f |dPη0 < ∞, then the function Z f (ω) exp{η τ T (ω)}h(ω)dν(ω)

is infinitely often differentiable in a neighborhood of η0 , and the derivatives may be computed by differentiation under the integral sign. Using Theorem 2.1(ii) and the result in Example 2.5, we obtain that the m.g.f. of the binomial distribution Bi(p, n) is ψη (t) = exp{n log(1 + eη+t ) − n log(1 + eη )} n 1 + eη et = 1 + eη = (1 − p + pet )n ,

since p = eη /(1 + eη ). Definition 2.3 (Location-scale families). Let P be a known probability measure on (Rk , B k ), V ⊂ Rk , and Mk be a collection of k × k symmetric positive definite matrices. The family {P(µ,Σ) : µ ∈ V, Σ ∈ Mk } is called a location-scale family (on Rk ), where P(µ,Σ) (B) = P Σ−1/2 (B − µ) ,

(2.10)

B ∈ Bk ,

Σ−1/2 (B − µ) = {Σ−1/2 (x − µ) : x ∈ B} ⊂ Rk , and Σ−1/2 is the inverse of the “square root” matrix Σ1/2 satisfying Σ1/2 Σ1/2 = Σ. The parameters µ and Σ1/2 are called the location and scale parameters, respectively.

The following are some important examples of location-scale families. The family {P(µ,Ik ) : µ ∈ Rk } is called a location family, where Ik is the k × k identity matrix. The family {P(0,Σ) : Σ ∈ Mk } is called a scale family. In some cases, we consider a location-scale family of the form {P(µ,σ2 Ik ) : µ ∈ Rk , σ > 0}. If X1 , ..., Xk are i.i.d. with a common distribution in the location-scale family {P(µ,σ2 ) : µ ∈ R, σ > 0}, then the joint distribution of the vector (X1 , ..., Xk ) is in the location-scale family {P(µ,σ2 Ik ) : µ ∈ V, σ > 0} with V = {(x, ..., x) ∈ Rk : x ∈ R}.

100

2. Fundamentals of Statistics

A location-scale family can be generated as follows. Let X be a random k-vector having a distribution P . Then the distribution of Σ1/2 X + µ is P(µ,Σ) . On the other hand, if X is a random k-vector whose distribution is in the location-scale family (2.10), then the distribution DX + c is also in the same family, provided that Dµ + c ∈ V and DΣDτ ∈ Mk . Let F be the c.d.f. of P . Then the c.d.f. of P(µ,Σ) is F Σ−1/2 (x − µ) , x ∈ Rk . If F has a Lebesgue p.d.f. f , then the Lebesgue p.d.f. of P(µ,Σ) is Det(Σ−1/2 )f Σ−1/2 (x − µ) , x ∈ Rk (Proposition 1.8). Many families of distributions in Table 1.2 (§1.3.1) are location, scale, or location-scale families. For example, the family of exponential distributions E(a, θ) is a location-scale family on R with location parameter a and scale parameter θ; the family of uniform distributions U (0, θ) is a scale family on R with a scale parameter θ. The k-dimensional normal family discussed in Example 2.4 is a location-scale family on Rk .

2.2 Statistics, Sufficiency, and Completeness Let us assume now that our data set is a realization of a sample X (a random vector) from an unknown population P on a probability space.

2.2.1 Statistics and their distributions A measurable function of X, T (X), is called a statistic if T (X) is a known value whenever X is known, i.e., the function T is a known function. Statistical analyses are based on various statistics, for various purposes. Of course, X itself is a statistic, but it is a trivial statistic. The range of a nontrivial statistic T (X) is usually simpler than that of X. For example, X may be a random n-vector and T (X) may be a random p-vector with a p much smaller than n. This is desired since T (X) simplifies the original data. From a probabilistic point of view, the “information” within the statistic T (X) concerning the unknown distribution of X is contained in the σfield σ(T (X)). To see this, assume that S is any other statistic for which σ(S(X)) = σ(T (X)). Then, by Lemma 1.2, S is a measurable function of T , and T is a measurable function of S. Thus, once the value of S (or T ) is known, so is the value of T (or S). That is, it is not the particular values of a statistic that contain the information, but the generated σ-field of the statistic. Values of a statistic may be important for other reasons. Note that σ(T (X)) ⊂ σ(X) and the two σ-fields are the same if and only if T is one-to-one. Usually σ(T (X)) simplifies σ(X), i.e., a statistic provides a “reduction” of the σ-field.

2.2. Statistics, Sufficiency, and Completeness

101

Any T (X) is a random element. If the distribution of X is unknown, then the distribution of T may also be unknown, although T is a known function. Finding the form of the distribution of T is one of the major problems in statistical inference and decision theory. Since T is a transformation of X, tools we learn in Chapter 1 for transformations may be useful in finding the distribution or an approximation to the distribution of T (X). Example 2.8. Let X1 , ..., Xn be i.i.d. random variables having a common ¯ and sample distribution P and X = (X1 , ..., Xn ). The sample mean X variance S 2 defined in (2.1) and (2.2), respectively, are two commonly used ¯ and S 2 ? statistics. Can we find the joint or the marginal distributions of X It depends on how much we know about P . ¯ and S 2 . Assume that P has a First, let us consider the moments of X finite mean denoted by µ. Then ¯ = µ. EX R ¯ = xdPθ = µ(θ) If P is in a parametric family {Pθ : θ ∈ Θ}, then E X for some function µ(·). Even if the form of µ is known, µ(θ) may still be unknown when θ is unknown. Assume now that P has a finite variance denoted by σ 2 . Then ¯ = σ 2 /n, Var(X) which equals σ 2 (θ)/n for some function σ 2 (·) if P is in a parametric family. With a finite σ 2 = Var(X1 ), we can also obtain that ES 2 = σ 2 . ¯ 3 and Cov(X, ¯ S 2 ), and with a With a finite E|X1 |3 , we can obtain E(X) 4 2 finite E|X1 | , we can obtain Var(S ) (exercise). ¯ If P is in a parametric family, we Next, consider the distribution of X. ¯ See Example 1.20 and some exercises can often find the distribution of X. ¯ is N (µ, σ 2 /n) if P is N (µ, σ 2 ); nX ¯ has the gamma in §1.6. For example, X distribution Γ(n, θ) if P is the exponential distribution E(0, θ). If P is not in a parametric family, then it is usually hard to find the exact form of the ¯ One can, however, use the CLT (§1.5.4) to obtain an distribution of X. ¯ Applying Corollary 1.2 (for the approximation to the distribution of X. case of k = 1), we obtain that √ ¯ − µ) →d N (0, σ 2 ) n(X

¯ can be approximated by N (µ, σ 2 /n), and, by (1.100), the distribution of X 2 where µ and σ are the mean and variance of P , respectively, and are assumed to be finite. ¯ the distribution of S 2 is harder to obtain. Assuming Compared to X, 2 that P is N (µ, σ ), one can show that (n − 1)S 2 /σ 2 has the chi-square

102

2. Fundamentals of Statistics

distribution χ2n−1 (see Example 2.18). An approximate distribution for ¯ and S 2 S 2 can be obtained from the approximate joint distribution of X discussed next. ¯ Under the assumption that P is N (µ, σ 2 ), it can be shown that X and S 2 are independent (Example 2.18). Hence, the joint distribution of ¯ S 2 ) is the product of the marginal distributions of X ¯ and S 2 given in the (X, previous discussion. Without the normality assumption, an approximate joint distribution can be obtained as follows. Assume again that µ = EX1 , σ 2 = Var(X1 ), and E|X1 |4 are finite. Let Yi = (Xi − µ, (Xi − µ)2 ), i = 1, ..., n. Then Y1 , ..., Yn are i.i.d. random 2-vectors with EY1 = (0, σ 2 ) and variance-covariance matrix E(X1 − µ)3 σ2 . Σ= E(X1 − µ)3 E(X1 − µ)4 − σ 4 Pn ¯ − µ, S˜2 ), where S˜2 = n−1 Pn (Xi − µ)2 . Note that Y¯ = n−1 i=1 Yi = (X i=1 Applying the CLT (Corollary 1.2) to Yi ’s, we obtain that √ ¯ − µ, S˜2 − σ 2 ) →d N2 (0, Σ). n(X Since

i n h ˜2 ¯ − µ)2 S − (X n−1 ¯ and X →a.s. µ (the SLLN, Theorem 1.13), an application of Slutsky’s theorem (Theorem 1.11) leads to √ ¯ − µ, S 2 − σ 2 ) →d N2 (0, Σ). n(X S2 =

Example 2.9 (Order statistics). Let X = (X1 , ..., Xn ) with i.i.d. random components and let X(i) be the ith smallest value of X1 , ..., Xn . The statistics X(1) , ..., X(n) are called the order statistics, which is a set of very useful statistics in addition to the sample mean and variance in the previous example. Suppose that Xi has a c.d.f. F having a Lebesgue p.d.f. f . Then the joint Lebesgue p.d.f. of X(1) , ..., X(n) is x1 < x2 < · · · < xn n!f (x1 )f (x2 ) · · · f (xn ) g(x1 , x2 , ..., xn ) = 0 otherwise. The joint Lebesgue p.d.f. of X(i) and X(j) , 1 ≤ i < j ≤ n, is ( i−1 j−i−1 n−j n![F (x)]

gi,j (x, y) =

[F (y)−F (x)] [1−F (y)] (i−1)!(j−i−1)!(n−j)!

f (x)f (y)

0

and the Lebesgue p.d.f. of X(i) is gi (x) =

n! [F (x)]i−1 [1 − F (x)]n−i f (x). (i − 1)!(n − i)!

x 0 and dP/dν > 0 a.e. ν on C. Then there exists a sequence {Ci } ⊂ C P such that ν(Ci ) → supC∈C ν(C). Let C0 be the union of all Ci ’s ∞ and Q = i=1 ci Pi , where Pi is the probability measure corresponding to Ci . Then C0 ∈ C (exercise). Suppose now that Q(A) = 0. Let P ∈ P0 and B = {x : dP/dν > 0}. Since Q(A ∩ C0 ) = 0, ν(A ∩ C0 ) = 0 and P (A ∩ C0 ) = 0. Then P (A) = P (A ∩ C0c ∩ B). If P (A ∩ C0c ∩ B) > 0, then ν(C0 ∪(A∩C0c ∩B)) > ν(C0 ), which contradicts ν(C0 ) = supC∈C ν(C) since A ∩ C0c ∩ B and therefore C0 ∪ (A ∩ C0c ∩ B) is in C. Thus, P (A) = 0 for all P ∈ P0 . Theorem 2.2 (The factorization theorem). Suppose that X is a sample from P ∈ P and P is a family of probability measures on (Rn , B n ) dominated by a σ-finite measure ν. Then T (X) is sufficient for P ∈ P if and only if there are nonnegative Borel functions h (which does not depend on P ) on (Rn , B n ) and gP (which depends on P ) on the range of T such that dP (x) = gP T (x) h(x). dν

(2.11)

2.2. Statistics, Sufficiency, and Completeness

105

Proof. (i) Suppose that T is sufficient for P ∈ P. Then, for any A ∈ B n , P (A|T ) does not depend on P . Let Q be the probability measure in Lemma 2.1. By Fubini’s theorem and the result in Exercise 35 of §1.6, Q(A ∩ B) = =

∞ X j=1 ∞ X

cj Pj (A ∩ B) cj

j=1

=

Z

Z X ∞

P (A|T )dPj B

cj P (A|T )dPj

B j=1

=

Z

P (A|T )dQ

B

for any B ∈ σ(T ). Hence, P (A|T ) = EQ (IA |T ) a.s. Q, where EQ (IA |T ) denotes the conditional expectation of IA given T w.r.t. Q. Let gP (T ) be the Radon-Nikodym derivative dP/dQ on the space (Rn , σ(T ), Q). From Propositions 1.7 and 1.10, Z P (A) = P (A|T )dP Z = EQ (IA |T )gP (T )dQ Z = EQ [IA gP (T )|T ]dQ Z dQ gP (T ) = dν dν A for any A ∈ B n . Hence, (2.11) holds with h = dQ/dν. (ii) Suppose that (2.11) holds. Then ∞ X ∞ dP dP X dPi = = gP (T ) ci gPi (T ) a.s. Q, dQ dν dν i=1 i=1

(2.12)

where the second equality follows from the result in Exercise 35 of §1.6. Let A ∈ σ(X) and P ∈ P. The sufficiency of T follows from P (A|T ) = EQ (IA |T ) a.s. P ,

(2.13)

where EQ (IA |T ) is given in part (i) of the proof. This is because EQ (IA |T ) does not vary with P ∈ P, and result (2.13) and Theorem 1.7 imply that the conditional distribution of X given T is determined by EQ (IA |T ), A ∈ σ(X). By the definition of conditional probability, (2.13) follows from Z Z IA dP = EQ (IA |T )dP (2.14) B

B

106

2. Fundamentals of Statistics

for any B ∈ σ(T ). Let B ∈ σ(T ). By (2.12), dP/dQ is a Borel function of T . Then, by Proposition 1.7(i), Proposition 1.10(vi), and the definition of the conditional expectation, the right-hand side of (2.14) is equal to Z Z Z dP dP dP dQ = T dQ = dQ, EQ (IA |T ) EQ IA IA dQ dQ dQ B B B which equals the left-hand side of (2.14). This proves (2.14) for any B ∈ σ(T ) and completes the proof.

If P is an exponential family with p.d.f.’s given by (2.4) and X(ω) = ω, then we can apply Theorem 2.2 with gθ (t) = exp{[η(θ)]τ t − ξ(θ)} and conclude that T is a sufficient statistic for θ ∈ Θ. In Example P 2.10 the joint distribution of X is in an exponential family with T (X) = ni=1 Xi . Hence, we can conclude that T is sufficient for θ ∈ (0, 1) without computing the conditional distribution of X given T . Example 2.11 (Truncation families). Let φ(x) be a positive Borel function Rb on (R, B) such that a φ(x)dx < ∞ for any a and b, −∞ < a < b < ∞. Let θ = (a, b), Θ = {(a, b) ∈ R2 : a < b}, and fθ (x) = c(θ)φ(x)I(a,b) (x), i−1 hR b . Then {fθ : θ ∈ Θ}, called a truncation where c(θ) = a φ(x)dx family, is a parametric family dominated by the Lebesgue measure on R. Let X1 , ..., Xn be i.i.d. random variables having the p.d.f. fθ . Then the joint p.d.f. of X = (X1 , ..., Xn ) is n Y

i=1

fθ (xi ) = [c(θ)]n I(a,∞) (x(1) )I(−∞,b) (x(n) )

n Y

φ(xi ),

(2.15)

i=1

where x(i) is the ith smallest value of x1 , ..., xn . LetQT (X) = (X(1) , X(n) ), n gθ (t1 , t2 ) = [c(θ)]n I(a,∞) (t1 )I(−∞,b) (t2 ), and h(x) = i=1 φ(xi ). By (2.15) and Theorem 2.2, T (X) is sufficient for θ ∈ Θ. Example 2.12 (Order statistics). Let X = (X1 , ..., Xn ) and X1 , ..., Xn be i.i.d. random variables having a distribution P ∈ P, where P is the family of distributions on R having Lebesgue p.d.f.’s. Let X(1) , ..., X(n) be the order statistics given in Example 2.9. Note that the joint p.d.f. of X is f (x1 ) · · · f (xn ) = f (x(1) ) · · · f (x(n) ). Hence, T (X) = (X(1) , ..., X(n) ) is sufficient for P ∈ P. The order statistics can be shown to be sufficient even when P is not dominated by any σ-finite measure, but Theorem 2.2 is not applicable (see Exercise 31 in §2.6).

107

2.2. Statistics, Sufficiency, and Completeness

There are many sufficient statistics for a given family P. In fact, if T is a sufficient statistic and T = ψ(S), where ψ is measurable and S is another statistic, then S is sufficient. This is obvious from Theorem 2.2 if the population has a p.d.f., but it can be proved directly Pm from PnDefinition 2.4 (Exercise 25). For instance, in Example 2.10, ( i=1 Xi , i=m+1 Xi ) is sufficient for θ, where m is any fixed integer between 1 and n. If T is sufficient and T = ψ(S) with a measurable ψ that is not one-to-one, then σ(T ) ⊂ σ(S) and T is more useful than S, since T provides a further reduction of the data (or σ-field) without loss of information. Is there a sufficient statistic that provides “maximal” reduction of the data? Before introducing the next concept, we need the following notation. If a statement holds except for outcomes in an event A satisfying P (A) = 0 for all P ∈ P, then we say that the statement holds a.s. P. Definition 2.5 (Minimal sufficiency). Let T be a sufficient statistic for P ∈ P. T is called a minimal sufficient statistic if and only if, for any other statistic S sufficient for P ∈ P, there is a measurable function ψ such that T = ψ(S) a.s. P. If both T and S are minimal sufficient statistics, then by definition there is a one-to-one measurable function ψ such that T = ψ(S) a.s. P. Hence, the minimal sufficient statistic is unique in the sense that two statistics that are one-to-one measurable functions of each other can be treated as one statistic. Example 2.13. Let X1 , ..., Xn be i.i.d. random variables from Pθ , the uniform distribution U (θ, θ + 1), θ ∈ R. Suppose that n > 1. The joint Lebesgue p.d.f. of (X1 , ..., Xn ) is fθ (x) =

n Y

i=1

I(θ,θ+1) (xi ) = I(x(n) −1,x(1) ) (θ),

x = (x1 , ..., xn ) ∈ Rn ,

where x(i) denotes the ith smallest value of x1 , ..., xn . By Theorem 2.2, T = (X(1) , X(n) ) is sufficient for θ. Note that x(1) = sup{θ : fθ (x) > 0} and x(n) = 1 + inf{θ : fθ (x) > 0}. If S(X) is a statistic sufficient for θ, then by Theorem 2.2, there are Borel functions h and gθ such that fθ (x) = gθ (S(x))h(x). For x with h(x) > 0, x(1) = sup{θ : gθ (S(x)) > 0} and x(n) = 1 + inf{θ : gθ (S(x)) > 0}. Hence, there is a measurable function ψ such that T (x) = ψ(S(x)) when h(x) > 0. Since h > 0 a.s. P, we conclude that T is minimal sufficient.

108

2. Fundamentals of Statistics

Minimal sufficient statistics exist under weak assumptions, e.g., P contains distributions on Rk dominated by a σ-finite measure (Bahadur, 1957). The next theorem provides some useful tools for finding minimal sufficient statistics. Theorem 2.3. Let P be a family of distributions on Rk . (i) Suppose that P0 ⊂ P and a.s. P0 implies a.s. P. If T is sufficient for P ∈ P and minimal sufficient for P ∈ P0 , then T is minimal sufficient for P ∈ P. (ii) Suppose that PPcontains p.d.f.’s f0 , f1 , f2 , ..., w.r.t. a σ-finite meaP∞ ∞ sure. Let f∞ (x) = i=0 ci fi (x), where ci > 0 for all i and i=0 ci = 1, and let Ti (X) = fi (x)/f∞ (x) when f∞ (x) > 0, i = 0, 1, 2, .... Then T (X) = (T0 , T1 , T2 , ...) is minimal sufficient for P ∈ P. Furthermore, if {x : fi (x) > 0} ⊂ {x : f0 (x) > 0} for all i, then we may replace f∞ by f0 , in which case T (X) = (T1 , T2 , ...) is minimal sufficient for P ∈ P. (iii) Suppose that P contains p.d.f.’s fP w.r.t. a σ-finite measure and that there exists a sufficient statistic T (X) such that, for any possible values x and y of X, fP (x) = fP (y)φ(x, y) for all P implies T (x) = T (y), where φ is a measurable function. Then T (X) is minimal sufficient for P ∈ P. Proof. (i) If S is sufficient for P ∈ P, then it is also sufficient for P ∈ P0 and, therefore, T = ψ(S) a.s. P0 holds for a measurable function ψ. The result follows from the assumption that a.s. P0 implies a.s. P. (ii) Note that f∞ > 0 a.s. P. Let gi (T ) = Ti , i = 0, 1, 2, .... Then fi (x) = gi (T (x))f∞ (x) a.s. P. By Theorem 2.2, T is sufficient for P ∈ P. Suppose that S(X) is another sufficient statistic. By Theorem 2.2, there are Borel functions h and Pg˜∞i such that fi (x) = g˜i (S(x))h(x), i = 0, 1, 2, .... Then Ti (x) = g˜i (S(x))/ j=0 cj g˜j (S(x)) for x’s satisfying f∞ (x) > 0. By Definition 2.5, T is minimal sufficient for P ∈ P. The proof for the case where f∞ is replaced by f0 is the same. (iii) From Bahadur (1957), there exists a minimal sufficient statistic S(X). The result follows if we can show that T (X) = ψ(S(X)) a.s. P for a measurable function ψ. By Theorem 2.2, there are Borel functions gP and h such that fP (x) = gP (S(x))h(x) for all P . Let A = {x : h(x) = 0}. Then P (A) = 0 for all P . For x and y such that S(x) = S(y), x 6∈ A and y 6∈ A, fP (x) = gP (S(x))h(x) = gP (S(y))h(x)h(y)/h(y) = fP (y)h(x)/h(y) for all P . Hence T (x) = T (y). This shows that there is a function ψ such that T (x) = ψ(S(x)) except for x ∈ A. It remains to show that ψ is measurable. Since S is minimal sufficient, g(T (X)) = S(X) a.s. P for a measurable function g. Hence g is one-to-one and ψ = g −1 . The measurability of ψ follows from Theorem 3.9 in Parthasarathy (1967).

109

2.2. Statistics, Sufficiency, and Completeness

Example 2.14. Let P = {fθ : θ ∈ Θ} be an exponential family with p.d.f.’s fθ given by (2.4) and X(ω) = ω. Suppose that there exists Θ0 = {θ0 , θ1 , ..., θp } ⊂ Θ such that the vectors ηi = η(θi ) − η(θ0 ), i = 1, ..., p, are linearly independent in Rp . (This is true if the family is of full rank.) We have shown that T (X) is sufficient for θ ∈ Θ. We now show that T is in fact minimal sufficient for θ ∈ Θ. Let P0 = {fθ : θ ∈ Θ0 }. Note that the set {x : fθ (x) > 0} does not depend on θ. It follows from Theorem 2.3(ii) with f∞ = fθ0 that S(X) = exp{η1τ T (x) − ξ1 }, ..., exp{ηpτ T (x) − ξp }

is minimal sufficient for θ ∈ Θ0 , where ξi = ξ(θi ) − ξ(θ0 ). Since ηi ’s are linearly independent, there is a one-to-one measurable function ψ such that T (X) = ψ(S(X)) a.s. P0 . Hence, T is minimal sufficient for θ ∈ Θ0 . It is easy to see that a.s. P0 implies a.s. P. Thus, by Theorem 2.3(i), T is minimal sufficient for θ ∈ Θ. The results in Examples 2.13 and 2.14 can also be proved by using Theorem 2.3(iii) (Exercise 32). The sufficiency (and minimal sufficiency) depends on the postulated family P of populations (statistical models). Hence, it may not be a useful concept if the proposed statistical model is wrong or at least one has some doubts about the correctness of the proposed model. From the examples in this section and some exercises in §2.6, one can find that for a wide ¯ in (2.1), S 2 in (2.2), (X(1) , X(n) ) in variety of models, statistics such as X Example 2.11, and the order statistics in Example 2.9 are sufficient. Thus, using these statistics for data reduction and summarization does not lose any information when the true model is one of those models but we do not know exactly which model is correct.

2.2.3 Complete statistics A statistic V (X) is said to be ancillary if its distribution does not depend on the population P and first-order ancillary if E[V (X)] is independent of P . A trivial ancillary statistic is the constant statistic V (X) ≡ c ∈ R. If V (X) is a nontrivial ancillary statistic, then σ(V (X)) ⊂ σ(X) is a nontrivial σ-field that does not contain any information about P . Hence, if S(X) is a statistic and V (S(X)) is a nontrivial ancillary statistic, it indicates that σ(S(X)) contains a nontrivial σ-field that does not contain any information about P and, hence, the “data” S(X) may be further reduced. A sufficient statistic T appears to be most successful in reducing the data if no nonconstant function of T is ancillary or even first-order ancillary. This leads to the following concept of completeness.

110

2. Fundamentals of Statistics

Definition 2.6 (Completeness). A statistic T (X) is said to be complete for P ∈ P if and only if, for any Borel f , E[f (T )] = 0 for all P ∈ P implies f (T ) = 0 a.s. P. T is said to be boundedly complete if and only if the previous statement holds for any bounded Borel f . A complete statistic is boundedly complete. If T is complete (or boundedly complete) and S = ψ(T ) for a measurable ψ, then S is complete (or boundedly complete). Intuitively, a complete and sufficient statistic should be minimal sufficient, which was shown by Lehmann and Scheff´e (1950) and Bahadur (1957) (see Exercise 48). However, a minimal sufficient statistic is not necessarily complete; for example, the minimal sufficient statistic (X(1) , X(n) ) in Example 2.13 is not complete (Exercise 47). Proposition 2.1. If P is in an exponential family of full rank with p.d.f.’s given by (2.6), then T (X) is complete and sufficient for η ∈ Ξ. Proof. We have shown that T is sufficient. Suppose that there is a function f such that E[f (T )] = 0 for all η ∈ Ξ. By Theorem 2.1(i), Z f (t) exp{η τ t − ζ(η)}dλ = 0 for all η ∈ Ξ, where λ is a measure on (Rp , B p ). Let η0 be an interior point of Ξ. Then Z Z τ τ (2.16) f+ (t)eη t dλ = f− (t)eη t dλ for all η ∈ N (η0 ), where N (η0 ) = {η ∈ Rp : kη − η0 k < ǫ} for some ǫ > 0. In particular, Z Z τ τ f+ (t)eη0 t dλ = f− (t)eη0 t dλ = c. τ

τ

If c = 0, then f = 0 a.e. λ. If c > 0, then c−1 f+ (t)eη0 t and c−1 f− (t)eη0 t are p.d.f.’s w.r.t. λ and (2.16) implies that their m.g.f.’s are the same in a τ τ neighborhood of 0. By Theorem 1.6(ii), c−1 f+ (t)eη0 t = c−1 f− (t)eη0 t , i.e., f = f+ − f− = 0 a.e. λ. Hence T is complete. Proposition 2.1 is useful for finding a complete and sufficient statistic when the family of distributions is an exponential family of full rank. Example 2.15. Suppose that X1 , ..., Xn are i.i.d. random variables having the N (µ, σ 2 ) distribution, µ ∈ R, σ > 0. From Example 2.6, the joint Pn p.d.f. of X1 , ...,P Xn is (2π)−n/2 exp {η1 T1 + η2 T2 − nζ(η)}, where T = 1 i=1 Xi , n T2 = − i=1 Xi2 , and η = (η1 , η2 ) = σµ2 , 2σ1 2 . Hence, the family of distributions for X = (X1 , ..., Xn ) is a natural exponential family of full rank (Ξ = R × (0, ∞)). By Proposition 2.1, T (X) = (T1 , T2 ) is complete and sufficient for η. Since there is a one-to-one correspondence between η

2.2. Statistics, Sufficiency, and Completeness

111

and θ = (µ, σ 2 ), T is also complete and sufficient for θ. It can be shown that any one-to-one measurable function of a complete and sufficient statistic ¯ S 2 ) is complete and is also complete and sufficient (exercise). Thus, (X, 2 ¯ sufficient for θ, where X and S are the sample mean and variance given by (2.1) and (2.2), respectively. The following examples show how to find a complete statistic for a nonexponential family. Example 2.16. Let X1 , ..., Xn be i.i.d. random variables from Pθ , the uniform distribution U (0, θ), θ > 0. The largest order statistic, X(n) , is complete and sufficient for θ ∈ (0, ∞). The sufficiency of X(n) follows from the fact that the joint Lebesgue p.d.f. of X1 , ..., Xn is θ−n I(0,θ) (x(n) ). From Example 2.9, X(n) has the Lebesgue p.d.f. (nxn−1 /θn )I(0,θ) (x) on R. Let f be a Borel function on [0, ∞) such that E[f (X(n) )] = 0 for all θ > 0. Then Z

θ

f (x)xn−1 dx = 0 for all θ > 0.

0

Let G(θ) be the left-hand side of the previous equation. Applying the result of differentiation of an integral (see, e.g., Royden (1968, §5.3)), we obtain that G′ (θ) = f (θ)θn−1 a.e. m+ , where m+ is the Lebesgue measure on ([0, ∞), B[0,∞) ). Since G(θ) = 0 for all θ > 0, f (θ)θn−1 = 0 a.e. m+ and, hence, f (x) = 0 a.e. m+ . Therefore, X(n) is complete and sufficient for θ ∈ (0, ∞). Example 2.17. In Example 2.12, we showed that the order statistics T (X) = (X(1) , ..., X(n) ) of i.i.d. random variables X1 , ..., Xn is sufficient for P ∈ P, where P is the family of distributions on R having Lebesgue p.d.f.’s. We now show that T (X) is also complete for P ∈ P. Let P0 be the family of Lebesgue p.d.f.’s of the form f (x) = C(θ1 , ..., θn ) exp{−x2n + θ1 x + θ2 x2 + · · · + θn xn }, R where θj ∈ R and C(θ1 , ..., θn ) is a normalizing constant such that f (x)dx = 1. Then P0 ⊂ P and P0 is an exponential family of full rank. Note that the joint distribution of X = (X1 , ..., Xn ) is also in an exponential family of full rank. Thus, by Proposition Pn 2.1, U = (U1 , ..., Un ) is a complete statistic for P ∈ P0 , where Uj = i=1 Xij . Since a.s. P0 implies a.s. P, U (X) is also complete for P ∈ P. The result follows if we can show that there Pnis a one-to-onePcorrespondenceP between T (X) and U (X). Let V1 = i<j Xi Xj , i=1 Xi , V2 = V3 = i<j 0. It can be easily shown that the family {N (µ, σ 2 ) : µ ∈ R} is an exponential family of full rank with natural parameter η = µ/σ 2 . By Proposition 2.1, the sample ¯ in (2.1) is complete and sufficient for η (and µ). Let S 2 be the mean X Pn ¯ 2 , where sample variance given by (2.2). SincePS 2 = (n − 1)−1 i=1 (Zi − Z) n 2 −1 2 Zi = Xi −µ is N (0, σ ) and Z¯ = n Z , S is an ancillary statistic (σ 2 i=1 i 2 ¯ and S are independent w.r.t. N (µ, σ 2 ) is known). By Basu’s theorem, X 2 ¯ and S 2 are independent w.r.t. N (µ, σ 2 ) with µ ∈ R. Since σ is arbitrary, X 2 for any µ ∈ R and σ > 0. ¯ and S 2 , we now show that (n − 1)S 2 /σ 2 Using the independence of X has the chi-square distribution χ2n−1 . Note that 2 2 ¯ n X −µ (n − 1)S 2 X Xi − µ n + = . σ σ2 σ i=1 ¯ − µ)2 /σ 2 has the chiFrom the properties of the normal distributions, n(X P 2 −1/2 square distribution χ1 with the m.g.f. (1 − 2t) and ni=1 (Xi − µ)2 /σ 2

113

2.3. Statistical Decision Theory

has the chi-square distribution χ2n with the m.g.f. (1 − 2t)−n/2 , t < 1/2. By ¯ and S 2 , the m.g.f. of (n − 1)S 2 /σ 2 is the independence of X (1 − 2t)−n/2 /(1 − 2t)−1/2 = (1 − 2t)−(n−1)/2

for t < 1/2. This is the m.g.f. of the chi-square distribution χ2n−1 and, therefore, the result follows.

2.3 Statistical Decision Theory In this section, we describe some basic elements in statistical decision theory. More developments are given in later chapters.

2.3.1 Decision rules, loss functions, and risks Let X be a sample from a population P ∈ P. A statistical decision is an action that we take after we observe X, for example, a conclusion about P or a characteristic of P . Throughout this section, we use A to denote the set of allowable actions. Let FA be a σ-field on A. Then the measurable space (A, FA ) is called the action space. Let X be the range of X and FX be a σ-field on X. A decision rule is a measurable function (a statistic) T from (X, FX ) to (A, FA ). If a decision rule T is chosen, then we take the action T (X) ∈ A whence X is observed. The construction or selection of decision rules cannot be done without any criterion about the performance of decision rules. In statistical decision theory, we set a criterion using a loss function L, which is a function from P × A to [0, ∞) and is Borel on (A, FA ) for each fixed P ∈ P. If X = x is observed and our decision rule is T , then our “loss” (in making a decision) is L(P, T (x)). The average loss for the decision rule T , which is called the risk of T , is defined to be Z RT (P ) = E[L(P, T (X))] = L(P, T (x))dPX (x). (2.17) X

The loss and risk functions are denoted by L(θ, a) and RT (θ) if P is a parametric family indexed by θ. A decision rule with small loss is preferred. But it is difficult to compare L(P, T1 (X)) and L(P, T2 (X)) for two decision rules, T1 and T2 , since both of them are random. For this reason, the risk function (2.17) is introduced and we compare two decision rules by comparing their risks. A rule T1 is as good as another rule T2 if and only if RT1 (P ) ≤ RT2 (P ) for any P ∈ P,

(2.18)

and is better than T2 if and only if (2.18) holds and RT1 (P ) < RT2 (P ) for at least one P ∈ P. Two decision rules T1 and T2 are equivalent if and only

114

2. Fundamentals of Statistics

if RT1 (P ) = RT2 (P ) for all P ∈ P. If there is a decision rule T∗ that is as good as any other rule in ℑ, a class of allowable decision rules, then T∗ is said to be ℑ-optimal (or optimal if ℑ contains all possible rules). Example 2.19. Consider the measurement problem in Example 2.1. Suppose that we need a decision on the value of θ ∈ R, based on the sample X = (X1 , ..., Xn ). If Θ is all possible values of θ, then it is reasonable to consider the action space (A, FA ) = (Θ, BΘ ). An example of a decision rule ¯ the sample mean defined by (2.1). A common loss function is T (X) = X, in this problem is the squared error loss L(P, a) = (θ − a)2 , a ∈ A. Then ¯ is the squared deviation between X ¯ and θ. the loss for the decision rule X Assuming that the population has mean µ and variance σ 2 < ∞, we obtain ¯ the following risk function for X: ¯ 2 RX¯ (P ) = E(θ − X) ¯ 2 + E(E X ¯ − X) ¯ 2 = (θ − E X) 2 ¯ + Var(X) ¯ = (θ − E X) 2

= (µ − θ) +

2

σ n

,

(2.19) (2.20)

¯ in Examwhere result (2.20) follows from the results for the moments of X ple 2.8. If θ is in fact the mean of the population, then the first term on the right-hand side of (2.20) is 0 and the risk is an increasing function of the population variance σ 2 and a decreasing function of the sample size n. Consider another decision rule T1 (X) = (X(1) + X(n) )/2. However, RT1 (P ) does not have an explicit form if there is no further assumption on ¯ (or T1 ) is the population P . Suppose that P ∈ P. Then, for some P, X ¯ ¯ nor T1 is better than T1 (or X) (exercise), whereas for some P, neither X better than the other. A different loss function may also be considered. For example, L(P, a) = |θ − a|, which is called the absolute error loss. However, RX¯ (P ) and RT1 (P ) do not have explicit forms unless P is of some specific form. The problem in Example 2.19 is a special case of a general problem called estimation, in which the action space is the set of all possible values of a population characteristic ϑ to be estimated. In an estimation problem, a decision rule T is called an estimator and result (2.19) holds with θ = ϑ and ¯ replaced by any estimator with a finite variance. The following example X describes another type of important problem called hypothesis testing. Example 2.20. Let P be a family of distributions, P0 ⊂ P, and P1 = {P ∈ P : P 6∈ P0 }. A hypothesis testing problem can be formulated as that of deciding which of the following two statements is true: H0 : P ∈ P0

versus

H1 : P ∈ P 1 .

(2.21)

115

2.3. Statistical Decision Theory

Here, H0 is called the null hypothesis and H1 is called the alternative hypothesis. The action space for this problem contains only two elements, i.e., A = {0, 1}, where 0 is the action of accepting H0 and 1 is the action of rejecting H0 . A decision rule is called a test. Since a test T (X) is a function from X to {0, 1}, T (X) must have the form IC (X), where C ∈ FX is called the rejection region or critical region for testing H0 versus H1 . A simple loss function for this problem is the 0-1 loss: L(P, a) = 0 if a correct decision is made and 1 if an incorrect decision is made, i.e., L(P, j) = 0 for P ∈ Pj and L(P, j) = 1 otherwise, j = 0, 1. Under this loss, the risk is P (T (X) = 1) = P (X ∈ C) P ∈ P0 RT (P ) = P (T (X) = 0) = P (X 6∈ C) P ∈ P1 . See Figure 2.2 on page 127 for an example of a graph of RT (θ) for some T and P in a parametric family. The 0-1 loss implies that the loss for two types of incorrect decisions (accepting H0 when P ∈ P1 and rejecting H0 when P ∈ P0 ) are the same. In some cases, one might assume unequal losses: L(P, j) = 0 for P ∈ Pj , L(P, 0) = c0 when P ∈ P1 , and L(P, 1) = c1 when P ∈ P0 . In the following example the decision problem is neither an estimation nor a testing problem. Another example is given in Exercise 93 in §2.6. Example 2.21. A hazardous toxic waste site requires clean-up when the true chemical concentration θ in the contaminated soil is higher than a given level θ0 ≥ 0. Because of the limitation in resources, we would like to spend our money and efforts more in those areas that pose high risk to public health. In a particular area where soil samples are obtained, we would like to take one of these three actions: a complete clean-up (a1 ), a partial clean-up (a2 ), and no clean-up (a3 ). Then A = {a1 , a2 , a3 }. Suppose that the cost for a complete clean-up is c1 and for a partial clean-up is c2 < c1 ; the risk to public health is c3 (θ − θ0 ) if θ > θ0 and 0 if θ ≤ θ0 ; a complete clean-up can reduce the toxic concentration to an amount ≤ θ0 , whereas a partial clean-up can only reduce a fixed amount of the toxic concentration, i.e., the chemical concentration becomes θ−t after a partial clean-up, where t is a known constant. Then the loss function is given by L(θ, a) θ ≤ θ0 θ 0 < θ ≤ θ0 + t θ > θ0 + t

a1 c1 c1 c1

a2 c2 c2 c2 + c3 (θ − θ0 − t)

a3 0 c3 (θ − θ0 ) c3 (θ − θ0 )

The risk function can be calculated once the decision rule is specified. We discuss this example again in Chapter 4.

116

2. Fundamentals of Statistics

Sometimes it is useful to consider randomized decision rules. Examples are given in §2.3.2, Chapters 4 and 6. A randomized decision rule is a function δ on X × FA such that, for every A ∈ FA , δ(·, A) is a Borel function and, for every x ∈ X, δ(x, ·) is a probability measure on (A, FA ). To choose an action in A when a randomized rule δ is used, we need to simulate a pseudorandom element of A according to δ(x, ·). Thus, an alternative way to describe a randomized rule is to specify the method of simulating the action from A for each x ∈ X. If A is a subset of a Euclidean space, for example, then the result in Theorem 1.7(ii) can be applied. Also, see §7.2.3. A nonrandomized decision rule T previously discussed can be viewed as a special randomized decision rule with δ(x, {a}) = I{a} (T (x)), a ∈ A, x ∈ X. Another example of a randomized rule is a discrete distribution δ(x, ·) assigning probability pj (x) to a nonrandomized decision rule Tj (x), j = 1, 2, ..., in which case the rule δ can be equivalently defined as a rule taking value Tj (x) with probability pj (x). See Exercise 64 for an example. The loss function for a randomized rule δ is defined as Z L(P, δ, x) = L(P, a)dδ(x, a), A

which reduces to the same loss function we discussed when δ is a nonrandomized rule. The risk of a randomized rule δ is then Z Z Rδ (P ) = E[L(P, δ, X)] = L(P, a)dδ(x, a)dPX (x). (2.22) X

A

2.3.2 Admissibility and optimality Consider a given decision problem with a given loss L(P, a). Definition 2.7 (Admissibility). Let ℑ be a class of decision rules (randomized or nonrandomized). A decision rule T ∈ ℑ is called ℑ-admissible (or admissible when ℑ contains all possible rules) if and only if there does not exist any S ∈ ℑ that is better than T (in terms of the risk). If a decision rule T is inadmissible, then there exists a rule better than T . Thus, T should not be used in principle. However, an admissible decision rule is not necessarily good. For example, in an estimation problem a silly estimator T (X) ≡ a constant may be admissible (Exercise 71). The relationship between the admissibility and the optimality defined in §2.3.1 can be described as follows. If T∗ is ℑ-optimal, then it is ℑ-admissible; if T∗ is ℑ-optimal and T0 is ℑ-admissible, then T0 is also ℑ-optimal and is equivalent to T∗ ; if there are two ℑ-admissible rules that are not equivalent, then there does not exist any ℑ-optimal rule.

2.3. Statistical Decision Theory

117

Suppose that we have a sufficient statistic T (X) for P ∈ P. Intuitively, our decision rule should be a function of T , based on the discussion in §2.2.2. This is not true in general, but the following result indicates that this is true if randomized decision rules are allowed. Proposition 2.2. Suppose that A is a subset of Rk . Let T (X) be a sufficient statistic for P ∈ P and let δ0 be a decision rule. Then δ1 (t, A) = E[δ0 (X, A)|T = t],

(2.23)

which is a randomized decision rule depending only on T , is equivalent to δ0 if Rδ0 (P ) < ∞ for any P ∈ P. Proof. Note that δ1 defined by (2.23) is a decision rule since δ1 does not depend on the unknown P by the sufficiency of T . From (2.22), Z Rδ1 (P ) = E L(P, a)dδ1 (X, a) A Z L(P, a)dδ0 (X, a) T =E E Z A L(P, a)dδ0 (X, a) =E A

= Rδ0 (P ),

where the proof of the second equality is left to the reader. Note that Proposition 2.2 does not imply that δ0 is inadmissible. Also, if δ0 is a nonrandomized rule, δ1 (t, A) = E[IA (δ0 (X))|T = t] = P (δ0 (X) ∈ A|T = t) is still a randomized rule, unless δ0 (X) = h(T (X)) a.s. P for some Borel function h (Exercise 75). Hence, Proposition 2.2 does not apply to situations where randomized rules are not allowed. The following result tells us when nonrandomized rules are all we need and when decision rules that are not functions of sufficient statistics are inadmissible. Theorem 2.5. Suppose that A is a convex subset of Rk and that for any P ∈ P, L(P, a) is a convex function of a. R (i) Let δ be a randomized R rule satisfying A kakdδ(x, a) < ∞ for any x ∈ X and let T1 (x) = A adδ(x, a). Then L(P, T1 (x)) ≤ L(P, δ, x) (or L(P, T1 (x)) < L(P, δ, x) if L is strictly convex in a) for any x ∈ X and P ∈ P. (ii) (Rao-Blackwell theorem). Let T be a sufficient statistic for P ∈ P, T0 ∈ Rk be a nonrandomized rule satisfying EkT0 k < ∞, and T1 = E[T0 (X)|T ]. Then RT1 (P ) ≤ RT0 (P ) for any P ∈ P. If L is strictly convex in a and T0 is not a function of T , then T0 is inadmissible.

118

2. Fundamentals of Statistics

The proof of Theorem 2.5 is an application of Jensen’s inequality (1.47) and is left to the reader. The concept of admissibility helps us to eliminate some decision rules. However, usually there are still too many rules left after the elimination of some rules according to admissibility and sufficiency. Although one is typically interested in a ℑ-optimal rule, frequently it does not exist, if ℑ is either too large or too small. The following examples are illustrations. Example 2.22. Let X1 , ..., Xn be i.i.d. random variables from a population P ∈ P that is the family of populations having finite mean µ and variance σ 2 . Consider the estimation of µ (A = R) under the squared error loss. It can be shown that if we let ℑ be the class of all possible estimators, then there is no ℑ-optimal rule (exercise). Next, P let ℑ1 be the class of all linear n functions in X = (X1 , ..., Xn ), i.e., T (X) = i=1 ci Xi with known ci ∈ R, i = 1, ..., n. It follows from (2.19) and the discussion after Example 2.19 that !2 n n X X RT (P ) = µ2 ci − 1 + σ 2 c2i . (2.24) i=1

i=1

Pn

We now show that there does not exist T∗ = i=1 c∗i Xi such that RT∗ (P ) ≤ RT (P ) for any P ∈ P and T ∈ ℑ1 . If there is such a T∗ , then (c∗1 , ..., c∗n ) is a minimum of the function of (c1 , ..., cn ) on the right-hand side of (2.24). Then c∗1 , ..., c∗n must be the same and equal to µ2 /(σ 2 +nµ2 ), which depends on P . Hence T∗ is not a statistic. This shows that there is no ℑ1 -optimal rule. Pn Consider now a subclass ℑ2 ⊂ ℑ1 with ci ’s satisfying P i=1 ci = 1. From P n (2.24), RT (P ) = σ 2 i=1 c2i if T ∈ ℑ2 . Minimizing σ 2 ni=1 c2i subject to Pn −1 for all i. Thus, the i=1 ci = 1 leads to an optimal solution of ci = n ¯ sample mean X is ℑ2 -optimal. There may not be any optimal rule if we consider a small class of decision ¯ then one rules. For example, if ℑ3 contains all the rules in ℑ2 except X, can show that there is no ℑ3 -optimal rule. Example 2.23. Assume that the sample X has the binomial distribution Bi(θ, n) with an unknown θ ∈ (0, 1) and a fixed integer n > 1. Consider the hypothesis testing problem described in Example 2.20 with H0 : θ ∈ (0, θ0 ] versus H1 : θ ∈ (θ0 , 1), where θ0 ∈ (0, 1) is a fixed value. Suppose that we are only interested in the following class of nonrandomized decision rules: ℑ = {Tj : j = 0, 1, ..., n − 1}, where Tj (X) = I{j+1,...,n} (X). From Example 2.20, the risk function for Tj under the 0-1 loss is RTj (θ) = P (X > j)I(0,θ0 ] (θ) + P (X ≤ j)I(θ0 ,1) (θ).

119

2.3. Statistical Decision Theory

For any integers k and j, 0 ≤ k < j ≤ n − 1, −P (k < X ≤ j) < 0 RTj (θ) − RTk (θ) = P (k < X ≤ j) > 0

0 < θ ≤ θ0 θ0 < θ < 1.

Hence, neither Tj nor Tk is better than the other. This shows that every Tj is ℑ-admissible and, thus, there is no ℑ-optimal rule. In view of the fact that an optimal rule often does not exist, statisticians adopt the following two approaches to choose a decision rule. The first approach is to define a class ℑ of decision rules that have some desirable properties (statistical and/or nonstatistical) and then try to find the best rule in ℑ. In Example 2.22, for instance, any estimator T in ℑ2 has the property that T is linear in X and E[T (X)] = µ. In a general estimation problem, we can use the following concept. Definition 2.8 (Unbiasedness). In an estimation problem, the bias of an estimator T (X) of a real-valued parameter ϑ of the unknown population is defined to be bT (P ) = E[T (X)] − ϑ (which is denoted by bT (θ) when P is in a parametric family indexed by θ). An estimator T (X) is said to be unbiased for ϑ if and only if bT (P ) = 0 for any P ∈ P. Thus, ℑ2 in Example 2.22 is the class of unbiased estimators linear in X. In Chapter 3, we discuss how to find a ℑ-optimal estimator when ℑ is the class of unbiased estimators or unbiased estimators linear in X. Another class of decision rules can be defined after we introduce the concept of invariance. Definition 2.9 Let X be a sample from P ∈ P. (i) A class G of one-to-one transformations of X is called a group if and only if gi ∈ G implies g1◦g2 ∈ G and gi−1 ∈ G. (ii) We say that P is invariant under G if and only if g¯(PX ) = Pg(X) is a one-to-one transformation from P onto P for each g ∈ G. (iii) A decision problem is said to be invariant if and only if P is invariant under G and the loss L(P, a) is invariant in the sense that, for every g ∈ G and every a ∈ A, there exists a unique g(a) ∈ A such that L(PX , a) = L Pg(X) , g(a) . (Note that g(X) and g(a) are different functions in general.) (iv) A decision rule T (x) is said to be invariant if and only if, for every g ∈ G and every x ∈ X, T (g(x)) = g(T (x)). Invariance means that our decision is not affected by one-to-one transformations of data. In a problem where the distribution of X is in a location-scale family

120

2. Fundamentals of Statistics

P on Rk , we often consider location-scale transformations of data X of the form g(X) = AX + c, where c ∈ C ⊂ Rk and A ∈ T , a class of invertible k × k matrices. Assume that if Ai ∈ T , i = 1, 2, then A−1 ∈ T and i A1 A2 ∈ T , and that if ci ∈ C, i = 1, 2, then −ci ∈ C and Ac1 + c2 ∈ C for any A ∈ T . Then the collection of all transformations is a group. A special case is given in the following example. Example 2.24. Let X have i.i.d. components from a population in a location family P = {Pµ : µ ∈ R}. Consider the location transformation gc (X) = X +cJk , where c ∈ R and Jk is the k-vector whose components are all equal to 1. The group of transformation is G = {gc : c ∈ R}, which is a location-scale transformation group with T = {Ik } and C = {cJk : c ∈ R}. P is invariant under G with g¯c (Pµ ) = Pµ+c . For estimating µ under the loss L(µ, a) = L(µ − a), where L(·) is a nonnegative Borel function, the decision problem is invariant with gc (a) = a + c. A decision rule T is invariant if and only if T (x + cJk ) = T (x) + c for every x ∈ Rk and c ∈ R. An example of an invariant decision rule is T (x) = lτ x for some l ∈ Rk with lτ Jk = 1. Note that T (x) = lτ x with lτ Jk = 1 is in the class ℑ2 in Example 2.22. In §4.2 and §6.3, we discuss the problem of finding a ℑ-optimal rule when ℑ is a class of invariant decision rules. The second approach to finding a good decision rule is to consider some characteristic RT of RT (P ), for a given decision rule T , and then minimize RT over T ∈ ℑ. The following are two popular ways to carry out this idea. The first one is to consider an average of RT (P ) over P ∈ P: rT (Π) =

Z

RT (P )dΠ(P ),

P

where Π is a known probability measure on (P, FP ) with an appropriate σ-field FP . rT (Π) is called the Bayes risk of T w.r.t. Π. If T∗ ∈ ℑ and rT∗ (Π) ≤ rT (Π) for any T ∈ ℑ, then T∗ is called a ℑ-Bayes rule (or Bayes rule when ℑ contains all possible rules) w.r.t. Π. The second method is to consider the worst situation, i.e., supP ∈P RT (P ). If T∗ ∈ ℑ and supP ∈P RT∗ (P ) ≤ supP ∈P RT (P ) for any T ∈ ℑ, then T∗ is called a ℑ-minimax rule (or minimax rule when ℑ contains all possible rules). Bayes and minimax rules are discussed in Chapter 4. Example 2.25. We usually try to find a Bayes rule or a minimax rule in a parametric problem where P = Pθ for a θ ∈ Rk . Consider the special case of k = 1 and L(θ, a) = (θ − a)2 , the squared error loss. Note that rT (Π) =

Z

R

E[θ − T (X)]2 dΠ(θ),

121

2.3. Statistical Decision Theory

which is equivalent to E[θ − T (X)]2 , where θ is a random variable having the distribution Π and, given θ = θ, the conditional distribution of X is Pθ . Then, the problem can be viewed as a prediction problem for θ using functions of X. Using the result in Example 1.22, the best predictor is E(θ|X), which is the ℑ-Bayes rule w.r.t. Π with ℑ being the class of rules T (X) satisfying E[T (X)]2 < ∞ for any θ. As a more specific example, let X = (X1 , ..., Xn ) with i.i.d. components having the N (µ, σ 2 ) distribution with an unknown µ = θ ∈ R and a known σ 2 , and let Π be the N (µ0 , σ02 ) distribution with known µ0 and σ02 . Then the conditional distribution of θ given X = x is N (µ∗ (x), c2 ) with µ∗ (x) =

σ2 nσ02 µ + x¯ 0 nσ02 + σ 2 nσ02 + σ 2

and

c2 =

σ02 σ 2 nσ02 + σ 2

(2.25)

(exercise). The Bayes rule w.r.t. Π is E(θ|X) = µ∗ (X). ¯ is ℑ-minimax In this special case we can show that the sample mean X with ℑ being the collection of all decision rules. For any decision rule T , Z sup RT (θ) ≥ RT (θ)dΠ(θ) θ∈R ZR ≥ Rµ∗ (θ)dΠ(θ) R = E [θ − µ∗ (X)]2 = E E{[θ − µ∗ (X)]2 |X} = E(c2 )

= c2 , where µ∗ (X) is the Bayes rule given in (2.25) and c2 is also given in (2.25). Since this result is true for any σ02 > 0 and c2 → σ 2 /n as σ02 → ∞, sup RT (θ) ≥

θ∈R

σ2 = sup RX¯ (θ), n θ∈R

¯ under the squared error loss where the equality holds because the risk of X ¯ is minimax. is, by (2.20), σ 2 /n and independent of θ = µ. Thus, X A minimax rule in a general case may be difficult to obtain. It can be seen that if both µ and σ 2 are unknown in the previous discussion, then sup θ∈R×(0,∞)

RX¯ (θ) = ∞,

(2.26)

¯ cannot be minimax unless (2.26) holds with where θ = (µ, σ 2 ). Hence X ¯ X replaced by any decision rule T , in which case minimaxity becomes meaningless.

122

2. Fundamentals of Statistics

2.4 Statistical Inference The loss function plays a crucial role in statistical decision theory. Loss functions can be obtained from a utility analysis (Berger, 1985), but in many problems they have to be determined subjectively. In statistical inference, we make an inference about the unknown population based on the sample X and inference procedures without using any loss function, although any inference procedure can be cast in decision-theoretic terms as a decision rule. There are three main types of inference procedures: point estimators, hypothesis tests, and confidence sets.

2.4.1 Point estimators The problem of estimating an unknown parameter related to the unknown population is introduced in Example 2.19 and the discussion after Example 2.19 as a special statistical decision problem. In statistical inference, however, estimators of parameters are derived based on some principle (such as the unbiasedness, invariance, sufficiency, substitution principle, likelihood principle, Bayesian principle, etc.), not based on a loss or risk function. Since confidence sets are sometimes also called interval estimators or set estimators, estimators of parameters are called point estimators. In Chapters 3 through 5, we consider how to derive a “good” point estimator based on some principle. Here we focus on how to assess performance of point estimators. ˜ ⊂ R be a parameter to be estimated, which is a function of Let ϑ ∈ Θ the unknown population P or θ if P is in a parametric family. An estimator ˜ First, one has to realize that any estimator T (X) is a statistic with range Θ. of ϑ is subject to an estimation error T (x) − ϑ when we observe X = x. This is not just because T (X) is random. In some problems T (x) never equals ϑ. A trivial example is when T (X) has a continuous c.d.f. so that P (T (X) = ϑ) = 0. As a nontrivial example, let X1 , ..., Xn be i.i.d. binary random variables (also called Bernoulli variables) with P (Xi = 1) = p and ¯ is shown to be a good estimator P (Xi = 0) = 1 − p. The sample mean X of ϑ = p in later chapters, but x ¯ never equals ϑ if ϑ is not one of j/n, j = 0, 1, ..., n. Thus, we cannot assess the performance of T (X) by the values of T (x) with particular x’s and it is also not worthwhile to do so. The bias bT (P ) and unbiasedness of a point estimator T (X) is defined in Definition 2.8. Unbiasedness of T (X) means that the mean of T (X) is equal to ϑ. An unbiased estimator T (X) can be viewed as an estimator without “systematic” error, since, on the average, it does not overestimate (i.e., bT (P ) > 0) or underestimate (i.e., bT (P ) < 0). However, an unbiased

123

2.4. Statistical Inference

estimator T (X) may have large positive and negative errors T (x)−ϑ, x ∈ X, although theseRerrors cancel each other in the calculation of the bias, which is the average [T (x) − ϑ]dPX (x). Hence, for an unbiased estimator T (X), it is desired that the values of T (x) be highly concentrated around ϑ. The variance of T (X) is commonly used as a measure of the dispersion of T (X). The mean squared error (mse) of T (X) as an estimator of ϑ is defined to be mseT (P ) = E[T (X) − ϑ]2 = [bT (P )]2 + Var(T (X)),

(2.27)

which is denoted by mseT (θ) if P is in a parametric family. mseT (P ) is equal to the variance Var(T (X)) if and only if T (X) is unbiased. Note that the mse is simply the risk of T in statistical decision theory under the squared error loss. In addition to the variance and the mse, the following are other measures of dispersion that are often used in point estimation problems. The first one is the mean absolute error of an estimator T (X) defined to be E|T (X) − ϑ|. The second one is the probability of falling outside a stated distance of ϑ, i.e., P (|T (X) − ϑ| ≥ ǫ) with a fixed ǫ > 0. Again, these two measures of dispersion are risk functions in statistical decision theory with loss functions |ϑ − a| and I(ǫ,∞) (|ϑ − a|), respectively. For the bias, variance, mse, and mean absolute error, we have implicitly assumed that certain moments of T (X) exist. On the other hand, the dispersion measure P (|T (X)−ϑ| ≥ ǫ) depends on the choice of ǫ. It is possible that some estimators are good in terms of one measure of dispersion, but not in terms of other measures of dispersion. The mse, which is a function of bias and variance according to (2.27), is mathematically easy to handle and, hence, is used the most often in the literature. In this book, we use the mse to assess and compare point estimators unless otherwise stated. Examples 2.19 and 2.22 provide some examples of estimators and their biases, variances, and mse’s. The following are two more examples. Example 2.26. Consider the life-time testing problem in Example 2.2. Let X1 , ..., Xn be i.i.d. from an unknown c.d.f. F . Suppose that the parameter of interest is ϑ = 1 − F (t) for a fixed t > 0. If F is not in a parametric family, then a nonparametric estimator of F (t) is the empirical c.d.f. n

Fn (t) =

1X I(−∞,t] (Xi ), n i=1

t ∈ R.

(2.28)

Since I(−∞,t] (X1 ), ..., I(−∞,t] (Xn ) are i.i.d. binary random variables with P (I(−∞,t] (Xi ) = 1) = F (t), the random variable nFn (t) has the binomial distribution Bi(F (t), n). Consequently, Fn (t) is an unbiased estimator of

124

2. Fundamentals of Statistics

F (t) and Var(Fn (t)) = mseFn (t) (P ) = F (t)[1 − F (t)]/n. Since any linear combination of unbiased estimators is unbiased for the same linear combination of the parameters (by the linearity of expectations), an unbiased estimator of ϑ is U (X) = 1 − Fn (t), which has the same variance and mse as Fn (t). The estimator U (X) = 1 − Fn (t) can be improved in terms of the mse if there is further information about F . Suppose that F is the c.d.f. of the exponential distribution E(0, θ) with an unknown θ > 0. Then ¯ is sufficient for θ > 0. Since ϑ = e−t/θ . From §2.2.2, the sample mean X the squared error loss is strictly convex, an application of Theorem 2.5(ii) ¯ (Rao-Blackwell theorem) shows that the estimator T (X) = E[1 − Fn (t)|X], which is also unbiased, is better than U (X) in terms of the mse. Figure 2.1 shows graphs of the mse’s of U (X) and T (X), as functions of θ, in the special case of n = 10, t = 2, and F (x) = (1 − e−x/θ )I(0,∞) (x). Example 2.27. Consider the sample survey problem in Example 2.3 a Pwith N constant selection probability p(s) and univariate yi . Let ϑ = Y = i=1 yi , P the population total. We now show that the estimator Yˆ = N i∈s yi is n an unbiased estimator of Y . Let a = 1 if i ∈ s and a = 0 otherwise. Thus, i i PN a y . Since p(s) is constant, E(a ) = P (a = 1) = n/N and Yˆ = N i i i=1 i i n ! N N N X N X N X E(Yˆ ) = E ai y i = yi E(ai ) = yi = Y. n i=1 n i=1 i=1 Note that

Var(ai ) = E(ai ) − [E(ai )]2 = and for i 6= j,

n n 1− N N

Cov(ai , aj ) = P (ai = 1, aj = 1) − E(ai )E(aj ) =

n2 n(n − 1) − 2. N (N − 1) N

Hence, the variance or the mse of Yˆ is ! N X N2 ˆ Var(Y ) = 2 Var ai y i n i=1 N 2 X X N = 2 yi2 Var(ai ) + 2 yi yj Cov(ai , aj ) n i=1 1≤i<j≤N N X X n 2 N 1− yi2 − yi yj = n N N −1 i=1 1≤i<j≤N

2 N n X Y N2 1− yi − . = n(N − 1) N i=1 N

125

0.02

mse of T mse of U

0.0

0.01

mse

0.03

2.4. Statistical Inference

0

2

4

6

8

10

q

Figure 2.1: mse’s of U (X) and T (X) in Example 2.26

2.4.2 Hypothesis tests The basic elements of a hypothesis testing problem are described in Example 2.20. In statistical inference, tests for a hypothesis are derived based on some principles similar to those given in an estimation problem. Chapter 6 is devoted to deriving tests for various types of hypotheses. Several key ideas are discussed here. To test the hypotheses H0 versus H1 given in (2.21), there are only two types of statistical errors we may commit: rejecting H0 when H0 is true (called the type I error) and accepting H0 when H0 is wrong (called the type II error). In statistical inference, a test T , which is a statistic from X to {0, 1}, is assessed by the probabilities of making two types of errors: αT (P ) = P (T (X) = 1)

P ∈ P0

(2.29)

and 1 − αT (P ) = P (T (X) = 0)

P ∈ P1 ,

(2.30)

which are denoted by αT (θ) and 1 − αT (θ) if P is in a parametric family indexed by θ. Note that these are risks of T under the 0-1 loss in statistical decision theory. However, an optimal decision rule (test) does not exist even for a very simple problem with a very simple class of tests (Example 2.23).

126

2. Fundamentals of Statistics

That is, error probabilities in (2.29) and (2.30) cannot be minimized simultaneously. Furthermore, these two error probabilities cannot be bounded simultaneously by a fixed α ∈ (0, 1) when we have a sample of a fixed size. Therefore, a common approach to finding an “optimal” test is to assign a small bound α to one of the error probabilities, say αT (P ), P ∈ P0 , and then to attempt to minimize the other error probability 1 − αT (P ), P ∈ P1 , subject to sup αT (P ) ≤ α. (2.31) P ∈P0

The bound α is called the level of significance. The left-hand side of (2.31) is called the size of the test T . Note that the level of significance should be positive, otherwise no test satisfies (2.31) except the silly test T (X) ≡ 0 a.s. P. Example 2.28. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 . Consider the hypotheses H0 : µ ≤ µ0

versus

H1 : µ > µ0 ,

¯ is sufficient for where µ0 is a fixed constant. Since the sample mean X µ ∈ R, it is reasonable to consider the following class of tests: Tc (X) = ¯ i.e., H0 is rejected (accepted) if X ¯ > c (X ¯ ≤ c), where c ∈ R is I(c,∞) (X), a fixed constant. Let Φ be the c.d.f. of N (0, 1). Then, by the property of the normal distributions, √ n(c − µ) αTc (µ) = P (Tc (X) = 1) = 1 − Φ . (2.32) σ Figure 2.2 provides an example of a graph of two types of error probabilities, with µ0 = 0. Since Φ(t) is an increasing function of t, √ n(c − µ0 ) . sup αTc (µ) = 1 − Φ σ P ∈P0 In fact, it is also true that sup [1 − αTc (µ)] = Φ

P ∈P1

√ n(c − µ0 ) . σ

If we would like to use an α as the level of significance, then the most effective way is to choose a cα (a test Tcα (X)) such that α = sup αTcα (µ), P ∈P0

in which case cα must satisfy 1−Φ

√ n(cα − µ0 ) = α, σ

127

0.6 0.4 0.0

0.2

error probability

0.8

1.0

2.4. Statistical Inference

-2

-1

0

1

2

m

Figure 2.2: Error probabilities in Example 2.28 √ i.e., cα = σz1−α / n + µ0 , where za = Φ−1 (a). In Chapter 6, it is shown that for any test T (X) satisfying (2.31), 1 − αT (µ) ≥ 1 − αTcα (µ),

µ > µ0 .

The choice of a level of significance α is usually somewhat subjective. In most applications there is no precise limit to the size of T that can be tolerated. Standard values, such as 0.10, 0.05, or 0.01, are often used for convenience. For most tests satisfying (2.31), a small α leads to a “small” rejection region. It is good practice to determine not only whether H0 is rejected or accepted for a given α and a chosen test Tα , but also the smallest possible level of significance at which H0 would be rejected for the computed Tα (x), i.e., α ˆ = inf{α ∈ (0, 1) : Tα (x) = 1}. Such an α ˆ , which depends on x and the chosen test and is a statistic, is called the p-value for the test Tα . Example 2.29. Consider the problem in Example 2.28. Let us calculate the p-value for Tcα . Note that α=1−Φ

√ √ n(cα − µ0 ) n(¯ x − µ0 ) >1−Φ σ σ

128

2. Fundamentals of Statistics

if and only if x ¯ > cα (or Tcα (x) = 1). Hence √ n(¯ x − µ0 ) = inf{α ∈ (0, 1) : Tcα (x) = 1} = α 1−Φ ˆ (x) σ is the p-value for Tcα . It turns out that Tcα (x) = I(0,α) (ˆ α(x)). With the additional information provided by p-values, using p-values is typically more appropriate than using fixed-level tests in a scientific problem. However, a fixed level of significance is unavoidable when acceptance or rejection of H0 implies an imminent concrete decision. For more discussions about p-values, see Lehmann (1986) and Weerahandi (1995). In Example 2.28, the equality in (2.31) can always be achieved by a suitable choice of c. This is, however, not true in general. In Example 2.23, for instance, it is possible to find an α such that sup P (Tj (X) = 1) 6= α

0j 1 Tj,q (X) = q X =j 0 X < j,

where j = 0, 1, ..., n − 1 and q ∈ [0, 1]. Then

αTj,q (θ) = P (X > j) + qP (X = j)

0 < θ ≤ θ0

and 1 − αTj,q (θ) = P (X < j) + (1 − q)P (X = j)

θ0 < θ < 1.

It can be shown that for any α ∈ (0, 1), there exist an integer j and q ∈ (0, 1) such that the size of Tj,q is α (exercise).

2.4. Statistical Inference

129

2.4.3 Confidence sets Let ϑ be a k-vector of unknown parameters related to the unknown popk ulation P ∈ P and C(X) ∈ BΘ ˜ depending only on the sample X, where k ˜ Θ ∈ B is the range of ϑ. If inf P (ϑ ∈ C(X)) ≥ 1 − α,

P ∈P

(2.33)

where α is a fixed constant in (0, 1), then C(X) is called a confidence set for ϑ with level of significance 1 − α. The left-hand side of (2.33) is called the confidence coefficient of C(X), which is the highest possible level of significance for C(X). A confidence set is a random element that covers the unknown ϑ with certain probability. If (2.33) holds, then the coverage probability of C(X) is at least 1−α, although C(x) either covers or does not cover ϑ whence we observe X = x. The concepts of level of significance and confidence coefficient are very similar to the level of significance and size in hypothesis testing. In fact, it is shown in Chapter 7 that some confidence sets are closely related to hypothesis tests. Consider a real-valued ϑ. If C(X) = [ϑ(X), ϑ(X)] for a pair of realvalued statistics ϑ and ϑ, then C(X) is called a confidence interval for ϑ. If C(X) = (−∞, ϑ(X)] (or [ϑ(X), ∞)), then ϑ (or ϑ) is called an upper (or a lower) confidence bound for ϑ. A confidence set (or interval) is also called a set (or an interval) estimator of ϑ, although it is very different from a point estimator (discussed in §2.4.1). Example 2.31. Consider Example 2.28. Suppose that a confidence inter¯ and ϑ(X), ¯ val for ϑ = µ is needed. Again, we only need to consider ϑ(X) ¯ since the sample mean X is sufficient. Consider confidence intervals of the ¯ − c, X ¯ + c], where c ∈ (0, ∞) is fixed. Note that form [X √ ¯ − c, X ¯ + c] = P |X ¯ − µ| ≤ c = 1 − 2Φ − nc/σ , P µ ∈ [X ¯ −c, X ¯ +c] which is independent of µ. Hence, the confidence coefficient of [X √ is 1 − 2Φ (− nc/σ), which is an increasing function of c and converges to 1 as c → ∞ or 0 as c → 0. Thus, confidence coefficients are positive but less ¯ X] ¯ and (−∞, ∞). We can than 1 except for silly confidence intervals [X, choose a confidence interval with an arbitrarily large confidence coefficient, but the chosen confidence interval may be so wide that it is practically useless. ¯ − c, X ¯ + c] has confidence coefficient 0 If σ2 is also unknown, then [X and, therefore, is not a good inference procedure. In such a case a different confidence interval for µ with positive confidence coefficient can be derived (Exercise 97 in §2.6).

130

2. Fundamentals of Statistics

This example tells us that a reasonable approach is to choose a level of significance 1 − α ∈ (0, 1) (just like the level of significance in hypothesis testing) and a confidence interval or set satisfying √ (2.33). In Example 2.31, when σ2 is known and c is chosen to be σz1−α/2 / n, where za = Φ−1 (a), ¯ − c, X ¯ + c] is exactly the confidence coefficient of the confidence interval [X 1 − α for any fixed α ∈ (0, 1). This is desirable since, for all confidence intervals satisfying (2.33), the one with the shortest interval length is preferred. For a general confidence interval [ϑ(X), ϑ(X)], its length is ϑ(X)−ϑ(X), which may be random. We may consider the expected (or average) length E[ϑ(X)−ϑ(X)]. The confidence coefficient and expected length are a pair of good measures of performance of confidence intervals. Like the two types of error probabilities of a test in hypothesis testing, however, we cannot maximize the confidence coefficient and minimize the length (or expected length) simultaneously. A common approach is to minimize the length (or expected length) subject to (2.33). For an unbounded confidence interval, its length is ∞. Hence we have to define some other measures of performance. For an upper (or a lower) confidence bound, we may consider the distance ϑ(X) − ϑ (or ϑ − ϑ(X)) or its expectation. To conclude this section, we discuss an example of a confidence set for a two-dimensional parameter. General discussions about how to construct and assess confidence sets are given in Chapter 7. Example 2.32. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with both µ ∈ R and σ 2 > 0 unknown. Let θ = (µ, σ2 ) and α ∈ (0, 1) be ¯ be the sample mean and S 2 be the sample variance. Since given. Let X 2 ¯ S ) is sufficient (Example 2.15), we focus on C(X) that is a function of (X, ¯ S 2 ). From Example 2.18, X ¯ and S 2 are independent and (n − 1)S 2 /σ 2 (X, √ ¯ has the chi-square distribution χ2n−1 . Since n(X − µ)/σ has the N (0, 1) distribution (Exercise 43 in §1.6), ¯ −µ √ X √ P −˜ cα ≤ ≤ c˜α = 1 − α, σ/ n √ where c˜α = Φ−1 1+ 21−α (verify). Since the chi-square distribution χ2n−1 is a known distribution, we can always find two constants c1α and c2α such that √ (n − 1)S 2 P c1α ≤ = 1 − α. ≤ c 2α 2 σ Then ¯ −µ X (n − 1)S 2 √ ≤ c˜α , c1α ≤ = 1 − α, ≤ c P −˜ cα ≤ 2α σ/ n σ2

131

4 0

2

variance

6

8

2.5. Asymptotic Criteria and Inference

-4

-2

0

2

4

mean

Figure 2.3: A confidence set for θ in Example 2.32 or P

¯ − µ)2 (n − 1)S 2 (n − 1)S 2 n(X ≤ σ2 , ≤ σ2 ≤ 2 c˜α c2α c1α

= 1 − α.

(2.34)

The left-hand side of (2.34) defines a set in the range of θ = (µ, σ2 ) bounded by two straight lines, σ 2 = (n − 1)S 2 /ciα , i = 1, 2, and a curve σ 2 = ¯ −µ)2 /˜ n(X c2α (see the shadowed part of Figure 2.3). This set is a confidence set for θ with confidence coefficient 1 − α, since (2.34) holds for any θ.

2.5 Asymptotic Criteria and Inference We have seen that in statistical decision theory and inference, a key to the success of finding a good decision rule or inference procedure is being able to find some moments and/or distributions of various statistics. Although many examples are presented (including those in the exercises in §2.6), there are more cases in which we are not able to find exactly the moments or distributions of given statistics, especially when the problem is not parametric (see, e.g., the discussions in Example 2.8). In practice, the sample size n is often large, which allows us to approximate the moments and distributions of statistics that are impossible

132

2. Fundamentals of Statistics

to derive, using the asymptotic tools discussed in §1.5. In an asymptotic analysis, we consider a sample X = (X1 , ..., Xn ) not for fixed n, but as a member of a sequence corresponding to n = n0 , n0 + 1, ..., and obtain the limit of the distribution of an appropriately normalized statistic or variable Tn (X) as n → ∞. The limiting distribution and its moments are used as approximations to the distribution and moments of Tn (X) in the situation with a large but actually finite n. This leads to some asymptotic statistical procedures and asymptotic criteria for assessing their performances, which are introduced in this section. The asymptotic approach is not only applied to the situation where no exact method is available, but also used to provide an inference procedure simpler (e.g., in terms of computation) than that produced by the exact approach (the approach considering a fixed n). Some examples are given in later chapters. In addition to providing more theoretical results and/or simpler inference procedures, the asymptotic approach requires less stringent mathematical assumptions than does the exact approach. The mathematical precision of the optimality results obtained in statistical decision theory, for example, tends to obscure the fact that these results are approximations in view of the approximate nature of the assumed models and loss functions. As the sample size increases, the statistical properties become less dependent on the loss functions and models. However, a major weakness of the asymptotic approach is that typically no good estimates for the precision of the approximations are available and, therefore, we cannot determine whether a particular n in a problem is large enough to safely apply the asymptotic results. To overcome this difficulty, asymptotic results are frequently used in combination with some numerical/empirical studies for selected values of n to examine the finite sample performance of asymptotic procedures.

2.5.1 Consistency A reasonable point estimator is expected to perform better, at least on the average, if more information about the unknown population is available. With a fixed model assumption and sampling plan, more data (larger sample size n) provide more information about the unknown population. Thus, it is distasteful to use a point estimator Tn which, if sampling were to continue indefinitely, could possibly have a nonzero estimation error, although the estimation error of Tn for a fixed n may never equal 0 (see the discussion in §2.4.1). Definition 2.10 (Consistency of point estimators). Let X = (X1 , ..., Xn ) be a sample from P ∈ P and Tn (X) be a point estimator of ϑ for every n. (i) Tn (X) is called consistent for ϑ if and only if Tn (X) →p ϑ w.r.t. any

2.5. Asymptotic Criteria and Inference

133

P ∈ P. (ii) Let {an } be a sequence of positive constants diverging to ∞. Tn (X) is called an -consistent for ϑ if and only if an [Tn (X) − ϑ] = Op (1) w.r.t. any P ∈ P. (iii) Tn (X) is called strongly consistent for ϑ if and only if Tn (X) →a.s. ϑ w.r.t. any P ∈ P. (iv) Tn (X) is called Lr -consistent for ϑ if and only if Tn (X) →Lr ϑ w.r.t. any P ∈ P for some fixed r > 0. Consistency is actually a concept relating to a sequence of estimators, {Tn , n = n0 , n0 + 1, ...}, but we usually just say “consistency of Tn ” for simplicity. Each of the four types of consistency in Definition 2.10 describes the convergence of Tn (X) to ϑ in some sense, as n → ∞. In statistics, consistency according to Definition 2.10(i), which is sometimes called weak consistency since it is implied by any of the other three types of consistency, is the most useful concept of convergence of Tn to ϑ. L2 -consistency is also called consistency in mse, which is the most useful type of Lr -consistency. Example 2.33. Let X1 , ..., Xn be i.i.d. from P ∈ P. If ϑ = µ, which is the mean of P and is assumed to be finite, then by the SLLN (Theorem ¯ is strongly consistent for µ and, therefore, is 1.13), the sample mean X also consistent for µ. If we further assume that the variance of P is finite, ¯ is consistent in mse and is √n-consistent. With the finite then by (2.20), X variance assumption, the sample variance S 2 is strongly consistent for the variance of P , according to the SLLN. Pn Consider estimators of the form Tn = i=1 cni Xi , where {cni } is a double array of constants. If P has a finite variance, by (2.24), Tn Pn Pthen n is consistent in mse if and only if i=1 cni → 1 and i=1 c2ni → 0. If we only assumePthe existence of the mean of P , then Tn with cni = ci /n satn isfying n−1 i=1 ci → 1 and supi |ci | < ∞ is strongly consistent (Theorem 1.13(ii)). One or a combination of the law of large numbers, the CLT, Slutsky’s theorem (Theorem 1.11), and the continuous mapping theorem (Theorems 1.10 and 1.12) are typically applied to establish consistency of point estimators. In particular, Theorem 1.10 implies that if Tn is (strongly) consistent for ϑ and g is a continuous function of ϑ, then g(Tn ) is (strongly) consistent ¯ 2 is strongly for g(ϑ). For example, in Example 2.33 the point estimator X ¯ 2 is √n-consistent under the assumption consistent for µ2 . To show that X that P has a finite variance σ 2 , we can use the identity √ √ ¯ 2 − µ2 ) = n(X ¯ − µ)(X ¯ + µ) n(X

¯ is √n-consistent for µ and X ¯ + µ = Op (1). (Note that and the fact that X

134

2. Fundamentals of Statistics

¯ 2 may not be consistent in mse since we do not assume that P has a finite X √ ¯2 2 fourth moment.) Alternatively, we can use the fact that n( √ X − µ ) →d 2 2 N (0, 4µ σ ) (by the CLT and Theorem 1.12) to show the n-consistency ¯ 2. of X The following example shows another way to establish consistency of some point estimators. Example 2.34. Let X1 , ..., Xn be i.i.d. from an unknown P with a continuous c.d.f. F satisfying F (θ) = 1 for some θ ∈ R and F (x) < 1 for any x < θ. Consider the largest order statistic X(n) . For any ǫ > 0, F (θ −ǫ) < 1 and n P (|X(n) − θ| ≥ ǫ) = P (X(n) ≤ θ − ǫ) = [F (θ − ǫ)] , which imply (according to Theorem 1.8(v)) X(n) →a.s. θ, i.e., X(n) is strongly consistent for θ. If we assume that F (i) (θ−), the ith-order lefthand derivative of F at θ, exists and vanishes for any i ≤ m and that F (m+1) (θ−) exists and is nonzero, where m is a nonnegative integer, then

(−1)m F (m+1) (θ−) (θ − X(n) )m+1 + o |θ − X(n) |m+1 a.s. (m + 1)! This result and the fact that P n[1 − F (X(n) )] ≥ s = (1 − s/n)n imply −1 that (θ − X(n) )m+1 = Op (n−1 ), i.e., X(n) is n(m+1) -consistent. If m = 0, then X(n) is n-consistent, which is the most common situation. If m = 1, √ −1 then X(n) is n-consistent. The limiting distribution of n(m+1) (X(n) − θ) can be derived as follows. Let (m+1)−1 (−1)m (m + 1)! hn (θ) = . nF (m+1) (θ−)

1 − F (X(n) ) =

For t ≤ 0, by Slutsky’s theorem, lim P

n→∞

X(n) − θ ≤ t = lim P n→∞ hn (θ)

θ − X(n) hn (θ)

m+1

m+1

≥ (−t)

= lim P n[1 − F (X(n) )] ≥ (−t)m+1 n→∞ n = lim 1 − (−t)m+1 /n n→∞

= e−(−t)

m+1

!

.

It can be seen from the previous examples that there are many consistent estimators. Like the admissibility in statistical decision theory, consistency is a very essential requirement in the sense that any inconsistent estimators should not be used, but a consistent estimator is not necessarily good. Thus, consistency should be used together with one or a few more criteria.

2.5. Asymptotic Criteria and Inference

135

We now discuss a situation in which finding a consistent estimator is crucial. Suppose that an estimator Tn of ϑ satisfies cn [Tn (X) − ϑ] →d σY,

(2.35)

where Y is a random variable with a known distribution, σ > 0 is an unknown parameter, of constants; for example, in √ ¯and {cn } is a sequence Example 2.33, n(X − µ) →d N (0, σ 2 ); in Example 2.34, (2.35) holds −1 −1 with cn = n(m+1) and σ = [(−1)m (m + 1)!/F (m+1) (θ−)](m+1) . If a consistent estimator σ ˆn of σ can be found, then, by Slutsky’s theorem, σn →d Y cn [Tn (X) − ϑ]/ˆ σn by and, thus, we may approximate the distribution of cn [Tn (X) − ϑ]/ˆ the known distribution of Y .

2.5.2 Asymptotic bias, variance, and mse Unbiasedness as a criterion for point estimators is discussed in §2.3.2 and §2.4.1. In some cases, however, there is no unbiased estimator (Exercise 84 in §2.6). Furthermore, having a “slight” bias in some cases may not be a bad idea (see Exercise 63 in §2.6). Let Tn (X) be a point estimator of ϑ for every n. If ETn exists for every n and limn→∞ E(Tn − ϑ) = 0 for any P ∈ P, then Tn is said to be approximately unbiased. There are many reasonable point estimators whose expectations are not well defined. For example, consider i.i.d. (X1 , Y1 ), ..., (Xn , Yn ) from a bivariate normal distribution with µx = EX1 and µy = EY1 6= 0. Let ¯ Y¯ , the ratio of two sample means. Then ETn is ϑ = µx /µy and Tn = X/ not defined for any n. It is then desirable to define a concept of asymptotic bias for point estimators whose expectations are not well defined. Definition 2.11. (i) Let ξ, ξ1 , ξ2 , ... be random variables and {an } be a sequence of positive numbers satisfying an → ∞ or an → a > 0. If an ξn →d ξ and E|ξ| < ∞, then Eξ/an is called an asymptotic expectation of ξn . (ii) Let Tn be a point estimator of ϑ for every n. An asymptotic expectation of Tn − ϑ, if it exists, is called an asymptotic bias of Tn and denoted by ˜bT (P ) (or ˜bT (θ) if P is in a parametric family). If limn→∞ ˜bT (P ) = 0 for n n n any P ∈ P, then Tn is said to be asymptotically unbiased. Like the consistency, the asymptotic expectation (or bias) is a concept relating to sequences {ξn } and {Eξ/an } (or {Tn } and {˜bTn (P )}). Note that the exact bias bTn (P ) is not necessarily the same as ˜bTn (P ) when both of them exist (Exercise 115 in §2.6). The following result shows that the asymptotic expectation defined in Definition 2.11 is essentially unique.

136

2. Fundamentals of Statistics

Proposition 2.3. Let {ξn } be a sequence of random variables. Suppose that both Eξ/an and Eη/bn are asymptotic expectations of ξn defined according to Definition 2.11(i). Then, one of the following three must hold: (a) Eξ = Eη = 0; (b) Eξ 6= 0, Eη = 0, and bn /an → 0; or Eξ = 0, Eη 6= 0, and an /bn → 0; (c) Eξ 6= 0, Eη 6= 0, and (Eξ/an )/(Eη/bn ) → 1. Proof. According to Definition 2.11(i), an ξn →d ξ and bn ξn →d η. (i) If both ξ and η have nondegenerate c.d.f.’s, then the result follows from Exercise 129 of §1.6. (ii) Suppose that ξ has a nondegenerate c.d.f. but η is a constant. If η 6= 0, then by Theorem 1.11(iii), an /bn → ξ/η, which is impossible since ξ has a nondegenerate c.d.f. If η = 0, then by Theorem 1.11(ii), bn /an → 0. (iii) Suppose that both ξ and η are constants. If ξ = η = 0, the result follows. If ξ 6= 0 and η = 0, then bn /an → 0. If ξ 6= 0 and η 6= 0, then bn /an → η/ξ. If Tn is a consistent estimator of ϑ, then Tn = ϑ + op (1) and, by Definition 2.11(ii), Tn is asymptotically unbiased, although Tn may not be approximately unbiased; in fact, g(Tn ) is asymptotically unbiased for g(ϑ) for ¯ Y¯ , Tn →a.s. µx /µy any continuous function g. For the example of Tn = X/ by the SLLN and Theorem 1.10. Hence Tn is asymptotically unbiased, although ETn may not be defined. In Example 2.34, X(n) has the asymptotic −1 bias ˜bX(n) (P ) = hn (θ)EY , which is of order n−(m+1) . ¯ 2 and ϑ = µ2 in When an (Tn − ϑ) →d Y with EY = 0 (e.g., Tn = X Example 2.33), a more precise order of the asymptotic bias of Tn may be obtained (for comparing different estimators in terms of their asymptotic biases). Suppose that there is a sequence of random variables {ηn } such that an ηn →d Y and a2n (Tn − ϑ − ηn ) →d W, (2.36) where Y and W are random variables with finite means, EY = 0 and ˜ EW 6= 0. Then we may define a−2 n to be the order of bTn (P ) or define EW/a2n to be the a−2 order asymptotic bias of T . However, ηn in (2.36) n n may not be unique. Some regularity conditions have to be imposed so that the order of asymptotic bias of Tn can be uniquely defined. In the following we focus on the case where X1 , ..., Xn are i.i.d. random k-vectors. Suppose that Tn has the following expansion: n n n 1 1X 1 XX , Tn − ϑ = φ(Xi ) + 2 ψ(Xi , Xj ) + op n i=1 n i=1 j=1 n

(2.37)

where φ and ψ are functions that may depend on P , Eφ(X1 ) = 0, E[φ(X1 )]2 < ∞, ψ(x, y) = ψ(y, x), Eψ(x, X1 ) = 0 for all x, E[ψ(Xi , Xj )]2 < ∞, i ≤ j, and Eψ(X1 , X1 ) 6= 0. From the result for V-statistics in §3.5.3 (Theorem

2.5. Asymptotic Criteria and Inference

3.16 and Exercise 113 in §3.6), n

137

n

1 XX ψ(Xi , Xj ) →d W, n i=1 j=1 where W is a random variable with P EW = Eψ(X1 , X1 ). Hence (2.36) √ n holds with an = n and ηn = n−1 i=1 φ(Xi ). Consequently, we can define Eψ(X1 , X1 )/n to be the n−1 order asymptotic bias of Tn . Examples of estimators that have expansion (2.37) are provided in §3.5.3 and §5.2.1. In the following we consider the special case of functions of sample means. ¯ Let PnX1 , ..., Xn be i.i.d. random k-vectors with finite Σ k= Var(X1 ), X = ¯ where g is a function on R that is secondn−1 i=1 Xi , and Tn = g(X), order differentiable at µ = EX1 ∈ Rk . Consider Tn as an estimator of ϑ = g(µ). Using Taylor’s expansion, we obtain expansion (2.37) with φ(x) = [∇g(µ)]τ (x − µ) and ψ(x, y) = (x − µ)τ ∇2 g(µ)(y − µ)/2, where ∇g is the kvector of partial derivatives of g and ∇2 g is the k ×k matrix of second-order partial derivatives of g. By the CLT and Theorem 1.10(iii), n

n

τ 2 1 XX n ¯ ¯ − µ) →d ZΣ ∇ g(µ)ZΣ , − µ)τ ∇2 g(µ)(X ψ(Xi , Xj ) = (X n i=1 j=1 2 2

where ZΣ = Nk (0, Σ). Thus,

tr ∇2 g(µ)Σ E[ZΣτ ∇2 g(µ)ZΣ ] = (2.38) 2n 2n ¯ where tr(A) denotes the is the n−1 order asymptotic bias of Tn = g(X), trace of the matrix A. Note that the quantity in (2.38) is the same as the ¯ obtained under a much more leading term in the exact bias of Tn = g(X) stringent condition on the derivatives of g (Lehmann, 1983, Theorem 2.5.1). Example 2.35. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = p, where p ∈ (0, 1) is unknown. Consider first the estimation ¯ = p(1−p)/n, the n−1 order asymptotic bias of of ϑ = p(1−p). Since Var(X) ¯ − X) ¯ according to (2.38) with g(x) = x(1 − x) is −p(1 − p)/n. On Tn = X(1 ¯ − X)] ¯ = EX ¯ − EX ¯2 = the other hand, a direct computation shows E[X(1 2 ¯ ¯ = p(1 − p) − p(1 − p)/n. Hence, the exact bias of Tn p − (E X) − Var(X) is the same as the n−1 order asymptotic bias. Consider next the estimation of ϑ = p−1 . In this case, there is no ¯ −1 . Then, an unbiased estimator of p−1 (Exercise 84 in §2.6). Let Tn = X −1 n order asymptotic bias of Tn according to (2.38) with g(x) = x−1 is (1 − p)/(p2 n). On the other hand, ETn = ∞ for every n. Like the bias, the mse of an estimator Tn of ϑ, mseTn (P ) = E(Tn − ϑ)2 , is not well defined if the second moment of Tn does not exist. We now

138

2. Fundamentals of Statistics

define a version of asymptotic mean squared error (amse) and a measure of assessing different point estimators of a common parameter. Definition 2.12. Let Tn be an estimator of ϑ for every n and {an } be a sequence of positive numbers satisfying an → ∞ or an → a > 0. Assume that an (Tn − ϑ) →d Y with 0 < EY 2 < ∞. (i) The asymptotic mean squared error of Tn , denoted by amseTn (P ) or amseTn (θ) if P is in a parametric family indexed by θ, is defined to be the asymptotic expectation of (Tn − ϑ)2 , i.e., amseTn (P ) = EY 2 /a2n . The asymptotic variance of Tn is defined to be σT2 n (P ) = Var(Y )/a2n . (ii) Let Tn′ be another estimator of ϑ. The asymptotic relative efficiency of Tn′ w.t.r. Tn is defined to be eTn′ ,Tn (P ) = amseTn (P )/amseTn′ (P ). (iii) Tn is said to be asymptotically more efficient than Tn′ if and only if lim supn eTn′ ,Tn (P ) ≤ 1 for any P and < 1 for some P . The amse and asymptotic variance are the same if and only if EY = 0. By Proposition 2.3, the amse or the asymptotic variance of Tn is essentially unique and, therefore, the concept of asymptotic relative efficiency in Definition 2.12(ii)-(iii) is well defined. 2 2 2 In Example 2.33, amseX¯ 2 (P ) = σX ¯ 2 (P ) = 4µ σ /n. In Example 2.34, 2 2 σX(n) (P ) = [hn (θ)] Var(Y ) and amseX(n) (P ) = [hn (θ)]2 EY 2 . When both mseTn (P ) and mseTn′ (P ) exist, one may compare Tn and Tn′ by evaluating the relative efficiency mseTn (P )/mseTn′ (P ). However, this comparison may be different from the one using the asymptotic relative efficiency in Definition 2.12(ii), since the mse and amse of an estimator may be different (Exercise 115 in §2.6). The following result shows that when the exact mse of Tn exists, it is no smaller than the amse of Tn . It also provides a condition under which the exact mse and the amse are the same. Proposition 2.4. Let Tn be an estimator of ϑ for every n and {an } be a sequence of positive numbers satisfying an → ∞ or an → a > 0. Suppose that an (Tn − ϑ) →d Y with 0 < EY 2 < ∞. Then (i) EY 2 ≤ lim inf n E[a2n (Tn − ϑ)2 ] and (ii) EY 2 = limn→∞ E[a2n (Tn − ϑ)2 ] if and only if {a2n (Tn − ϑ)2 } is uniformly integrable. Proof. (i) By Theorem 1.10(iii), min{a2n (Tn − ϑ)2 , t} →d min{Y 2 , t} for any t > 0. Since min{a2n (Tn − ϑ)2 , t} is bounded by t, lim E(min{a2n (Tn − ϑ)2 , t}) = E(min{Y 2 , t})

n→∞

139

2.5. Asymptotic Criteria and Inference

(Theorem 1.8(viii)). Then EY 2 = lim E(min{Y 2 , t}) t→∞

= lim lim E(min{a2n (Tn − ϑ)2 , t}) t→∞ n→∞

= lim inf E(min{a2n (Tn − ϑ)2 , t}) t,n

≤ lim inf E[a2n (Tn − ϑ)2 ], n

where the third equality follows from the fact that E(min{a2n (Tn − ϑ)2 , t}) is nondecreasing in t for any fixed n. (ii) The result follows from Theorem 1.8(viii). Example 2.36. Let X1 , ..., Xn be i.i.d. from the Poisson distribution P (θ) with an unknown θ > 0. Consider the estimation of ϑ = P (Xi = 0) = e−θ . Let T1n = Fn (0), where Fn is the empirical c.d.f. defined √ in (2.28). Then T1n is unbiased and has mseT1n (θ) = e−θ (1−e−θ )/n. Also, n(T1n −ϑ) →d N (0, e−θ (1 − e−θ )) by the CLT. Thus, in this case amseT1n (θ) = mseT1n (θ). ¯

−1/n

−1) Next, consider T2n = e−X . Note that ET2n = enθ(e . Hence −θ nb (θ) → θe /2. Using Theorem 1.12 and the CLT, we can show that T 2n √ n(T2n − ϑ) →d N (0, e−2θ θ). By Definition 2.12(i), amseT2n (θ) = e−2θ θ/n. Thus, the asymptotic relative efficiency of T1n w.r.t. T2n is

eT1n ,T2n (θ) = θ/(eθ − 1), which is always less than 1. This shows that T2n is asymptotically more efficient than T1n . ¯ of The result for T2n in Example 2.36 is a special case (with Un = X) the following general result. Theorem 2.6. Let g be a function on Rk that is differentiable at θ ∈ Rk and let Un be a k-vector of statistics satisfying an (Un − θ) →d Y for a random k-vector Y with 0 < EkY k2 < ∞ and a sequence of positive numbers {an } satisfying an → ∞. Let Tn = g(Un ) be an estimator of ϑ = g(θ). Then, the amse and asymptotic variance of Tn are, respectively, E{[∇g(θ)]τ Y }2 /a2n and [∇g(θ)]τ Var(Y )∇g(θ)/a2n .

2.5.3 Asymptotic inference Statistical inference based on asymptotic criteria and approximations is called asymptotic statistical inference or simply asymptotic inference. We have previously considered asymptotic estimation. We now focus on asymptotic hypothesis tests and confidence sets.

140

2. Fundamentals of Statistics

Definition 2.13. Let X = (X1 , ..., Xn ) be a sample from P ∈ P and Tn (X) be a test for H0 : P ∈ P0 versus H1 : P ∈ P1 . (i) If lim supn αTn (P ) ≤ α for any P ∈ P0 , then α is an asymptotic significance level of Tn . (ii) If limn→∞ supP ∈P0 αTn (P ) exists, then it is called the limiting size of Tn . (iii) Tn is called consistent if and only if the type II error probability converges to 0, i.e., limn→∞ [1 − αTn (P )] = 0, for any P ∈ P1 . (iv) Tn is called Chernoff-consistent if and only if Tn is consistent and the type I error probability converges to 0, i.e., limn→∞ αTn (P ) = 0, for any P ∈ P0 . Tn is called strongly Chernoff-consistent if and only if Tn is consistent and the limiting size of Tn is 0. Obviously if Tn has size (or significance level) α for all n, then its limiting size (or asymptotic significance level) is α. If the limiting size of Tn is α ∈ (0, 1), then for any ǫ > 0, Tn has size α + ǫ for all n ≥ n0 , where n0 is independent of P . Hence Tn has level of significance α + ǫ for any n ≥ n0 . However, if P0 is not a parametric family, it is likely that the limiting size of Tn is 1 (see, e.g., Example 2.37). This is the reason why we consider the weaker requirement in Definition 2.13(i). If Tn has asymptotic significance level α, then for any ǫ > 0, αTn (P ) < α + ǫ for all n ≥ n0 (P ) but n0 (P ) depends on P ∈ P0 ; and there is no guarantee that Tn has significance level α + ǫ for any n. The consistency in Definition 2.13(iii) only requires that the type II error probability converge to 0. We may define uniform consistency to be limn→∞ supP ∈P1 [1 − αTn (P )] = 0, but it is not satisfied in most problems. If α ∈ (0, 1) is a pre-assigned level of significance for the problem, then a consistent test Tn having asymptotic significance level α is called asymptotically correct, and a consistent test having limiting size α is called strongly asymptotically correct. The Chernoff-consistency (or strong Chernoff-consistency) in Definition 2.13(iv) requires that both types of error probabilities converge to 0. Mathematically, Chernoff-consistency (or strong Chernoff-consistency) is better than asymptotic correctness (or strongly asymptotic correctness). After all, both types of error probabilities should decrease to 0 if sampling can be continued indefinitely. However, if α is chosen to be small enough so that error probabilities smaller than α can be practically treated as 0, then the asymptotic correctness (or strongly asymptotic correctness) is enough, and is probably preferred, since requiring an unnecessarily small type I error probability usually results in an unnecessary increase in the type II error probability, as the following example illustrates. Example 2.37. Consider the testing problem H0 : µ ≤ µ0 versus H1 :

2.5. Asymptotic Criteria and Inference

141

µ > µ0 based on i.i.d. X1 , ..., Xn with EX1 = µ ∈ R. If each Xi has the 2 N (µ, σ 2 ) distribution with √ a known σ , then the test Tcα given in Example 2.28 with cα = σz1−α / n + µ0 and α ∈ (0, 1) has size α (and, therefore, limiting size α). It also follows from (2.32) that for any µ > µ0 , √ n(µ0 − µ) →0 (2.39) 1 − αTcα (µ) = Φ z1−α + σ as n → ∞. This shows that Tcα is consistent and, hence, is strongly asymptotically correct. Note that the convergence in (2.39) is not uniform in µ > µ0 , but is uniform in µ > µ1 for any fixed µ1 > µ0 . Since the size of Tcα is α for all n, Tcα is not Chernoff-consistent. A strongly Chernoff-consistent test can be obtained as follows. Let √ αn = 1 − Φ( nan ), (2.40) √ where an ’s are positive numbers satisfying an → 0 and nan → ∞. Let Tn be Tcα with α = αn for each n. Then, Tn has size αn . Since αn → 0, The limiting size of Tn is 0. On the other hand, (2.39) still holds with α replaced by αn . This follows from the fact that √ n(µ0 − µ) √ µ0 − µ = n an + → −∞ z1−αn + σ σ for any µ > µ0 . Hence Tn is strongly Chernoff-consistent. However, if αn < α, then, from the left-hand side of (2.39), 1 − αTcα (µ) < 1 − αTn (µ) for any µ > µ0 . We now consider the case where the population P is not in a parametric family. We still assume that σ 2 = Var(Xi ) is known. Using the CLT, we can show that for µ > µ0 , √ n(µ0 − µ) = 0, lim [1 − αTcα (µ)] = lim Φ z1−α + n→∞ n→∞ σ i.e., Tcα is still consistent. For µ ≤ µ0 ,

√ n(µ0 − µ) , lim αTcα (µ) = 1 − lim Φ z1−α + n→∞ n→∞ σ

which equals α if µ = µ0 and 0 if µ < µ0 . Thus, the asymptotic significance level of Tcα is α. Combining these two results, we know that Tcα is asymptotically correct. However, if P contains all possible populations on R with finite second moments, then one can show that the limiting size of Tcα is 1 (exercise). For αn defined by (2.40), we can show that Tn = Tcα with α = αn is Chernoff-consistent (exercise). But Tn is not strongly Chernoffconsistent if P contains all possible populations on R with finite second moments.

142

2. Fundamentals of Statistics

Definition 2.14. Let X = (X1 , ..., Xn ) be a sample from P ∈ P, ϑ be a k-vector of parameters related to P , and C(X) be a confidence set for ϑ. (i) If lim inf n P (ϑ ∈ C(X)) ≥ 1 − α for any P ∈ P, then 1 − α is an asymptotic significance level of C(X). (ii) If limn→∞ inf P ∈P P (ϑ ∈ C(X)) exists, then it is called the limiting confidence coefficient of C(X). Note that the asymptotic significance level and limiting confidence coefficient of a confidence set are very similar to the asymptotic significance level and limiting size of a test, respectively. Some conclusions are also similar. For example, in a parametric problem one can often find a confidence set having limiting confidence coefficient 1 − α ∈ (0, 1), which implies that for any ǫ > 0, the confidence coefficient of C(X) is 1 − α − ǫ for all n ≥ n0 , where n0 is independent of P ; in a nonparametric problem the limiting confidence coefficient of C(X) might be 0, whereas C(X) may have asymptotic significance level 1 − α ∈ (0, 1), but for any fixed n, the confidence coefficient of C(X) might be 0. √ The confidence interval in Example 2.31 with c = σz1−α/2 / n and the confidence set in Example 2.32 have confidence coefficient 1 − α for any n and, therefore, have limiting confidence coefficient 1 − α. If we drop the normality assumption and assume EXi4 < ∞, then these confidence sets have asymptotic significance level 1−α; their limiting confidence coefficients may be 0 (exercise).

2.6 Exercises 1. Consider Example 2.3. Suppose that p(s) is constant. Show that Xi and Xj , i 6= j, are not uncorrelated and, hence, X1 , ..., Xn are not independent. Furthermore, when yi ’s are either 0 or 1, show that Pn Z = X i has a hypergeometric distribution and compute the i=1 mean of Z. 2. Consider Example 2.3. Suppose that we do not require that the elements in s be distinct, i.e., we consider sampling with replacement. Define a probability measure p and a sample (X1 , ..., Xn ) such that (2.3) holds. If p(s) is constant, are X1 , ..., Xn independent? If p(s) is constant and Pnyi ’s are either 0 or 1, what are the distribution and mean of Z = i=1 Xi ?

3. Show that {Pθ : θ ∈ Θ} is an exponential family and find its canonical form and natural parameter space, when (a) Pθ is the Poisson distribution P (θ), θ ∈ Θ = (0, ∞); (b) Pθ is the negative binomial distribution N B(θ, r) with a fixed r,

2.6. Exercises

143

θ ∈ Θ = (0, 1); (c) Pθ is the exponential distribution E(a, θ) with a fixed a, θ ∈ Θ = (0, ∞); (d) Pθ is the gamma distribution Γ(α, γ), θ = (α, γ) ∈ Θ = (0, ∞) × (0, ∞); (e) Pθ is the beta distribution B(α, β), θ = (α, β) ∈ Θ = (0, 1)×(0, 1); (f) Pθ is the Weibull distribution W (α, θ) with a fixed α > 0, θ ∈ Θ = (0, ∞). 4. Show that the family of exponential distributions E(a, θ) with two unknown parameters a and θ is not an exponential family. 5. Show that the family of negative binomial distributions N B(p, r) with two unknown parameters p and r is not an exponential family. 6. Show that the family of Cauchy distributions C(µ, σ) with two unknown parameters µ and σ is not an exponential family. 7. Show that the family of Weibull distributions W (α, θ) with two unknown parameters α and θ is not an exponential family. 8. Is the family of log-normal distributions LN (µ, σ 2 ) with two unknown parameters µ and σ 2 an exponential family? 9. Show that the family of double exponential distributions DE(µ, θ) with two unknown parameters µ and θ is not an exponential family, but the family of double exponential distributions DE(µ, θ) with a fixed µ and an unknown parameter θ is an exponential family. 10. Show that the k-dimensional normal family discussed in Example 2.4 is an exponential family. Identify the functions T , η, ξ, and h. 11. Obtain the variance-covariance matrix for (X1 , ..., Xk ) in Example 2.7, using (a) Theorem 2.1(ii) and (b) direct computation. 12. Show that the m.g.f. of the gamma distribution Γ(α, γ) is (1 − γt)−α , t < γ −1 , using Theorem 2.1(ii). 13. A discrete random variable X with P (X = x) = γ(x)θx /c(θ), x = 0, 1, 2, ..., P∞ where γ(x) ≥ 0, θ > 0, and c(θ) = x=0 γ(x)θx , is called a random variable with a power series distribution. (a) Show that {γ(x)θx /c(θ) : θ > 0} is an exponential family. (b) Suppose that X1 , ..., XP n are i.i.d. with a power series distribution n γ(x)θx /c(θ). Show that i=1 Xi has the power series distribution γn (x)θx /[c(θ)]n , where γn (x) is the coefficient of θx in the power series expansion of [c(θ)]n .

144

2. Fundamentals of Statistics

14. Let X be a random variable with a p.d.f. fθ in an exponential family {Pθ : θ ∈ Θ} and let A be a Borel set. Show that the distribution of X truncated on A (i.e., the conditional distribution of X given X ∈ A) has a p.d.f. fθ IA /Pθ (A) that is in an exponential family. 15. Let {P(µ,Σ) : µ ∈ Rk , Σ ∈ Mk } be a location-scale family on Rk . Suppose that P(0,Ik ) has a Lebesgue p.d.f. that is always positive and that the mean and variance-covariance matrix of P(0,Ik ) are 0 and Ik , respectively. Show that the mean and variance-covariance matrix of P(µ,Σ) are µ and Σ, respectively. 16. Show that if the distribution of a positive random variable X is in a scale family, then the distribution of log X is in a location family. 17. Let X be a random variable having the gamma distribution Γ(α, γ) with a known α and an unknown γ > 0 and let Y = σ log X. (a) Show that if σ > 0 is unknown, then the distribution of Y is in a location-scale family. (b) Show that if σ > 0 is known, then the distribution of Y is in an exponential family. 18. Let X1 , ..., Xn be i.i.d. random variables having a finite E|X1 |4 and ¯ and S 2 be the sample mean and variance defined by (2.1) and let X ¯ 3 ), Cov(X, ¯ S 2 ), and Var(S 2 ) in terms of µk = (2.2). Express E(X k ¯ and S 2 are EX1 , k = 1, 2, 3, 4. Find a condition under which X uncorrelated. 19. Let X1 , ..., Xn be i.i.d. random variables having the gamma distribution Γ(α, γx ) and Y1 , ..., Yn be i.i.d. random variables having the gamma distribution Γ(α, γy ), where α > 0, γx > 0, and γy > 0. Assume that Xi ’s and Yi ’s are independent. Derive the distribution of ¯ Y¯ , where X ¯ and Y¯ are the sample means based on the statistic X/ Xi ’s and Yi ’s, respectively. 20. Let X1 , ..., Xn be i.i.d. random variables having the exponential distribution E(a, θ), a ∈ R, and θ > 0. Show that the smallest order statistic, X(1) , has the exponential distribution E(a, θ/n) and that Pn 2 i=1 (Xi − X(1) )/θ has the chi-square distribution χ22n−2 .

21. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors. Suppose that X1 has the Cauchy distribution C(0, 1) and given X1 = x, Y1 has ¯ and Y¯ be the Cauchy distribution C(βx, 1), where β ∈ R. Let X the sample means based on Xi ’s and Yi ’s, respectively. Obtain the ¯ and Y¯ /X. ¯ marginal distributions of Y¯ , Y¯ − β X,

145

2.6. Exercises

22. Let Xi = (Yi , Zi ), i = 1, ..., n, be i.i.d. random 2-vectors. The sample correlation coefficient is defined to be T (X) =

n X 1 ¯ p (Yi − Y¯ )(Zi − Z), (n − 1) SY2 SZ2 i=1

P Pn Pn −1 2 −1 ¯ ¯ 2 where Y¯ = n−1 ni=1 i=1 (Yi−Y ) , PYni , Z = n ¯ 2 i=1 Zi , SY = (n−1) 2 −1 and SZ = (n−1) i=1 (Zi − Z) . (a) Assume that E|Yi |4 < ∞ and E|Zi |4 < ∞. Show that √ n[T (X) − ρ] →d N (0, c2 ), where ρ is the correlation coefficient between Y1 and Z1 and c is a constant depending on some unknown parameters. (b) Assume that Yi and Zi are independently distributed as N (µ1 , σ12 ) and N (µ2 , σ22 ), respectively. Show that T has the Lebesgue p.d.f. Γ n−1 2 (1 − t2 )(n−4)/2 I(−1,1) (t). f (t) = √ πΓ n−2 2 (c) Assume the conditions in (b). Obtain the result in (a) using Scheff´e’s theorem (Proposition 1.18).

23. Let X1 , ..., X√ with EX14 < ∞, P T = (Y, Z), n be i.i.d. random variables Pn n −1 −1 2 and T1 = Y / Z, where Y = n |X | and Z = n i i=1 i=1 Xi . √ √ (a) Show that n(T − θ) →d N2 (0, Σ) and n(T1 − ϑ) →d N (0, c2 ). Identify θ, Σ, ϑ, and c2 in terms of moments of X1 . (b) Repeat (a) when X1 has the normal distribution N (0, σ 2 ). (c) Repeat (a) when X1 has the double exponential distribution D(0, σ). 24. Prove the claims in Example 2.9 for the distributions related to order statistics. 25. Show that if T is a sufficient statistic and T = ψ(S), where ψ is measurable and S is another statistic, then S is sufficient. 26. In the proof of Lemma 2.1, show that C0 ∈ C. Also, prove Lemma 2.1 when P is dominated by a σ-finite measure. 27. Let X1 , ..., Xn be i.i.d. random variables from Pθ ∈ {Pθ : θ ∈ Θ}. In the following cases, find a sufficient statistic for θ ∈ Θ that has the same dimension as θ. (a) Pθ is the Poisson distribution P (θ), θ ∈ (0, ∞). (b) Pθ is the negative binomial distribution N B(θ, r) with a known r, θ ∈ (0, 1).

146

2. Fundamentals of Statistics

(c) Pθ is the exponential distribution E(0, θ), θ ∈ (0, ∞). (d) Pθ is the gamma distribution Γ(α, γ), θ = (α, γ) ∈ (0, ∞)×(0, ∞). (e) Pθ is the beta distribution B(α, β), θ = (α, β) ∈ (0, 1) × (0, 1). (f) Pθ is the log-normal distribution LN (µ, σ 2 ), θ = (µ, σ 2 ) ∈ R × (0, ∞). (g) Pθ is the Weibull distribution W (α, θ) with a known α > 0, θ ∈ (0, ∞). 28. Let X1 , ..., Xn be i.i.d. random variables from P(a,θ) , where (a, θ) ∈ R2 is a parameter. Find a two-dimensional sufficient statistic for (a, θ) in the following cases. (a) P(a,θ) is the exponential distribution E(a, θ), a ∈ R, θ ∈ (0, ∞). (b) P(a,θ) is the Pareto distribution P a(a, θ), a ∈ (0, ∞), θ ∈ (0, ∞). 29. In Example 2.11, show that X(1) (or X(n) ) is sufficient for a (or b) if we consider a subfamily {f(a,b) : a < b} with a fixed b (or a). 30. Let X and Y be two random variables such that Y has the binomial distribution Bi(π, N ) and, given Y = y, X has the binomial distribution Bi(p, y). (a) Suppose that p ∈ (0, 1) and π ∈ (0, 1) are unknown and N is known. Show that (X, Y ) is minimal sufficient for (p, π). (b) Suppose that π and N are known and p ∈ (0, 1) is unknown. Show whether X is sufficient for p and whether Y is sufficient for p. 31. Let X1 , ..., Xn be i.i.d. random variables having a distribution P ∈ P, where P is the family of distributions on R having continuous c.d.f.’s. Let T = (X(1) , ..., X(n) ) be the vector of order statistics. Show that, given T , the conditional distribution of X = (X1 , ..., Xn ) is a discrete distribution putting probability 1/n! on each of the n! points (Xi1 , ..., Xin ) ∈ Rn , where {i1 , ..., in } is a permutation of {1, ..., n}; hence, T is sufficient for P ∈ P. 32. In Example 2.13 and Example 2.14, show that T is minimal sufficient for θ by using Theorem 2.3(iii). 33. A coin has probability p of coming up heads and 1 − p of coming up tails, where p ∈ (0, 1). The first stage of an experiment consists of tossing this coin a known total of M times and recording X, the number of heads. In the second stage, the coin is tossed until a total of X + 1 tails have come up. The number Y of heads observed in the second stage along the way to getting the X + 1 tails is then recorded. This experiment is repeated independently a total of n times and the two counts (Xi , Yi ) for the ith experiment are recorded, i = 1, ..., n. Obtain a statistic that is minimal sufficient for p and derive its distribution.

2.6. Exercises

147

34. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. n o 4 − ξ(θ) , fθ (x) = exp − x−µ σ

where θ = (µ, σ) ∈ Θ = R × (0, ∞). Show that P = {Pθ : θ ∈ Θ} is an exponential family, where P Pθ nis the joint Pn distribution Pn of3 X P1 ,n..., X4n, 2 and that the statistic T = i=1 Xi , i=1 Xi , i=1 Xi , i=1 Xi is minimal sufficient for θ ∈ Θ.

35. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. fθ (x) = (2θ)−1 I(0,θ) (x) + I(2θ,3θ) (x) . Find a minimal sufficient statistic for θ ∈ (0, ∞).

36. Let X1 , ..., Xn be i.i.d. random variables having the Cauchy distribution C(µ, σ) with unknown µ ∈ R and σ > 0. Show that the vector of order statistics is minimal sufficient for (µ, σ). 37. Let X1 , ..., Xn be i.i.d. random variables having the double exponential distribution DE(µ, θ) with unknown µ ∈ R and θ > 0. Show that the vector of order statistics is minimal sufficient for (µ, θ). 38. Let X1 , ..., Xn be i.i.d. random variables having the Weibull distribution W (α, θ) with unknown α > 0 and θ > 0. Show that the vector of order statistics is minimal sufficient for (α, θ). 39. Let X1 , ..., Xn be i.i.d. random variables having the beta distribution B(β, β) with an unknown β > 0. Find a minimal sufficient statistic for β. 40. Let X1 , ..., Xn be i.i.d. random variables having a population P in a parametric family indexed by (θ, j), where θ > 0, j = 1, 2, and n ≥ 2. When j = 1, P is the N (0, θ2 ) distribution. When j = 2, PPis the double Pn exponential distribution DE(0, θ). Show that T = n ( i=1 Xi2 , i=1 |Xi |) is minimal sufficient for (θ, j).

41. Let X1 , ..., Xn be i.i.d. random variables having a population P in a parametric family indexed by (θ, j), where θ ∈ (0, 1), j = 1, 2, and n ≥ 2. When j = 1, P is the Poisson distribution P (θ). When j = 2, P is the binomial distribution Bi(θ, 1). Pn (a) Show that T = i=1 Xi is not sufficient for (θ, j). (b) Find a two-dimensional minimal sufficient statistic for (θ, j).

42. Let X be a sample from P ∈ P = {fθ,j : θ ∈ Θ, j = 1, ..., k}, where fθ,j ’s are p.d.f.’s w.r.t. a common σ-finite measure and Θ is a set of parameters. Assume that {x : fθ,j (x) > 0} ⊂ {x : fθ,k (x) > 0} for all

148

2. Fundamentals of Statistics

θ and j = 1, ..., k − 1. Suppose that for each fixed j, T = T (X) is a statistic sufficient for θ. (a) Obtain a k-dimensional statistic that is sufficient for (θ, j). (b) Derive a sufficient condition under which T is minimal sufficient for (θ, j). 43. A box has an unknown odd number of balls labeled consecutively as −θ, −(θ − 1), ..., −2, −1, 0, 1, 2, ..., (θ − 1), θ, where θ is an unknown nonnegative integer. A simple random sample X1 , ..., Xn is taken without replacement, where Xi is the label on the ith ball selected and n < 2θ + 1. (a) Find a statistic that is minimal sufficient for θ and derive its distribution. (b) Show that the minimal sufficient statistic in (a) is also complete. 44. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. θ−1 e−(x−θ)/θ I(θ,∞) (x), where θ > 0 is an unknown parameter. (a) Find a statistic that is minimal sufficient for θ. (b) Show whether the minimal sufficient statistic in (a) is complete. 45. Let X1 , ..., Xn (n ≥ 2) be i.i.d. random variables having the normal distribution N (θ, 2) when θ = 0 and the normal distribution N (θ, 1) ¯ is a complete when θ ∈ R and θ 6= 0. Show that the sample mean X statistic for θ but it is not a sufficient statistic for θ. 46. Let X be a random variable with a distribution Pθ in {Pθ : θ ∈ Θ}, fθ be the p.d.f. of Pθ w.r.t. a measure ν, A be an event, and PA = {fθ IA /Pθ (A) : θ ∈ Θ}. (a) Show that if T (X) is sufficient for Pθ ∈ P, then it is sufficient for Pθ ∈ PA . (b) Show that if T is sufficient and complete for Pθ ∈ P, then it is complete for Pθ ∈ PA . 47. Show that (X(1) , X(n) ) in Example 2.13 is not complete. 48. Let T be a complete (or boundedly complete) and sufficient statistic. Suppose that there is a minimal sufficient statistic S. Show that T is minimal sufficient and S is complete (or boundedly complete). 49. Let T and S be two statistics such that S = ψ(T ) for a measurable ψ. Show that (a) if T is complete, then S is complete; (b) if T is complete and sufficient and ψ is one-to-one, then S is complete and sufficient; (c) the results in (a) and (b) still hold if the completeness is replaced by the bounded completeness.

149

2.6. Exercises

50. Find complete and sufficient statistics for the families in Exercises 27 and 28. 51. Show that (X(1) , X(n) ) in Example 2.11 is complete. 52. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors having the following Lebesgue p.d.f. p fθ (x, y) = (2πγ 2 )−1 I(0,γ) (x − a)2 + (y − b)2 , (x, y) ∈ R2 , where θ = (a, b, γ) ∈ R2 × (0, ∞). (a) If a = 0 and b = 0, find a complete and sufficient statistic for γ. (b) If all parameters are unknown, show that the convex hull of the sample points is a sufficient statistic for θ.

53. Let X be a discrete random variable with θ fθ (x) = (1 − θ)2 θx−1 0

p.d.f. x=0 x = 1, 2, ... otherwise,

where θ ∈ (0, 1). Show that X is boundedly complete, but not complete.

54. Show that the sufficient statistic T in Example 2.10 is also complete without using Proposition 2.1. 55. Let Y1 , ..., Yn be i.i.d. random variables having the Lebesgue p.d.f. λxλ−1 I(0,1) (x) with an unknown λ > 0 and let Z1 , ..., Zn be i.i.d. discrete random variables having the power series distribution given in Exercise 13 with an unknown θ > 0. Assume that Yi ’s and Zj ’s are independent. Let Xi = Yi + Zi , i = 1, ..., n. Find a complete and sufficient statistic for the unknown parameter (θ, λ) based on the sample X = (X1 , ..., Xn ). 56. Suppose that (X1 , Y1 ), ..., (Xn , Yn ) are i.i.d. random 2-vectors and 2 Xi and Yi are independently distributed as N (µ, σX ) and N (µ, σY2 ), 2 2 ¯ and respectively, with θ = (µ, σX , σY ) ∈ R × (0, ∞) × (0, ∞). Let X 2 SX be the sample mean and variance given by (2.1) and (2.2) for Xi ’s and Y¯ and SY2 be the sample mean and variance for Yi ’s. Show that ¯ Y¯ , S 2 , S 2 ) is minimal sufficient for θ but T is not boundedly T = (X, X Y complete. 57. Let X1 , ..., Xn be i.i.d. from the N (θ, θ2 ) distribution, where θ > 0 is a parameter. Find a minimal sufficient statistic for θ and show whether it is complete.

150

2. Fundamentals of Statistics

58. Suppose that (X1 , Y1 ), ..., (Xn , Yn ) are i.i.d. random 2-vectors having the normal distribution with EX1 = EY1 = 0, Var(X1 ) = Var(Y1 ) = 1, and Cov(X1 , Y1 ) = θ ∈ (−1, 1). (a) Find a minimal sufficient statistic for θ. (b) Show whether the minimal sufficient statistic in (a) is complete or not. Pn Pn (c) Prove that T1 = i=1 Xi2 and T2 = i=1 Yi2 are both ancillary but (T1 , T2 ) is not ancillary. 59. Let X1 , ..., Xn be i.i.d. random variables having the exponential distribution E(a, θ). Pn (a) Show that i=1 (Xi − X(1) ) and X(1) are independent for any (a, θ). (b) Show that Zi = (X(n)P− X(i) )/(X(n) − X(n−1) ), i = 1, ..., n − 2, n are independent of (X(1) , i=1 (Xi − X(1) )).

60. Let X1 , ..., Xn be i.i.d. random having the gamma distriP variablesP bution Γ(α, γ). Show that ni=1 Xi and ni=1 [log Xi − log X(1) ] are independent for any (α, γ).

61. Let X1 , ..., Xn be i.i.d. random variables having the uniform distribution on the interval (a, b), where −∞ < a < b < ∞. Show that (X(i) − X(1) )/(X(n) − X(1) ), i = 2, ..., n − 1, are independent of (X(1) , X(n) ) for any a and b. 62. Consider Example 2.19. Assume that n > 2. ¯ is better than T1 if P = N (θ, σ 2 ), θ ∈ R, σ > 0. (a) Show that X ¯ if P is the uniform distribution on (b) Show that T1 is better than X 1 1 the interval (θ − 2 , θ + 2 ), θ ∈ R. ¯ nor T1 is better than the (c) Find a family P for which neither X other. 63. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution, where µ ∈ R and σ > 0. Consider the estimation of σ 2 with the squared error loss. 2 2 Show that n−1 n S is better than S , the sample variance. Can you 2 find an estimator of the form cS with a nonrandom c such that it is 2 better than n−1 n S ? 64. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = θ ∈ (0, 1). Consider estimating θ with the squared error loss. Calculate the risks of the following estimators: ¯ (the sample mean) and (a) the nonrandomized estimators X if more than half of Xi ’s are 0 0 T0 (X) = 1 if more than half of Xi ’s are 1 1 if exactly half of Xi ’s are 0; 2

151

2.6. Exercises

(b) the randomized estimators ¯ X T1 (X) = T0 and T2 (X) =

¯ X 1 2

with probability with probability

1 2 1 2

¯ with probability X ¯ with probability 1 − X.

65. Let X1 , ..., Xn be i.i.d. random variables having the exponential distribution E(0, θ), θ ∈ (0, ∞). Consider estimating θ with the squared ¯ and cX(1) , where error loss. Calculate the risks of the sample mean X ¯ better than cX(1) for some c? c is a positive constant. Is X 66. Consider the estimation of an unknown parameter θ ≥ 0 under the squared error loss. Show that if T and U are two estimators such that T ≤ U and RT (P ) < RU (P ), then RT+ (P ) < RU+ (P ), where RT (P ) is the risk of an estimator T and T+ denotes the positive part of T . 67. Let X1 , ..., Xn be i.i.d. random variables having the exponential distribution E(0, θ), θ ∈ (0, ∞). Consider the hypotheses H0 : θ ≤ θ 0

versus H1 : θ > θ0 ,

where θ0 > 0 is a fixed constant. Obtain the risk function (in terms ¯ under the 0-1 loss. of θ) of the test rule Tc (X) = I(c,∞) (X), 68. Let X1 , ..., Xn be i.i.d. random variables having the Cauchy distribution C(µ, σ) with unknown µ ∈ R and σ > 0. Consider the hypotheses H0 : µ ≤ µ0

versus H1 : µ > µ0 ,

where µ0 is a fixed constant. Obtain the risk function of the test rule ¯ under the 0-1 loss. Tc (X) = I(c,∞) (X), 69. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = θ, where θ ∈ (0, 1) is unknown and n is an even integer. Consider the problem of testing H0 : θ ≤ 0.5 versus H1 : θ > 0.5 with action space {0, 1} (0 means H0 is accepted and 1 means H1 is accepted). Let the loss function be L(θ, a) = 0 if Hj is true and a = j, j = 0, 1; L(θ, 0) = C0 when θ > 0.5; and L(θ, 1) = C1 when θ ≤ 0.5, where C0 > C1 > 0 are some constants. Calculate the risk function of the following randomized test (decision rule): if more than half of Xi ’s are 0 0 T = 1 if more than half of Xi ’s are 1 1 if exactly half of Xi ’s are 0. 2

152

2. Fundamentals of Statistics

70. Consider Example 2.21. Suppose that our decision rule, based on a sample X = (X1 , ..., Xn ) with i.i.d. components from the N (θ, 1) distribution with an unknown θ > 0, is ¯ b1 < X a1 ¯ ≤ b1 T (X) = b0 < X a 2 ¯ a3 X ≤ b0 . Express the risk of T in terms of θ.

71. Consider an estimation problem with P = {Pθ : θ ∈ Θ} (a parametric family), A = Θ, and the squared error loss. If θ0 ∈ Θ satisfies that Pθ ≪ Pθ0 for any θ ∈ Θ, show that the estimator T ≡ θ0 is admissible. 72. Let ℑ be a class of decision rules. A subclass ℑ0 ⊂ ℑ is called ℑcomplete if and only if, for any T ∈ ℑ and T 6∈ ℑ0 , there is a T0 ∈ ℑ0 that is better than T , and ℑ0 is called ℑ-minimal complete if and only if ℑ0 is ℑ-complete and no proper subclass of ℑ0 is ℑ-complete. Show that if a ℑ-minimal complete class exists, then it is exactly the class of ℑ-admissible rules. 73. Let X1 , ..., Xn be i.i.d. random variables having a distribution P ∈ P. Assume that EX12 < ∞. Consider estimating µ = EX1 under the squared error loss. ¯ + b is inadmissible, where (a) Show that any estimator of the form aX ¯ X is the sample mean, a and b are constants, and a > 1. ¯ + b is inadmissible, where (b) Show that any estimator of the form X b 6= 0 is a constant. 74. Consider an estimation problem with ϑ ∈ [c, d] ⊂ R, where c and d are known. Suppose that the action space is A ⊃ [c, d] and the loss function is L(|ϑ − a|), where L(·) is an increasing function on [0, ∞). Show that any decision rule T with P (T (X) 6∈ [c, d]) > 0 for some P ∈ P is inadmissible. k 75. Suppose that the action space is (Ω, BΩ ), where Ω ∈ B k . Let X be a sample from P ∈ P, δ0 (X) be a nonrandomized rule, and T be a sufficient statistic for P ∈ P. Show that if E[IA (δ0 (X))|T ] is a k nonrandomized rule, i.e., E[IA (δ0 (X))|T ] = IA (h(T )) for any A ∈ BΩ , where h is a Borel function, then δ0 (X) = h(T (X)) a.s. P .

76. Let T , δ0 , and δ1 be as given in the statement of Proposition 2.2. Show that Z Z L(P, a)dδ1 (X, a) = E L(P, a)dδ0 (X, a) T a.s. P . A

A

153

2.6. Exercises

77. Prove Theorem 2.5. 78. In Exercise 64, use Theorem 2.5 to find decision rules that are better than Tj , j = 0, 1, 2. 79. In Exercise 65, use Theorem 2.5 to find a decision rule better than cX(1) . 80. Consider Example 2.22. (a) Show that there is no optimal rule if ℑ contains all possible estimators. (Hint: consider constant estimators.) (b) Find a ℑ2 -optimal rule if X1 , ..., Xn are independent random variables having a common mean µ and Var(Xi ) = σ 2 /ai with known ai , i = 1, ..., n. (c) Find a ℑ2 -optimal rule if X1 , ..., Xn are identically distributed but are correlated with a common correlation coefficient ρ. 81. Let Xij = µ + ai + ǫij , i = 1, ..., m, j = 1, ..., n, where ai ’s and ǫij ’s 2 are independent random variables, ai is N (0, σa2 ), ǫij is N (0, Pσne ), and 2 2 −1 ¯ µ, σa , and σe are unknown parameters. Define Xi = n j=1 Xij , ¯ i − X) ¯ 2 , and MSE ¯ = m−1 Pm X ¯ i , MSA = n(m − 1)−1 Pm (X X i=1 P i=1 m Pn ¯ i )2 . Assume that m(n − 1) > 4. = m−1 (n − 1)−1 i=1 j=1 (Xij − X Consider the following class of estimators of θ = σa2 /σe2 : 1 MSA ˆ θ(δ) = (1 − δ) −1 :δ ∈R . n MSE (a) Show that MSA and MSE are independent. ˆ (b) Obtain a δ ∈ R such that θ(δ) is unbiased for θ. ˆ (c) Show that the risk of θ(δ) under the squared error loss is a function of (δ, θ). (d) Show that there is a constant δ ∗ such that for any fixed θ, the risk ˆ is strictly decreasing in δ for δ < δ ∗ and strictly increasing for of θ(δ) δ > δ∗. (e) Show that the unbiased estimator of θ derived in (b) is inadmissible. 82. Let T0 (X) be an unbiased estimator of ϑ in an estimation problem. Show that any unbiased estimator of ϑ is of the form T (X) = T0 (X)− U (X), where U (X) is an “unbiased estimator” of 0. 83. Let X be a discrete random variable with P (X = −1) = p,

P (X = k) = (1 − p)2 pk ,

k = 0, 1, 2, ...,

where p ∈ (0, 1) is unknown. (a) Show that U (X) is an unbiased estimator of 0 if and only if U (k) =

154

2. Fundamentals of Statistics

ak for all k = −1, 0, 1, 2, ... and some a. (b) Show that T0 (X) = I{0} (X) is unbiased for ϑ = (1 − p)2 and that, under the squared error loss, T0 is a ℑ-optimal rule, where ℑ is the class of all unbiased estimators of ϑ. (c) Show that T0 (X) = I{−1} (X) is unbiased for ϑ = p and that, under the squared error loss, there is no ℑ-optimal rule, where ℑ is the class of all unbiased estimators of ϑ. 84. (Nonexistence of an unbiased estimator). Let X be a random variable having the binomial distribution Bi(p, n) with an unknown p ∈ (0, 1) and a known n. Consider the problem of estimating ϑ = p−1 . Show that there is no unbiased estimator of ϑ. 85. Let X1 , ..., Xn be i.i.d. random variables having the normal distribution N (θ, 1), where θ = 0 or 1. Consider the estimation of θ. (a) Let ℑ be the class of nonrandomized rules (estimators), i.e., estimators that take values 0 and 1 only. Show that there does not exist any unbiased estimator of θ in ℑ. (b) Find an estimator in ℑ that is approximately unbiased. 86. Let X1 , ..., Xn be i.i.d. from the Poisson distribution P (θ) with an ¯ unknown θ > 0. Find the bias and mse of Tn = (1 − a/n)nX as an −aθ estimator of ϑ = e , where a 6= 0 is a known constant.

87. Let X1 , ..., Xn be i.i.d. (n ≥ 3) from N (µ, σ 2 ), where µ > 0 and σ > 0 ¯ are unknown parameters. Let T1 = X/S be an estimator of µ/σ and 2 2 ¯ ¯ and S 2 are the sample mean T2 = X be an estimator of µ , where X and variance, respectively. Calculate the mse’s of T1 and T2 . 88. Consider a location family {Pµ : µ ∈ Rk } on Rk , where Pµ = P(µ,Ik ) is given in (2.10). Let l0 ∈ Rk be a fixed vector and L(P, a) = L(kµ − ak), where a ∈ A = Rk and L(·) is a nonnegative Borel function on [0, ∞). Show that the family is invariant and the decision problem is invariant under the transformation g(X) = X + cl0 , c ∈ R. Find an invariant decision rule. 89. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with unknown µ ∈ R and σ 2 > 0. Consider the scale transformation aX, a ∈ (0, ∞). (a) For estimating σ 2 under the loss function L(P, a) = (1 − a/σ 2 )2 , show that the problem is invariant and that the sample variance S 2 is invariant. (b) For testing H0 : µ ≤ 0 versus H1 : µ > 0 under the loss |µ| µ I(0,∞) (µ) and L(P, 1) = I(−∞,0] (µ), σ σ show p that the problem is invariant and any test that is a function of ¯ S 2 /n is invariant. X/ L(P, 0) =

2.6. Exercises

155

90. Let X1 , ..., Xn be i.i.d. random variables having the c.d.f. F (x − θ), where F is symmetric aboutP 0 and θ ∈ R is unknown. n (a) Show that the c.d.f. of i=1 wi X(i) − θ is symmetric about 0, where X(i) is the ith Pnorder statistic and wi ’s are constants satisfying wi = wn−i+1 and Pn i=1 wi = 1. (b) Show that i=1 wi X(i) in (a) is unbiased for θ if the mean of F exists. Pn Pn (c) Show that i=1 wi X(i) is location invariant when i=1 wi = 1.

91. In Example 2.25, show that the conditional distribution of θ given X = x is N (µ∗ (x), c2 ) with µ∗ (x) and c2 given by (2.25). 92. A median of a random variable Y (or its distribution) is any value m such that P (Y ≤ m) ≥ 12 and P (Y ≥ m) ≥ 12 . (a) Show that the set of medians is a closed interval [m0 , m1 ]. (b) Suppose that E|Y | < ∞. If c is not a median of Y , show that E|Y − c| ≥ E|Y − m| for any median m of Y . (c) Let X be a sample from Pθ , where θ ∈ Θ ⊂ R. Consider the estimation of θ under the absolute error loss function |a − θ|. Let Π be a given distribution on Θ with finite mean. Find the ℑ-Bayes rule w.r.t. Π, where ℑ is the class of all rules. 93. (Classification). Let X be a sample having a p.d.f. fj (x) w.r.t. a σfinite measure ν, where j is unknown and j ∈ {1, ..., J} with a known integer J ≥ 2. Consider a decision problem in which the action space A = {1, ..., J} and the loss function is 0 if a = j L(j, a) = 1 if a 6= j. (a) Let ℑ be the class of all nonrandomized decision rules. Obtain the risk of a δ ∈ ℑ. (b) Let Π be a probability measure on {1, ..., J} with Π({j}) = πj , j = 1, ..., J. Obtain the Bayes risk of δ ∈ ℑ w.r.t. Π. (c) Obtain a ℑ-Bayes rule w.r.t. Π in (b). (d) Assume that J = 2, π1 = π2 = 0.5, and fj (x) = φ(x − µj ), where φ(x) is the p.d.f. of the standard normal distribution and µj , j = 1, 2, are known constants. Obtain the Bayes rule in (c) and compute the Bayes risk. (e) Obtain the risk and the Bayes risk (w.r.t. Π in (b)) of a randomized decision rule. (f) Obtain a Bayes rule w.r.t. Π. (g) Obtain a minimax rule. 94. Let θˆ be an unbiased estimator of an unknown θ ∈ R. (a) Under the squared error loss, show that the estimator θˆ + c is not

156

2. Fundamentals of Statistics

minimax unless supθ RT (θ) = ∞ for any estimator T , where c 6= 0 is a known constant. (b) Under the squared error loss, show that the estimator cθˆ is not minimax unless supθ RT (θ) = ∞ for any estimator T , where c ∈ (0, 1) is a known constant. (c) Consider the loss function L(θ, a) = (a − θ)2 /θ2 (assuming θ 6= 0). Show that θˆ is not minimax unless supθ RT (θ) = ∞ for any T . 95. Let X be a binary observation with P (X = 1) = θ1 or θ2 , where 0 < θ1 < θ2 < 1 are known values. Consider the estimation of θ with action space {a1 , a2 } and loss function L(θi , aj ) = lij , where l21 ≥ l12 > l11 = l22 = 0. For a decision rule δ(X), the vector (Rδ (θ1 ), Rδ (θ2 )) is defined to be its risk point. (a) Show that the set of risk points of all decision rules is the convex hull of the set of risk points of all nonrandomized rules. (b) Find a minimax rule. (c) Let Π be a distribution on {θ1 , θ2 }. Obtain the class of all Bayes rules w.r.t. Π. Discuss when there is a unique Bayes rule. 96. Consider the decision problem in Example 2.23. (a) Let Π be the uniform distribution on (0, 1). Show that a ℑ-Bayes rule w.r.t. Π is Tj ∗ (X), where j ∗ is the largest integer in {0, 1, ..., n−1} such that Bj+1,n−j+1 (θ0 ) ≥ 21 and Ba,b (·) denotes the c.d.f. of the beta distribution B(a, b). (b) Derive a ℑ-minimax rule. 97. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with unknown µ ∈ R and σ 2 > 0. To test the hypotheses H0 : µ ≤ µ0

versus

H1 : µ > µ0 ,

where µ0 is a fixed constant, consider p a test of the form Tc (X) = ¯ − µ0 )/ S 2 /n and c is a fixed constant. I(c,∞) (Tµ0 ), where Tµ0 = (X (a) Find the size of Tc . (Hint: Tµ0 has the t-distribution tn−1 .) (b) If α is a given level of significance, find a cα such that Tcα has size α. (c) Compute the p-value for Tcp α derived in (b). p ¯ − cα S 2 /n, X ¯ + cα S 2 /n] is a confidence (d) Find a cα such that [X interval for µ with confidence coefficient 1 − α. What is the expected interval length? 98. In Exercise 67, calculate the size of Tc (X); find a cα such that Tcα has size α, a given level of significance; and find the p-value for Tcα . 99. In Exercise 68, assume that σ is known. Calculate the size of Tc (X); find a cα such that Tcα has size α, a given level of significance; and find the p-value for Tcα .

2.6. Exercises

157

100. Let α ∈ (0, 1) be given and Tj,q (X) be the test given in Example 2.30. Show that there exist integer j and q ∈ (0, 1) such that the size of Tj,q is α. 101. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(a, θ) with unknown a ∈ R andPθ > 0. Let α ∈ (0, 1) be given. (a) Using T1 (X) = ni=1 (Xi − X(1) ), construct a confidence interval for θ with confidence coefficient 1 − α and find the expected interval length. (b) Using T1 (X) and T2 (X) = X(1) , construct a confidence interval for a with confidence coefficient 1 − α and find the expected interval length. (c) Using the method in Example 2.32, construct a confidence set for the two-dimensional parameter (a, θ) with confidence coefficient 1−α. 102. Suppose that X is a sample and a statistic T (X) has a distribution in a location family {Pµ : µ ∈ R}. Using T (X), derive a confidence interval for µ with level of significance 1 − α and obtain the expected interval length. Show that if the c.d.f. of T (X) is continuous, then we can always find a confidence interval for µ with confidence coefficient 1 − α for any α ∈ (0, 1). 103. Let X = (X1 , ..., Xn ) be a sample from Pθ , where θ ∈ {θ1 , ..., θk } with a fixed integer k. Let Tn (X) be an estimator of θ with range {θ1 , ..., θk }. (a) Show that Tn (X) is consistent if and only if Pθ (Tn (X) = θ) → 1. (b) Show that if Tn (X) is consistent, then it is an -consistent for any {an }. 104. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (θ − 21 , θ + 12 ), where θ ∈ R is unknown. Show that (X(1) + X(n) )/2 is strongly consistent for θ and also consistent in mse. 105. Let X1 , ..., Xn be i.i.d. from a population with the Lebesgue p.d.f. fθ (x) = 2−1 (1 + θx)I(−1,1) (x), where θ ∈ (−1, 1) is an unknown√parameter. Find a consistent estimator of θ. Is your estimator nconsistent? 106. Let X1 , ..., Xn be i.i.d. observations. Suppose that Tn is an unbiased estimator of ϑ based on X1 , ..., Xn such that for any n, Var(Tn ) < ∞ and Var(Tn ) ≤ Var(Un ) for any other unbiased estimator Un of ϑ based on X1 , ..., Xn . Show that Tn is consistent in mse. 107. Consider the Bayes rule √ µ∗ (X) in Example 2.25. Show that µ∗ (X) is a strongly consistent, n-consistent, and L2 -consistent estimator of µ. What is the order of the bias of µ∗ (X) as an estimator of µ?

158

2. Fundamentals of Statistics

108. In Exercise 21, show that ¯ is an inconsistent estimator of β; (a) Y¯ /X ˆ (b) β = Z(m) is a consistent estimator of β, where m = n/2 when n is even, m = (n + 1)/2 when n is odd, and Z(i) is the ith smallest value of Yi /Xi , i = 1, ..., n. 109. Show that the estimator T0 of θ in Exercise 64 is inconsistent. 110. Let g1 , g2 ,... be continuous functions on (a, b) ⊂ R such that gn (x) → g(x) uniformly for x in any closed subinterval of (a, b). Let Tn be a consistent estimator of θ ∈ (a, b). Show that gn (Tn ) is consistent for ϑ = g(θ). 111. Let X1 , ..., Xn be i.i.d. from P with unknown mean µ ∈ R and variance σ 2 > 0, and let g(µ) = 0 if µ 6= 0 and g(0) = 1. Find a consistent estimator of ϑ = g(µ). 112. Establish results for the smallest order statistic X(1) (based on i.i.d. random variables X1 , ..., Xn ) similar to those in Example 2.34. 113. (Consistency for finite population). In Example 2.27, show that Yˆ →p Y as n → N for any fixed N and population. Is Yˆ still consistent if sampling is with replacement? 114. Assume that Xi = θti + ei , i = 1, ..., n, where θ ∈ Θ is an unknown parameter, Θ is a closed subset of R, ei ’s are i.i.d. on the interval [−τ, τ ] with some unknown τ > 0 and Eei = 0, and ti ’s are fixed constants. Let Tn = Sn (θ˜n ) = min Sn (γ), γ∈Θ

where Sn (γ) = 2 max |Xi − γti |/ i≤n

p 1 + γ 2.

(a) Assume that supi |ti | < ∞ and supi ti − inf i ti > 2τ . Show that the sequence {θ˜n , n = 1, 2, ...} is bounded a.s. (b) Let θn ∈ Θ, n = 1, 2, .... If θn → θ, show that Sn (θn ) − Sn (θ) = O(|θn − θ|) a.s. (c) Under the conditions in (a), show that Tn is a strongly consistent estimator of ϑ = minγ∈Θ S(γ), where S(γ) = limn→∞ Sn (γ) a.s. ¯ be 115. Let X1 , ..., Xn be i.i.d. random variables with EX12 < ∞ and X the sample mean. Consider the estimation of µ = EX1 . ¯ + ξn /√n, where ξn is a random variable satisfying (a) Let Tn = X ξn = 0 with probability 1 − n−1 and ξn = n3/2 with probability n−1 .

2.6. Exercises

159

(P ) for any P . Show that bTn (P ) 6= ˜bTn√ ¯ + ηn / n, where ηn is a random variable that is (b) Let Tn = X independent of X1 , ..., Xn and equals 0 with probability 1 − 2n−1 and √ ± n with probability n−1 . Show that amseTn (P ) = amseX¯ (P ) = mseX¯ (P ) and mseTn (P ) > amseTn (P ) for any P . 116. Let X1 , ..., Xn be i.i.d. random variables with finite θ = EX1 and Var(X θ > 0 is unknown. Consider the estimation of √1 ) = θ, where √ ¯ ¯ and S 2 are the ¯ and T2n = X/S, where X ϑ = θ. Let T1n = X sample mean and sample variance. (a) Obtain the n−1 order asymptotic biases of T1n and T2n according to (2.38). (b) Obtain the asymptotic relative efficiency of T1n w.r.t. T2n . 117. Let X1 , ..., Xn be i.i.d. according to N (µ, 1) with an unknown µ ∈ R. Let ϑ = P (X1 ≤ c) for a fixed constant c. Consider the following estimators of ϑ: T1n = Fn (c), where Fn is the empirical c.d.f. defined ¯ where Φ is the c.d.f. of N (0, 1). in (2.28), and T2n = Φ(c − X), −1 (a) Find the n order asymptotic bias of T2n according to (2.38). (b) Find the asymptotic relative efficiency of T1n w.r.t. T2n . 118. Let X1 , ..., Xn be i.i.d. from the N (0, σ 2 ) distribution with an unknown σ > 0. Consider estimation of ϑ = σ.PFind the asymptotic p theP relative efficiency of π/2 ni=1 |Xi |/n w.r.t. ( ni=1 Xi2 /n)1/2 .

119. Let X1 , ..., Xn be i.i.d. from P with EX14 < ∞ and unknown mean µ ∈ R and variance σ 2 > 0. Consider the estimation of ϑ = µ2 and ¯ 2 , T2n = X ¯ 2 − S 2 /n, T3n = the following three estimators: T1n = X 2 ¯ max{0, T2n }, where X and S are the sample mean and variance. Show that the amse’s of Tjn , j = 1, 2, 3, are the same when µ 6= 0 but may be different when µ = 0. Which estimator is the best in terms of the asymptotic relative efficiency when µ = 0? 120. Prove Theorem 2.6. 121. Let X1 , ..., Xn be EXi = µ, Var(Xi ) = 1, and EXi4 < ∞. Pni.i.d. with −1 2 −1 ¯2 Let T1n = n be estimators of i=1 Xi − 1 and T2n = X − n 2 ϑ=µ . (a) Find the asymptotic relative efficiency of T1n w.r.t. T2n . (b) Show that eT1n ,T2n (P ) ≤ 1 if the c.d.f. of Xi − µ is symmetric about 0 and µ 6= 0. (c) Find a distribution P for which eT1n ,T2n (P ) > 1.

122. Let X1 , ..., Xn be i.i.d. binary random variables with unknown p = P (Xi = 1) ∈ (0, 1). Consider the estimation of p. Let a and b be two positive constants. Find the asymptotic relative efficiency of the ¯ ¯ estimator (a + nX)/(a + b + n) w.r.t. X.

160

2. Fundamentals of Statistics

123. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with an unknown µ ∈ R and a ¯ be the sample mean and T2 = µ∗ (X) be the known σ 2 . Let T1 = X Bayes estimator given in (2.25). Assume that EX14 < ∞. (a) Calculate the exact mse of both estimators. Can you conclude that one estimator is better than the other in terms of the mse? (b) Find the asymptotic relative efficiency of T1 w.r.t. T2 . 124. In Example 2.37, show that (a) the limiting size of Tcα is 1 if P contains all possible populations on R with finite second moments; (b) Tn = Tcα with α = αn (given by (2.40)) is Chernoff-consistent; (c) Tn in (b) is not strongly Chernoff-consistent if P contains all possible populations on R with finite second moments. 125. Let X1 , ..., Xn be i.i.d. with unknown mean µ ∈ R and variance σ 2 > 0. For testing H0 : µ ≤ µ0 versus H1 : µ > µ0 , consider the test Tcα obtained in Exercise 97(b). (a) Show that Tcα has asymptotic significance level α and is consistent. (b) Find a test that is Chernoff-consistent. 126. Consider the test Tj in Example 2.23. For each n, find a j = jn such that Tjn has asymptotic significance level α ∈ (0, 1). 127. Show that the test Tcα in Exercise 98 is consistent, but Tcα in Exercise 99 is not consistent. 128. In Example 2.31, suppose that we drop the normality assumption but assume that µ = EXi and σ 2 = Var(Xi ) are finite. (a) Show that when σ 2 is known, the asymptotic significance level ¯ − cα , X ¯ + cα ] is 1 − α, where cα = of the confidence interval [X √ −1 σz1−α/2 / n and za = Φ (a). (b) Show that when σ 2 is known, the limiting confidence coefficient of the interval in (a) might be 0 if P contains all possible populations on R. (c) Show that the confidence interval in Exercise 97(d) has asymptotic significance level 1 − α. 129. Let X1 , ..., Xn be i.i.d. with unknown mean µ ∈ R and variance σ 2 > 0. Assume that EX14 < ∞. Using the sample variance S 2 , construct a confidence interval for σ 2 that has asymptotic significance level 1 − α. 130. Consider the sample correlation coefficient T defined in Exercise 22. Construct a confidence interval for ρ that has asymptotic significance level 1 − α, assuming that (Yi , Zi ) is normally distributed. (Hint: show that the asymptotic variance of T is (1 − ρ2 )2 .)

Chapter 3

Unbiased Estimation Unbiased or asymptotically unbiased estimation plays an important role in point estimation theory. Unbiasedness of point estimators is defined in §2.3.2. In this chapter, we discuss in detail how to derive unbiased estimators and, more importantly, how to find the best unbiased estimators in various situations. Although an unbiased estimator (even the best unbiased estimator if it exists) is not necessarily better than a slightly biased estimator in terms of their mse’s (see Exercise 63 in §2.6), unbiased estimators can be used as “building blocks” for the construction of better estimators. Furthermore, one may give up the exact unbiasedness, but cannot give up asymptotic unbiasedness since it is necessary for consistency (see §2.5.2). Properties and the construction of asymptotically unbiased estimators are studied in the last part of this chapter.

3.1 The UMVUE Let X be a sample from an unknown population P ∈ P and ϑ be a realvalued parameter related to P . Recall that an estimator T (X) of ϑ is unbiased if and only if E[T (X)] = ϑ for any P ∈ P. If there exists an unbiased estimator of ϑ, then ϑ is called an estimable parameter. Definition 3.1. An unbiased estimator T (X) of ϑ is called the uniformly minimum variance unbiased estimator (UMVUE) if and only if Var(T (X)) ≤ Var(U (X)) for any P ∈ P and any other unbiased estimator U (X) of ϑ. Since the mse of any unbiased estimator is its variance, a UMVUE is ℑ-optimal in mse with ℑ being the class of all unbiased estimators. One 161

162

3. Unbiased Estimation

can similarly define the uniformly minimum risk unbiased estimator in statistical decision theory when we use an arbitrary loss instead of the squared error loss that corresponds to the mse.

3.1.1 Sufficient and complete statistics The derivation of a UMVUE is relatively simple if there exists a sufficient and complete statistic for P ∈ P. Theorem 3.1 (Lehmann-Scheff´e theorem). Suppose that there exists a sufficient and complete statistic T (X) for P ∈ P. If ϑ is estimable, then there is a unique unbiased estimator of ϑ that is of the form h(T ) with a Borel function h. (Two estimators that are equal a.s. P are treated as one estimator.) Furthermore, h(T ) is the unique UMVUE of ϑ. This theorem is a consequence of Theorem 2.5(ii) (Rao-Blackwell theorem). One can easily extend this theorem to the case of the uniformly minimum risk unbiased estimator under any loss function L(P, a) that is strictly convex in a. The uniqueness of the UMVUE follows from the completeness of T (X). There are two typical ways to derive a UMVUE when a sufficient and complete statistic T is available. The first one is solving for h when the distribution of T is available. The following are two typical examples. Example 3.1. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (0, θ), θ > 0. Let ϑ = g(θ), where g is a differentiable function on (0, ∞). Since the sufficient and complete statistic X(n) has the Lebesgue p.d.f. nθ−n xn−1 I(0,θ) (x), an unbiased estimator h(X(n) ) of ϑ must satisfy θn g(θ) = n

Z

θ

h(x)xn−1 dx

for all θ > 0.

0

Differentiating both sizes of the previous equation and applying the result of differentiation of an integral (Royden (1968, §5.3)) lead to nθn−1 g(θ) + θn g ′ (θ) = nh(θ)θn−1 . Hence, the UMVUE of ϑ is h(X(n) ) = g(X(n) ) + n−1 X(n) g ′ (X(n) ). In particular, if ϑ = θ, then the UMVUE of θ is (1 + n−1 )X(n) . Example 3.2. Let X1 , ..., Xn be i.i.d. from Pn the Poisson distribution P (θ) with an unknown θ > 0. Then T (X) = i=1 Xi is sufficient and complete for θ > 0 and has the Poisson distribution P (nθ). P Suppose that ϑ = g(θ), ∞ j where g is a smooth function such that g(x) = j=0 aj x , x > 0. An

163

3.1. The UMVUE

unbiased estimator h(T ) of ϑ must satisfy ∞ X h(t)nt t=0

t!

θt = enθ g(θ) ∞ X nk

θk

∞ X

aj θ j k! j=0 k=0 ∞ X X n k aj θt = k! t=0 =

j,k:j+k=t

for any θ > 0. Thus, a comparison of coefficients in front of θt leads to h(t) =

t! nt

X

j,k:j+k=t

n k aj , k!

i.e., h(T ) is the UMVUE of ϑ. In particular, if ϑ = θr for some fixed integer r ≥ 1, then ar = 1 and ak = 0 if k 6= r and ( 0 t 0 and I(t,∞) (X1 ) is unbiased for ϑ, ¯ = P (X1 > t|X) ¯ T (X) = E[I(t,∞) (X1 )|X] ¯ is availis the UMVUE of ϑ. If the conditional distribution of X1 given X ¯ directly. But the following techable, then we can calculate P (X1 > t|X) nique can be applied to avoid the derivation of conditional distributions. ¯ and X ¯ are independent. By By Basu’s theorem (Theorem 2.4), X1 /X Proposition 1.10(vii), ¯ =x ¯ > t/X| ¯X ¯ =x ¯ > t/¯ P (X1 > t|X ¯) = P (X1 /X ¯) = P (X1 /X x).

164

3. Unbiased Estimation

To compute this unconditional probability, we need the distribution of ! X n n X X1 + X1 Xi = X1 Xi . i=1

i=2

Using Pn the transformation technique discussed in §1.3.1 and the fact that is independent of X1 and has a gamma distribution, we obtain i=2 Xi P n that X1 / i=1 Xi has the Lebesgue p.d.f. (n − 1)(1 − x)n−2 I(0,1) (x). Hence n−1 Z 1 t ¯ =x P (X1 > t|X ¯) = (n − 1) (1 − x)n−2 dx = 1 − n¯ x t/(n¯ x) and the UMVUE of ϑ is n−1 t T (X) = 1 − ¯ . nX We now show more examples of applying these two methods to find UMVUE’s. Example 3.4. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ ∈ R ¯ S 2 ) is sufficient and comand σ 2 > 0. From Example 2.18, T = (X, 2 2 ¯ plete for θ = (µ, σ ) and X and (n − 1)S /σ 2 are independent and have the N (µ, σ 2 /n) and chi-square distribution χ2n−1 , respectively. Using the ¯ the method of solving for h directly, we find that the UMVUE for µ is X; 2 2 2 r ¯ UMVUE of µ is X − S /n; the UMVUE for σ with r > 1 − n is kn−1,r S r , where nr/2 Γ(n/2) kn,r = r/2 n+r 2 Γ 2 ¯ (exercise); and the UMVUE of µ/σ is kn−1,−1 X/S, if n > 2. Suppose that ϑ satisfies P (X1 ≤ ϑ) = p with a fixed p ∈ (0, 1). Let Φ be the c.d.f. of the standard normal distribution. Then ϑ = µ + σΦ−1 (p) ¯ + kn−1,1 SΦ−1 (p). and its UMVUE is X Let c be a fixed constant and ϑ = P (X1 ≤ c) = Φ c−µ . We can σ find the UMVUE of ϑ using the method of conditioning and the technique used in Example 3.3. Since I(−∞,c) (X1 ) is an unbiased estimator of ϑ, the UMVUE of ϑ is E[I(−∞,c) (X1 )|T ] = P (X1 ≤ c|T ). By Basu’s theorem, ¯ ¯ S 2 ). the ancillary statistic Z(X) = (X1 − X)/S is independent of T = (X, Then, by Proposition 1.10(vii), ¯ c−X 2 2 P X1 ≤ c|T = (¯ T = (¯ x, s ) x, s ) = P Z ≤ S c−x ¯ . =P Z≤ s

165

3.1. The UMVUE

It can be shown that Z has the Lebesgue p.d.f. √ (n/2)−2 nΓ n−1 nz 2 2 f (z) = √ 1− I(0,(n−1)/√n) (|z|) (3.1) (n − 1)2 π(n − 1)Γ n−2 2

(exercise). Hence the UMVUE of ϑ is Z (c−X)/S ¯ P (X1 ≤ c|T ) = f (z)dz √

(3.2)

−(n−1)/ n

with f given by (3.1). Suppose that we would like to estimate ϑ = σ1 Φ′ c−µ , the Lebesgue σ p.d.f. of X1 evaluated at a fixed c, where Φ′ is the first-order derivative ¯ = x¯ and S 2 = s2 is of Φ. By (3.2), the conditional p.d.f. of X1 given X x−¯ x −1 2 ¯ s f s . Let fT be the joint p.d.f. of T = (X, S ). Then Z Z ¯ c − x¯ c−X 1 1 f fT (t)dt = E f . ϑ= s s S S Hence the UMVUE of ϑ is 1 f S

¯ c−X . S

Example 3.5. Let X1 , ..., Xn be i.i.d. from a power series distribution (see Exercise 13 in §2.6), i.e., P (Xi = x) = γ(x)θx /c(θ),

x = 0, 1, 2, ...,

with a known function γ(x) ≥ 0 and an unknown parameter θ > 0. It turns out that the joint distribution of X = (X1 , ..., Xn ) isP in an exponential famn ily with a sufficient and complete statistic T (X) = i=1 Xi . Furthermore, the distribution of T is also in a power series family, i.e., P (T = t) = γn (t)θt /[c(θ)]n ,

t = 0, 1, 2, ...,

where γn (t) is the coefficient of θt in the power series expansion of [c(θ)]n (Exercise 13 in §2.6). This result can help us to find the UMVUE of ϑ = g(θ). For example, by comparing both sides of ∞ X

h(t)γn (t)θt = [c(θ)]n−p θr ,

t=0

we conclude that the UMVUE of θr /[c(θ)]p is ( 0 T 1,

168

3. Unbiased Estimation

where c and b are some constants. From the previous discussion, E[h(X(n) )U (X(n) )] = 0,

θ ≥ 1.

Since E[h(X(n) )] = θ, we obtain that θ = cP (X(n) ≤ 1) + bE[X(n) I(1,∞) (X(n) )] = cθ−n + [bn/(n + 1)](θ − θ−n ).

Thus, c = 1 and b = (n + 1)/n. The UMVUE of θ is then 1 0 ≤ X(n) ≤ 1 T = X(n) > 1. (1 + n−1 )X(n) This estimator is better than (1 + n−1 )X(n) , which is the UMVUE when Θ = (0, ∞) and does not make use of the information about θ ≥ 1. Example 3.8. Let X be a sample (of size 1) from the uniform distribution U (θ − 12 , θ + 12 ), θ ∈ R. We now apply Theorem 3.2 to show that there is no UMVUE of ϑ = g(θ) for any nonconstant function g. Note that an unbiased estimator U (X) of 0 must satisfy Z θ+ 12 U (x)dx = 0 for all θ ∈ R. θ− 12

Differentiating both sizes of the previous equation and applying the result of differentiation of an integral lead to U (x) = U (x + 1) a.e. m, where m is the Lebesgue measure on R. If T is a UMVUE of g(θ), then T (X)U (X) is unbiased for 0 and, hence, T (x)U (x) = T (x+1)U (x+1) a.e. m, where U (X) is any unbiased estimator of 0. Since this is true for all U , T (x) = T (x + 1) a.e. m. Since T is unbiased for g(θ), Z θ+ 21 g(θ) = T (x)dx for all θ ∈ R. θ− 12

Differentiating both sizes of the previous equation and applying the result of differentiation of an integral, we obtain that g ′ (θ) = T θ + 12 − T θ − 12 = 0 a.e. m. As a consequence of Theorem 3.2, we have the following useful result. Corollary 3.1. (i) Let Tj be a UMVUE of ϑj , j = 1, ..., k, where k is a Pk Pk fixed positive integer. Then j=1 cj Tj is a UMVUE of ϑ = j=1 cj ϑj for any constants c1 , ..., ck . (ii) Let T1 and T2 be two UMVUE’s of ϑ. Then T1 = T2 a.s. P for any P ∈ P.

169

3.1. The UMVUE

3.1.3 Information inequality Suppose that we have a lower bound for the variances of all unbiased estimators of ϑ and that there is an unbiased estimator T of ϑ whose variance is always the same as the lower bound. Then T is a UMVUE of ϑ. Although this is not an effective way to find UMVUE’s (compared with the methods introduced in §3.1.1 and §3.1.2), it provides a way of assessing the performance of UMVUE’s. The following result provides such a lower bound in some cases. Theorem 3.3 (Cram´er-Rao lower bound). Let X = (X1 , ..., Xn ) be a sample from P ∈ P = {Pθ : θ ∈ Θ}, where Θ is an open set in Rk . Suppose that T (X) is an estimator with E[T (X)] = g(θ) being a differentiable function of θ; Pθ has a p.d.f. fθ w.r.t. a measure ν for all θ ∈ Θ; and fθ is differentiable as a function of θ and satisfies Z Z ∂ ∂ h(x)fθ (x)dν = h(x) fθ (x)dν, θ ∈ Θ, (3.3) ∂θ ∂θ for h(x) ≡ 1 and h(x) = T (x). Then ∂ τ ∂ Var(T (X)) ≥ ∂θ g(θ) [I(θ)]−1 ∂θ g(θ), where

I(θ) = E

τ ∂ ∂ log fθ (X) log fθ (X) ∂θ ∂θ

(3.4)

(3.5)

is assumed to be positive definite for any θ ∈ Θ. Proof. We prove the univariate case (k = 1) only. The proof for the multivariate case (k > 1) is left to the reader. When k = 1, (3.4) reduces to [g ′ (θ)]2 Var(T (X)) ≥ (3.6) 2 . ∂ E ∂θ log fθ (X)

From inequality (1.37), we only need to show that

2 ∂ ∂ E log fθ (X) = Var log fθ (X) ∂θ ∂θ and

∂ log fθ (X) . g (θ) = Cov T (X), ∂θ ′

These two results are consequences of condition (3.3). The k × k matrix I(θ) in (3.5) is called the Fisher information matrix. The greater I(θ) is, the easier it is to distinguish θ from neighboring values

170

3. Unbiased Estimation

and, therefore, the more accurately θ can be estimated. In fact, if the equality in (3.6) holds for an unbiased estimator T (X) of g(θ) (which is then a UMVUE), then the greater I(θ) is, the smaller Var(T (X)) is. Thus, I(θ) is a measure of the information that X contains about the unknown θ. The inequalities in (3.4) and (3.6) are called information inequalities. The following result is helpful in finding the Fisher information matrix. Proposition 3.1. (i) Let X and Y be independent with the Fisher information matrices IX (θ) and IY (θ), respectively. Then, the Fisher information about θ contained in (X, Y ) is IX (θ) + IY (θ). In particular, if X1 , ..., Xn are i.i.d. and I1 (θ) is the Fisher information about θ contained in a single Xi , then the Fisher information about θ contained in X1 , ..., Xn is nI1 (θ). (ii) Suppose that X has the p.d.f. fθ that is twice differentiable in θ and that (3.3) holds with h(x) ≡ 1 and fθ replaced by ∂fθ /∂θ. Then ∂2 log fθ (X) . (3.7) I(θ) = −E ∂θ∂θτ Proof. Result (i) follows from the independence of X and Y and the definition of the Fisher information. Result (ii) follows from the equality τ ∂2 ∂2 ∂ ∂ ∂θ∂θ τ fθ (X) − log f log f log f (X) = (X) (X) . θ θ θ ∂θ∂θτ fθ (X) ∂θ ∂θ The following example provides a formula for the Fisher information matrix for many parametric families with a two-dimensional parameter θ. Example 3.9. Let X1 , ..., Xn be i.i.d. with the Lebesgue p.d.f. σ1 f x−µ , σ where f (x) > 0 and f ′ (x) exists for all x ∈ R, µ ∈ R, and σ > 0 (a location-scale family). Let θ = (µ, σ). Then, the Fisher information about θ contained in X1 , ..., Xn is (exercise) R [f ′ (x)]2 R f ′ (x)[xf ′ (x)+f (x)] dx f (x) dx f (x) n I(θ) = 2 . σ R [xf ′ (x)+f (x)]2 R f ′ (x)[xf ′ (x)+f (x)] dx dx f (x) f (x) Note that I(θ) depends on the particular parameterization. If θ = ψ(η) and ψ is differentiable, then the Fisher information that X contains about η is h i ∂ ∂η ψ(η)I(ψ(η))

∂ ∂η ψ(η)

τ

.

However, it is easy to see that the Cram´er-Rao lower bound in (3.4) or (3.6) is not affected by any one-to-one reparameterization.

171

3.1. The UMVUE

If we use inequality (3.4) or (3.6) to find a UMVUE T (X), then we obtain a formula for Var(T (X)) at the same time. On the other hand, the Cram´er-Rao lower bound in (3.4) or (3.6) is typically not sharp. Under some regularity conditions, the Cram´er-Rao lower bound is attained if and only if fθ is in an exponential family; see Propositions 3.2 and 3.3 and the discussion in Lehmann (1983, p. 123). Some improved information inequalities are available (see, e.g., Lehmann (1983, Sections 2.6 and 2.7)). Proposition 3.2. Suppose that the distribution of X is from an exponential family {fθ : θ ∈ Θ}, i.e., the p.d.f. of X w.r.t. a σ-finite measure is fθ (x) = exp [η(θ)]τ T (x) − ξ(θ) c(x) (3.8)

(see §2.1.3), where Θ is an open subset of Rk . (i) The regularity condition (3.3) is satisfied for any h with E|h(X)| < ∞ and (3.7) holds. (ii) If I(η) is the Fisher information matrix for the natural parameter η, then the variance-covariance matrix Var(T ) = I(η). (iii) If I(ϑ) is the Fisher information matrix for the parameter ϑ = E[T (X)], then Var(T ) = [I(ϑ)]−1 . Proof. (i) This is a direct consequence of Theorem 2.1. (ii) From (2.6), the p.d.f. under the natural parameter η is fη (x) = exp {η τ T (x) − ζ(η)} c(x). From Theorem 2.1 and result (1.54) in §1.3.3, E[T (X)] = result follows from ∂ ∂η

log fη (x) = T (x) −

(iii) Since ϑ = E[T (X)] = I(η) =

∂ϑ ∂η I(ϑ)

∂ ∂η ζ(η).

The

∂ ∂η ζ(η).

∂ ∂η ζ(η), ∂ϑ ∂η

τ

=

∂2 ∂η∂η τ

ζ(η)I(ϑ)

h

∂2 ∂η∂η τ

By Theorem 2.1, result (1.54), and the result in (ii), I(η). Hence

ζ(η)

∂2 ∂η∂η τ

iτ

.

ζ(η) = Var(T ) =

I(ϑ) = [I(η)]−1 I(η)[I(η)]−1 = [I(η)]−1 = [Var(T )]−1 . A direct consequence of Proposition 3.2(ii) is that the variance of any linear function of T in (3.8) attains the Cram´er-Rao lower bound. The following result gives a necessary condition for Var(U (X)) of an estimator U (X) to attain the Cram´er-Rao lower bound.

172

3. Unbiased Estimation

Proposition 3.3. Assume that the conditions in Theorem 3.3 hold with T (X) replaced by U (X) and that Θ ⊂ R. (i) If Var(U (X)) attains the Cram´er-Rao lower bound in (3.6), then a(θ)[U (X) − g(θ)] = g ′ (θ)

∂ log fθ (X) a.s. Pθ ∂θ

for some function a(θ), θ ∈ Θ. (ii) Let fθ and T be given by (3.8). If Var(U (X)) attains the Cram´er-Rao lower bound, then U (X) is a linear function of T (X) a.s. Pθ , θ ∈ Θ. Example 3.10. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 . Let fµ be the joint distribution of X = (X1 , ..., Xn ). Then ∂ ∂µ

log fµ (X) =

n X i=1

(Xi − µ)/σ 2 .

¯ attains the Cram´er-Rao lower Thus, I(µ) = n/σ 2 . It is obvious that Var(X) ¯2 = bound in (3.6). Consider now the estimation of ϑ = µ2 . Since E X 2 2 2 2 ¯ ¯ µ + σ /n, the UMVUE of ϑ is h(X) = X − σ /n. A straightforward calculation shows that ¯ = Var(h(X))

4µ2 σ 2 2σ 4 + 2 . n n

On the other hand, the Cram´er-Rao lower bound in this case is 4µ2 σ 2 /n. ¯ does not attain the Cram´er-Rao lower bound. The difHence Var(h(X)) ference is 2σ 4 /n2 . Condition (3.3) is a key regularity condition for the results in Theorem 3.3 and Proposition 3.3. If fθ is not in an exponential family, then (3.3) has to be checked. Typically, it does not hold if the set {x : fθ (x) > 0} depends on θ (Exercise 37). More discussions can be found in Pitman (1979).

3.1.4 Asymptotic properties of UMVUE’s UMVUE’s are typically consistent (see Exercise 106 in §2.6). If there is an unbiased estimator of ϑ whose mse is of the order a−2 n , where {an } is a sequence of positive numbers diverging to ∞, then the UMVUE of ϑ (if it exists) has an mse of order a−2 n and is an -consistent. For instance, in Example 3.3, the mse of U (X) = 1 − Fn (t) is Fθ (t)[1 − Fθ (t)]/n; hence the √ UMVUE T (X) is n-consistent and its mse is of the order n−1 . UMVUE’s are exactly unbiased so that there is no need to discuss their asymptotic biases. Their variances (or mse’s) are finite, but amse’s can be

173

3.1. The UMVUE

used to assess their performance if the exact forms of mse’s are difficult to obtain. In many cases, although the variance of a UMVUE Tn does not attain the Cram´er-Rao lower bound, the limit of the ratio of the amse (or mse) of Tn over the Cram´er-Rao lower bound (if it is not 0) is 1. For instance, in Example 3.10, ¯ 2 − σ 2 /n) σ2 Var(X =1+ 2 →1 the Cram´er-Rao lower bound 2µ n if µ 6= 0. In general, under the conditions in Theorem 3.3, if Tn (X) is unbiased for g(θ) and if, for any θ ∈ Θ, ∂ τ ∂ Tn (X) − g(θ) = ∂θ g(θ) [I(θ)]−1 ∂θ log fθ (X) [1 + op (1)] a.s. Pθ , (3.9) then

amseTn (θ) = the Cram´er-Rao lower bound

(3.10)

whenever the Cram´er-Rao lower bound is not 0. Note that the case of zero Cram´er-Rao lower bound is not of interest since a zero lower bound does not provide any information on the performance of estimators. n−1 Consider the UMVUE Tn = 1 − ntX¯ of e−t/θ in Example 3.3. Using the fact that log(1 − x) = − we obtain that

∞ X xj j=1

j

|x| ≤ 1,

,

¯ Tn − e−t/X = Op n−1 .

Using Taylor’s expansion, we obtain that ¯ ¯ − θ)[1 + op (1)], e−t/X − e−t/θ = g ′ (θ)(X where g(θ) = e−t/θ . On the other hand,

∂ ¯ − θ. [I(θ)]−1 ∂θ log fθ (X) = X

Hence (3.9) and (3.10) hold. Note that the exact variance of Tn is not easy to obtain. In this example, it can be shown that {n[Tn − g(θ)]2 } is uniformly integrable and, therefore, lim nVar(Tn ) = lim n[amseTn (θ)]

n→∞

n→∞

= lim n[g ′ (θ)]2 [I(θ)]−1 =

n→∞ 2 −2t/θ

t e θ2

.

It is shown in Chapter 4 that if (3.10) holds, then Tn is asymptotically optimal in some sense. Hence UMVUE’s satisfying (3.9), which is often true, are asymptotically optimal, although they may be improved in terms of the exact mse’s.

174

3. Unbiased Estimation

3.2 U-Statistics Let X1 , ..., Xn be i.i.d. from an unknown population P in a nonparametric family P. In Example 3.6 we argued that if the vector of order statistic is sufficient and complete for P ∈ P, then a symmetric unbiased estimator of any estimable ϑ is the UMVUE of ϑ. In a large class of problems, parameters to be estimated are of the form ϑ = E[h(X1 , ..., Xm )] with a positive integer m and a Borel function h that is symmetric and satisfies E|h(X1 , ..., Xm )| < ∞ for any P ∈ P. It is easy to see that a symmetric unbiased estimator of ϑ is Un =

−1 X n h(Xi1 , ..., Xim ), m c

P where c denotes the summation over the elements {i1 , ..., im } from {1, ..., n}.

n m

(3.11)

combinations of m distinct

Definition 3.2. The statistic Un in (3.11) is called a U -statistic with kernel h of order m.

3.2.1 Some examples The use of U-statistics is an effective way of obtaining unbiased estimators. In nonparametric problems, U-statistics are often UMVUE’s, whereas in parametric problems, U-statistics can be used as initial estimators to derive more efficient estimators. If m = 1, Un in (3.11) is simply a type of sample mean. Examples include the empirical c.d.f. (2.28) evaluated at a particular t and the sample Pn moments n−1 i=1 Xik for a positive integer k. We now consider some examples with m > 1. Consider the estimation of ϑ = µm , where µ = EX1 and m is a positive integer. Using h(x1 , ..., xm ) = x1 · · · xm , we obtain the following U-statistic unbiased for ϑ = µm : −1 X n Xi1 · · · Xim . Un = m c Consider next the estimation of ϑ = σ 2 = Var(X1 ). Since σ 2 = [Var(X1 ) + Var(X2 )]/2 = E[(X1 − X2 )2 /2],

(3.12)

175

3.2. U-Statistics

we obtain the following U-statistic with kernel h(x1 , x2 ) = (x1 − x2 )2 /2: ! n X X (Xi − Xj )2 1 2 2 2 ¯ = = S2, Un = X − nX n(n − 1) 2 n − 1 i=1 i 1≤i<j≤n

which is the sample variance in (2.2). In some cases, we would like to estimate ϑ = E|X1 − X2 |, a measure of concentration. Using kernel h(x1 , x2 ) = |x1 − x2 |, we obtain the following U-statistic unbiased for ϑ = E|X1 − X2 |: Un =

2 n(n − 1)

X

1≤i<j≤n

|Xi − Xj |,

which is known as Gini’s mean difference. Let ϑ = P (X1 + X2 ≤ 0). Using kernel h(x1 , x2 ) = I(−∞,0] (x1 + x2 ), we obtain the following U-statistic unbiased for ϑ: Un =

2 n(n − 1)

X

I(−∞,0] (Xi + Xj ),

1≤i<j≤n

which is known as the one-sample Wilcoxon statistic. Let Tn = Tn (X1 , ..., Xn ) be a given statistic and let r and d be two positive integers such that r + d = n. For any s = {i1 , ..., ir } ⊂ {1, ..., n}, define Tr,s = Tr (Xi1 , ..., Xir ), which is the statistic Tn computed after Xi , i 6∈ s, are deleted from the original sample. Let −1 X n 2 r Un = (3.13) d (Tr,s − Tn ) . r c Then Un is a U-statistic with kernel hn (x1 , ..., xr ) = dr [Tr (x1 , ..., xr ) − Tn (x1 , ..., xn )]2 . Unlike the kernels in the previous examples, the kernel in this example depends on n. The order of the kernel, r, may also depend on n. The statistic Un in (3.13) is known as the delete-d jackknife variance estimator for Tn (see, e.g., Shao and Tu (1995)), since it is often true that E[hn (X1 , ..., Xr )] ≈ Var(Tn ). ¯ then nUn in (3.13) is exactly the same as It can be shown that if Tn = X, 2 the sample variance S (exercise).

176

3. Unbiased Estimation

3.2.2 Variances of U-statistics If E[h(X1 , ..., Xm )]2 < ∞, then the variance of Un in (3.11) with kernel h has an explicit form. To derive Var(Un ), we need some notation. For k = 1, ..., m, let hk (x1 , ..., xk ) = E[h(X1 , ..., Xm )|X1 = x1 , ..., Xk = xk ] = E[h(x1 , ..., xk , Xk+1 , ..., Xm )]. Note that hm = h. It can be shown that hk (x1 , ..., xk ) = E[hk+1 (x1 , ..., xk , Xk+1 )].

(3.14)

˜ k = hk − E[h(X1 , ..., Xm )], h

(3.15)

Define

˜=˜ k = 1, ..., m, and h hm . Then, for any Un defined by (3.11), −1 X n ˜ i1 , ..., Xim ). h(X Un − E(Un ) = m c

(3.16)

Theorem 3.4 (Hoeffding’s theorem). For a U-statistic Un given by (3.11) with E[h(X1 , ..., Xm )]2 < ∞, −1 X m m n−m n ζk , Var(Un ) = k m−k m k=1

where ζk = Var(hk (X1 , ..., Xk )). Proof. Consider two sets {i1 , ..., im } and {j1 , ..., jm } of m distinct integers from {1, ..., n} with exactly k integers in common. The number of distinct n m n−m ˜ choices of two such sets is m k m−k . By the symmetry of hm and independence of X1 , ..., Xn , ˜ j , ..., Xj )] = ζk ˜ i , ..., Xi )h(X E[h(X 1 m 1 m for k = 1, ..., m (exercise). Then, by (3.16), −2 X X n ˜ i , ..., Xi )h(X ˜ j , ..., Xj )] E[h(X Var(Un ) = 1 m 1 m m c c −2 X m n m n−m n ζk . = m k m−k m k=1

This proves the result.

(3.17)

177

3.2. U-Statistics

Corollary 3.2. Under the condition of Theorem 3.4, 2 (i) mn ζ1 ≤ Var(Un ) ≤ m n ζm ; (ii) (n + 1)Var(Un+1 ) ≤ nVar(Un ) for any n > m; (iii) For any fixed m and k = 1, ..., m, if ζj = 0 for j < k and ζk > 0, then 2 k! m ζk 1 k Var(Un ) = . + O nk nk+1 It follows from Corollary 3.2 that a U-statistic Un as an estimator of its mean is consistent in mse (under the finite second moment assumption on h). In fact, for any fixed m, if ζj = 0 for j < k and ζk > 0, then the mse of Un is of the order n−k and, therefore, Un is nk/2 -consistent. Example 3.11. Consider first h(x1 , x2 ) = x1 x2 , which leads to a U˜ 1 (x1 ) = statistic unbiased for µ2 , µ = EX1 . Note that h1 (x1 ) = µx1 , h ˜ 1 (X1 )]2 = µ2 Var(X1 ) = µ2 σ 2 , h(x ˜ 1 , x2 ) = x1 x2 − µ2 , µ(x1 − µ), ζ1 = E[h and ζ2 = Var(X1 X2 ) = E(X1 X2 )2 − µ4 = (µ2 + σ 2 )2 − µ4 . By Theorem −1 P 3.4, for Un = n2 1≤i<j≤n Xi Xj , −1 2 n−2 n 2 n−2 Var(Un ) = ζ1 + ζ2 1 1 2 2 0 2 2(n − 2)µ2 σ 2 + (µ2 + σ 2 )2 − µ4 = n(n − 1) 2σ 4 4µ2 σ 2 + . = n n(n − 1)

¯ 2 − σ 2 /n in Example 3.10, which is the UMVUE Comparing Un with X under the normality and known σ 2 assumption, we find that ¯ 2 − σ 2 /n) = Var(Un ) − Var(X

2σ 4 . n2 (n − 1)

Next, consider h(x1 , x2 ) = I(−∞,0] (x1 + x2 ), which leads to the onesample Wilcoxon statistic. Note that h1 (x1 ) = P (x1 + X2 ≤ 0) = F (−x1 ), where F is the c.d.f. of P . Then ζ1 = Var(F (−X1 )). Let ϑ = E[h(X1 , X2 )]. Then ζ2 = Var(h(X1 , X2 )) = ϑ(1 − ϑ). Hence, for Un being the one-sample Wilcoxon statistic, Var(Un ) =

2 [2(n − 2)ζ1 + ϑ(1 − ϑ)] . n(n − 1)

If F is continuous and symmetric about 0, then ζ1 can be simplified as ζ1 = Var(F (−X1 )) = Var(1 − F (X1 )) = Var(F (X1 )) =

1 12 ,

178

3. Unbiased Estimation

since F (X1 ) has the uniform distribution on [0, 1]. Finally, consider h(x1 , x2 ) = |x1 − x2 |, which leads to Gini’s mean difference. Note that Z h1 (x1 ) = E|x1 − X2 | = |x1 − y|dP (y),

and

ζ1 = Var(h1 (X1 )) = where ϑ = E|X1 − X2 |.

Z Z

2 |x − y|dP (y) dP (x) − ϑ2 ,

3.2.3 The projection method Since P is nonparametric, the exact distribution of any U-statistic is hard to derive. In this section, we study asymptotic distributions of U-statistics by using the method of projection. Definition 3.3. Let Tn be a given statistic based on X1 , ..., Xn . The projection of Tn on kn random elements Y1 , ..., Ykn is defined to be Tˇn = E(Tn ) +

kn X i=1

[E(Tn |Yi ) − E(Tn )].

Let ψn (Xi ) = E(Tn |Xi ). If Tn is symmetric (as a function of X1 , ..., Xn ), then ψn (X1 ), ..., ψn (Xn ) are i.i.d. with mean E[ψn (Xi )] = E[E(Tn |Xi )] = E(Tn ). If E(Tn2 ) < ∞ and Var(ψn (Xi )) > 0, then n

X 1 p [ψn (Xi ) − E(Tn )] →d N (0, 1) nVar(ψn (X1 )) i=1

(3.18)

by the CLT. Let Tˇn be the projection of Tn on X1 , ..., Xn . Then Tn − Tˇn = Tn − E(Tn ) −

n X i=1

[ψn (Xi ) − E(Tn )].

(3.19)

If we can show that Tn − Tˇn has a negligible order of magnitude, then we can derive the asymptotic distribution of Tn by using (3.18)-(3.19) and Slutsky’s theorem. The order of magnitude of Tn − Tˇn can be obtained with the help of the following lemma. Lemma 3.1. Let Tn be a symmetric statistic with Var(Tn ) < ∞ for every n and Tˇn be the projection of Tn on X1 , ..., Xn . Then E(Tn ) = E(Tˇn ) and E(Tn − Tˇn )2 = Var(Tn ) − Var(Tˇn ).

179

3.2. U-Statistics Proof. Since E(Tn ) = E(Tˇn ), E(Tn − Tˇn )2 = Var(Tn ) + Var(Tˇn ) − 2Cov(Tn , Tˇn ). From Definition 3.3 with Yi = Xi and kn = n, Var(Tˇn ) = nVar(E(Tn |Xi )). The result follows from Cov(Tn , Tˇn ) = E(Tn Tˇn ) − [E(Tn )]2 = nE[Tn E(Tn |Xi )] − n[E(Tn )]2

= nE{E[Tn E(Tn |Xi )|Xi ]} − n[E(Tn )]2 = nE{[E(Tn |Xi )]2 } − n[E(Tn )]2 = nVar(E(Tn |Xi )) = Var(Tˇn ).

This method of deriving the asymptotic distribution of Tn is known as the method of projection and is particularly effective for U-statistics. For a U-statistic Un given by (3.11), one can show (exercise) that n

X ˜ 1 (Xi ), ˇn = E(Un ) + m U h n i=1

(3.20)

ˇn is the projection of Un on X1 , ..., Xn and ˜h1 is defined by (3.15). where U Hence ˇn ) = m2 ζ1 /n Var(U and, by Corollary 3.2 and Lemma 3.1, ˇn )2 = O(n−2 ). E(Un − U If ζ1 > 0, then (3.18) holds with ψn (Xi ) = mh1 (Xi ), which leads to the result in Theorem 3.5(i) stated later. ˜ 1 ≡ 0 and we have to use another projection of Un . If ζ1 = 0, then h Suppose that ζ1 = · · · = ζk−1 = 0 and ζk > 0 for an integer k > 1. ˇkn of Un on n random vectors {Xi1 , ..., Xi }, Consider the projection U k k 1 ≤ i1 < · · · < ik ≤ n. We can establish a result similar to that in Lemma 3.1 (exercise) and show that ˇn )2 = O(n−(k+1) ). E(Un − U Also, see Serfling (1980, §5.3.4). With these results, we obtain the following theorem.

180

3. Unbiased Estimation

Theorem 3.5. Let Un be given by (3.11) with E[h(X1 , ..., Xm )]2 < ∞. (i) If ζ1 > 0, then √ n[Un − E(Un )] →d N (0, m2 ζ1 ). (ii) If ζ1 = 0 but ζ2 > 0, then ∞

m(m − 1) X λj (χ21j − 1), n[Un − E(Un )] →d 2 j=1

(3.21)

where χ21j ’s are i.i.d. random variables having the chi-square distribution χ21 P∞ and λj ’s are some constants (which may depend on P ) satisfying j=1 λ2j = ζ2 . We have actually proved Theorem 3.5(i). A proof for Theorem 3.5(ii) is given in Serfling (1980, §5.5.2). One may derive results for the cases where ζ2 = 0, but the case of either ζ1 > 0 or ζ2 > 0 is the most interesting case in applications. If ζ1 > 0, it follows from Theorem 3.5(i) and Corollary 3.2(iii) that amseUn (P ) = m2 ζ1 /n = Var(Un ) + O(n−2 ). By Proposition 2.4(ii), {n[Un − E(Un )]2 } is uniformly integrable. If ζ1 = 0 but ζ2 > 0, it follows from Theorem 3.5(ii) that amseUn (P ) = EY 2 /n2 , where Y denotes the random variable on the right-hand side of (3.21). The following result provides the value of EY 2 . Lemma 3.2. Let Y be the random variable on the right-hand side of 2 2 ζ2 . (3.21). Then EY 2 = m (m−1) 2 Proof. Define k

Yk =

m(m − 1) X λj (χ21j − 1), 2 j=1

k = 1, 2, ....

It can be shown (exercise) that {Yk2 } is uniformly integrable. Since Yk →d Y as k → ∞, limk→∞ EYk2 = EY 2 (Theorem 1.8(viii)). Since χ21j ’s are independent chi-square random variables with Eχ21j = 1 and Var(χ21j ) = 2, EYk = 0 for any k and k m2 (m − 1)2 X 2 λj Var(χ21j ) 4 j=1 k m2 (m − 1)2 X 2 = λj 2 4 j=1

EYk2 =

→

m2 (m − 1)2 ζ2 . 2

181

3.2. U-Statistics

It follows from Corollary 3.2(iii) and Lemma 3.2 that amseUn (P ) = = Var(Un ) + O(n−3 ) if ζ1 = 0. Again, by Proposition 2.4(ii), the sequence {n2 [Un − E(Un )]2 } is uniformly integrable. We now apply in Example 3.11. For P Theorem 3.5 to the2 U-statistics 2 2 Un = n(n−1) X X , ζ = µ σ . Thus, if µ 6= 0, the result in i j 1 1≤i<j≤n Theorem 3.5(i) holds with ζ1 = µ2 σ 2 . If µ = 0, then ζ1 = 0, ζ2 = σ 4 > 0, and Theorem 3.5(ii) applies. However, it is not convenient to use Theorem 3.5(ii) to find the limiting distribution of Un . We may derive this limiting distribution using the following technique, which is further discussed in §3.5. By the CLT and Theorem 1.10, m2 (m−1)2 ζ2 /n2 2

¯ 2 /σ 2 →d χ2 nX 1 when µ = 0, where χ21 is a random variable having the chi-square distribution χ21 . Note that n

¯2 nX 1 X 2 (n − 1)Un = 2 X + . 2 σ σ n i=1 i σ2

By the SLLN, leads to

1 σ2 n

Pn

i=1

Xi2 →a.s. 1. An application of Slutsky’s theorem nUn /σ 2 →d χ21 − 1.

Since µ = 0, this implies that the right-hand side of (3.21) is σ 2 (χ21 − 1), i.e., λ1 = σ 2 and λj = 0 when j > 1. For the one-sample Wilcoxon statistic, ζ1 = Var(F (−X1 )) > 0 unless F is degenerate. Similarly, for Gini’s mean difference, ζ1 > 0 unless F is degenerate. Hence Theorem 3.5(i) applies to these two cases. Theorem 3.5 does not apply to Un defined by (3.13) if r, the order of the kernel, depends on n and diverges to ∞ as n → ∞. We consider the simple case where n 1X Tn = ψ(Xi ) + Rn (3.22) n i=1 for some Rn satisfying E(Rn2 ) = o(n−1 ). Note that (3.22) is satisfied for Tn being a U-statistic (exercise). Assume that r/d is bounded. Let Sψ2 = P P (n − 1)−1 ni=1 [ψ(Xi ) − n−1 ni=1 ψ(Xi )]2 . Then nUn = Sψ2 + op (1)

(3.23)

(exercise). Under (3.22), if 0 < E[ψ(Xi )]2 < ∞, then amseTn (P ) = E[ψ(Xi )]2 /n. Hence, the jackknife estimator Un in (3.13) provides a consistent estimator of amseTn (P ), i.e., Un /amseTn (P ) →p 1.

182

3. Unbiased Estimation

3.3 The LSE in Linear Models One of the most useful statistical models for non-i.i.d. data in applications is the general linear model Xi = β τ Z i + ε i ,

i = 1, ..., n,

(3.24)

where Xi is the ith observation and is often called the ith response; β is a p-vector of unknown parameters, p < n; Zi is the ith value of a pvector of explanatory variables (or covariates); and ε1 , ..., εn are random errors. Our data in this case are (X1 , Z1 ), ..., (Xn , Zn ) (εi ’s are not observed). Throughout this book Zi ’s are considered to be nonrandom or given values of a random p-vector, in which case our analysis is conditioned on Z1 , ..., Zn . Each εi can be viewed as a random measurement error in measuring the unknown mean of Xi when the covariate vector is equal to Zi . The main parameter of interest is β. More specific examples of model (3.24) are provided in this section. Other examples and examples of data from model (3.24) can be found in many standard books for linear models, for example, Draper and Smith (1981) and Searle (1971).

3.3.1 The LSE and estimability Let X = (X1 , ..., Xn ), ε = (ε1 , ..., εn ), and Z be the n × p matrix whose ith row is the vector Zi , i = 1, ..., n. Then, a matrix form of model (3.24) is X = Zβ + ε.

(3.25)

Definition 3.4. Suppose that the range of β in model (3.25) is B ⊂ Rp . A least squares estimator (LSE) of β is defined to be any βˆ ∈ B such that ˆ 2 = min kX − Zbk2 . kX − Z βk b∈B

(3.26)

For any l ∈ Rp , lτ βˆ is called an LSE of lτ β. Throughout this book, we consider B = Rp unless otherwise stated. Differentiating kX − Zbk2 w.r.t. b, we obtain that any solution of Z τ Zb = Z τ X

(3.27)

is an LSE of β. If the rank of the matrix Z is p, in which case (Z τ Z)−1 exists and Z is said to be of full rank, then there is a unique LSE, which is βˆ = (Z τ Z)−1 Z τ X.

(3.28)

183

3.3. The LSE in Linear Models

If Z is not of full rank, then there are infinitely many LSE’s of β. It can be shown (exercise) that any LSE of β is of the form βˆ = (Z τ Z)− Z τ X, τ

−

where (Z Z)

(3.29) τ

is called a generalized inverse of Z Z and satisfies Z τ Z(Z τ Z)− Z τ Z = Z τ Z.

Generalized inverse matrices are not unique unless Z is of full rank, in which case (Z τ Z)− = (Z τ Z)−1 and (3.29) reduces to (3.28). To study properties of LSE’s of β, we need some assumptions on the distribution of X. Since Zi ’s are nonrandom, assumptions on the distribution of X can be expressed in terms of assumptions on the distribution of ε. Several commonly adopted assumptions are stated as follows. Assumption A1: ε is distributed as Nn (0, σ 2 In ) with an unknown σ 2 > 0. Assumption A2: E(ε) = 0 and Var(ε) = σ 2 In with an unknown σ 2 > 0. Assumption A3: E(ε) = 0 and Var(ε) is an unknown matrix. Assumption A1 is the strongest and implies a parametric model. We may assume a slightly more general assumption that ε has the Nn (0, σ 2 D) distribution with unknown σ 2 but a known positive definite matrix D. Let D−1/2 be the inverse of the square root matrix of D. Then model (3.25) with assumption A1 holds if we replace X, Z, and ε by the transformed ˜ = D−1/2 X, Z˜ = D−1/2 Z, and ε˜ = D−1/2 ε, respectively. A variables X similar conclusion can be made for assumption A2. Under assumption A1, the distribution of X is Nn (Zβ, σ 2 In ), which is in an exponential family P with parameter θ = (β, σ 2 ) ∈ Rp × (0, ∞). However, if the matrix Z is not of full rank, then P is not identifiable (see §2.1.2), since Zβ1 = Zβ2 does not imply β1 = β2 . Suppose that the rank of Z is r ≤ p. Then there is an n × r submatrix Z∗ of Z such that Z = Z∗ Q (3.30) and Z∗ is of rank r, where Q is a fixed r × p matrix. Then Zβ = Z∗ Qβ and P is identifiable if we consider the reparameterization β˜ = Qβ. Note that the new parameter β˜ is in a subspace of Rp with dimension r. In many applications, we are interested in estimating some linear functions of β, i.e., ϑ = lτ β for some l ∈ Rp . From the previous discussion, however, estimation of lτ β is meaningless unless l = Qτ c for some c ∈ Rr so that ˜ lτ β = cτ Qβ = cτ β.

184

3. Unbiased Estimation

The following result shows that lτ β is estimable if l = Qτ c, which is also necessary for lτ β to be estimable under assumption A1. Theorem 3.6. Assume model (3.25) with assumption A3. (i) A necessary and sufficient condition for l ∈ Rp being Qτ c for some c ∈ Rr is l ∈ R(Z) = R(Z τ Z), where Q is given by (3.30) and R(A) is the smallest linear subspace containing all rows of A. (ii) If l ∈ R(Z), then the LSE lτ βˆ is unique and unbiased for lτ β. (iii) If l 6∈ R(Z) and assumption A1 holds, then lτ β is not estimable. Proof. (i) Note that a ∈ R(A) if and only if a = Aτ b for some vector b. If l = Qτ c, then l = Qτ c = Qτ Z∗τ Z∗ (Z∗τ Z∗ )−1 c = Z τ [Z∗ (Z∗τ Z∗ )−1 c]. Hence l ∈ R(Z). If l ∈ R(Z), then l = Z τ ζ for some ζ and l = (Z∗ Q)τ ζ = Qτ c with c = Z∗τ ζ. (ii) If l ∈ R(Z) = R(Z τ Z), then l = Z τ Zζ for some ζ and by (3.29), ˆ = E[lτ (Z τ Z)− Z τ X] E(lτ β) = ζ τ Z τ Z(Z τ Z)− Z τ Zβ = ζ τ Z τ Zβ = lτ β. If β¯ is any other LSE of β, then, by (3.27), ¯ = ζ τ (Z τ X − Z τ X) = 0. lτ βˆ − lτ β¯ = ζ τ (Z τ Z)(βˆ − β) (iii) Under assumption A1, if there is an estimator h(X, Z) unbiased for lτ β, then Z h(x, Z)(2π)−n/2 σ −n exp − 2σ1 2 kx − Zβk2 dx. lτ β = Rn

Differentiating w.r.t. β and applying Theorem 2.1 lead to Z h(x, Z)(2π)−n/2 σ −n−2 (x − Zβ) exp − 2σ1 2 kx − Zβk2 dx, l = Zτ Rn

which implies l ∈ R(Z).

Theorem 3.6 shows that LSE’s are unbiased for estimable parameters lτ β. If Z is of full rank, then R(Z) = Rp and, therefore, lτ β is estimable for any l ∈ Rp .

185

3.3. The LSE in Linear Models

Example 3.12 (Simple linear regression). Let β = (β0 , β1 ) ∈ R2 and Zi = (1, ti ), ti ∈ R, i = 1, ..., n. Then model (3.24) or (3.25) is called a simple linear regression model. It turns out that Pn t n Pn Pni=1 2i Zτ Z = . i=1 ti i=1 ti

This matrix is invertible if and only if some ti ’s are different. Thus, if some ti ’s are different, then the unique unbiased LSE of lτ β for any l ∈ R2 is lτ (Z τ Z)−1 Z τ X, which has the normal distribution if assumption A1 holds. The result can be easily extended to the case of polynomial regression of order p in which β = (β0 , β1 , ..., βp−1 ) and Zi = (1, ti , ..., tp−1 ). i Example 3.13 (One-way ANOVA). Suppose that n = positive integers n1 , ..., nm and that

Pm

j=1

nj with m

Xi = µj + εi , i = kj−1 + 1, ..., kj , j = 1, ..., m, P where k0 = 0, kj = jl=1 nl , j = 1, ..., m, and (µ1 , ..., µm ) = β. Let Jm be the m-vector of ones. Then the matrix Z in this case is a block diagonal matrix with Jnj as the jth diagonal column. Consequently, Z τ Z is an m × m diagonal matrix whose jth diagonal element is nj . Thus, Z τ Z is invertible and the unique LSE of β is the m-vector whose jth component Pkj is n−1 j i=kj−1 +1 Xi , j = 1, ..., m. Sometimes it is more convenient to use the following notation: Xij = Xki−1 +j , εij = εki−1 +j ,

j = 1, ..., ni , i = 1, ..., m,

and µi = µ + αi ,

i = 1, ..., m.

Then our model becomes Xij = µ + αi + εij ,

j = 1, ..., ni , i = 1, ..., m,

(3.31)

which is called a one-way analysis of variance (ANOVA) model. Under model (3.31), β = (µ, α1 , ..., αm ) ∈ Rm+1 . The matrix Z under model (3.31) is not of full rank (exercise). An LSE of β under model (3.31) is ¯ X ¯ 1· − X, ¯ ..., X ¯ m· − X ¯ , βˆ = X, ¯ is still the sample mean of Xij ’s and X ¯ i· is the sample mean of the where X ith group {Xij , j = 1, ..., ni }. The problem of finding the form of l ∈ R(Z) under model (3.31) is left as an exercise.

The notation used in model (3.31) allows us to generalize the one-way ANOVA model to any s-way ANOVA model with a positive integer s under

186

3. Unbiased Estimation

the so-called factorial experiments. The following example is for the twoway ANOVA model. Example 3.14 (Two-way balanced ANOVA). Suppose that Xijk = µ + αi + βj + γij + εijk ,

i = 1, ..., a, j = 1, ..., b, k = 1, ..., c, (3.32)

where a, b, and c are some positive integers. Model (3.32) is called a twoway balanced ANOVA model. If we view model (3.32) as a special case of model (3.25), then the parameter vector β is β = (µ, α1 , ..., αa , β1 , ..., βb , γ11 , ..., γ1b , ..., γa1 , ..., γab ).

(3.33)

One can obtain the matrix Z and show that it is n × p, where n = abc and p = 1 + a + b + ab, and is of rank ab < p (exercise). It can also be shown (exercise) that an LSE of β is given by the right-hand side of (3.33) with µ, ¯ ··· , αi , βj , and γij replaced by µ ˆ, α ˆ i , βˆj , and γˆij , respectively, where µ ˆ=X ¯ i·· − X ¯ ··· , βˆj = X ¯ ·j· − X ¯ ··· , γˆij = X ¯ ij· − X ¯ i·· − X ¯ ·j· + X ¯ ··· , and a dot α ˆi = X is used to denote averaging over the indicated subscript, e.g., a

c

XX ¯ ·j· = 1 Xijk X ac i=1 k=1

with a fixed j.

3.3.2 The UMVUE and BLUE We now study UMVUE’s in model (3.25) with assumption A1. Theorem 3.7. Consider model (3.25) with assumption A1. (i) The LSE lτ βˆ is the UMVUE of lτ β for any estimable lτ β. ˆ 2 , where r is the rank ˆ 2 = (n − r)−1 kX − Z βk (ii) The UMVUE of σ 2 is σ of Z. Proof. (i) Let βˆ be an LSE of β. By (3.27), ˆ τ Z(βˆ − β) = (X τ Z − X τ Z)(βˆ − β) = 0 (X − Z β) and, hence, kX − Zβk2 = kX − Z βˆ + Z βˆ − Zβk2 ˆ 2 + kZ βˆ − Zβk2 = kX − Z βk

ˆ 2 − 2β τ Z τ X + kZβk2 + kZ βk ˆ 2. = kX − Z βk

Using this result and assumption A1, we obtain the following joint Lebesgue p.d.f. of X: o n τ τ ˆ 2 ˆ 2 βk kZβk2 . − (2πσ 2 )−n/2 exp β σZ2 x − kx−Z βk2σ+kZ 2 2σ2

187

3.3. The LSE in Linear Models

By Proposition 2.1 and the fact that Z βˆ = Z(Z τ Z)− Z τ X is a function of ˆ 2 ) is complete and sufficient for θ = (β, σ 2 ). Note Z τ X, (Z τ X, kX − Z βk ˆ that β is a function of Z τ X and, hence, a function of the complete sufficient statistic. If lτ β is estimable, then lτ βˆ is unbiased for lτ β (Theorem 3.6) and, hence, lτ βˆ is the UMVUE of lτ β. ˆ 2 + kZ βˆ − Zβk2 and E(Z β) ˆ = Zβ (ii) From kX − Zβk2 = kX − Z βk (Theorem 3.6), ˆ 2 = E(X − Zβ)τ (X − Zβ) − E(β − β) ˆ τ Z τ Z(β − β) ˆ EkX − Z βk ˆ = tr Var(X) − Var(Z β) = σ 2 [n − tr Z(Z τ Z)− Z τ Z(Z τ Z)− Z τ ] = σ 2 [n − tr (Z τ Z)− Z τ Z ].

Since each row of Z ∈ R(Z), Z βˆ does not depend on the choice of (Z τ Z)− in βˆ = (Z τ Z)− Z τ X (Theorem 3.6). Hence, we can evaluate tr((Z τ Z)− Z τ Z) using a particular (Z τ Z)− . From the theory of linear algebra, there exists a p × p matrix C such that CC τ = Ip and Λ 0 , C τ (Z τ Z)C = 0 0 where Λ is an r × r diagonal matrix whose diagonal elements are positive. Then, a particular choice of (Z τ Z)− is −1 0 Λ τ − Cτ (Z Z) = C (3.34) 0 0 and τ

−

τ

(Z Z) Z Z = C

Ir 0

0 0

Cτ

whose trace is r. Hence σ ˆ 2 is the UMVUE of σ 2 , since it is a function of the complete sufficient statistic and ˆ 2 = σ2 . Eσ ˆ 2 = (n − r)−1 EkX − Z βk In general, ˆ = lτ (Z τ Z)− Z τ Var(ε)Z(Z τ Z)− l. Var(lτ β)

(3.35)

If l ∈ R(Z) and Var(ε) = σ 2 In (assumption A2), then the use of the genˆ = σ 2 lτ (Z τ Z)− l, which eralized inverse matrix in (3.34) leads to Var(lτ β) attains the Cram´er-Rao lower bound under assumption A1 (Proposition 3.2).

188

3. Unbiased Estimation

ˆ 2 is called The vector X − Z βˆ is called the residual vector and kX − Z βk the sum of squared residuals and is denoted by SSR. The estimator σ ˆ 2 is then equal to SSR/(n − r). Since X − Z βˆ = [In − Z(Z τ Z)− Z τ ]X and lτ βˆ = lτ (Z τ Z)− Z τ X are linear in X, they are normally distributed under assumption A1. Also, using the generalized inverse matrix in (3.34), we obtain that [In − Z(Z τ Z)− Z τ ]Z(Z τ Z)− = Z(Z τ Z)− − Z(Z τ Z)− Z τ Z(Z τ Z)− = 0, which implies that σ ˆ 2 and lτ βˆ are independent (Exercise 58 in §1.6) for any τ estimable l β. Furthermore, [Z(Z τ Z)− Z τ ]2 = Z(Z τ Z)− Z τ (i.e., Z(Z τ Z)− Z τ is a projection matrix) and SSR = X τ [In − Z(Z τ Z)− Z τ ]X. The rank of Z(Z τ Z)− Z τ is tr(Z(Z τ Z)− Z τ ) = r. Similarly, the rank of the projection matrix In − Z(Z τ Z)− Z τ is n − r. From X τ X = X τ [Z(Z τ Z)− Z τ ]X + X τ [In − Z(Z τ Z)− Z τ ]X and Theorem 1.5 (Cochran’s theorem), SSR/σ 2 has the chi-square distribution χ2n−r (δ) with δ = σ −2 β τ Z τ [In − Z(Z τ Z)− Z τ ]Zβ = 0. Thus, we have proved the following result. Theorem 3.8. Consider model (3.25) with assumption A1. For any esˆ 2 are independent; the timable parameter lτ β, the UMVUE’s lτ βˆ and σ τˆ τ 2 τ τ − distribution of l β is N (l β, σ l (Z Z) l); and (n − r)ˆ σ 2 /σ 2 has the chi2 square distribution χn−r . Example 3.15. In Examples 3.12-3.14, UMVUE’s of estimable lτ β are the ˆ under assumption A1. In Example 3.13, LSE’s lτ β, SSR =

ni m X X i=1 j=1

¯ i· )2 ; (Xij − X

in Example 3.14, if c > 1, SSR =

b X c a X X i=1 j=1 k=1

¯ ij· )2 . (Xijk − X

189

3.3. The LSE in Linear Models

ˆ 2 under assumption A2, i.e., withWe now study properties of lτ βˆ and σ out the normality assumption on ε. From Theorem 3.6 and the proof of Theorem 3.7(ii), lτ βˆ (with an l ∈ R(Z)) and σ ˆ 2 are still unbiased without ˆ 2 optimal beyond the normality assumption. In what sense are lτ βˆ and σ ˆ Some disbeing unbiased? We have the following result for the LSE lτ β. 2 cussion about σ ˆ can be found, for example, in Rao (1973, p. 228). Theorem 3.9. Consider model (3.25) with assumption A2. (i) A necessary and sufficient condition for the existence of a linear unbiased estimator of lτβ (i.e., an unbiased estimator that is linear in X) is l ∈ R(Z). (ii) (Gauss-Markov theorem). If l ∈ R(Z), then the LSE lτ βˆ is the best linear unbiased estimator (BLUE) of lτ β in the sense that it has the minimum variance in the class of linear unbiased estimators of lτ β. Proof. (i) The sufficiency has been established in Theorem 3.6. Suppose now a linear function of X, cτ X with c ∈ Rn , is unbiased for lτ β. Then lτ β = E(cτ X) = cτ EX = cτ Zβ.

Since this equality holds for all β, l = Z τ c, i.e., l ∈ R(Z). (ii) Let l ∈ R(Z) = R(Z τ Z). Then l = (Z τ Z)ζ for some ζ and lτ βˆ = ζ τ (Z τ Z)βˆ = ζ τ Z τ X by (3.27). Let cτ X be any linear unbiased estimator of lτ β. From the proof of (i), Z τ c = l. Then Cov(ζ τ Z τ X, cτ X − ζ τ Z τ X) = E(X τ Zζcτ X) − E(X τ Zζζ τ Z τ X) = σ 2 tr(Zζcτ ) + β τ Z τ Zζcτ Zβ − σ 2 tr(Zζζ τ Z τ ) − β τ Z τ Zζζ τ Z τ Zβ = σ 2 ζ τ l + (lτ β)2 − σ 2 ζ τ l − (lτ β)2

= 0. Hence

Var(cτ X) = Var(cτ X − ζ τ Z τ X + ζ τ Z τ X) = Var(cτ X − ζ τ Z τ X) + Var(ζ τ Z τ X) + 2Cov(ζ τ Z τ X, cτ X − ζ τ Z τ X) ˆ = Var(cτ X − ζ τ Z τ X) + Var(lτ β) τˆ ≥ Var(l β).

3.3.3 Robustness of LSE’s Consider now model (3.25) under assumption A3. An interesting question is under what conditions on Var(ε) is the LSE of lτ β with l ∈ R(Z) ˆ considered still the BLUE. If lτ βˆ is still the BLUE, then we say that lτ β, as a BLUE, is robust against violation of assumption A2. In general, a

190

3. Unbiased Estimation

statistical procedure having certain properties under an assumption is said to be robust against violation of the assumption if and only if the statistical procedure still has the same properties when the assumption is (slightly) violated. For example, the LSE of lτ β with l ∈ R(Z), as an unbiased estimator, is robust against violation of assumption A1 or A2, since the LSE is unbiased as long as E(ε) = 0, which can be always assumed without loss of generality. On the other hand, the LSE as a UMVUE may not be robust against violation of assumption A1 (see §3.5). Theorem 3.10. Consider model (3.25) with assumption A3. The following are equivalent. (a) lτ βˆ is the BLUE of lτ β for any l ∈ R(Z). ˆ τ X) = 0 for any l ∈ R(Z) and any η such that E(η τ X) = 0. (b) E(lτ βη τ (c) Z Var(ε)U = 0, where U is a matrix such that Z τ U = 0 and R(U τ ) + R(Z τ ) = Rn . (d) Var(ε) = ZΛ1 Z τ + U Λ2 U τ for some Λ1 and Λ2 . (e) The matrix Z(Z τ Z)− Z τ Var(ε) is symmetric. Proof. We first show that (a) and (b) are equivalent, which is an analogue of Theorem 3.2(i). Suppose that (b) holds. Let l ∈ R(Z). If cτ X is unbiased for lτ β, then E(η τ X) = 0 with η = c − Z(Z τ Z)− l. Hence ˆ Var(cτ X) = Var(cτ X − lτ βˆ + lτ β)

ˆ = Var(cτ X − lτ (Z τ Z)− Z τ X + lτ β) ˆ = Var(η τ X + lτ β) ˆ + 2Cov(η τ X, lτ β) ˆ = Var(η τ X) + Var(lτ β) ˆ + 2E(lτ βη ˆ τ X) = Var(η τ X) + Var(lτ β) ˆ = Var(η τ X) + Var(lτ β) ˆ ≥ Var(lτ β).

Suppose now that there are l ∈ R(Z) and η such that E(η τ X) = 0 but ˆ τ X) 6= 0. Let ct = tη + Z(Z τ Z)− l. From the previous proof, δ = E(lτ βη ˆ + 2δt. Var(cτt X) = t2 Var(η τ X) + Var(lτ β)

ˆ This As long as δ 6= 0, there exists a t such that Var(cτt X) < Var(lτ β). τˆ shows that l β cannot be a BLUE and, therefore, (a) implies (b). We next show that (b) implies (c). Suppose that (b) holds. Since l ∈ R(Z), l = Z τ γ for some γ. Let η ∈ R(U τ ). Then E(η τ X) = η τ Zβ = 0 and, hence, ˆ τ X) = E[γ τ Z(Z τ Z)− Z τ XX τ η] = γ τ Z(Z τ Z)− Z τ Var(ε)η. 0 = E(lτ βη Since this equality holds for all l ∈ R(Z), it holds for all γ. Thus, Z(Z τ Z)− Z τ Var(ε)U = 0,

3.3. The LSE in Linear Models

191

which implies Z τ Z(Z τ Z)− Z τ Var(ε)U = Z τ Var(ε)U = 0, since Z τ Z(Z τ Z)− Z τ = Z τ . Thus, (c) holds. To show that (c) implies (d), we need to use the following facts from the theory of linear algebra: there exists a nonsingular matrix C such that Var(ε) = CC τ and C = ZC1 + U C2 for some matrices Cj (since R(U τ ) + R(Z τ ) = Rn ). Let Λ1 = C1 C1τ , Λ2 = C2 C2τ , and Λ3 = C1 C2τ . Then Var(ε) = ZΛ1 Z τ + U Λ2 U τ + ZΛ3 U τ + U Λτ3 Z τ (3.36) and Z τ Var(ε)U = Z τ ZΛ3 U τ U , which is 0 if (c) holds. Hence, (c) implies 0 = Z(Z τ Z)− Z τ ZΛ3 U τ U (U τ U )− U τ = ZΛ3 U τ , which with (3.36) implies (d). If (d) holds, then Z(Z τ Z)− Z τ Var(ε) = ZΛ1 Z τ , which is symmetric. Hence (d) implies (e). To complete the proof, we need to show that (e) implies (b), which is left as an exercise. As a corollary of this theorem, the following result shows when the UMVUE’s in model (3.25) with assumption A1 are robust against the violation of Var(ε) = σ 2 In . Corollary 3.3. Consider model (3.25) with a full rank Z, ε = Nn (0, Σ), and an unknown positive definite matrix Σ. Then lτ βˆ is a UMVUE of lτ β for any l ∈ Rp if and only if one of (b)-(e) in Theorem 3.10 holds. Example 3.16. Consider model (3.25) with β replaced by a random vector β that is independent of ε. Such a model is called a linear model with random coefficients. Suppose that Var(ε) = σ 2 In and E(β) = β. Then X = Zβ + Z(β − β) + ε = Zβ + e,

(3.37)

where e = Z(β − β) + ε satisfies E(e) = 0 and Var(e) = ZVar(β)Z τ + σ 2 In . Since Z(Z τ Z)− Z τ Var(e) = ZVar(β)Z τ + σ 2 Z(Z τ Z)− Z τ is symmetric, by Theorem 3.10, the LSE lτ βˆ under model (3.37) is the BLUE for any lτ β, l ∈ R(Z). If Z is of full rank and ε is normal, then, by Corollary 3.3, lτ βˆ is the UMVUE of lτ β for any l ∈ Rp .

192

3. Unbiased Estimation

Example 3.17 (Random effects models). Suppose that Xij = µ + Ai + eij ,

j = 1, ..., ni , i = 1, ..., m,

(3.38)

where µ ∈ R is an unknown parameter, Ai ’s are i.i.d. random variables having mean 0 and variance σa2 , eij ’s are i.i.d. random errors with mean 0 and variance σ 2 , and Ai ’s and eij ’s are independent. Model (3.38) is called a one-way random effects model and Ai ’s are unobserved random effects. Let εij = Ai + eij . Then (3.38) is a special case of the general model (3.25) with Var(ε) = σa2 Σ + σ 2 In , τ where Σ is a block diagonal matrix whose ith block Pm is Jni Jni and Jτ k is−theτ kvector of ones. Under this model, Z = Jn , n = i=1 ni , and Z(Z Z) Z = n−1 Jn Jnτ . Note that

n1 Jn1 Jnτ1 n1 Jn2 Jnτ1 Jn Jnτ Σ = ······ n1 Jnm Jnτ1

n2 Jn1 Jnτ2 n2 Jn2 Jnτ2 ······ n2 Jnm Jnτ2

· · · nm Jn1 Jnτm · · · nm Jn2 Jnτm , ··· ······ · · · nm Jnm Jnτm

which is symmetric if and only if n1 = n2 = · · · = nm . Since Jn Jnτ Var(ε) is symmetric if and only if Jn Jnτ Σ is symmetric, a necessary and sufficient condition for the LSE of µ to be the BLUE is that all ni ’s are the same. This condition is also necessary and sufficient for the LSE of µ to be the UMVUE when εij ’s are normal. In some cases, we are interested in some (not all) linear functions of β. For example, consider lτ β with l ∈ R(H), where H is an n × p matrix such that R(H) ⊂ R(Z). We have the following result. Proposition 3.4. Consider model (3.25) with assumption A3. Suppose that H is a matrix such that R(H) ⊂ R(Z). A necessary and sufficient condition for the LSE l τ βˆ to be the BLUE of lτ β for any l ∈ R(H) is H(Z τ Z)− Z τ Var(ε)U = 0, where U is the same as that in (c) of Theorem 3.10. Example 3.18. Consider model (3.25) with assumption A3 and Z = (H1 H2 ), where H1τ H2 = 0. Suppose that under the reduced model X = H1 β1 + ε, lτ βˆ1 is the BLUE for any lτ β1 , l ∈ R(H1 ), and that under the reduced model X = H2 β2 + ε,

193

3.3. The LSE in Linear Models

lτ βˆ2 is not a BLUE for some lτ β2 , l ∈ R(H2 ), where β = (β1 , β2 ) and βˆj ’s are LSE’s under the reduced models. Let H = (H1 0) be n × p. Note that H(Z τ Z)− Z τ Var(ε)U = H1 (H1τ H1 )− H1τ Var(ε)U, which is 0 by Theorem 3.10 for the U given in (c) of Theorem 3.10, and Z(Z τ Z)− Z τ Var(ε)U = H2 (H2τ H2 )− H2τ Var(ε)U, which is not 0 by Theorem 3.10. This implies that some LSE lτ βˆ is not a BLUE of lτ β but lτ βˆ is the BLUE of lτ β if l ∈ R(H). Finally, we consider model (3.25) with Var(ε) being a diagonal matrix whose ith diagonal element is σi2 , i.e., εi ’s are uncorrelated but have unequal variances. A straightforward calculation shows that condition (e) in Theorem 3.10 holds if and only if, for all i 6= j, σi2 6= σj2 only when hij = 0, where hij is the (i, j)th element of the projection matrix Z(Z τ Z)− Z τ . Thus, an LSE is not a BLUE in general, although it is still unbiased for estimable lτ β. Suppose that the unequal variances of εi ’s are caused by some small perturbations, i.e., εi = ei + ui , where Var(ei ) = σ 2 , Var(ui ) = δi , and ei and ui are independent so that σi2 = σ 2 + δi . From (3.35), ˆ = lτ (Z τ Z)− Var(lτ β)

n X

σi2 Zi Ziτ (Z τ Z)− l.

i=1

If δi = 0 for all i (no perturbations), then assumption A2 holds and lτ βˆ ˆ = σ 2 lτ (Z τ Z)− l. Suppose is the BLUE of any estimable lτ β with Var(lτ β) 2 that 0 < δi ≤ σ δ. Then ˆ ≤ (1 + δ)σ 2 lτ (Z τ Z)− l. Var(lτ β) This indicates that the LSE is robust in the sense that its variance increases slightly when there is a slight violation of the equal variance assumption (small δ).

3.3.4 Asymptotic properties of LSE’s We consider first the consistency of the LSE lτ βˆ with l ∈ R(Z) for every n. Theorem 3.11. Consider model (3.25) with assumption A3. Suppose that supn λ+ [Var(ε)] < ∞, where λ+ [A] is the largest eigenvalue of the matrix A, and that limn→∞ λ+ [(Z τ Z)− ] = 0. Then lτ βˆ is consistent in mse for

194

3. Unbiased Estimation

any l ∈ R(Z). Proof. The result follows from the fact that lτ βˆ is unbiased and ˆ = lτ (Z τ Z)− Z τ Var(ε)Z(Z τ Z)− l Var(lτ β) ≤ λ+ [Var(ε)]lτ (Z τ Z)− l. Without the normality assumption on ε, the exact distribution of lτ βˆ is very hard to obtain. The asymptotic distribution of lτ βˆ is derived in the following result. Theorem 3.12. Consider model (3.25) with assumption A3. Suppose that 0 < inf n λ− [Var(ε)], where λ− [A] is the smallest eigenvalue of the matrix A, and that lim max Ziτ (Z τ Z)− Zi = 0. (3.39) n→∞ 1≤i≤n

Pk Suppose further that n = j=1 mj for some integers k, mj , j = 1, ..., k, with mj ’s bounded by a fixed integer m, ε = (ξ1 , ..., ξk ), ξj ∈ Rmj , and ξj ’s are independent. (i) If supi E|εi |2+δ < ∞, then for any l ∈ R(Z), q ˆ →d N (0, 1). lτ (βˆ − β) Var(lτ β) (3.40) (ii) Suppose that when mi = mj , 1 ≤ i < j ≤ k, ξi and ξj have the same distribution. Then result (3.40) holds for any l ∈ R(Z). Proof. Let l ∈ R(Z). Then lτ (Z τ Z)− Z τ Zβ − lτ β = 0 and l (βˆ − β) = lτ (Z τ Z)− Z τ ε = τ

k X

cτnj ξj ,

j=1

where cnj is the mj -vector whose components are lτ (Z τ Z)− Zi , i = kj−1 + P 1, ..., kj , k0 = 0, and kj = jt=1 mt , j = 1, ..., k. Note that k X j=1

kcnj k2 = lτ (Z τ Z)− Z τ Z(Z τ Z)− l = lτ (Z τ Z)− l.

Also, max kcnj k2 ≤ m max [lτ (Z τ Z)− Zi ]2

1≤j≤k

1≤i≤n

≤ mlτ (Z τ Z)− l max Ziτ (Z τ Z)− Zi , 1≤i≤n

(3.41)

195

3.4. Unbiased Estimators in Survey Problems

which, together with (3.41) and condition (3.39), implies that X k kcnj k2 = 0. lim max kcnj k2 n→∞

1≤j≤k

j=1

The results then follow from Corollary 1.3.

Under the conditions of Theorem 3.12, Var(ε) is a diagonal block matrix with Var(ξj ) as the jth diagonal block, which includes the case of independent εi ’s as a special case. The following lemma tells us how to check condition (3.39). Lemma 3.3. The following are sufficient conditions for (3.39). (a) λ+ [(Z τ Z)− ] → 0 and Znτ (Z τ Z)− Zn → 0, as n → ∞. (b) There is an increasing sequence {an } such that an → ∞, an /an+1 → 1, and Z τ Z/an converges to a positive definite matrix. P P If n−1 ni=1 t2i → c and n−1 ni=1 ti → d in the simple linear regression model (Example 3.12), where c is positive and c > d2 , then condition (b) in Lemma 3.3 is satisfied with an = n and, therefore, Theorem 3.12 applies. In the one-way ANOVA model (Example 3.13), max Ziτ (Z τ Z)− Zi = λ+ [(Z τ Z)− ] = max n−1 j .

1≤i≤n

1≤j≤m

Hence conditions related to Z in Theorem 3.12 are satisfied if and only if minj nj → ∞. Some similar conclusions can be drawn in the two-way ANOVA model (Example 3.14).

3.4 Unbiased Estimators in Survey Problems In this section, we consider unbiased estimation for another type of noni.i.d. data often encountered in applications: survey data from finite populations. A description of the problem is given in Example 2.3 of §2.1.1. Examples and a fuller account of theoretical aspects of survey sampling can be found, for example, in Cochran (1977) and S¨arndal, Swensson, and Wretman (1992).

3.4.1 UMVUE’s of population totals We use the same notation as in Example 2.3. Let X = (X1 , ..., Xn ) be a sample from a finite population P = {y1 , ..., yN } with P (X1 = yi1 , ..., Xn = yin ) = p(s)/n!,

196

3. Unbiased Estimation

where s = {i1 , ..., in } is a subset of distinct elements of {1, ..., N } and p is a selection probability measure. We consider univariate yi , although most of our conclusions are valid for the case of multivariate P yi . In many survey problems the parameter to be estimated is Y = N i=1 yi , the population total. ¯ =NP In Example 2.27, it is shown that Yˆ = N X i∈s yi is unbiased for n Y if p(s) is constant (simple random sampling); a formula of Var(Yˆ ) is also given. We now show that Yˆ is in fact the UMVUE of Y under simple QN random sampling. Let Y be the range of yi , θ = (y1 , ..., yN ) and Θ = i=1 Y. Under simple random sampling, the population under consideration is a parametric family indexed by θ ∈ Θ. Theorem 3.13 (Watson-Royall theorem). (i) If p(s) > 0 for all s, then the vector of order statistics X(1) ≤ · · · ≤ X(n) is complete for θ ∈ Θ. (ii) Under simple random sampling, the vector of order statistics is sufficient for θ ∈ Θ. (iii) Under simple random sampling, for any estimable function of θ, its unique UMVUE is the unbiased estimator g(X1 , ..., Xn ), where g is symmetric in its n arguments. Proof. (i) Let h(X) be a function of the order statistics. Then h is symmetric in its n arguments. We need to show that if X E[h(X)] = p(s)h (yi1 , ..., yin ) /n! = 0 (3.42) s={i1 ,...,in }⊂{1,...,N } for all θ ∈ Θ, then h(yi1 , ..., yin ) = 0 for all yi1 , ..., yin . First, suppose that all N elements of θ are equal to a ∈ Y. Then (3.42) implies h(a, ..., a) = 0. Next, suppose that N − 1 elements in θ are equal to a and one is b > a. Then (3.42) reduces to q1 h(a, ..., a) + q2 h(a, ..., a, b), where q1 and q2 are some known numbers in (0, 1). Since h(a, ..., a) = 0 and q2 6= 0, h(a, ..., a, b) = 0. Using the same argument, we can show that h(a, ..., a, b, ..., b) = 0 for any k a’s and n − k b’s. Suppose next that elements of θ are equal to a, b, or c, a < b < c. Then we can show that h(a, ..., a, b, ..., b, c, ..., c) = 0 for any k a’s, l b’s, and n−k −l c’s. Continuing inductively, we see that h(y1 , ..., yn ) = 0 for all possible y1 , ..., yn . This completes the proof of (i). (ii) The result follows from the factorization theorem (Theorem 2.2), the fact that p(s) is constant under simple random sampling, and P (X1 = yi1 , ..., Xn = yin ) = P (X(1) = y(i1 ) , ..., X(n) = y(in ) )/n!, where y(i1 ) ≤ · · · ≤ y(in ) are the ordered values of yi1 , ..., yin . (iii) The result follows directly from (i) and (ii).

3.4. Unbiased Estimators in Survey Problems

197

It is interesting to note the following two issues. (1) Although we have a parametric problem under simple random sampling, the sufficient and complete statistic is the same as that in a nonparametric problem (Example 2.17). (2) For the completeness of the order statistics, we do not need the assumption of simple random sampling. ¯ is unbiased for Y . Since Yˆ Example 3.19. From Example 2.27, Yˆ = N X is symmetric in its arguments, it is the UMVUE of Y . We now derive the UMVUE for Var(Yˆ ). From Example 2.27, Var(Yˆ ) = where σ2 =

N2 n 2 1− σ , n N

(3.43)

2 N 1 X Y yi − . N − 1 i=1 N

It can be shown (exercise) that E(S 2 ) = σ 2 , where S 2 is the usual sample variance !2 n X ˆ 1 X 1 Y 2 2 ¯ = yi − S = (Xi − X) . n − 1 i=1 n − 1 i∈s N 2 2 n Since S 2 is symmetric in its arguments, Nn 1 − N S is the UMVUE of Var(Yˆ ). Simple random sampling is simple and easy to use, but it is inefficient unless the population is fairly homogeneous w.r.t. the yi ’s. A sampling plan often used in practice is the stratified sampling plan, which can be described as follows. The population P is divided into nonoverlapping subpopulations P1 , ..., PH called strata; a sample is drawn from each stratum Ph , independently across the strata. There are many reasons for stratification: (1) it may produce a gain in precision in parameter estimation when a heterogeneous population is divided into strata, each of which is internally homogeneous; (2) sampling problems may differ markedly in different parts of the population; and (3) administrative considerations may also lead to stratification. More discussions can be found, for example, in Cochran (1977). In stratified sampling, if a simple random sample (without replacement), Xh = (Xh1 , ..., Xhnh ), is drawn from each stratum, where nh is the sample size in stratum h, then the joint distribution of X = (X1 , ..., XH ) is in a parametric family indexed by θ = (θ1 , ..., θH ), where θh = (yi , i ∈ Ph ), h = QNh Yh , where 1, ..., H. Let Yh be the range of yi ’s in stratum h and Θh = i=1 QH Nh is the size of Ph . We assume that the parameter space is Θ = i=1 Θh . The following result is similar to Theorem 3.13.

198

3. Unbiased Estimation

Theorem 3.14. Let X be a sample obtained using the stratified simple random sampling plan described previously. (i) For each h, let Zh be the vector of the ordered values of the sample in stratum h. Then (Z1 , ..., ZH ) is sufficient and complete for θ ∈ Θ. (ii) For any estimable function of θ, its unique UMVUE is the unbiased estimator g(X) that is symmetric in its first n1 arguments, symmetric in its second n2 arguments,..., and symmetric in its last nH arguments. Example 3.20. Consider the estimation of the population total Y based on a sample X = (Xhi , i = 1, ..., nh , h = 1, ..., H) obtained by stratified simple random sampling. Let Yh be the population total of the hth stratum and ¯ h· is the sample mean of the sample from stratum ¯ h· , where X let Yˆh = Nh X h, h = 1, ..., H. From Example 2.27, each Yˆh is an unbiased estimator of Yh . Let nh H H X X X Nh Xhi . Yˆh = Yˆst = nh i=1 h=1

h=1

Then, by Theorem 3.14, Yˆst is the UMVUE of Y . Since Yˆ1 , ..., YˆH are independent, it follows from (3.43) that Var(Yˆst ) =

H X Nh2 nh 1− σh2 , nh Nh

(3.44)

h=1

P where σh2 = (Nh − 1)−1 i∈Ph (yi − Yh /Nh )2 . An argument similar to that in Example 3.19 shows that the UMVUE of Var(Yˆst ) is 2 Sst

H X Nh2 nh 1− Sh2 , = nh Nh

(3.45)

h=1

where Sh2 is the usual sample variance based on Xh1 , ..., Xhnh . It is interesting to compare the mse of the UMVUE Yˆst with the mse of the UMVUE Yˆ under simple random sampling (Example 3.19). Let σ 2 be given in (3.43). Then (N − 1)σ 2 =

H X

(Nh − 1)σh2 +

h=1

H X

h=1

Nh (µh − µ)2 ,

where µh = Yh /Nh is the population mean of the hth stratum and µ = Y /N is the overall population mean. By (3.43), (3.44), and (3.45), Var(Yˆ ) ≥ Var(Yˆst ) if and only if H X

h=1

N 2 Nh n(N −1)

1−

n N

H h X Nh2 2 (µh − µ) ≥ nh 1 − h=1

nh Nh

−

N 2 (Nh −1) n(N −1)

1−

n N

i 2 σh .

199

3.4. Unbiased Estimators in Survey Problems

This means that stratified simple random sampling is better than simple nh n random sampling if the deviations µh − µ are sufficiently large. If N ≡N h (proportional allocation), then this condition simplifies to H X

h=1

H X Nh σh2 , 1− Nh (µh − µ) ≥ N 2

(3.46)

h=1

which is usually true when µh ’s are different and some Nh ’s are large. Note that the variances Var(Yˆ ) and Var(Yˆst ) are w.r.t. different sampling plans under which Yˆ and Yˆst are obtained.

3.4.2 Horvitz-Thompson estimators If some elements of the finite population P are groups (called clusters) of subunits, then sampling from P is cluster sampling. Cluster sampling is used often because of administrative convenience or economic considerations. Although sometimes the first intention may be to use the subunits as sampling units, it is found that no reliable list of the subunits in the population is available. For example, in many countries there are no complete lists of the people or houses in a region. From the maps of the region, however, it can be divided into units such as cities or blocks in the cities. In cluster sampling, one may greatly increase the precision of estimation by using sampling with probability proportional to cluster size. Thus, unequal probability sampling is often used. Suppose that a sample of clusters is obtained. If subunits within a selected cluster give similar results, then it may be uneconomical to measure them all. A sample of the subunits in any chosen cluster may be selected. This is called two-stage sampling. One can continue this process to have a multistage sampling (e.g., cities → blocks → houses → people). Of course, at each stage one may use stratified sampling and/or unequal probability sampling. When the sampling plan is complex, so is the structure of the observations. We now introduce a general method of deriving unbiased estimators of population totals, which are called Horvitz-Thompson estimators. Theorem 3.15. Let X = {yi , i ∈ s} denote a sample from P = {y1 , ..., yN } that is selected, without replacement, by some method. Define πi = probability that i ∈ s,

i = 1, ..., N.

(i) (Horvitz-Thompson). If πi > 0 for i = 1, ..., N and πi is known when P i ∈ s, then Yˆht = i∈s yi /πi is an unbiased estimator of the population

200

3. Unbiased Estimation

total Y . (ii) Define πij = probability that i ∈ s and j ∈ s,

i = 1, ..., N, j = 1, ..., N.

Then Var(Yˆht ) =

N X 1 − πi i=1

=

πi

N N X X

i=1 j=i+1

yi2 + 2

N N X X πij − πi πj yi yj πi πj i=1 j=i+1

(πi πj − πij )

yj yi − πi πj

2

.

(3.47)

(3.48)

Proof. (i) Let ai = 1 if i ∈ s and ai = 0 if i 6∈ s, i = 1, ..., N . Then E(ai ) = πi and ! N N X X ai y i ˆ = yi = Y. E(Yht ) = E πi i=1 i=1 (ii) Since a2i = ai , Var(ai ) = E(ai ) − [E(ai )]2 = πi (1 − πi ). For i 6= j, Cov(ai , aj ) = E(ai aj ) − E(ai )E(aj ) = πij − πi πj . Then Var(Yˆht ) = Var

N X ai y i i=1

=

N X i=1

=

N N X X yi2 yi yj Var(ai ) + 2 Cov(ai , aj ) πi2 π π i=1 j=i+1 i j

N X 1 − πi i=1

πi

!

πi

yi2

N N X X πij − πi πj +2 yi yj . πi πj i=1 j=i+1

Hence (3.47) follows. To show (3.48), note that N X

πi = n

i=1

which implies X

and

X

j=1,...,N,j6=i

πij = (n − 1)πi ,

(πij − πi πj ) = (n − 1)πi − πi (n − πi ) = −πi (1 − πi ).

j=1,...,N,j6=i

201

3.4. Unbiased Estimators in Survey Problems

Hence N X 1 − πi

πi

i=1

yi2 =

N X

X

(πi πj − πij )

i=1 j=1,...,N,j6=i

=

N N X X

i=1 j=i+1

(πi πj − πij )

yi2 πi2

yj2 yi2 + πi2 πj2

!

and, by (3.47), Var(Yˆht ) =

N N X X

i=1 j=i+1

=

N N X X

i=1 j=i+1

yj2 2yi yj yi2 + 2− 2 πi πj πi πj

(πij − πi πj ) (πi πj − πij )

yi yj − πi πj

2

!

.

Using the same idea, we can obtain unbiased estimators of Var(Yˆht ). Suppose that πij > 0 for all i and j and πij is known when i ∈ s and j ∈ s. By (3.47), an unbiased estimator of Var(Yˆht ) is v1 =

X 1 − πi i∈s

πi2

yi2 + 2

X X

πij − πi πj yi yj . πi πj πij i∈s j∈s,j>i

(3.49)

By (3.48), an unbiased estimator of Var(Yˆht ) is v2 =

X X

πi πj − πij πij i∈s j∈s,j>i

yi yj − πi πj

2

.

(3.50)

Variance estimators v1 and v2 may not be the same in general, but they are the same in some special cases (Exercise 92). A more serious problem is that they may take negative values. Some discussions about deriving better estimators of Var(Yˆht ) are provided in Cochran (1977, Chapter 9A). Some special cases of Theorem 3.15 are considered as follows. Under simple random sampling, πi = n/N . Thus, Yˆ in Example 3.19 is the Horvitz-Thompson estimator. Under stratified simple random sampling, πi = nh /Nh if unit i is in stratum h. Hence, the estimator Yˆst in Example 3.20 is the Horvitz-Thompson estimator. Suppose now each yi ∈ P is a cluster, i.e., yi = (yi1 , ..., yiMi ), where Mi is the size ofP the ith cluster, i = 1, ..., N . The total number of units in N P is then M = i=1 Mi . Consider a single-stage sampling plan, i.e., if yi is selected, then every yij is observed. If simple random sampling is used,

202

3. Unbiased Estimation

then πi = k/N , where k is the first-stage sample size (the total sample size Pk is n = i=1 Mi ), and the Horvitz-Thompson estimator is Mi N XX N X Yˆs = yij = Yi , k i∈s j=1 k i∈s 1 1

where s1 is the index set of first-stage sampled clusters and Yi is the total of the ith cluster. In this case, Var(Yˆs ) =

N2 k(N − 1)

N 2 k X Y 1− Yi − . N i=1 N

If the selection probability is proportional to the cluster size, then πi = kMi /M and the Horvitz-Thompson estimator is Mi M X 1 X M X Yi yij = Yˆpps = k i∈s Mi j=1 k i∈s Mi 1 1

whose variance is given by (3.47) or (3.48). Usually Var(Yˆpps ) is smaller than Var(Yˆs ); see the discussions in Cochran (1977, Chapter 9A). Consider next a two-stage sampling in which k first-stage clusters are selected and a simple random sample of size mi is selected from each sampled cluster yi , where sampling is independent across clusters. If the first-stage sampling plan is simple random sampling, then πi = kmi /(N Mi ) and the Horvitz-Thompson estimator is N X Mi X Yˆs = yij , k i∈s mi j∈s 1 2i where s2i denotes the second-stage sample from cluster i. If the first-stage selection probability is proportional to the cluster size, then πi = kmi /M and the Horvitz-Thompson estimator is M X 1 X Yˆpps = yij . k i∈s mi j∈s 1 2i Finally, let us consider another popular sampling method called systematic sampling. Suppose that P = {y1 , ..., yN } and the population size N = nk for two integers n and k. To select a sample of size n, we first draw a j randomly from {1, ..., k}. Our sample is then {yj , yj+k , yj+2k , ..., yj+(n−1)k }.

203

3.4. Unbiased Estimators in Survey Problems

Systematic sampling is used mainly because it is easier to draw a systematic sample and often easier to execute without mistakes. It is also likely that systematic sampling provides more efficient point estimators than simple random sampling or even stratified sampling, since the sample units are spread more evenly over the population. Under systematic sampling, πi = k −1 for every i and the Horvitz-Thompson estimator of the population total is n X yj+(t−1)k . Yˆsy = k t=1

The unbiasedness of this estimator is a direct consequence of Theorem 3.15, but it can be easily shown as follows. Since j takes value i ∈ {1, ..., k} with probability k −1 , ! n k X N X X 1 E(Yˆsy ) = k yi+(t−1)k = yi = Y. k i=1 t=1 i=1

The variance of Yˆsy is simply Var(Yˆsy ) =

k N2 X (µi − µ)2 , k i=1

Pn Pk where µi = n−1 t=1 yi+(t−1)k and µ = k −1 i=1 µi = Y /N . Let σ 2 be given in (3.43) and k

2 σsy =

Then (N − 1)σ 2 = n Thus, and

n

XX 1 (yi+(t−1)k − µi )2 . k(n − 1) i=1 t=1 k X i=1

(µi − µ)2 +

n k X X i=1 t=1

(yi+(t−1)k − µi )2 .

2 (N − 1)σ 2 = N −1 Var(Yˆsy ) + k(n − 1)σsy 2 Var(Yˆsy ) = N (N − 1)σ 2 − N (N − k)σsy .

Since the variance of the Horvitz-Thompson estimator of the population total under simple random sampling is, by (3.43), N2 n 2 1− σ = N (k − 1)σ 2 , n N

the Horvitz-Thompson estimator under systematic sampling has a smaller 2 variance if and only if σsy > σ2 .

204

3. Unbiased Estimation

3.5 Asymptotically Unbiased Estimators As we discussed in §2.5, we often need to consider biased but asymptotically unbiased estimators. A large and useful class of such estimators are smooth functions of some exactly unbiased estimators such as UMVUE’s, U-statistics, LSE’s, and Horvitz-Thompson estimators. Some other methods of constructing asymptotically unbiased estimators are also introduced in this section.

3.5.1 Functions of unbiased estimators If the parameter to be estimated is ϑ = g(θ) with a vector-valued parameter θ and Un is a vector of unbiased estimators of components of θ (i.e., EUn = θ), then Tn = g(Un ) is often asymptotically unbiased for ϑ. Assume that g is differentiable and cn (Un − θ) →d Y . Then amseTn (P ) = E{[∇g(θ)]τ Y }2 /c2n

(Theorem 2.6). Hence, Tn has a good performance in terms of amse if Un is optimal in terms of mse (such as the UMVUE). The following are some examples. Example 3.21 (Ratio estimators). Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors with EX1 = µx and EY1 = µy . Consider the estimation of ¯ the ratio of two population means: ϑ = µy /µx (µx 6= 0). Note that (Y¯ , X), the vector of sample means, is unbiased for (µy , µx ). The sample means are UMVUE’s under some statistical models (§3.1 and §3.2) and are BLUE’s ¯ Assume in general (Example 2.22). The ratio estimator is Tn = Y¯ /X. that σx2 = Var(X1 ), σy2 = Var(Y1 ), and σxy = Cov(X1 , Y1 ) exist. A direct calculation shows that the n−1 order asymptotic bias of Tn according to (2.38) is 2 ˜bT (P ) = ϑσx − σxy n µ2x n (verify). Using the CLT and the delta-method (Corollary 1.1), we obtain that ! σy2 − 2ϑσxy + ϑ2 σx2 √ n(Tn − ϑ) →d N 0, µ2x (verify), which implies amseTn (P ) =

σy2 − 2ϑσxy + ϑ2 σx2 . µ2x n

In some problems, we are not interested in the ratio, but the use of a ratio estimator to improve an estimator of a marginal mean. For example,

205

3.5. Asymptotically Unbiased Estimators

suppose that µx is known and we are interested in estimating µy . Consider the following estimator: ¯ x. µ ˆy = (Y¯ /X)µ Note that µ ˆy is not unbiased; its n−1 order asymptotic bias according to (2.38) is 2 ˜bµˆ (P ) = ϑσx − σxy ; y µx n and

σy2 − 2ϑσxy + ϑ2 σx2 . n ˆy is asympComparing µ ˆy with the unbiased estimator Y¯ , we find that µ totically more efficient if and only if amseµˆy (P ) =

2ϑσxy > ϑ2 σx2 , which means that µ ˆy is a better estimator if and only if the correlation between X1 and Y1 is large enough to pay off the extra variability caused ¯ by using µx /X. Another example related to a bivariate sample is the sample correlation coefficient defined in Exercise 22 in §2.6. Example 3.22. Consider a polynomial regression of order p: Xi = β τ Zi + εi ,

i = 1, ..., n,

), and εi ’s are i.i.d. with where β = (β0 , β1 , ..., βp−1 ), Zi = (1, ti , ..., tp−1 i mean 0 and variance σ 2 > 0. Suppose that the parameter to be estimated is tβ ∈ T ⊂ R such that p−1 X

βj tjβ

= max

j=0

t∈T

p−1 X

βj t j .

j=0

Note that tβ = g(β) for some function g. Let βˆ be the LSE of β. Then the ˆ is asymptotically unbiased and its amse can be derived estimator tˆβ = g(β) under some conditions (Exercise 98). Example 3.23. In the study of the reliability of a system component, we assume that Xij = θ τi z(tj ) + εij ,

i = 1, ..., k, j = 1, ..., m.

Here Xij is the measurement of the ith sample component at time tj ; z(t) is a q-vector whose components are known functions of the time t; θ i ’s

206

3. Unbiased Estimation

are unobservable random q-vectors that are i.i.d. from Nq (θ, Σ), where θ and Σ are unknown; εij ’s are i.i.d. measurement errors with mean zero and variance σ 2 ; and θi ’s and εij ’s are independent. As a function of t, θτ z(t) is the degradation curve for a particular component and θτ z(t) is the mean degradation curve. Suppose that a component will fail to work if θτ z(t) < η, a given critical value. Assume that θτ z(t) is always a decreasing function of t. Then the reliability function of a component is τ θ z(t) − η τ R(t) = P (θ z(t) > η) = Φ , s(t) p where s(t) = [z(t)]τ Σz(t) and Φ is the standard normal distribution function. For a fixed t, estimators of R(t) can be obtained by estimating θ and Σ, since Φ is a known function. It can be shown (exercise) that the BLUE of θ is the LSE ¯ θˆ = (Z τ Z)−1 Z τ X, where Z is the m × q matrix whose jth row is the vector z(tj ), Xi = ¯ is the sample mean of Xi ’s. The estimation of Σ is (Xi1 , ..., Xim ), and X more difficult. It can be shown (exercise) that a consistent (as k → ∞) estimator of Σ is k X τ −1 ˆ= 1 ¯ ¯ τ Σ (Z τ Z)−1 Z τ (Xi − X)(X −σ ˆ 2 (Z τ Z)−1 , i − X) Z(Z Z) k i=1

where

k

σ ˆ2 =

X 1 [X τ Xi − Xiτ Z(Z τ Z)−1 Z τ Xi ]. k(m − q) i=1 i

Hence an estimator of R(t) is

ˆ =Φ R(t) where sˆ(t) =

θˆτ z(t) − η sˆ(t)

!

,

q ˆ [z(t)]τ Σz(t).

If we define Yi1 = Xiτ Z(Z τ Z)−1 z(t), Yi2 = [Xiτ Z(Z τ Z)−1 z(t)]2 , Yi3 = [Xiτ Xi − Xiτ Z(Z τ Z)−1 Z τ Xi ]/(m − q), and Yi = (Yi1 , Yi2 , Yi3 ), then it is ˆ apparent that R(t) can be written as g(Y¯ ) for a function ! y1 − η . g(y1 , y2 , y3 ) = Φ p y2 − y12 − y3 [z(t)]τ (Z τ Z)−1 z(t)

Suppose that εij has a finite fourth moment, which implies the existence of ˆ can be derived (exercise). Var(Yi ). The amse of R(t)

3.5. Asymptotically Unbiased Estimators

207

3.5.2 The method of moments The method of moments is the oldest method of deriving point estimators. It almost always produces some asymptotically unbiased estimators, although they may not be the best estimators. Consider a parametric problem where X1 , ..., Xn are i.i.d. random variables from Pθ , θ ∈ Θ ⊂ Rk , and E|X1 |k < ∞. Let µj = EX1j be the jth moment of P and let n 1X j µ ˆj = X n i=1 i

be the jth sample moment, which is an unbiased estimator of µj , j = 1, ..., k. Typically, µj = hj (θ), j = 1, ..., k, (3.51)

for some functions hj on Rk . By substituting µj ’s on the left-hand side of ˆ i.e., θˆ (3.51) by the sample moments µ ˆj , we obtain a moment estimator θ, satisfies ˆ j = 1, ..., k, µ ˆ j = hj (θ), which is a sample analogue of (3.51). This method of deriving estimators is called the method of moments. Note that an important statistical principle, the substitution principle, is applied in this method. ˆ If the inverse Let µ ˆ = (ˆ µ1 , ..., µ ˆk ) and h = (h1 , ..., hk ). Then µ ˆ = h(θ). −1 µ). function h exists, then the unique moment estimator of θ is θˆ = h−1 (ˆ ˆ ˆ = h(θ) When h−1 does not exist (i.e., h is not one-to-one), any solution of µ ˆ is a moment estimator of θ; if possible, we always choose a solution θ in the parameter space Θ. In some cases, however, a moment estimator does not exist (see Exercise 111). Assume that θˆ = g(ˆ µ) for a function g. If h−1 exists, then g = h−1 . If g is continuous at µ = (µ1 , ..., µk ), then θˆ is strongly consistent for θ, since µ ˆj →a.s. µj by the SLLN. If g is differentiable at µ and E|X1 |2k < ∞, then θˆ is asymptotically normal, by the CLT and Theorem 1.12, and amseθˆ(θ) = n−1 [∇g(µ)]τ Vµ ∇g(µ), where Vµ is a k × k matrix whose (i, j)th element is µi+j − µi µj . Furthermore, it follows from (2.38) that the n−1 order asymptotic bias of θˆ is (2n)−1 tr ∇2 g(µ)Vµ .

Example 3.24. Let X1 , ..., Xn be i.i.d. from a population Pθ indexed by the parameter θ = (µ, σ 2 ), where µ = EX1 ∈ R and σ 2 = Var(X1 ) ∈ (0, ∞). This includes cases such as the family of normal distributions,

208

3. Unbiased Estimation

double exponential distributions, or logistic distributions (Table 1.2, page 20). Since EX1 = µ and EX12 = Var(X1 ) + (EX1 )2 = σ 2 + µ2 , setting µ ˆ1 = µ and µ ˆ2 = σ 2 + µ2 we obtain the moment estimator ! n n−1 2 1X 2 ˆ ¯ ¯ ¯ S . = X, θ = X, (Xi − X) n i=1 n ¯ is unbiased, but n−1 S 2 is not. If Xi is normal, then θˆ is suffiNote that X n cient and is nearly the same as an optimal estimator such as the UMVUE. On the other hand, if Xi is from a double exponential or logistic distribution, then θˆ is not sufficient and can often be improved. Consider now the estimation of σ 2 when we know that µ = 0. Obviously we cannot use the equation µ ˆ1 = µ to solve the problem.PUsing µ ˆ2 = µ2 = n σ 2 , we obtain the moment estimator σ ˆ2 = µ ˆ2 = n−1 i=1 Xi2 . This is still a good estimator when Xi is normal, but is not a function of sufficient statistic when Xi is from a double exponential distribution. For the double exponential case one can argue that we should first make a transformation Yi = |Xi | and then obtain the moment estimator based on the transformed data.P The moment estimator of σ 2 based on the transformed data is Y¯ 2 = n −1 2 2 (n i=1 |Xi |) , which is sufficient for σ . Note that this estimator can also be obtained based on absolute moment equations. Example 3.25. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (θ1 , θ2 ), −∞ < θ1 < θ2 < ∞. Note that EX1 = (θ1 + θ2 )/2 and EX12 = (θ12 + θ22 + θ1 θ2 )/3. ˆ2 = EX12 and substituting θ1 in the second equaSetting µ ˆ1 = EX1 and µ tion by 2ˆ µ1 − θ2 (the first equation), we obtain that µ1 − θ2 )θ2 = 3ˆ µ2 , (2ˆ µ1 − θ2 )2 + θ22 + (2ˆ which is the same as ˆ1 )2 = 3(ˆ µ2 − µ ˆ21 ). (θ2 − µ Since θ2 > EX1 , we obtain that q q ¯ + 3(n−1) S 2 ˆ1 + 3(ˆ µ2 − µ ˆ21 ) = X θˆ2 = µ n and

θˆ1 = µ ˆ1 −

q q ¯ − 3(n−1) S 2 . 3(ˆ µ2 − µ ˆ21 ) = X n

209

3.5. Asymptotically Unbiased Estimators

These estimators are not functions of the sufficient and complete statistic (X(1) , X(n) ). Example 3.26. Let X1 , ..., Xn be i.i.d. from the binomial distribution Bi(p, k) with unknown parameters k ∈ {1, 2, ...} and p ∈ (0, 1). Since EX1 = kp and EX12 = kp(1 − p) + k 2 p2 , we obtain the moment estimators ˆ21 − µ ˆ2 )/ˆ µ1 = 1 − pˆ = (ˆ µ1 + µ and

n−1 2 ¯ n S /X

¯ kˆ = µ ˆ21 /(ˆ µ1 + µ ˆ21 − µ ˆ2 ) = X/(1 −

n−1 2 ¯ n S /X).

The estimator pˆ is in the range of (0, 1). But kˆ may not be an integer. It can be improved by an estimator that is kˆ rounded to the nearest positive integer. Example 3.27. Suppose that X1 , ..., Xn are i.i.d. from the Pareto distribution P a(a, θ) with unknown a > 0 and θ > 2 (Table 1.2, page 20). Note that EX1 = θa/(θ − 1) and EX12 = θa2 /(θ − 2). From the moment equation, (θ−1)2 θ(θ−2)

Note that

(θ−1)2 θ(θ−2)

−1=

1 θ(θ−2) .

=µ ˆ 2 /ˆ µ21 .

Hence

µ2 − µ ˆ21 ). θ(θ − 2) = µ ˆ21 /(ˆ Since θ > 2, there is a unique solution in the parameter space: q q n ¯2 ˆ2 /(ˆ µ2 − µ ˆ21 ) = 1 + 1 + n−1 X /S 2 θˆ = 1 + µ and

µ ˆ 1 (θˆ − 1) ˆ q qθ . ¯ 1+ n X ¯ 2 /S 2 1 + 1+ =X n−1

a ˆ=

n ¯2 2 n−1 X /S

.

210

3. Unbiased Estimation

The method of moments can also be applied to nonparametric problems. Consider, for example, the estimation of the central moments cj = E(X1 − µ1 )j , Since cj =

j = 2, ..., k.

j X j (−µ1 )t µj−t , t t=0

the moment estimator of cj is cˆj =

j X j ¯ tµ (−X) ˆj−t , t t=0

where µ ˆ0 = 1. It can be shown (exercise) that n

cˆj =

1X ¯ j, (Xi − X) n i=1

j = 2, ..., k,

(3.52)

which are sample central moments. From the SLLN, cˆj ’s are strongly consistent. If E|X1 |2k < ∞, then √ n (ˆ c2 − c2 , ..., cˆk − ck ) →d Nk−1 (0, D) (3.53) (exercise), where the (i, j)th element of the (k − 1) × (k − 1) matrix D is ci+j+2 − ci+1 cj+1 − (i + 1)ci cj+2 − (j + 1)ci+2 cj + (i + 1)(j + 1)ci cj c2 .

3.5.3 V-statistics Let X1 , ..., Xn be i.i.d. from P . For every U-statistic Un defined in (3.11) as an estimator of ϑ = E[h(X1 , ..., Xm )], there is a closely related V-statistic defined by n n X 1 X Vn = m ··· h(Xi1 , ..., Xim ). (3.54) n i =1 i =1 1

m

As an estimator of ϑ, Vn is biased; but the bias is small asymptotically as the following results show. For a fixed sample size n, Vn may be better than Un in terms of their mse’s. Consider, for example, the kernel h(x1 , x2 ) = (x1 − x2 )2 /2 in §3.2.1, which leads to ϑ = σ 2 = Var(X1 ) and Un = S 2 , the sample variance. The corresponding V-statistic is n n 1 X X (Xi − Xj )2 1 = 2 n2 i=1 j=1 2 n

X

1≤i<j≤n

(Xi − Xj )2 =

n−1 2 S , n

211

3.5. Asymptotically Unbiased Estimators

which is the moment estimator of σ 2 discussed in Example 3.24. In Exercise 2 2 63 in §2.6, n−1 n S is shown to have a smaller mse than S when Xi is normally distributed. Of course, there are situations where U-statistics are better than their corresponding V-statistics. The following result provides orders of magnitude of the bias and variance of a V-statistic as an estimator of ϑ. Proposition 3.5. Let Vn be defined by (3.54). (i) Assume that E|h(Xi1 , ..., Xim )| < ∞ for all 1 ≤ i1 ≤ · · · ≤ im ≤ m. Then the bias of Vn satisfies bVn (P ) = O(n−1 ). (ii) Assume that E[h(Xi1 , ..., Xim )]2 < ∞ for all 1 ≤ i1 ≤ · · · ≤ im ≤ m. Then the variance of Vn satisfies Var(Vn ) = Var(Un ) + O(n−2 ), where Un is given by (3.11). Proof. (i) Note that h Un − Vn = 1 −

n!

nm (n−m)!

i

(Un − Wn ),

(3.55)

where Wn is the average of all terms h(Xi1 , ..., Xim ) with at least one equality im = il , m 6= l. The result follows from E(Un − Wn ) = O(1). (ii) The result follows from E(Un − Wn )2 = O(1), E[Wn (Un − ϑ)] = O(n−1 ) (exercise), and (3.55). To study the asymptotic behavior of a V-statistic, we consider the following representation of Vn in (3.54): m X m Vnj , Vn = j j=1 where Vnj = ϑ +

n n X 1 X · · · gj (Xi1 , ..., Xij ) nj i =1 i =1 1

is a “V-statistic” with

j

gj (x1 , ..., xj ) = hj (x1 , ..., xj ) − +

X

1≤i1 0, then √ n(Vn − ϑ) →d N (0, m2 ζ1 ). (ii) If ζ1 = 0 but ζ2 = Var(h2 (X1 , X2 )) > 0, then n(Vn − ϑ) →d

∞

m(m − 1) X λj χ21j , 2 j=1

where χ21j ’s and λj ’s are the same as those in (3.21). Result (3.57) and Theorem 3.16 imply that Vn has expansion (2.37) −1 and, therefore, the nP order asymptotic bias of Vn is E[g2 (X1 , X1 )]/n = ∞ nEVn2 = m(m − 1) j=1 λj /(2n) (exercise). Theorem 3.16 shows that if ζ1 > 0, then the amse’s of Un and Vn are the same. If ζ1 = 0 but ζ2 > 0, then an argument similar to that in the proof of Lemma 3.2 leads to 2 ∞ m2 (m − 1)2 ζ2 m2 (m − 1)2 X amseVn (P ) = + λj 2n2 4n2 j=1 2 ∞ X m (m − 1) = amseUn (P ) + λj 4n2 j=1 2

2

(see Lemma 3.2). Hence Un is asymptotically more efficient than Vn , unless P ∞ j=1 λj = 0. Technically, the proof of the asymptotic results for Vn also requires moment conditions stronger than those for Un . Example 3.28. Consider the estimation ofP µ2 , where µ = EX1 . From the 1 results in §3.2, the U-statistic Un = n(n−1) 1≤i<j≤n Xi Xj is unbiased for

3.5. Asymptotically Unbiased Estimators

213

¯ 2 . If µ 6= 0, then ζ1 6= 0 µ2 . The corresponding V-statistic is simply Vn = X and the asymptotic relative efficiency of Vn w.r.t. Un is 1. If µ = 0, then nVn →d σ 2 χ21

and

nUn →d σ 2 (χ21 − 1),

where χ21 is a random variable having the chi-square distribution χ21 . Hence the asymptotic relative efficiency of Vn w.r.t. Un is E(χ21 − 1)2 /E(χ21 )2 = 2/3.

3.5.4 The weighted LSE In linear model (3.25), the unbiased LSE of lτ β may be improved by a slightly biased estimator when Var(ε) is not σ 2 In and the LSE is not BLUE. Assume that Z in (3.25) is of full rank so that every lτ β is estimable. For simplicity, let us denote Var(ε) by V . If V is known, then the BLUE ˘ where of lτ β is lτ β, β˘ = (Z τ V −1 Z)−1 Z τ V −1 X

(3.58)

(see the discussion after the statement of assumption A3 in §3.3.1). If V is unknown and Vˆ is an estimator of V , then an application of the substitution principle leads to a weighted least squares estimator βˆw = (Z τ Vˆ −1 Z)−1 Z τ Vˆ −1 X.

(3.59)

The weighted LSE is not linear in X and not necessarily unbiased for β. If the distribution of ε is symmetric about 0 and Vˆ remains unchanged when ε changes to −ε (Examples 3.29 and 3.30), then the distribution of βˆw − β is symmetric about 0 and, if E βˆw is well defined, βˆw is unbiased for β. In such a case the LSE lτ βˆ may not be a UMVUE (when ε is normal), since ˆ Var(lτ βˆw ) may be smaller than Var(lτ β). Asymptotic properties of the weighted LSE depend on the asymptotic behavior of Vˆ . We say that Vˆ is consistent for V if and only if kVˆ −1 V − In kmax →p 0,

(3.60)

where kAkmax = maxi,j |aij | for a matrix A whose (i, j)th element is aij . Theorem 3.17. Consider model (3.25) with a full rank Z. Let β˘ and βˆw be defined by (3.58) and (3.59), respectively, with a Vˆ consistent in the sense of (3.60). Assume the conditions in Theorem 3.12. Then lτ (βˆw − β)/an →d N (0, 1),

214

3. Unbiased Estimation

where l ∈ Rp , l 6= 0, and ˘ = lτ (Z τ V −1 Z)−1 l. a2n = Var(lτ β) Proof. Using the same argument as in the proof of Theorem 3.12, we obtain that lτ (β˘ − β)/an →d N (0, 1). By Slutsky’s theorem, the result follows from lτ βˆw − lτ β˘ = op (an ). Define and

(3.61)

ξn = lτ (Z τ Vˆ −1 Z)−1 Z τ (Vˆ −1 − V −1 )ε ζn = lτ [(Z τ Vˆ −1 Z)−1 − (Z τ V −1 Z)−1 ]Z τ V −1 ε.

Then

lτ βˆw − lτ β˘ = ξn + ζn .

Let Bn = (Z τ Vˆ −1 Z)−1 Z τ V −1 Z − Ip and Cn = Vˆ 1/2p V −1 Vˆ 1/2 − In . By (3.60), kCn kmax = op (1). For any matrix A, denote tr(Aτ A) by kAk. Then kBn k2 = k(Z τ Vˆ −1 Z)−1 Z τ Vˆ −1/2 Cn Vˆ −1/2 Zk2 = tr (Z τ Vˆ −1 Z)−1 (Z τ Vˆ −1/2 Cn Vˆ −1/2 Z)2 (Z τ Vˆ −1 Z)−1 ≤ kCn k2max tr (Z τ Vˆ −1 Z)−1 (Z τ Vˆ −1 Z)2 (Z τ Vˆ −1 Z)−1 = op (1)tr(Ip ).

This proves that kBn kmax = op (1). Let An = V 1/2 Vˆ −1 V 1/2 − In . Using inequality (1.37) and the previous results, we obtain that ξn2 = [lτ (Z τ Vˆ −1 Z)−1 Z τ V −1/2 An V −1/2 ε]2 ≤ εV −1 ετ lτ (Z τ Vˆ −1 Z)−1 Z τ V −1/2 A2n V −1/2 Z(Z τ Vˆ −1 Z)−1 l ≤ Op (1)kAn k2 lτ (Z τ Vˆ −1 Z)−1 Z τ V −1 Z(Z τ Vˆ −1 Z)−1 l max

= op (1)lτ (Bn + Ip )2 (Z τ V −1 Z)−1 l

= op (an ). Since Ek(Z τ V −1 Z)−1/2 Z τ V −1 εk2 = p, k(Z τ V −1 Z)−1/2 Z τ V −1 εk = Op (1). Define B1n = (Z τ Vˆ −1 Z)1/2 Bn (Z τ Vˆ −1 Z)−1/2 . Then B1n = (Z τ Vˆ −1 Z)−1/2 Z τ Vˆ −1/2 Cn Vˆ −1/2 Z(Z τ Vˆ −1 Z)−1/2 ≤ kCn kmax (Z τ Vˆ −1 Z)−1/2 Z τ Vˆ −1 Z(Z τ Vˆ −1 Z)−1/2 = op (1)Ip .

215

3.5. Asymptotically Unbiased Estimators Let B2n = (Z τ V −1 Z)1/2 (Z τ Vˆ −1 Z)−1/2 . Since kB2n k2 = tr (Z τ V −1 Z)1/2 (Z τ Vˆ −1 Z)−1 (Z τ V −1 Z)1/2 = tr (Z τ Vˆ −1 Z)−1 Z τ V −1 Z = tr(Bn + Ip ) = p + op (1),

we obtain that τ k = op (1). kB2n B1n B2n

Then ζn2 = [lτ Bn (Z τ V −1 Z)−1 Z τ V −1 ε]2 τ = [lτ (Z τ V −1 Z)−1/2 B2n B1n B2n (Z τ V −1 Z)−1/2 Z τ V −1 ε]2 τ 2 ≤ lτ (Z τ V −1 Z)−1 lkB2n B1n B2n k k(Z τ V −1 Z)−1/2 Z τ V −1 εk2 = op (a2n ).

This proves (3.61) and thus completes the proof. Theorem 3.17 shows that as long as Vˆ is consistent in the sense of (3.60), ˘ which is the BLUE the weighted LSE βˆw is asymptotically as efficient as β, ˘ attains the if V is known. If V is known and ε is normal, then Var(lτ β) Cram´er-Rao lower bound (Proposition 3.2) and, thus, (3.10) holds with Tn = lτ βˆw . By Theorems 3.12 and 3.17, the asymptotic relative efficiency of the LSE lτ βˆ w.r.t. the weighted LSE lτ βˆw is lτ (Z τ V −1 Z)−1 l , Z(Z τ Z)−1 l

lτ (Z τ Z)−1 Z τ V

which is always less than 1 and equals 1 if lτ βˆ is a BLUE (in which case ˘ βˆ = β). Finding a consistent Vˆ is possible when V has a certain type of structure. We consider three examples. Example 3.29. Consider model (3.25). Suppose that V = Var(ε) is a block diagonal matrix with the ith diagonal block σ 2 Imi + Ui ΣUiτ ,

i = 1, ..., k,

(3.62)

where mi ’s are integers bounded by a fixed integer m, σ 2 > 0 is an unknown parameter, Σ is a q×q unknown nonnegative definite matrix, Ui is an mi ×q

216

3. Unbiased Estimation

full rank matrix whose columns are in R(Wi ), q < inf i mi , and Wi is the p × mi matrix such that Z τ = ( W1 W2 ... Wk ). Under (3.62), a consistent Vˆ can be obtained if we can obtain consistent estimators of σ 2 and Σ. Let X = (Y1 , ..., Yk ), where Yi is an mi -vector, and let Ri be the matrix whose columns are linearly independent rows of Wi . Then σ ˆ2 =

k 1 X τ Y [Imi − Ri (Riτ Ri )−1 Riτ ]Yi n − kq i=1 i

(3.63)

is an unbiased estimator of σ 2 . Assume that Yi ’s are independent and that supi E|εi |2+δ < ∞ for some δ > 0. Then σ ˆ 2 is consistent for σ 2 (exercise). τˆ Let ri = Yi − Wi β and k X τ ˆ= 1 ˆ 2 (Uiτ Ui )−1 . (Ui Ui )−1 Uiτ ri riτ Ui (Uiτ Ui )−1 − σ Σ k i=1

(3.64)

ˆ is consistent for Σ in the sense that It can be shown (exercise) that Σ ˆ ˆ − Σk →p 0 (see Exercise 116). kΣ − Σkmax →p 0 or, equivalently, kΣ Example 3.30. Suppose that V is a block diagonal matrix with the ith diagonal block matrix Vmi , i = 1, ..., k, where Vt is an unknown t × t matrix and mi ∈ {1, ..., m} with a fixed positive integer m. Thus, we need to obtain consistent estimators of at most m different matrices V1 , ..., Vm . It can be shown (exercise) that the following estimator is consistent for Vt when kt → ∞ as k → ∞: 1 X Vˆt = ri riτ , kt

t = 1, ..., m,

i∈Bt

where ri is the same as that in Example 3.29, Bt is the set of i’s such that mi = t, and kt is the number of i’s in Bt . Example 3.31. Suppose that V is diagonal with the ith diagonal element σi2 = ψ(Zi ), where ψ is an unknown function. The simplest case is ψ(t) = θ0 + θ1 v(Zi ) for a known function v and some unknown θ0 and θ1 . One can then obtain a consistent estimator Vˆ by using the LSE of θ0 and θ1 under the “model” Eri2 = θ0 + θ1 v(Zi ), i = 1, ..., n, (3.65) where ri = Xi − Ziτ βˆ (exercise). If ψ is nonlinear or nonparametric, some results are given in Carroll (1982) and M¨ uller and Stadrm¨ uller (1987). Finally, if Vˆ is not consistent (i.e., (3.60) does not hold), then the weighted LSE lτ βˆw can still be consistent and asymptotically normal, but

217

3.6. Exercises

its asymptotic variance is not lτ (Z τ V −1 Z)−1 l; in fact, lτ βˆw may not be asymptotically as efficient as the LSE lτ βˆ (Carroll and Cline, 1988; Chen and Shao, 1993). For example, if kVˆ −1 U − In kmax →p 0, where U is positive definite, 0 < inf n λ− [U ] ≤ supn λ+ [U ] < ∞, and U 6= V (i.e., Vˆ is inconsistent for V ), then, using the same argument as that in the proof of Theorem 3.17, we can show (exercise) that lτ (βˆw − β)/bn →d N (0, 1)

(3.66)

for any l 6= 0, where b2n = lτ (Z τ U −1 Z)−1 Z τ U −1 V U −1 Z(Z τ U −1 Z)−1 l. Hence, the asymptotic relative efficiency of the LSE lτ βˆ w.r.t. lτ βˆw can be less than 1 or larger than 1.

3.6 Exercises 1. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = p ∈ (0, 1). (a) Find the UMVUE of pm , m ≤ n. (b) Find the UMVUE of P (X1 + · · · + Xm = k), where m and k are positive integers ≤ n. (c) Find the UMVUE of P (X1 + · · · + Xn−1 > Xn ). 2. Let X1 , ..., Xn be i.i.d. having the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 > 0. (a) Find the UMVUE’s of µ3 and µ4 . d (b) Find the UMVUE’s of P (X1 ≤ t) and dt P (X1 ≤ t) with a fixed t ∈ R. 3. In Example 3.4, (a) show that the UMVUE of σ r is kn−1,r S r , where r > 1 − n; ¯ (b) prove that (X1 − X)/S has the p.d.f. given by (3.1); ¯ (c) show that (X1 − X)/S →d N (0, 1) by using (i) the SLLN and (ii) Scheff´e’s theorem (Proposition 1.18). 4. Let X1 , ..., Xm be i.i.d. having the N (µx , σx2 ) distribution and let Y1 , ..., Yn be i.i.d. having the N (µy , σy2 ) distribution. Assume that Xi ’s and Yj ’s are independent. (a) Assume that µx ∈ R, µy ∈ R, σx2 > 0, and σy2 > 0. Find the UMVUE’s of µx − µy and (σx /σy )r , where r > 0 and r < n. (b) Assume that µx ∈ R, µy ∈ R, and σx2 = σy2 > 0. Find the UMVUE’s of σx2 and (µx − µy )/σx .

218

3. Unbiased Estimation (c) Assume that µx = µy ∈ R, σx2 > 0, σy2 > 0, and σx2 /σy2 = γ is known. Find the UMVUE of µx . (d) Assume that µx = µy ∈ R, σx2 > 0, and σy2 > 0. Show that a UMVUE of µx does not exist. (e) Assume that µx ∈ R, µy ∈ R, σx2 > 0, and σy2 > 0. Find the UMVUE of P (X1 ≤ Y1 ). (f) Repeat (e) under the assumption that σx = σy .

5. Let X1 , ..., Xn be i.i.d. having the uniform distribution on the interval (θ1 −θ2 , θ1 +θ2 ), where θ1 ∈ R, θ2 > 0, and n > 2. Find the UMVUE’s of θj , j = 1, 2, and θ1 /θ2 . 6. Let X1 , ..., Xn be i.i.d. having the exponential distribution E(a, θ) with parameters θ > 0 and a ∈ R. (a) Find the UMVUE of a when θ is known. (b) Find the UMVUE of θ when a is known. (c) Find the UMVUE’s of θ and a. (d) Assume that θ is known. Find the UMVUE of P (X1 ≥ t) and d dt P (X1 ≥ t) for a fixed t > a. (e) Find the UMVUE of P (X1 ≥ t) for a fixed t > a. 7. Let X1 , ..., Xn be i.i.d. having the Pareto distribution P a(a, θ) with θ > 0 and a > 0. (a) Find the UMVUE of θ when a is known. (b) Find the UMVUE of a when θ is known. (c) Find the UMVUE’s of a and θ. 8. Consider Exercise 52(a) of §2.6. Find the UMVUE of γ. 9. Let X1 , ..., Xm be i.i.d. having the exponential distribution E(ax , θx ) with θx > 0 and ax ∈ R and Y1 , ..., Yn be i.i.d. having the exponential distribution E(ay , θy ) with θy > 0 and ay ∈ R. Assume that Xi ’s and Yj ’s are independent. (a) Find the UMVUE’s of ax − ay and θx /θy . (b) Suppose that θx = θy but it is unknown. Find the UMVUE’s of θx and (ax − ay )/θx . (c) Suppose that ax = ay but it is unknown. Show that a UMVUE of ax does not exist. (d) Suppose that n = m and ax = ay = 0 and that our sample is (Z1 , ∆1 ), ..., (Zn , ∆n ), where Zi = min{Xi , Yi } and ∆i = 1 if Xi ≥ Yi and 0 otherwise, i = 1, ..., n. Find the UMVUE of θx − θy . 10. Let X1 , ..., Xm be i.i.d. having the uniform distribution U (0, θx ) and Y1 , ..., Yn be i.i.d. having the uniform distribution U (0, θy ). Suppose that Xi ’s and Yj ’s are independent and that θx > 0 and θy > 0. Find the UMVUE of θx /θy when n > 1.

219

3.6. Exercises

11. Let X be a random variable having the negative binomial distribution N B(p, r) with an unknown p ∈ (0, 1) and a known r. (a) Find the UMVUE of pt , t < r. (b) Find the UMVUE of Var(X). (c) Find the UMVUE of log p. 12. Let X1 , ..., Xn be i.i.d. random variables having the Poisson distribution P (θ) truncated at 0, i.e., P (Xi = x) = (eθ − 1)−1 θx /x!, x = 1, 2, ..., θ > 0. Find the UMVUE of θ when n = 1, 2. 13. Let X be a random variable having the negative binomial distribution N B(p, r) truncated at r, where r is known and p ∈ (0, 1) is unknown. Let k be a fixed positive integer > r. For r = 1, 2, 3, find the UMVUE of pk . 14. Let X1 , ..., Xn be i.i.d. having the log-distribution L(p) with an unknown p ∈ (0, 1). Let k be a fixed positive integer. (a) For n = 1, 2, 3, find the UMVUE of pk . (b) For n = 1, 2, 3, find the UMVUE of P (X = k). 15. Consider Exercise 43 of §2.6. (a) Show that the estimator U = 2(|X1 | − 41 )I{X1 6=0} is unbiased for θ. (b) Derive the UMVUE of θ. 16. Derive the UMVUE of p in Exercise 33 of §2.6. 17. Derive the UMVUE’s of θ and λ in Exercise 55 of §2.6, based on data X1 , ..., Xn . 18. Suppose that (X0 , X1 , ..., Xk ) has the multinomial distribution in ExPk ample 2.7 with pi ∈ (0, 1), j=0 pj = 1. Find the UMVUE of pr00 · · · prkk , where rj ’s are nonnegative integers with r0 + · · · + rk ≤ n. 19. Let Y1 , ..., Yn be i.i.d. from the uniform distribution U (0, θ) with an unknown θ ∈ (1, ∞). (a) Suppose that we only observe if Yi ≥ 1 Yi Xi = i = 1, ..., n. 1 if Yi < 1, Derive a UMVUE of θ. (b) Suppose that we only observe if Yi ≤ 1 Yi Xi = 1 if Yi > 1,

i = 1, ..., n.

Derive a UMVUE of the probability P (Y1 > 1).

220

3. Unbiased Estimation

20. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors distributed as bivariate normal with EXi = EYi = βzi , Var(Xi ) = Var(Yi ) = σ 2 , and Cov(Xi , Yi ) = ρσ 2 , i = 1, ..., n, where β ∈ R, σ > 0, and ρ ∈ (−1, 1) are unknown parameters, and zi ’s are known constants. (a) Obtain a UMVUE of β and calculate its variance. (b) Obtain a UMVUE of σ 2 and calculate its variance. 21. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors from a population P ∈ P that is the family of all bivariate populations with Lebesgue p.d.f.’s. (a) Show that the set of n pairs (Xi , Yi ) ordered according to the value of their first coordinate constitutes a sufficient and complete statistic for P ∈ P. (b) A statistic T is a function of the complete and sufficient statistic if and only if T is invariant of the n pairs. Pnunder permutation ¯ i − Y¯ ) is the UMVUE of (c) Show that (n − 1)−1 i=1 (Xi − X)(Y Cov(X1 , Y1 ). (d) Find the UMVUE’s of P (Xi ≤ Yi ) and P (Xi ≤ Xj and Yi ≤ Yj ), i 6= j. 22. Let X1 , ..., Xn be i.i.d. from P ∈ P containing all symmetric c.d.f.’s with finite means and with Lebesgue p.d.f.’s on R. Show that there is no UMVUE of µ = EX1 when n > 1. 23. Prove Corollary 3.1. 24. Suppose that T is a UMVUE of an unknown parameter ϑ. Show that T k is a UMVUE of E(T k ), where k is any positive integer for which E(T 2k ) < ∞. 25. Consider the problem in Exercise 83 of §2.6. Use Theorem 3.2 to show that I{0} (X) is a UMVUE of (1 − p)2 and that there is no UMVUE of p. 26. Let X1 , ..., Xn be i.i.d. from a discrete distribution with P (Xi = θ − 1) = P (Xi = θ) = P (Xi = θ + 1) = 31 , where θ is an unknown integer. Show that no nonconstant function of θ has a UMVUE. 27. Let X be a random variable having the Lebesgue p.d.f. √ [(1 − θ) + θ/(2 x)]I(0,1) (x), where θ ∈ [0, 1]. Show that there is no UMVUE of θ.

3.6. Exercises

221

28. Let X be a discrete random variable with P (X = −1) = 2p(1 − p) and P (X = k) = pk (1 − p)3−k , k = 0, 1, 2, 3, where p ∈ (0, 1). (a) Determine whether there is a UMVUE of p. (b) Determine whether there is a UMVUE of p(1 − p). 29. Let X1 , ..., Xn be i.i.d. observations. Obtain a UMVUE of a in the following cases. (a) Xi has the exponential distribution E(a, θ) with a known θ and an unknown a ≤ 0. (b) Xi has the Pareto distribution P a(a, θ) with a known θ > 1 and an unknown a ∈ (0, 1]. 30. In Exercise 41 of §2.6, find a UMVUE of θ and show that it is unique a.s. 31. Prove Theorem 3.3 for the multivariate case (k > 1). 32. Let X be a single sample from Pθ . Find the Fisher information I(θ) in the following cases. (a) Pθ is the N (µ, σ 2 ) distribution with θ = µ ∈ R. (b) Pθ is the N (µ, σ 2 ) distribution with θ = σ 2 > 0. (c) Pθ is the N (µ, σ 2 ) distribution with θ = σ > 0. (d) Pθ is the N (σ, σ 2 ) distribution with θ = σ > 0. (e) Pθ is the N (µ, σ 2 ) distribution with θ = (µ, σ 2 ) ∈ R × (0, ∞). (f) Pθ is the negative binomial distribution N B(θ, r) with θ ∈ (0, 1). (g) Pθ is the gamma distribution Γ(α, γ) with θ = (α, γ) ∈ (0, ∞) × (0, ∞). (h) Pθ is the beta distribution B(α, β) with θ = (α, β) ∈ (0, 1)×(0, 1). 33. Find a function of θ for which the amount of information is independent of θ, when Pθ is (a) the Poisson distribution P (θ) with θ > 0; (b) the binomial distribution Bi(θ, r) with θ ∈ (0, 1); (c) the gamma distribution Γ(α, θ) with θ > 0. 34. Prove the result in Example 3.9. 35. Obtain the Fisher information matrix for a random variable with (a) the Cauchy distribution C(µ, σ), µ ∈ R, σ > 0; (b) the double exponential distribution DE(µ, θ), µ ∈ R, θ > 0; (c) the logistic distribution LG(µ, σ), µ ∈ R, σ > 0; (d) the c.d.f. Fr x−µ , where Fr is the c.d.f. of the t-distribution tr σ with a known r, µ ∈ R, σ > 0; (e) the Lebesgue p.d.f. fθ (x) = (1 − ǫ)φ(x − µ) + σǫ φ x−µ , θ = σ (µ, σ, ǫ) ∈ R × (0, ∞) × (0, 1), where φ is the standard normal p.d.f.

222

3. Unbiased Estimation

36. Let X be a sample having a p.d.f. satisfying the conditions in Theorem 3.3, where θ is a k-vector of unknown parameters, and let T (X) be a statistic. If T has a p.d.f. gθ satisfying the conditions in Theorem ∂ ∂ 3.3, then we define IT (θ) = E{ ∂θ log gθ (T )[ ∂θ log gθ (T )]τ } to be the Fisher information about θ contained in T . (a) Show that IX (θ) − IT (θ) is nonnegative definite, where IX (θ) is the Fisher information about θ contained in X. (b) Show that IX (θ) = IT (θ) if T is sufficient for θ. 37. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (0, θ) with θ > 0. (a) Show that condition (3.3) does not hold for h(X) = X(n) . (b) Show that the inequality in (3.6) does not hold for the UMVUE of θ. 38. Prove Proposition 3.3. 39. Let X be a single sample from the double exponential distribution DE(µ, θ) with µ = 0 and θ > 0. Find the UMVUE’s of the following parameters and, in each case, determine whether the variance of the UMVUE attains the Cram´er-Rao lower bound. (a) ϑ = θ; (b) ϑ = θr , where r > 1; (c) ϑ = (1 + θ)−1 . 40. Let X1 , ..., Xn be i.i.d. binary random variables with P (Xi = 1) = p ∈ (0, 1). ¯ − X)/(n ¯ (a) Show that the UMVUE of p(1 − p) is Tn = nX(1 − 1). (b) Show that Var(Tn ) does not attain the Cram´er-Rao lower bound. (c) Show that (3.10) holds. 41. Let X1 , ..., Xn be i.i.d. having the Poisson distribution P (θ) with θ > 0. Find the amse of the UMVUE of e−tθ with a fixed t > 0 and show that (3.10) holds. 42. Let X1 , ..., Xn be i.i.d. having the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 > 0. (a) Find the UMVUE of ϑ = etµ with a fixed t 6= 0. (b) Determine whether the variance of the UMVUE in (a) attains the Cram´er-Rao lower bound. (c) Show that (3.10) holds. 43. Show that if X1 , ..., Xn are i.i.d. binary random variables, Un in (3.12) equals P T (T − 1) · · · (T − m + 1)/[n(n − 1) · · · (n − m + 1)], where T = ni=1 Xi .

223

3.6. Exercises

¯ then Un in (3.13) is the same as the sample 44. Show that if Tn = X, 2 variance S in (2.2). Show that (3.23) holds for Tn given by (3.22) with E(Rn2 ) = o(n−1 ). 45. Prove (3.14), (3.16), and (3.17). 46. Let ζk be given in Theorem 3.4. Show that ζ1 ≤ ζ2 ≤ · · · ≤ ζm . 47. Prove Corollary 3.2. ˇn is also a U-statistic. 48. Prove (3.20) and show that Un − U

ˇ 49. Let Tn be a symmetric statistic with Var(Tn ) < ∞ for every n and Tn be the projection of Tn on nk random vectors {Xi1 , ..., Xik }, 1 ≤ i1 < · · · < ik ≤ n. Show that E(Tn ) = E(Tˇn ) and calculate E(Tn − Tˇn )2 .

50. Let Yk be defined in Lemma 3.2. Show that {Yk2 } is uniformly integrable. 51. Show that (3.22) with E(Rn2 ) = o(n−1 ) is satisfied for Tn being a U-statistic with E[h(X1 , ..., Xm )]2 < ∞. 52. Let S 2 be the sample variance given by (2.2), which is also a Ustatistic (§3.2.1). Find the corresponding h1 , h2 , ζ1 , and ζ2 . Discuss how to apply Theorem 3.5 to this case. 53. Let h(x1 , x2 , x3 ) = I(−∞,0) (x1 + x2 + x3 ). Define the U-statistic with this kernel and find hk and ζk , k = 1, 2, 3. 54. Let X1 , ..., Xn be i.i.d. random variables having finite µ = EX1 and µ ¯ = EX1−1 . Find a U-statistic that is an unbiased estimator of µ¯ µ and derive its variance and asymptotic distribution. 55. Show that βˆ is an LSE of β if and only if it is given by (3.29). 56. Obtain explicit forms for the LSE’s of βj , j = 0, 1, and SSR, under the simple linear regression model in Example 3.11, assuming that some ti ’s are different. 57. Consider the polynomial model Xi = β0 + β1 ti + β2 t2i + εi ,

i = 1, ..., n.

Find explicit forms for the LSE’s of βj , j = 0, 1, 2, and SSR, assuming that some ti ’s are different. 58. Suppose that Xij = αi + βtij + εij ,

i = 1, ..., a, j = 1, ..., b.

Find explicit forms for the LSE’s of β, αi , i = 1, ..., a, and SSR.

224

3. Unbiased Estimation

59. Consider the polynomial model Xi = β0 + β1 ti + β2 t2i + β3 t3i + εi ,

i = 1, ..., n,

where εi ’s are i.i.d. from N (0, σ 2 ). Suppose that n = 12, ti = −1, i = 1, ..., 4, ti = 0, i = 5, ..., 8, and ti = 1, i = 9, ..., 12. (a) Obtain the matrix Z τ Z when this polynomial model is considered as a special case of model (3.24). (b) Show whether the following parameters are estimable: β0 + β2 , β1 , β0 − β1 , β1 + β3 , and β0 + β1 + β2 + β3 . 60. Find the matrix Z, Z τ Z, and the form of l ∈ R(Z) under the one-way ANOVA model (3.31). 61. Obtain the matrix Z under the two-way balanced ANOVA model (3.32). Show that the rank of Z is ab. Verify the form of the LSE of β given in Example 3.14. Find the form of l ∈ R(Z). 62. Consider the following model as a special case of model (3.25): Xijk = µ + αi + βj + εijk ,

i = 1, ..., a, j = 1, ..., b, k = 1, ..., c.

Obtain the matrix Z, the parameter vector β, and the form of LSE’s of β. Discuss conditions under which l ∈ R(Z). 63. Under model (3.25) and assumption A1, find the UMVUE’s of (lτ β)2 , lτ β/σ, and (lτ β/σ)2 for an estimable lτ β. 64. Verify the formulas for SSR’s in Example 3.15. 65. Consider the one-way random effects model in Example 3.17. Assume that ni = n for all i and that Ai ’s and eij ’s are normally distributed. Show that the family of populations is an exponential family with ¯ ·· , SA = n Pm (X ¯ i· − X ¯ ·· )2 , and sufficient and complete statistics X i=1 Pm Pn 2 2 ¯ SE = i=1 j=1 (Xij − Xi· ) . Find the UMVUE’s of µ, σa , and σ 2 .

66. Consider model (3.25). Suppose that εi ’s are i.i.d. with Eεi = 0 and a Lebesgue p.d.f. σ −1 f (x/σ), where f is a known Lebesgue p.d.f. and σ > 0 is unknown. (a) Show that X is from a location-scale family given by (2.10). (b) Find the Fisher information about (β, σ) contained in Xi . (c) Find the Fisher information about (β, σ) contained in X.

67. Consider model (3.25) with assumption A2. Let c ∈ Rp . Show that if the equation c = Z τ y has a solution, then there is a unique solution y0 ∈ R(Z τ ) such that Var(y0τ X) ≤ Var(y τ X) for any other solution of c = Z τ y.

225

3.6. Exercises

68. Consider model (3.25). Show that the number of independent linear functions of X with mean 0 is n − r, where r is the rank of Z. ˆ which ˆ i = Z τ β, 69. Consider model (3.25) with assumption A2. Let X i is called the least squares prediction of Xi . Let hij be the (i, j)th element of Z(Z τ Z)− Z τ and hi = hii . Show that ˆ i ) = σ 2 hi ; (a) Var(X ˆ i ) = σ 2 (1 − hi ); (b) Var(Xi − X ˆi , X ˆ j ) = σ 2 hij ; (c) Cov(X ˆ i , Xj − X ˆ j ) = −σ 2 hij , i 6= j; (d) Cov(Xi − X ˆ ˆ (e) Cov(Xi , Xj − Xj ) = 0.

70. Consider model (3.25) with assumption A2. Let Z = (Z1 , Z2 ) and β = (β1 , β2 ), where Zj is n × pj and βj is a pj -vector, j = 1, 2. Assume that (Z1τ Z1 )−1 and [Z2τ Z2 − Z2τ Z1 (Z1τ Z1 )−1 Z1τ Z2 ]−1 exist. (a) Derive the LSE of β in terms of Z1 , Z2 , and X. (b) Let βˆ = (βˆ1 , βˆ2 ) be the LSE in (a). Calculate the covariance between βˆ1 and βˆ2 . (c) Suppose that it is known that β2 = 0. Let β˜1 be the LSE of β1 under the reduced model X = Z1 β1 + ε. Show that, for any l ∈ Rp1 , lτ β˜1 is better than lτ βˆ1 in terms of their mse’s. 71. Prove that (e) implies (b) in Theorem 3.10. 72. Show that (a) in Theorem 3.10 is equivalent to either (f) Var(ε)Z = ZB for some matrix B, or (g) R(Z τ ) is generated by r eigenvectors of Var(ε), where r is the rank of Z. 73. Prove Corollary 3.3. 74. Suppose that X = µJn + Hξ + e, where µ ∈ R is an unknown parameter, Jn is the n-vector of 1’s, H is an n × p known matrix of full rank, ξ is a random p-vector with E(ξ) = 0 and Var(ξ) = σξ2 Ip , e is a random n-vector with E(e) = 0 and Var(e) = σ 2 In , and ξ and e are independent. Show that the LSE of µ is the BLUE if and only if the row totals of HH τ are the same. 75. Consider a special case of model (3.25): Xij = µ + αi + βj + εij ,

i = 1, ..., a, j = 1, ..., b,

where µ, αi ’s, and βj ’s are unknown parameters, E(εij ) = 0, Var(εij ) = σ 2 , Cov(εij , εi′ j ′ ) = 0 if i 6= i′ , and Cov(εij , εij ′ ) = σ 2 ρ if j 6= j ′ . Show that the LSE of lτ β is the BLUE for any l ∈ R(Z).

226

3. Unbiased Estimation

76. Consider model (3.25) under assumption A3 with Var(ε) = a block diagonal matrix whose ith block diagonal Vi is ni ×ni and has a single eigenvalue λi with eigenvector Jni (the ni -vector of 1’s) and a repeated Pk eigenvalue ρi with multiplicity ni − 1, i = 1, ..., k, i=1 ni = n. Let U be the n × k matrix whose ith column is Ui , where U1 = (Jnτ1 , 0, ..., 0), U2 = (0, Jnτ2 , ..., 0),..., Uk = (0, 0, ..., Jnτk ). (a) If R(Z τ ) ⊂ R(U τ ) and λi ≡ λ, show that lτ βˆ is the BLUE of lτ β for any l ∈ R(Z). (b) If Z τ Ui = 0 for all i and ρi ≡ ρ, show that lτ βˆ is the BLUE of lτ β for any l ∈ R(Z). 77. Prove Proposition 3.4. 78. Show that the condition supn λ+ [Var(ε)] < ∞ is equivalent to the condition supi Var(εi ) < ∞. 79. Find a condition under which the mse of lτ βˆ is of the order n−1 . Apply it to problems in Exercises 56, 58, and 60-62. 80. Consider model (3.25) with i.i.d. ε1 , ..., εn having E(εi ) = 0 and ˆ i = Z τ βˆ and hi = Z τ (Z τ Z)− Zi . Var(εi ) = σ 2 . Let X i i (a) Show that for any ǫ > 0, ˆi − E X ˆ i | ≥ ǫ) ≥ min{P (εi ≥ ǫ/hi ), P (εi ≤ −ǫ/hi )}. P (|X (Hint: for independent random variables X and Y , P (|X + Y | ≥ ǫ) ≥ P (X ≥ ǫ)P (Y ≥ 0) + P (X ≤ −ǫ)P (Y < 0).) ˆi − EX ˆ i →p 0 if and only if hi → 0. (b) Show that X 81. Prove Lemma 3.3 and show that condition (a) is implied by {kZi k} being bounded and λ+ (Z τ Z)− → 0. 82. Consider the problem in Exercise 58. Suppose that {tij } is bounded. Find a condition under which (3.39) holds. 83. Under the two-way ANOVA models in Example 3.14 and Exercise 62, find sufficient conditions for (3.39). 84. Consider the one-way random effects model in Example 3.17. Assume that {ni } is bounded and E|eij |2+δ < ∞ for some δ > 0. Show that the LSE µ ˆ of µ is asymptotically normal and derive an explicit form of Var(ˆ µ). 85. Suppose that Xi = ρti + εi ,

i = 1, ..., n,

where ρ ∈ R is an unknown parameter, ti ’s are known and in (a, b), a and b are known positive constants, and εi ’s are independent random

3.6. Exercises

227

variables satisfying E(εi ) = 0, E|εi |2+δ < ∞ for some δ > 0, and Var(εi ) = σ 2 ti with an unknown σ 2 > 0. (a) Obtain the LSE of ρ. (b) Obtain the BLUE of ρ. (c) Show that both the LSE and BLUE are asymptotically normal and obtain the asymptotic relative efficiency of the BLUE w.r.t. the LSE. 86. In Example 3.19, show that E(S 2 ) = σ 2 given in (3.43). 87. Suppose that X = (X1 , ..., Xn ) is a simple random sample (without replacement) from a finite population P = {y1 , ..., yN } with univariate yi . (a) Show that a necessary condition for h(θ) to be estimable is that h is symmetric in its N arguments. (b) Find the UMVUE of Y m , where m is a fixed positive integer < n and Y is the population total. (c) Find the UMVUE of P (Xi ≤ Xj ), i 6= j. (d) Find the UMVUE of Cov(Xi , Xj ), i 6= j. 88. Prove Theorem 3.14. 89. Under stratified simple random sampling described in §3.4.1, show that the vector of ordered values of all Xhi ’s is neither sufficient nor complete for θ ∈ Θ. 90. Let P = {y1 , ..., yN } be a population with univariate yi . Define the P population c.d.f. by F (t) = N −1 N i=1 I(−∞,t] (yi ). Find the UMVUE of F (t) under (a) simple random sampling and (b) stratified simple random sampling. 91. Consider the estimation of F (t) in the previous exercise. Suppose that a sample of size n is selected with πi > 0. Find the Horvitz-Thompson estimator of F (t). Is it a c.d.f.? 92. Show that v1 in (3.49) and v2 in (3.50) are unbiased estimators of Var(Yˆht ). Prove that v1 = v2 under (a) simple random sampling and (b) stratified simple random sampling. 93. Consider the following two-stage stratified sampling plan. In the first stage, the population is stratified into H strata and kh clusters are selected from stratum h with probability proportional to cluster size, where sampling is independent across strata. In the second stage, a sample of mhi units is selected from sampled cluster i in stratum h, and sampling is independent across clusters. Find πi and the HorvitzThompson estimator Yˆht of the population total.

228

3. Unbiased Estimation

94. In the previous exercise, prove the unbiasedness of Yˆht directly (without using Theorem 3.15). 95. Under systematic sampling, show that Var(Yˆsy ) is equal to k 1 σ2 2 X 1− + N n nN i=1

X

1≤t 0. Consider the estimation of ϑ = EΦ(a + bX1 ), where Φ is the standard normal c.d.f. and a and b are known constants. Obtain an explicit ¯ S 2 ). form of a function g(µ, σ 2 ) = ϑ and the amse of ϑˆ = g(X, 101. Let X1 , ..., Xn be i.i.d. with mean µ, variance σ 2 , and finite µj = EX1j , ¯ j = 2, 3, 4. The sample coefficient of variation is defined to be S/X, 2 where S is the squared root sample variance S . √ of the ¯ − σ/µ) →d N (0, τ ) and obtain an (a) If µ 6= 0, show that n(S/X explicit formula of τ in terms of µ, σ 2 , and µj . ¯ →d [N (0, 1)]−1 . (b) If µ = 0, show that n−1/2 S/X 102. Prove (3.52) and (3.53). 103. Let X1 , ..., Xn be i.i.d. from P in a parametric family. Obtain moment estimators of parameters in the following cases. (a) P is the gamma distribution Γ(α, γ), α > 0, γ > 0. (b) P is the exponential distribution E(a, θ), a ∈ R, θ > 0. (c) P is the beta distribution B(α, β), α > 0, β > 0. (d) P is the log-normal distribution LN (µ, σ 2 ), µ ∈ R, σ > 0. (e) P is the uniform distribution U (θ − 12 , θ + 12 ), θ ∈ R. (f) P is the negative binomial distribution N B(p, r), p ∈ (0, 1), r = 1, 2,....

229

3.6. Exercises

(g) P is the log-distribution L(p), p ∈ (0, 1). (h) P is the log-normal distribution LN (µ, σ2 ), µ ∈ R, σ = 1. (i) P is the chi-square distribution χ2k with an unknown k = 1, 2, .... 104. Obtain moment estimators of λ and p in Exercise 55 of §2.6, based on data X1 , ..., Xn . 105. Obtain the asymptotic distributions of the moment estimators in Exercise 103(a), (c), (e), and (g), and the asymptotic relative efficiencies of moment estimators w.r.t. UMVUE’s in Exercise 103(b) and (h). 106. In Exercise 19(a), find a moment estimator of θ and derive its asymptotic distribution. In Exercise 19(b), obtain a moment estimator of θ−1 and its asymptotic relative efficiency w.r.t. the UMVUE of θ−1 . 107. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. fα,β (x) = αβ −α xα−1 I(0,β) (x), where α > 0 and β > 0 are unknown. (a) Obtain moment estimators of α and β. (b) Obtain the asymptotic distribution of the moment estimators of α and β derived in (a). 108. Let X1 , ..., Xn be i.i.d. from the following discrete distribution: P (X1 = 1) =

2(1 − θ) , 2−θ

P (X1 = 2) =

θ , 2−θ

where θ ∈ (0, 1) is unknown. (a) Obtain an estimator of θ using the method of moments. (b) Obtain the amse of the moment estimator in (a). 109. Let X1 , ..., Xn (n > 1) be i.i.d. from a population having the Lebesgue p.d.f. x−µ ǫ fθ (x) = (1 − ǫ)φ(x − µ) + φ , σ σ where φ is the standard normal p.d.f., θ = (µ, σ) ∈ R × (0, ∞) is unknown, and ǫ ∈ (0, 1) is a known constant. (a) Obtain an estimator of θ using the method of moments. (b) Obtain the asymptotic distribution of the moment estimator in part (a). 110. Let X1 , ..., Xn be i.i.d. random variables having the Lebesgue p.d.f. x>0 (θ1 + θ2 )−1 e−x/θ1 fθ1 ,θ2 (x) = x ≤ 0, (θ1 + θ2 )−1 ex/θ2 where θ1 > 0 and θ2 > 0 are unknown. (a) Obtain an estimator of (θ1 , θ2 ) using the method of moments.

230

3. Unbiased Estimation

(b) Obtain the asymptotic distribution of the moment estimator in part (a). 111. (Nonexistence of a moment estimator). Consider X1 , ..., Xn and the parametric family indexed by (θ, j) ∈ (0, 1) × {1, 2} in Exercise 41 of §2.6. Let hi (θ, j) = EX1i , i = 1, 2. Show that P (ˆ µi = hi (θ, j) has a solution) → 0 as n → ∞, when Xi ’s are from the Poisson distribution P (θ). 112. In the proof of Proposition 3.5, show that E[Wn (Un − ϑ)] = O(n−1 ). 113. Assume the conditions of Theorem 3.16. (a) Prove (3.56). P∞ (b) Show that E[g2 (X1 , X1 )]/n = nEVn2 = m(m − 1) j=1 λj /(2n).

114. Let X1 , ..., Xn be i.i.d.Rwith a c.d.f. F and Un and Vn be the U- and Vstatistics with kernel [I(−∞,y] (x1 ) − F0 (y)][I(−∞,y] (x2 ) − F0 (y)]dF0 , where F0 is a known c.d.f. (a) Obtain the asymptotic distributions of Un and Vn when F 6= F0 . (b) Obtain the asymptotic relative efficiency of Un w.r.t. Vn when F = F0 . 115. Let X1 , ..., Xn be i.i.d. with a c.d.f. F having a finite sixth moment. Consider the estimation of µ3 , where µ = EX1 . When µ = 0, find −1 P amseX¯ 3 (P )/amseUn (P ), where Un = n3 1≤i<j 0. Under the ˆ in (3.64) is consistent for Σ in the sense same conditions, show that Σ ˆ that kΣ − Σkmax →p 0. 118. In Example 3.30, show that Vˆt is consistent for Vt when kt → ∞ as k → ∞. 119. Show how to use equation (3.65) to obtain consistent estimators of θ0 and θ1 . 120. Prove (3.66) under the assumed conditions in §3.5.4.

Chapter 4

Estimation in Parametric Models In this chapter, we consider point estimation methods in parametric models. One such method, the moment method, has been introduced in §3.5.2. It is assumed in this chapter that the sample X is from a population in a parametric family P = {Pθ : θ ∈ Θ}, where Θ ⊂ Rk for a fixed integer k ≥ 1.

4.1 Bayes Decisions and Estimators Bayes rules are introduced in §2.3.2 as decision rules minimizing the average risk w.r.t. a given probability measure Π on Θ. Bayes rules, however, are optimal rules in the Bayesian approach, which is fundamentally different from the classical frequentist approach that we have been adopting.

4.1.1 Bayes actions In the Bayesian approach, θ is viewed as a realization of a random vector θ whose prior distribution is Π. The prior distribution is based on past experience, past data, or a statistician’s belief and thus may be very subjective. A sample X is drawn from Pθ = Px|θ , which is viewed as the conditional distribution of X given θ = θ. The sample X = x is then used to obtain an updated prior distribution, which is called the posterior distribution and can be derived as follows. By Theorem 1.7, the joint distribution of X and θ is a probability measure on X × Θ determined by Z P (A × B) = Px|θ (A)dΠ(θ), A ∈ BX , B ∈ BΘ , B

231

232

4. Estimation in Parametric Models

where X is the range of X. The posterior distribution of θ, given X = x, is the conditional distribution Pθ|x whose existence is guaranteed by Theorem 1.7 for almost all x ∈ X. When Px|θ has a p.d.f., the following result provides a formula for the p.d.f. of the posterior distribution Pθ|x . Theorem 4.1 (Bayes formula). Assume that P = {Px|θ : θ ∈ Θ} is dP

dominated by a σ-finite measure ν and fθ (x) = dνx|θ (x) is a Borel function on (X × RΘ, σ(BX × BΘ )). Let Π be a prior distribution on Θ. Suppose that m(x) = Θ fθ (x)dΠ > 0. (i) The posterior distribution Pθ|x ≪ Π and dPθ|x fθ (x) = . dΠ m(x) (ii) If Π ≪ λ and

dΠ dλ

= π(θ) for a σ-finite measure λ, then dPθ|x fθ (x)π(θ) = . dλ m(x)

(4.1)

Proof. Result (ii) follows from result (i) and Proposition 1.7(iii). To show (i), we first show that m(x) < ∞ a.e. ν. Note that Z Z Z Z Z m(x)dν = fθ (x)dΠdν = fθ (x)dνdΠ = 1, (4.2) X

X

Θ

Θ

X

where the second equality follows from Fubini’s theorem. Thus, m(x) is integrable w.r.t. ν and m(x) < ∞ a.e. ν. For x ∈ X with m(x) < ∞, define Z 1 P (B, x) = fθ (x)dΠ, B ∈ BΘ . m(x) B Then P (·, x) is a probability measure on Θ a.e. ν. By Theorem 1.7, it remains to show that P (B, x) = P (θ ∈ B|X = x). By Fubini’s theorem, P (B, ·) is a measurable function of x. Let Px,θ denote the “joint” distribution of (X, θ). For any A ∈ σ(X), Z Z Z IB (θ)dPx,θ = fθ (x)dΠdν A×Θ A B Z Z Z fθ (x) dΠ fθ (x)dΠ dν = A B m(x) Θ Z Z Z fθ (x) dΠ fθ (x)dνdΠ = m(x) ZΘ A B = P (B, x)dPx,θ , A×Θ

233

4.1. Bayes Decisions and Estimators

where the third equality follows from Fubini’s theorem. This completes the proof. Because of (4.2), m(x) is called the marginal p.d.f. of X w.r.t. ν. If m(x) = 0 for an x ∈ X, then fθ (x) = 0 a.s. Π. Thus, either x should be eliminated from X or the prior Π is incorrect and a new prior should be specified. Therefore, without loss of generality we may assume that the assumption of m(x) > 0 in Theorem 4.1 is always satisfied. If both X and θ are discrete and ν and λ are the counting measures, then (4.1) becomes P (θ = θ|X = x) = P

P (X = x|θ = θ)P (θ = θ) , θ∈Θ P (X = x|θ = θ)P (θ = θ)

which is the Bayes formula that appears in elementary probability. In the Bayesian approach, the posterior distribution Pθ|x contains all the information we have about θ and, therefore, statistical decisions and inference should be made based on Pθ|x , conditional on the observed X = x. In the problem of estimating θ, Pθ|x can be viewed as a randomized decision rule under the approach discussed in §2.3. Definition 4.1. Let A be an action space in a decision problem and L(θ, a) ≥ 0 be a loss function. For any x ∈ X, a Bayes action w.r.t. Π is any δ(x) ∈ A such that E[L(θ, δ(x))|X = x] = min E[L(θ, a)|X = x], a∈A

(4.3)

where the expectation is w.r.t. the posterior distribution Pθ|x . The existence and uniqueness of Bayes actions can be discussed under some conditions on the loss function and the action space. Proposition 4.1. Assume that the conditions in Theorem 4.1 hold; L(θ, a) is convex in a for each fixed θ; and for each x ∈ X, E[L(θ, a)|X = x] < ∞ for some a. (i) If A is a compact subset of Rp for some integer p ≥ 1, then a Bayes action δ(x) exists for each x ∈ X. (ii) If A = Rp and L(θ, a) tends to ∞ as kak → ∞ uniformly in θ ∈ Θ0 ⊂ Θ with Π(Θ0 ) > 0, then a Bayes action δ(x) exists for each x ∈ X. (iii) In (i) or (ii), if L(θ, a) is strictly convex in a for each fixed θ, then the Bayes action is unique. Proof. The convexity of the loss function implies the convexity and continuity of E[L(θ, a)|X = x] as a function of a with any fixed x. Then, the result in (i) follows from the fact that any continuous function on a compact

234

4. Estimation in Parametric Models

set attains its minimum. The result in (ii) follows from the fact that Z L(θ, a)dPθ|x = ∞ lim E[L(θ, a)|X = x] ≥ lim kak→∞

kak→∞

Θ0

under the assumed condition in (ii). Finally, the result in (iii) follows from the fact that E[L(θ, a)|X = x] is strictly convex in a for any fixed x under the assumed conditions. Other conditions on L under which a Bayes action exists can be found, for example, in Lehmann (1983, §1.6 and §4.1). Example 4.1. Consider the estimation of ϑ = g(θ) for some real-valued R function g such that Θ [g(θ)]2 dΠ < ∞. Suppose that A = the range of g(θ) and L(θ, a) = [g(θ) − a]2 (squared error loss). Using the same argument as in Example 1.22, we obtain the Bayes action R R g(θ)fθ (x)dΠ Θ g(θ)fθ (x)dΠ δ(x) = = ΘR , (4.4) m(x) f (x)dΠ Θ θ

which is the posterior expectation of g(θ), given X = x. More specifically, let us consider the case where g(θ) = θj for some integer j ≥ 1, fθ (x) = e−θ θx I{0,1,2,...} (x)/x! (the Poisson distribution) with θ > 0, and Π has a Lebesgue p.d.f. π(θ) = θα−1 e−θ/γ I(0,∞) (θ)/[Γ(α)γ α ] (the gamma distribution Γ(α, γ) with known α > 0 and γ > 0). Then, for x = 0, 1, 2, ..., fθ (x)π(θ) = c(x)θx+α−1 e−θ(γ+1)/γ I(0,∞) (θ), m(x)

(4.5)

where c(x) is some function of x. By using Theorem 4.1 and matching the right-hand side of (4.5) with that of the p.d.f. of the gamma distribution, we know that the posterior is the gamma distribution Γ(x + α, γ/(γ + 1)). Hence, without actually working out the integral m(x), we know that c(x) = (1 + γ −1 )x+α /Γ(x + α). Then Z ∞ θj+x+α−1 e−θ(γ+1)/γ dθ. δ(x) = c(x) 0

Note that the integrand is proportional to the p.d.f. of the gamma distribution Γ(j + x + α, γ/(γ + 1)). Hence δ(x) = c(x)Γ(j + x + α)/(1 + γ −1 )j+x+α = (j + x + α − 1) · · · (x + α)/(1 + γ −1 )j . In particular, δ(x) = (x + α)γ/(γ + 1) when j = 1.

235

4.1. Bayes Decisions and Estimators

An interesting phenomenon in Example 4.1 is that the prior and the posterior are in the same parametric family of distributions. Such a prior is called a conjugate prior. Under a conjugate prior, Bayes actions often have explicit forms (in x) when the loss function is simple. Whether a prior is conjugate involves a pair of families; one is the family P = {fθ : θ ∈ Θ} and the other is the family from which Π is chosen. Example 4.1 shows that the Poisson family and the gamma family produce conjugate priors. It can be shown (exercise) that many pairs of families in Table 1.1 (page 18) and Table 1.2 (pages 20-21) produce conjugate priors. In general, numerical methods have to be used in evaluating the integrals in (4.4) or Bayes actions under general loss functions. Even under a conjugate prior, the integral in (4.4) involving a general g may not have an explicit form. More discussions on the computation of Bayes actions are given in §4.1.4. As an example of deriving a Bayes action in a general decision problem, we consider Example 2.21. Example 4.2. Consider the decision problem in Example 2.21. Let Pθ|x be the posterior distribution of θ, given X = x. In this problem, A = {a1 , a2 , a3 }, which is compact in R. By Proposition 4.1, we know that there is a Bayes action if the mean of Pθ|x is finite. Let Eθ|x be the expectation w.r.t. Pθ|x . Since A contains only three elements, a Bayes action can be obtained by comparing j=1 c1 Eθ|x [L(θ, aj )] = j=2 c2 + c3 Eθ|x [ψ(θ, t)] j = 3, c3 Eθ|x [ψ(θ, 0)] where ψ(θ, t) = (θ − θ0 − t)I(θ0 +t,∞) (θ).

The minimization problem (4.3) is the same as the minimization problem Z Z L(θ, δ(x))fθ (x)dΠ = min L(θ, a)fθ (x)dΠ. (4.6) Θ

a∈A

Θ

The minimization problem (4.6) is still defined even if Π is not a probability measure but a σ-finite measure on Θ, in which case m(x) may not be finite. If Π(Θ) 6= 1, Π is called an improper prior. A prior with Π(Θ) = 1 is then called a proper prior. An action δ(x) that satisfies (4.6) with an improper prior is called a generalized Bayes action. The following is a reason why we need to discuss improper priors and generalized Bayes actions. In many cases, one has no past information and has to choose a prior subjectively. In such cases, one would like to select a noninformative prior that tries to treat all parameter values in Θ

236

4. Estimation in Parametric Models

equitably. A noninformative prior is often improper. We only provide one example here. For more detailed discussions of the use of improper priors, see Jeffreys (1939, 1948, 1961), Box and Tiao (1973), and Berger (1985). Example 4.3. Suppose that X = (X1 , ..., Xn ) and Xi ’s are i.i.d. from N (µ, σ 2 ), where µ ∈ Θ ⊂ R is unknown and σ 2 is known. Consider the estimation of ϑ = µ under the squared error loss. If Θ = [a, b] with −∞ < a < b < ∞, then a noninformative prior that treats all parameter values equitably is the uniform distribution on [a, b]. If Θ = R, however, the corresponding “uniform distribution” is the Lebesgue measure on R, which is an improper prior. If Π is the Lebesgue measure on R, then ) ( n Z ∞ X (xi − µ)2 2 −n/2 2 (2πσ ) dµ < ∞. µ exp − 2σ 2 −∞ i=1 By differentiating a in (2πσ 2 )−n/2

Z

∞

−∞

(

(µ − a)2 exp −

n X (xi − µ)2 i=1

2σ 2

)

dµ

P P and using the fact that ni=1 (xi − µ)2 = ni=1 (xi − x ¯)2 + n(¯ x − µ)2 , where x ¯ is the sample mean of the observations x1 , ..., xn , we obtain that R∞ µ exp −n(¯ x − µ)2 /(2σ 2 ) dµ −∞ =x ¯. δ(x) = R ∞ exp {−n(¯ x − µ)2 /(2σ 2 )} dµ −∞

Thus, the sample mean is a generalized Bayes action under the squared error loss. From Example 2.25 and Exercise 91 in §2.6, if Π is N (µ0 , σ02 ), then the Bayes action is µ∗ (x) in (2.25). Note that in this case x ¯ is a limit of µ∗ (x) as σ02 → ∞.

4.1.2 Empirical and hierarchical Bayes methods A Bayes action depends on the chosen prior that may depend on some parameters called hyperparameters. In §4.1.1, hyperparameters are assumed to be known. If hyperparameters are unknown, one way to solve the problem is to estimate them using data x1 , ..., xn ; the resulting Bayes action is called an empirical Bayes action. The simplest empirical Bayes method is to estimate prior parameters by viewing x = (x1 , ..., xn ) as a “sample” from the marginal distribution Z Px|ξ (A) = Px|θ (A)dΠθ|ξ , A ∈ BX , Θ

237

4.1. Bayes Decisions and Estimators

where Πθ|ξ is a prior depending on an unknown vector ξ of hyperparameters, or from the marginal p.d.f. m(x) in (4.2), if Px|θ has a p.d.f. fθ . The method of moments introduced in §3.5.3, for example, can be applied to estimate ξ. We consider an example. Example 4.4. Let X = (X1 , ..., Xn ) and Xi ’s be i.i.d. from N (µ, σ 2 ) with an unknown µ ∈ R and a known σ 2 . Consider the prior Πµ|ξ = N (µ0 , σ02 ) with ξ = (µ0 , σ02 ). To obtain a moment estimate of ξ, we need to calculate Z Z x1 m(x)dx and x21 m(x)dx, Rn

Rn

where x = (x1 , ..., xn ). These two integrals can be obtained without calculating m(x). Note that Z Z Z Z x1 m(x)dx = x1 fµ (x)dxdΠµ|ξ = µdΠµ|ξ = µ0 Rn

and Z

Rn

x21 m(x)dx =

Rn

Θ

Z Z Θ

Rn

x21 fµ (x)dxdΠµ|ξ = σ 2 +

R

Z

R

µ2 dΠµ|ξ = σ 2 +µ20 +σ02 .

Thus, by viewing x1 , ..., xn as a sample from m(x), we obtain the moment estimates n 1X µ ˆ0 = x¯ and σ ˆ02 = (xi − x¯)2 − σ 2 , n i=1

where x ¯ is the sample mean of xi ’s. Replacing µ0 and σ02 in formula (2.25) (Example 2.25) by µ ˆ0 and σ ˆ02 , respectively, we find that the empirical Bayes action under the squared error loss is simply the sample mean x¯ (which is a generalized Bayes action; see Example 4.3). Note that σ ˆ02 in Example 4.4 can be negative. Better empirical Bayes methods can be found, for example, in Berger (1985, §4.5). The following method, called the hierarchical Bayes method, is generally better than empirical Bayes methods. Instead of estimating hyperparameters, in the hierarchical Bayes approach we put a prior on hyperparameters. Let Πθ|ξ be a (first-stage) prior with a hyperparameter vector ξ and let Λ be a prior on Ξ, the range of ξ. Then the “marginal” prior for θ is defined by Z Π(B) = Πθ|ξ (B)dΛ(ξ), B ∈ BΘ . (4.7) Ξ

If the second-stage prior Λ also depends on some unknown hyperparameters, then one can go on to consider a third-stage prior. In most applications,

238

4. Estimation in Parametric Models

however, two-stage priors are sufficient, since misspecifying a second-stage prior is much less serious than misspecifying a first-stage prior (Berger, 1985, §4.6). In addition, the second-stage prior can be chosen to be noninformative (improper). Bayes actions can be obtained in the same way as before using the prior in (4.7). Thus, the hierarchical Bayes method is simply a Bayes method with a hierarchical prior. Empirical Bayes methods, however, deviate from the Bayes method since x1 , ..., xn are used to estimate hyperparameters. Suppose that X has a p.d.f. fθ (x) w.r.t. a σ-finite measure ν and Πθ|ξ has a p.d.f. πθ|ξ (θ) w.r.t. a σ-finite measure κ. Then the prior Π in (4.7) has a p.d.f. Z π(θ) = πθ|ξ (θ)dΛ(ξ) Ξ

w.r.t. κ and

m(x) =

Z Z Θ

fθ (x)πθ|ξ (θ)dΛdκ. Ξ

Let Pθ|x,ξ be the posterior distribution of θ given x and ξ (or ξ is assumed known) and Z mx|ξ (x) = fθ (x)πθ|ξ (θ)dκ, Θ

which is the marginal of X given ξ (or ξ is assumed known). Then the posterior distribution Pθ|x has a p.d.f. dPθ|x fθ (x)π(θ) = dκ m(x) Z fθ (x)πθ|ξ (θ) dΛ(ξ) = m(x) Ξ Z fθ (x)πθ|ξ (θ) mx|ξ (x) = dΛ(ξ) mx|ξ (x) m(x) Ξ Z dPθ|x,ξ dPξ|x , = dκ Ξ

where Pξ|x is the posterior distribution of ξ given x. Thus, under the estimation problem considered in Example 4.1, the (hierarchical) Bayes action is Z δ(x) = δ(x, ξ)dPξ|x , (4.8) Ξ

where δ(x, ξ) is the Bayes action when ξ is known. A result similar to (4.8) is given in Lemma 4.1.

Example 4.5. Consider Example 4.4 again. Suppose that one of the parameters in the first-stage prior N (µ0 , σ02 ), µ0 , is unknown and σ02 is

239

4.1. Bayes Decisions and Estimators

known. Let the second-stage prior for ξ = µ0 be the Lebesgue measure on R (improper prior). From Example 2.25, δ(x, ξ) =

σ2 nσ02 ξ+ x ¯. 2 +σ nσ02 + σ 2

nσ02

To obtain the Bayes action δ(x), it suffices to calculate Eξ|x (ξ), where the expectation is w.r.t. Pξ|x . Note that the p.d.f. of Pξ|x is proportional to Z ∞ o n 2 −µ)2 dµ. ψ(ξ) = exp − n(¯x2σ − (µ−ξ) 2 2 2σ 0

−∞

Using the properties of normal distributions, one can show that −1 2 ξ ξ2 n 1 n¯ x − 2σ2 ψ(ξ) = C1 exp 2σ2 + 2σ02 2σ2 + 2σ02 0 o n 2 xξ = C2 exp − 2(nσnξ2 +σ2 ) + nσn¯ 2 2 0 0 +σ o n 2 n(ξ−¯ x) = C3 exp − 2(nσ , 2 +σ2 ) 0

where C1 , C2 , and C3 are quantities not depending on ξ. Hence Eξ|x (ξ) = x¯. The (hierarchical) generalized Bayes action is then δ(x) =

σ2 nσ02 E (ξ) + x ¯=x ¯. ξ|x nσ02 + σ 2 nσ02 + σ 2

4.1.3 Bayes rules and estimators The discussion in §4.1.1 and §4.1.2 is more general than point estimation and adopts an approach that is different from the frequentist approach used in the rest of this book. In the frequentist approach, if a Bayes action δ(x) is a measurable function of x, then δ(X) is a nonrandomized decision rule. It can be shown (exercise) that δ(X) defined in Definition 4.1 (if it exists R for X = x ∈ A with Θ Pθ (A)dΠ = 1) also minimizes the Bayes risk Z RT (θ)dΠ rT (Π) = Θ

over all decision rules T (randomized or nonrandomized), where RT (θ) is the risk function of T defined in (2.22). Thus, δ(X) is a Bayes rule (§2.3.2). In an estimation problem, a Bayes rule is called a Bayes estimator. Generalized Bayes risks, generalized Bayes rules (or estimators), and empirical Bayes rules (or estimators) can be defined similarly. In view of the discussion in §2.3.2, even if we do not adopt the Bayesian approach, the method described in §4.1.1 can be used as a way of generating

240

4. Estimation in Parametric Models

decision rules. In this section, we study a Bayes rule or estimator in terms of its risk (and bias and consistency for a Bayes estimator). Bayes rules are typically admissible since, if there is a rule better than a Bayes rule, then that rule has the same Bayes risk as the Bayes rule and, therefore, is itself a Bayes rule. This actually proves part (i) of the following result. The proof of the other parts of the following result is left as an exercise. Theorem 4.2. In a decision problem, let δ(X) be a Bayes rule w.r.t. a prior Π. (i) If δ(X) is a unique Bayes rule, then δ(X) is admissible. (ii) If Θ is a countable set, the Bayes risk rδ (Π) < ∞, and Π gives positive probability to each θ ∈ Θ, then δ(X) is admissible. (iii) Let ℑ be the class of decision rules having continuous risk functions. If δ(X) ∈ ℑ, rδ (Π) < ∞, and Π gives positive probability to any open subset of Θ, then δ(X) is ℑ-admissible. Generalized Bayes rules or estimators are not necessarily admissible. Many generalized Bayes rules are limits of Bayes rules (see Examples 4.3 and 4.7). Limits of Bayes rules are often admissible (Farrell, 1968a,b). The following result shows a technique of proving admissibility using limits of generalized Bayes risks. Theorem 4.3. Suppose that Θ is an open set of Rk . In a decision problem, let ℑ be the class of decision rules having continuous risk functions. A decision rule T ∈ ℑ is ℑ-admissible if there exists a sequence {Πj } of (possibly improper) priors such that (a) the generalized Bayes risks rT (Πj ) are finite for all j; (b) for any θ0 ∈ Θ and η > 0, rT (Πj ) − rj∗ (Πj ) = 0, j→∞ Πj (Oθ0 ,η ) lim

where rj∗ (Πj ) = inf T ∈ℑ rT (Πj ) and Oθ0 ,η = {θ ∈ Θ : kθ − θ0 k < η} with Πj (Oθ0 ,η ) < ∞ for all j. Proof. Suppose that T is not ℑ-admissible. Then there exists T0 ∈ ℑ such that RT0 (θ) ≤ RT (θ) for all θ and RT0 (θ0 ) < RT (θ0 ) for a θ0 ∈ Θ. From the continuity of the risk functions, we conclude that RT0 (θ) < RT (θ) − ǫ for all θ ∈ Oθ0 ,η and some constants ǫ > 0 and η > 0. Then, for any j, rT (Πj ) − rj∗ (Πj ) ≥ rT (Πj ) − rT0 (Πj ) Z [RT (θ) − RT0 (θ)]dΠj (θ) ≥ Oθ0 ,η

≥ ǫΠj (Oθ0 ,η ),

which contradicts condition (b). Hence, T is ℑ-admissible.

241

4.1. Bayes Decisions and Estimators

Example 4.6. Consider Example 4.3 and the estimation of µ under the squared error loss. From Theorem 2.1, the risk function of any decision rule is continuous in µ if the risk is finite. We now apply Theorem 4.3 to show ¯ is admissible. Let Πj = N (0, j). Since RX¯ (µ) = that the sample mean X 2 2 σ /n, rX¯ (Πj ) = σ /n for any j. Hence, condition (a) in Theorem 4.3 is satisfied. From Example 2.25, the Bayes estimator w.r.t. Πj is δj (X) = nj ¯ nj+σ2 X (see formula (2.25)). Thus, Rδj (µ) = and rj∗ (Πj )

=

Z

σ 2 nj 2 + σ 4 µ2 (nj + σ 2 )2

Rδj (µ)dΠj =

σ2 j . nj + σ 2

For any Oµ0 ,η = {µ : |µ − µ0 | < η}, µ0 − η 2ηΦ′ (ξj ) µ0 + η √ √ √ −Φ = Πj (Oµ0 ,η ) = Φ j j j √ √ for some ξj satisfying (µ0 − η)/ j ≤ ξj ≤ (µ0 + η)/ j, where Φ is the standard normal c.d.f. and Φ′ is its derivative. Since Φ′ (ξj ) → Φ′ (0) = (2π)−1/2 , √ rX¯ (Πj ) − rj∗ (Πj ) σ4 j = →0 Πj (Oµ0 ,η ) 2ηΦ′ (ξj )n(nj + σ 2 ) as j → ∞. Thus, condition (b) in Theorem 4.3 is satisfied and, hence, the ¯ is admissible. sample mean X More results in admissibility can be found in §4.2 and §4.3. The following result concerns the bias of a Bayes estimator. Proposition 4.2. Let δ(X) be a Bayes estimator of ϑ = g(θ) under the squared error loss. Then δ(X) is not unbiased unless the Bayes risk rδ (Π) = 0. Proof. Suppose that δ(X) is unbiased, i.e., E[δ(X)|θ] = g(θ). Conditioning on θ and using Proposition 1.10, we obtain that E[g(θ)δ(X)] = E{g(θ)E[δ(X)|θ]} = E[g(θ)]2 . Since δ(X) = E[g(θ)|X], conditioning on X and using Proposition 1.10, we obtain that E[g(θ)δ(X)] = E{δ(X)E[g(θ)|X]} = E[δ(X)]2 . Then rδ (Π) = E[δ(X) − g(θ)]2 = E[δ(X)]2 + E[g(θ)]2 − 2E[g(θ)δ(X)] = 0.

242

4. Estimation in Parametric Models

Since rδ (Π) = 0 occurs usually in some trivial cases, a Bayes estimator is typically not unbiased. Hence, Proposition 4.2 can be used to check whether an estimator can be a Bayes estimator w.r.t. some prior under the squared error loss. However, a generalized Bayes estimator may be unbiased; see, for instance, Examples 4.3 and 4.7. Bayes estimators are usually consistent and approximately unbiased. In a particular problem, it is usually easy to check directly whether Bayes estimators are consistent and approximately unbiased (Examples 4.7-4.9), especially when Bayes estimators have explicit forms. Bayes estimators also have some other good asymptotic properties, which are studied in §4.5.3. Let us consider some examples. Example 4.7. Let X = (X1 , ..., Xn ) and Xi ’s be i.i.d. from the exponential distribution E(0, θ) with an unknown θ > 0. Let the prior be such that θ−1 has the gamma distribution Γ(α, γ) with known α > 0 and γ > 0. Then ¯ + γ −1 )−1 ) the posterior of ω = θ−1 is the gamma distribution Γ(n + α, (nX ¯ (verify), where X is the sample mean. Consider first the estimation of θ = ω −1 . The Bayes estimator of θ under the squared error loss is ¯ + γ −1 )n+α Z ∞ ¯ + γ −1 −1 (nX nX ¯ . δ(X) = ω n+α−2 e−(nX+γ )ω dω = Γ(n + α) n+α−1 0 The bias of δ(X) is nθ + γ −1 γ −1 − (α − 1)θ −θ = =O n+α−1 n+α−1

1 . n

¯ It is also easy to see that δ(X) is consistent. The UMVUE of θ is X. 2 ¯ ¯ Since Var(X) = θ /n, rX¯ (Π) > 0 for any Π and, hence, X is not a Bayes ¯ is the generalized Bayes estimator w.r.t. the estimator. In this case, X dΠ improper prior dω = I(0,∞) (ω) and is a limit of Bayes estimators δ(X) as α → 1 and γ → ∞ (exercise). The admissibility of δ(X) is considered in Exercises 32 and 80. Consider next the estimation of e−t/θ = e−tω (see Examples 2.26 and 3.3). The Bayes estimator under the squared error loss is ¯ + γ −1 )n+α Z ∞ −1 (nX ¯ δt (X) = ω n+α−1 e−(nX+γ +t)ω dω Γ(n + α) 0 −(n+α) t = 1+ ¯ . nX + γ −1 Again, this estimator is biased and it is easy to show that δt (X) is consistent as n → ∞. In this case, the UMVUE given in Example 3.3 is neither a Bayes estimator nor a limit of δt (X).

243

4.1. Bayes Decisions and Estimators

Example 4.8. Let X = (X1 , ..., Xn ) and Xi ’s be i.i.d. from N (µ, σ 2 ) with unknown µ ∈ R and σ 2 > 0. Let the prior for ω = (2σ 2 )−1 be the gamma distribution Γ(α, γ) with known α and γ and let the prior for µ be N (µ0 , σ02 /ω) (conditional on ω). Then the posterior p.d.f. of (µ, ω) is proportional to i o n h 2 ¯ − µ)2 + (µ−µ20 ) ω , ω (n+1)/2+α−1 exp − γ −1 + Y + n(X 2σ 0

Pn

¯ 2 and X ¯ is the sample mean. Note that where Y = i=1 (Xi − X) 2 ¯2 + ¯ + µ02 µ + nX ¯ − µ)2 + (µ−µ20 ) = n + 1 2 µ2 − 2 nX n(X 2σ 2σ 2σ 0

0

0

µ20 . 2σ02

Hence, the posterior p.d.f. of (µ, ω) is proportional to i o n h ω (n+1)/2+α−1 exp − γ −1 + W + n + 2σ1 2 (µ − ζ(X))2 ω , 0

where

ζ(X) =

¯+ nX n+

µ0 2σ02 1 2σ02

µ20 1 2 ¯ and W = Y + nX + 2 − n + 2 [ζ(X)]2 . 2σ0 2σ0

Thus, the posterior of ω is the gamma distribution Γ(n/2+α, (γ −1 +W )−1 ) and the posterior of µ (given ω and X) is N ζ(X), [(2n+σ0−2 )ω]−1 . Under the squared error loss, the Bayes estimator of µ is ζ(X) and the Bayes estimator of σ 2 = (2ω)−1 is (γ −1 +W )/(n+2α−2), provided that n+2α > 2. Apparently, these Bayes estimators are biased but the biases are of the order n−1 ; and they are consistent as n → ∞. To consider the last example, we need the following useful lemma whose proof is similar to the proof of result (4.8). Lemma 4.1. Suppose that X has a p.d.f. fθ (x) w.r.t. a σ-finite measure ν. Suppose that θ = (θ1 , θ2 ), θj ∈ Θj , and that the prior has a p.d.f. π(θ) = πθ1 |θ2 (θ1 )πθ2 (θ2 ), where πθ2 (θ2 ) is a p.d.f. w.r.t. a σ-finite measure ν2 on Θ2 and for any given θ2 , πθ1 |θ2 (θ1 ) is a p.d.f. w.r.t. a σ-finite measure ν1 on Θ1 . Suppose further that if θ2 is given, the Bayes estimator of h(θ1 ) = g(θ1 , θ2 ) under the squared error loss is δ(X, θ2 ). Then the Bayes estimator of g(θ1 , θ2 ) under the squared error loss is δ(X) with Z δ(x) = δ(x, θ2 )pθ2 |x (θ2 )dν2 , Θ2

where pθ2 |x (θ2 ) is the posterior p.d.f. of θ 2 given X = x.

244

4. Estimation in Parametric Models

Example 4.9. Consider a linear model Xij = β τ Zi + εij ,

j = 1, ..., ni , i = 1, ..., k,

where β ∈ Rp is unknown, Zi ’s are known vectors, εij ’s are independent, and εij is N (0, σi2 ), j = 1, ..., ni , i = 1, ..., k. Let X be the sample vector containing all Xij ’s. The parameter vector is then θ = (β, ω), where ω = (ω1 , ..., ωk ) and ωi = (2σi2 )−1 . Assume that the prior for θ has the Lebesgue p.d.f. k Y c π(β) ωiα e−ωi /γ , (4.9) i=1

where α > 0, γ > 0, and c > 0 are known constants and π(β) is a known Lebesgue p.d.f. on Rp . The posterior p.d.f. of θ is then proportional to h(X, θ) = π(β)

k Y

n /2+α −[γ −1 +vi (β)]ωi

ωi i

e

,

i=1

Pni (Xij − β τ Zi )2 . If β is known, the Bayes estimator of where vi (β) = j=1 σi2 under the squared error loss is Z 1 h(X, θ) γ −1 + vi (β) R dω = . 2ωi h(X, θ)dω 2α + ni By Lemma 4.1, the Bayes estimator of σi2 is Z −1 γ + vi (β) fβ|X (β)dβ, σ ˆi2 = 2α + ni

(4.10)

where fβ|X (β) ∝

Z

h(X, θ)dω

∝ π(β) ∝ π(β)

k Z Y

α+ni /2 −[γ −1 +vi (β)]ωi

ωi

e

dωi

i=1

k Y −(α+1+ni /2) −1 γ + vi (β)

(4.11)

i=1

is the posterior p.d.f. of β. The Bayes estimator of lτ β for any l ∈ Rp is then the posterior mean of lτ β w.r.t. the p.d.f. fβ|X (β). In this problem, Bayes estimators do not have explicit forms. A numerical method (such as one of those in §4.1.4) has to be used to evaluate Bayes estimators (see Example 4.10).

245

4.1. Bayes Decisions and Estimators

¯ i· and S 2 be the sample mean and variance of Xij , j = 1, ..., ni Let X i 2 (Si is defined to be 0 if ni = 1), and let σ02 = (2αγ)−1 (the prior mean of σi2 ). Then the Bayes estimator σ ˆi2 in (4.10) can be written as ni − 1 2 ni 2α σ2 + S + 2α + ni 0 2α + ni i 2α + ni

Z

¯ i· − β τ Zi )2 fβ|X (β)dβ. (X

(4.12)

The Bayes estimator in (4.12) is a weighted average of prior information, “within group” variation, and averaged squared “residuals”. If ni → ∞, then the first term in (4.12) converges to 0 and the second term in (4.12) is consistent and approximately unbiased for σi2 . Hence, the Bayes estimator σ ˆi2 is consistent and approximately unbiased for σi2 if the mean of the last term in (4.12) tends to 0, which is true under some conditions (see, e.g., Exercise 36). It is easy to see that σ ˆi2 is consistent and 2 approximately unbiased for σi w.r.t. the joint distribution of (X, θ), since the mean of the last term in (4.12) w.r.t. the joint distribution of (X, θ) is bounded by σ02 /ni .

4.1.4 Markov chain Monte Carlo As we discussed previously, Bayes actions or estimators have to be computed numerically in many applications. Typically we need to compute an integral of the form Z Ep (g) =

g(θ)p(θ)dν

Θ

with some function g, where p(θ) is a p.d.f. w.r.t. a σ-finite measure ν on (Θ, BΘ ) and Θ ⊂ Rk . For example, if g is an indicator function of A ∈ BΘ and p(θ) is the posterior p.d.f. of θ given X = x, then Ep (g) is the posterior probability of A; under the squared error loss, Ep (g) is the Bayes action (4.4) if p(θ) is the posterior p.d.f. There are many numerical methods for computing integrals Ep (g); see, for example, §4.5.3 and Berger (1985, §4.9). In this section, we discuss the Markov chain Monte Carlo (MCMC) methods, which are powerful numerical methods not only for Bayesian computations, but also for general statistical computing (see, e.g., §4.4.1). We start with the simple Monte Carlo method, which can be viewed as a special case of the MCMC. Suppose that we can generate i.i.d. θ(1) , ..., θ(m) from a p.d.f. h(θ) > 0 w.r.t. ν. By the SLLN (Theorem 1.13(ii)), as m → ∞, Z m X g(θ(j) )p(θ(j) ) g(θ)p(θ) ˆp (g) = 1 h(θ)dν = Ep (g). → E a.s. (j) m j=1 h(θ) h(θ ) Θ ˆp (g) can be used as a numerical approximation to Ep (g). The Hence E process of generating θ(j) according to h is called importance sampling and

246

4. Estimation in Parametric Models

h(θ) is called the importance function. More discussions on importance sampling can be found, for example, in Berger (1985), Geweke (1989), Shao (1989), and Tanner (1996). When p(θ) is intractable or complex, it is often difficult to choose a function h that is simple enough for importance ˆp (g) as well. sampling and results in a fast convergence of E The simple Monte Carlo method, however, may not work well when k, the dimension of Θ, is large. This is because, when k is large, the converˆp (g) requires a very large m; generating a random vector from gence of E a k-dimensional distribution is usually expensive, if not impossible. More sophisticated MCMC methods are different from the simple Monte Carlo in two aspects: generating random vectors can be done using distributions whose dimensions are much lower than k; and θ(1) , ..., θ(m) are not independent, but form a Markov chain. Let {Y (t) : t = 0, 1, ...} be a Markov chain (§1.4.4) taking values in Y ⊂ Rk . {Y (t) } is homogeneous if and only if P (Y (t+1) ∈ A|Y (t) ) = P (Y (1) ∈ A|Y (0) ) for any t. For a homogeneous Markov chain {Y (t) }, define P (y, A) = P (Y (1) ∈ A|Y (0) = y),

y ∈ Y, A ∈ BY ,

which is called the transition kernel of the Markov chain. Note that P (y, ·) is a probability measure for every y ∈ Y; P (·, A) is a Borel function for every A ∈ BY ; and the distribution of a homogeneous Markov chain is determined by P (y, A) and the distribution ofR Y (0) (initial distribution). Pm MCMC approximates an integral of the form Y g(y)p(y)dν by m−1 t=1 g(Y (t) ) with a Markov chain {Y (t) : t = 0, 1, ...}. The basic justification of the MCMC approximation is given in the following result. Theorem 4.4. Let p(y) be Ra p.d.f. on Y w.r.t. a σ-finite measure ν and g be a Borel function on Y with Y |g(y)|p(y)dν < ∞. Let {Y (t) : t = 0, 1, ...} be a homogeneous Markov chain taking values on Y ⊂ Rk with the transition kernel P (y, A). Then m

1 X g(Y (t) ) →a.s. m t=1

Z

g(y)p(y)dν

(4.13)

Y

and, as t → ∞, P t (y, A) = P (Y (t) ∈ A|Y (0) = y) →a.s.

Z

p(y)dν,

(4.14)

A

provided that (a) the Markov chain is aperiodic in the sense that there does not exist d ≥ 2

247

4.1. Bayes Decisions and Estimators

nonempty disjoint events A0 , ..., Ad−1 in BY such that for all i = 0, ..., d − 1 and all y ∈ Ai , P (y, Aj ) = 1 for j = i + 1 (mod d); R R(b) the Markov chain is p-invariant in the sense that P (y, A)p(y)dν = p(y)dν for all A ∈ BY ; A (c) the Markov chain is p-irreducible in the sense that for any y ∈ Y and any R A with A p(y)dν > 0, there exists a positive integer t such that P t (y, A) in (4.14) is positive; and recurrent in the sense that for any A with R(d) the Markov chain P∞is Harris (t) p(y)dν > 0, P ) = ∞|Y (0) = y = 1 for all y. t=1 IA (Y A

The proof of these results is beyond the scope of this book and, hence, is omitted. It can be found, for example, in Nummelin (1984), Chan (1993), and Tierney (1994). A homogeneous Markov chain satisfying conditions (a)-(d) in Theorem 4.4 is called ergodic with equilibrium distribution p. Result (4.13) means that the MCMC approximation is consistent and result (4.14) indicates that p is the limiting p.d.f. of the Markov chain. One of the key issues in MCMC is the choice of the kernel P (y, A). The first requirement on P (y, A) is that conditions (a)-(d) in Theorem 4.4 be satisfied. Condition (a) is usually easy to check for any given P (y, A). In the following, we consider two popular MCMC methods satisfying conditions (a)-(d). Gibbs sampler One way to construct a p-invariant homogeneous Markov chain is to use conditioning. Suppose that Y has the p.d.f. p(y). Let Yi (or yi ) be the ith component of Y (or y) and let Y−i (or y−i ) be the (k − 1)-vector containing all components of Y (or y) except Yi (or yi ). Then Pi (y−i , A) = P (Y ∈ A|Y−i = y−i ) is a transition kernel for any i. The MCMC method using this kernel is called the single-site Gibbs sampler. Note that Z Z Pi (y−i , A)p(y)dν = E[P (Y ∈ A|Y−i )] = P (Y ∈ A) = p(y)dν A

and, therefore, the chain with kernel Pi (y−i , A) is p-invariant. However, this chain is not p-irreducible since Pi (y−i , ·) puts all its mass on the set ψi−1 (y−i ), where ψi (y) = y−i . Gelfand and Smith (1990) considered a systematic scan Gibbs sampler whose kernel P (y, A) is a composite of k kernels Pi (y−i , A), i = 1, ..., k. More precisely, the chain is defined as follows. Given (t) (t−1) (t−1) (t) Y (t−1) = y (t−1) , we generate y1 from P1 (y2 , ..., yk , · ),..., yj from (t)

(t)

(t−1)

(t−1)

Pj (y1 , ..., yj−1 , yj+1 , ..., yk

(t)

(t)

(t)

, · ),..., yk from Pk (y1 , ..., yk−1 , · ). It can

248

4. Estimation in Parametric Models

be shown that this Markov chain is still p-invariant. We illustrate this with (1) (0) the case of k = 2. Note that Y1 is generated from P2 (y2 , ·), the con(0) (1) (0) ditional distribution of Y given Y2 = y2 . Hence (Y1 , Y2 ) has p.d.f. p. (1) (1) Similarly, we can show that Y (1) = (Y1 , Y2 ) has p.d.f. p. Thus, Z Z P (y, A)p(y)dν = P (Y (1) ∈ A|Y (0) = y)p(y)dν = E[P (Y (1) ∈ A|Y (0) )] = P (Y (1) ∈ A) Z = p(y)dν. A

This Markov chain is also p-irreducible and aperiodic if p(y) > 0 for all y ∈ Y; see, for example, Chan (1993). Finally, if p(y) > 0 for all y ∈ Y, then P (y, A) ≪ the distribution with p.d.f. p for all y and, by Corollary 1 of Tierney (1994), the Markov chain is Harris recurrent. Thus, Theorem 4.4 applies and (4.13) and (4.14) hold. The previous Gibbs sampler can obviously be extended to the case where yi ’s are subvectors (of possibly different dimensions) of y. Let us now return to Bayesian computation and consider the following example. Example 4.10. Consider Example 4.9. Under the given prior for θ = (β, ω), it is difficult to generate random vectors directly from the posterior p.d.f., given X = x (which does not have a familiar form). To apply a Gibbs sampler with y = θ, y1 = β, and y2 = ω, we need to generate random vectors from the posterior of β, given x and ω, and the posterior of ω, given x and β. From (4.9) and (4.11), the posterior of ω = (ω1 , ..., ωk ), given x and β, is a product of marginals of ωi ’s that are the gamma distributions Γ(α + 1 + ni /2, [γ −1 + vi (β)]−1 ), i = 1, ..., k. Assume now that π(β) ≡ 1 (noninformative prior for β). It follows from (4.9) that the posterior p.d.f. of β, given x and ω, is proportional to k Y

i=1

e−ωi vi (β) ∝ e−kW

1/2

Zβ−W 1/2 Xk2

,

where W is the diagonal on the diagonal Pk block matrix whose ith block 1/2 is ωi Ini . Let n = Zβ, given X i=1 ni . Then, the posterior of W and ω, is Nn (W 1/2 X, 2−1 In ) and the posterior of β, given X and ω, is Np ((Z τ W Z)−1 Z τ W X, 2−1 (Z τ W Z)−1 ) (Z τ W Z is assumed of full rank for simplicity), since β = [(Z τ W Z)−1 Z τ W 1/2 ]W 1/2 Zβ. Note that random generation using these two posterior distributions is fairly easy.

249

4.1. Bayes Decisions and Estimators

The Metropolis algorithm A large class of MCMC methods are obtained using the Metropolis algorithm (Metropolis et al., 1953). We introduce Hastings’ version of the algorithm. Let Q(y, A) be a transition kernel of a homogeneous Markov chain satisfying Z Q(y, A) =

q(y, z)dν(z)

A

for a measurable function q(y, z) ≥ 0 on Y × Y and R a σ-finite measure ν on (Y, BY ). Without loss of generality, assume that Y p(y)dν = 1 and that p is not concentrated on a single point. Define o n ( p(z)q(z,y) ,1 p(y)q(y, z) > 0 min p(y)q(y,z) α(y, z) = 1 p(y)q(y, z) = 0

and p(y, z) =

q(y, z)α(y, z) 0

y= 6 z y = z.

The Metropolis kernel P (y, A) is defined by Z p(y, z)dν(z) + r(y)διy (A), P (y, A) =

(4.15)

A

R where r(y) = 1 − p(y, z)dν(z) and διy is the point mass at y defined in (1.22). The corresponding Markov chain can be described as follows. If the chain is currently at a point Y (t) = y, then it generates a candidate value z for the next location Y (t+1) from Q(y, ·). With probability α(y, z), the chain moves to Y (t+1) = z. Otherwise, the chain remains at Y (t+1) = y. Note that this algorithm only depends on p(y) through p(y)/p(z). Thus, it can be used when p(y) is known up to a normalizing constant, which often occurs in Bayesian analysis. We now show that a Markov chain with a Metropolis kernel P (y, A) is p-invariant. First, by the definition of p(y, z) and α(y, z), p(y)p(y, z) = p(z)p(z, y)

for any y and z. Then, for any A ∈ BY , Z Z Z Z p(y, z)dν(z) p(y)dν(y) + r(y)διy (A)p(y)dν(y) P (y, A)p(y)dν = A Z Z Z = p(y, z)p(y)dν(y) dν(z) + r(y)p(y)dν(y) A A Z Z Z = p(z, y)p(z)dν(y) dν(z) + r(y)p(y)dν(y) A

A

250

4. Estimation in Parametric Models

=

Z

A

=

Z

[1 − r(z)]p(z)dν(z) +

Z

r(z)p(z)dν(z)

A

p(z)dν(z).

A

If a MarkovR chain with a Metropolis kernel defined by (4.15) is pirreducible and r(y)>0 p(y)dν > 0, then, by the results of Nummelin (1984, §2.4), the chain is aperiodic; by Corollary 2 of Tierney (1994), the chain is Harris recurrent. Hence, to apply Theorem 4.4 to a Markov chain with a Metropolis kernel, it suffices to show that the chain is p-irreducible. Lemma 4.2. Suppose that Q(y, A) is the transition kernel of a p-irreducible Markov chain and that either q(y, z) > 0 for all y and z or q(y, z) = q(z, y) for all y and z. Then the chain with the Metropolis kernel p(y, A) in (4.15) is p-irreducible. Proof. It can be shown (exercise) that if Q is any transition kernel of a homogeneous Markov chain, then Z Y Z Z t Qt (y, A) = ··· q(zn−j+1 , zn−j )dν(zn−j ), (4.16) A

j=1

R where zn = y, y ∈ Y, and A ∈ RBY . Let y ∈ Y, A ∈ BY with A p(z)dν > 0, and By = {z : α(y, z) = 1}. If A∩B c p(z)dν > 0, then y Z Z q(z, y)p(z) q(y, z)α(y, z)dν(z) = P (y, A) ≥ dν(z) > 0, p(y) A∩Byc A∩Byc

which follows from either Rq(z, y) > 0 or q(z, y) = q(y, z) > 0 on Byc . If R A∩B c p(z)dν = 0, then A∩By p(z)dν > 0. From the irreducibility of y

Q(y, A), there exists a t ≥ 1 such that Qt (y, A ∩ By ) > 0. Then, by (4.15) and (4.16), P t (y, A) ≥ P t (y, A ∩ By ) ≥ Qt (y, A ∩ By ) > 0.

Two examples of q(y, z) given by Tierney (1994) are q(y, z) = f (z − y) with a Lebesgue p.d.f. f on Rk , which corresponds to a random walk chain, and q(y, z) = f (z) with a p.d.f. f , which corresponds to an independence chain and is closely related to the importance sampling discussed earlier. Although the MCMC methods have been used over the last 50 years, the research on the theory of MCMC is still very active. Important topics include the choice of the transition kernel for MCMC; the rate of the convergence in (4.13); the choice of the Monte Carlo size m; and the estimation of the errors due to Monte Carlo. See more results and discussions in Tierney (1994), Basag et al. (1995), Tanner (1996), and the references therein.

251

4.2. Invariance

4.2 Invariance The concept of invariance is introduced in §2.3.2 (Definition 2.9). In this section, we study the best invariant estimators and their properties in one-parameter location families (§4.2.1), in one-parameter scale families (§4.2.2), and in general location-scale families (§4.2.3). Note that invariant estimators are also called equivariant estimators.

4.2.1 One-parameter location families Assume that the sample X = (X1 , ..., Xn ) has a joint distribution Pµ with a Lebesgue p.d.f. f (x1 − µ, ..., xn − µ), (4.17) where f is known and µ ∈ R is an unknown location parameter. The family P = {Pµ : µ ∈ R} is called a one-parameter location family, a special case of the general location-scale family described in Definition 2.3. It is invariant under the location transformations gc (X) = (X1 + c,..., Xn + c), c ∈ R. We consider the estimation of µ as a statistical decision problem with action space A = R and loss function L(µ, a). It is natural to consider the same transformation in the action space, i.e., if Xi is transformed to Xi + c, then our action a is transformed to a+ c. Consequently, the decision problem is invariant under location transformation if and only if L(µ, a) = L(µ + c, a + c)

for all c ∈ R,

which is equivalent to L(µ, a) = L(a − µ)

(4.18)

for a Borel function L(·) on R. According to Definition 2.9 (see also Example 2.24), an estimator T (decision rule) of µ is location invariant if and only if T (X1 + c, ..., Xn + c) = T (X1 , ..., Xn ) + c.

(4.19)

Many estimators of µ, such as the sample mean and weighted average of the order statistics, are location invariant. The following result provides a characterization of location invariant estimators. Proposition 4.3. Let T0 be a location invariant estimator of µ. Let di = xi − xn , i = 1, ..., n − 1, and d = (d1 , ..., dn−1 ). A necessary and sufficient condition for an estimator T to be location invariant is that there exists a Borel function u on Rn−1 (u ≡ a constant if n = 1) such that T (x) = T0 (x) − u(d)

for all x ∈ Rn .

(4.20)

252

4. Estimation in Parametric Models

Proof. It is easy to see that T given by (4.20) satisfies (4.19) and, therefore, is location invariant. Suppose that T is location invariant. Let u ˜(x) = T (x) − T0 (x) for any x ∈ Rn . Then u ˜(x1 + c, ..., xn + c) = T (x1 + c, ..., xn + c) − T0 (x1 + c, ..., xn + c) = T (x1 , ..., xn ) − T0 (x1 , ..., xn ) for all c ∈ R and xi ∈ R. Putting c = −xn leads to u ˜(x1 − xn , ..., xn−1 − xn , 0) = T (x) − T0 (x),

x ∈ Rn .

The result follows with u(d1 , ..., dn−1 ) = u˜(x1 − xn , ..., xn−1 − xn , 0). Therefore, once we have a location invariant estimator T0 of µ, any other location invariant estimator of µ can be constructed by taking the difference between T0 and a Borel function of the ancillary statistic D = (X1 − Xn , ..., Xn−1 − Xn ). The next result states an important property of location invariant estimators. Proposition 4.4. Let X be distributed with the p.d.f. given by (4.17) and let T be a location invariant estimator of µ under the loss function given by (4.18). If the bias, variance, and risk of T are well defined, then they are all constant (do not depend on µ). Proof. The result for the bias follows from Z bT (µ) = T (x)f (x1 − µ, ..., xn − µ)dx − µ Z = T (x1 + µ, ..., xn + µ)f (x)dx − µ Z = [T (x) + µ]f (x)dx − µ Z = T (x)f (x)dx. The proof of the result for variance or risk is left as an exercise. An important consequence of this result is that the problem of finding the best location invariant estimator reduces to comparing constants instead of risk functions. The following definition can be used not only for location invariant estimators, but also for general invariant estimators. Definition 4.2. Consider an invariant estimation problem in which all invariant estimators have constant risks. An invariant estimator T is called the minimum risk invariant estimator (MRIE) if and only if T has the smallest risk among all invariant estimators.

253

4.2. Invariance

Theorem 4.5. Let X be distributed with the p.d.f. given by (4.17) and consider the estimation of µ under the loss function given by (4.18). Suppose that there is a location invariant estimator T0 of µ with finite risk. Let D = (X1 − Xn , ..., Xn−1 − Xn ). (i) Assume that for each d there exists a u∗ (d) that minimizes h(d) = E0 [L(T0 (X) − u(d))|D = d] over all functions u, where the expectation E0 is calculated under the assumption that X has p.d.f. f (x1 , ..., xn ). Then an MRIE exists and is given by T∗ (X) = T0 (X) − u∗ (D). (ii) The function u∗ in (i) exists if L(t) is convex and not monotone; it is unique if L is strictly convex. (iii) If T0 and D are independent, then u∗ is a constant that minimizes E0 [L(T0 (X) − u)]. If, in addition, the distribution of T0 is symmetric about µ and L is convex and even, then u∗ = 0. Proof. By Theorem 1.7 and Propositions 4.3 and 4.4, RT (µ) = E0 [h(D)], where T (X) = T0 (X) − u(D). This proves part (i). If L is (strictly) convex and not monotone, then E0 [L(T0 (x)−a)|D = d] is (strictly) convex and not monotone in a (exercise). Hence lim|a|→∞ E0 [L(T0 (x) − a)|D = d] = ∞. This proves part (ii). The proof of part (iii) is left as an exercise. Theorem 4.6. Assume the conditions of Theorem 4.5 and that the loss is the squared error loss. (i) The unique MRIE of µ is R∞ tf (X1 − t, ..., Xn − t)dt T∗ (X) = R−∞ , ∞ f (X1 − t, ..., Xn − t)dt −∞

which is known as the Pitman estimator of µ. (ii) The MRIE of µ is unbiased. Proof. (i) Under the squared error loss,

u∗ (d) = E0 [T0 (X)|D = d]

(4.21)

(exercise). Let T0 (X) = Xn (the nth observation). Then Xn is location invariant. If there exists a location invariant estimator of µ with finite risk, then E0 (Xn |D = d) is finite a.s. P (exercise). By Proposition 1.8, when µ = 0, the joint Lebesgue p.d.f. of (D, Xn ) is f (d1 + xn , ..., dn−1 + xn , xn ), d = (d1 , ..., dn−1 ). The conditional p.d.f. of Xn given D = d is then f (d + xn , ..., dn−1 + xn , xn ) R∞ 1 f (d1 + t, ..., dn−1 + t, t)dt −∞

254

4. Estimation in Parametric Models

(see (1.61)). By Proposition 1.9, R∞

tf (d1 + t, ..., dn−1 + t, t)dt

E0 (Xn |D = d) = R−∞ ∞

−∞

R∞

= R−∞ ∞

f (d1 + t, ..., dn−1 + t, t)dt tf (x1 − xn + t, ..., xn−1 − xn + t, t)dt

f (x1 − xn + t, ..., xn−1 − xn + t, t)dt R∞ uf (x1 − u, ..., xn − u)du = xn − R−∞ ∞ −∞ f (x1 − u, ..., xn − u)du −∞

by letting u = xn −t. The result in (i) follows from T∗ (X) = Xn −E(Xn |D) (Theorem 4.5). (ii) Let b be the constant bias of T∗ (Proposition 4.4). Then T1 (X) = T∗ (X) − b is a location invariant estimator of µ and RT1 = E[T∗ (X) − b − µ]2 = Var(T∗ ) ≤ Var(T∗ ) + b2 = RT∗ . Since T∗ is the MRIE, b = 0, i.e., T∗ is unbiased. Theorem 4.6(ii) indicates that we only need to consider unbiased location invariant estimators in order to find the MRIE, if the loss is the squared error loss. In particular, a location invariant UMVUE is an MRIE. Example 4.11. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with an unknown ¯ is location invariant. Since X ¯ is the µ ∈ R and a known σ 2 . Note that X UMVUE of µ (§2.1), it is the MRIE under the squared error loss. Since the ¯ is symmetric about µ and X ¯ is independent of D (Basu’s distribution of X ¯ is an MRIE if L is convex theorem), it follows from Theorem 4.5(iii) that X and even. Example 4.12. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(µ, θ), where θ is known and µ ∈ R is unknown. Since X(1) − θ/n is location invariant and is the UMVUE of µ, it is the MRIE under the squared error loss. Note that X(1) is independent of D (Basu’s theorem). By Theorem 4.5(iii), an MRIE is of the form X(1) − u∗ with a constant u∗ . For the absolute error loss, X(1) − θ log 2/n is an MRIE (exercise). Example 4.13. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (µ − 12 , µ + 12 ) with an unknown µ ∈ R. Consider the squared error loss. Note that f (x1 − µ, ..., xn − µ) =

1 0

µ − 12 ≤ x(1) ≤ x(n) ≤ µ + otherwise.

1 2

255

4.2. Invariance

By Theorem 4.6(i), the MRIE of µ is T∗ (X) =

Z

X(1) + 12

tdt X(n) − 12

Z

X(1) + 12

dt =

X(n) − 12

X(1) + X(n) . 2

We end this section with a brief discussion of the admissibility of MRIE’s in a one-parameter location problem. Under the squared error loss, the MRIE (Pitman’s estimator) is admissible if there exists a location invariant estimator T0 with E|T0 (X)|3 < ∞ (Stein, 1959). Under a general loss function, an MRIE is admissible when it is a unique MRIE (under some other minor conditions). See Farrell (1964), Brown (1966), and Brown and Fox (1974) for further discussions.

4.2.2 One-parameter scale families Assume that the sample X = (X1 , ..., Xn ) has a joint distribution Pσ with a Lebesgue p.d.f. x1 xn 1 (4.22) σn f σ , ..., σ ,

where f is known and σ > 0 is an unknown scale parameter. The family P = {Pσ : σ > 0} is called a one-parameter scale family and is a special case of the general location-scale family in Definition 2.3. This family is invariant under the scale transformations gr (X) = rX, r > 0. We consider the estimation of σ h with A = [0, ∞), where h is a nonzero constant. The transformation gr induces the transformation gr (σ h ) = rh σ h . Hence, a loss function L is scale invariant if and only if L(rσ, rh a) = L(σ, a)

for all r > 0,

which is equivalent to L(σ, a) = L

a σh

(4.23)

for a Borel function L(·) on [0, ∞). An example of a loss function satisfying (4.23) is p a |a − σ h |p L(σ, a) = h − 1 = , (4.24) σ σ ph

where p ≥ 1 is a constant. However, the squared error loss does not satisfy (4.23). An estimator T of σ h is scale invariant if and only if T (rX1 , ..., rXn ) = rh T (X1 , ..., Xn ).

2 Examples of scale invariant estimators are √ the sample variance S (for h = 2), the sample standard deviation S = S 2 (for h = 1), the sample range

256

4. Estimation in Parametric Models

P ¯ X(n) − X(1) (for h = 1), and the sample mean deviation n−1 ni=1 |Xi − X| (for h = 1). The following result is an analogue of Proposition 4.3. Its proof is left as an exercise. Proposition 4.5. Let T0 be a scale invariant estimator of σ h . A necessary and sufficient condition for an estimator T to be scale invariant is that there exists a positive Borel function u on Rn such that T (x) = T0 (x)/u(z)

for all x ∈ Rn ,

where z = (z1 , ..., zn ), zi = xi /xn , i = 1, ..., n − 1, and zn = xn /|xn |. The next result is similar to Proposition 4.4. It applies to any invariant problem defined in Definition 2.9. We use the notation in Definition 2.9. Theorem 4.7. Let P be a family invariant under G (a group of transformations). Suppose that the loss function is invariant and T is an invariant decision rule. Then the risk function of T is a constant. The proof is left as an exercise. Note that a special case of Theorem 4.7 is that any scale invariant estimator of σ h has a constant risk and, therefore, an MRIE (Definition 4.2) of σ h usually exists. However, Proposition 4.4 is not a special case of Theorem 4.7, since the bias of a scale invariant estimator may not be a constant in general. For example, the bias of the sample standard deviation is a function of σ. The next result and its proof are analogues of those of Theorem 4.5. Theorem 4.8. Let X be distributed with the p.d.f. given by (4.22) and consider the estimation of σ h under the loss function given by (4.23). Suppose that there is a scale invariant estimator T0 of σ h with finite risk. Let Z = (Z1 , ..., Zn ) with Zi = Xi /Xn , i = 1, ..., n − 1, and Zn = Xn /|Xn |. (i) Assume that for each z there exists a u∗ (z) that minimizes E1 [L(T0 (X)/u(z))|Z = z] over all positive Borel functions u, where the conditional expectation E1 is calculated under the assumption that X has p.d.f. f (x1 , ..., xn ). Then, an MRIE exists and is given by T∗ (X) = T0 (X)/u∗ (Z). (ii) The function u∗ in (i) exists if γ(t) = L(et ) is convex and not monotone; it is unique if γ(t) is strictly convex.

257

4.2. Invariance

The loss function given by (4.24) satisfies the condition in Theorem 4.8(ii). A loss function corresponding to the squared error loss in this problem is the loss function (4.24) with p = 2. We have the following result similar to Theorem 4.6 (its proof is left as an exercise). Corollary 4.1. Under the conditions of Theorem 4.8 and the loss function (4.24) with p = 2, the unique MRIE of σ h is R ∞ n+h−1 t f (tX1 , ..., tXn )dt T0 (X)E1 [T0 (X)|Z] 0 = R∞ , T∗ (X) = 2 n+2h−1 E1 {[T0 (X)] |Z} t f (tX1 , ..., tXn )dt 0

which is known as the Pitman estimator of σ h .

Example 4.14. Let X1 , ...,P Xn be i.i.d. from N (0, σ 2 ) and consider the esn 2 timation of σ . Then T0 = i=1 Xi2 is scale invariant. By Basu’s theorem, T0 is independent of Z. Hence u∗ in Theorem 4.8 is a constant minimizing E1 [L(T0 /u)] over u > 0. When the loss is given by (4.24) with p = 2, by Corollary 4.1, the MRIE (Pitman’s estimator) is n

T∗ (X) =

T0 (X)E1 [T0 (X)] 1 X 2 = X , 2 E1 [T0 (X)] n + 2 i=1 i

since T0 has the chi-square distribution χ2n when σ = 1. Note that the UMVUE of σ 2 is T0 /n, which is different from the MRIE. Example 4.15. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (0, σ) and consider the estimation of σ. By Basu’s theorem, the scale invariant estimator X(n) is independent of Z. Hence u∗ in Theorem 4.8 is a constant minimizing E1 [L(X(n) /u)] over u > 0. When the loss is given by (4.24) with p = 2, by Corollary 4.1, the MRIE (Pitman’s estimator) is T∗ (X) =

X(n) E1 X(n) (n + 2)X(n) . = 2 E1 X(n) n+1

4.2.3 General location-scale families Assume that X = (X1 , ..., Xn ) has a joint distribution Pθ with a Lebesgue p.d.f. x1 −µ xn −µ 1 , (4.25) σn f σ , ..., σ

where f is known, θ = (µ, σ) ∈ Θ, and Θ = R × (0, ∞). The family P = {Pθ : θ ∈ Θ} is a location-scale family defined by Definition 2.3 and is invariant under the location-scale transformations of the form gc,r (X) = (rX1 + c, ..., rXn + c), c ∈ R, r > 0, which induce similar transformations on Θ: gc,r (θ) = (rµ + c, rσ), c ∈ R, r > 0.

258

4. Estimation in Parametric Models

Consider the estimation of σ h with a fixed h 6= 0 under the loss function (4.23), which is invariant under the location-scale transformations gc,r . An estimator T of σ h is location-scale invariant if and only if T (rX1 + c, ..., rXn + c) = rh T (X1 , ..., Xn ).

(4.26)

By Theorem 4.7, any location-scale invariant T has a constant risk. Letting r = 1 in (4.26), we obtain that T (X1 + c, ..., Xn + c) = T (X1 , ..., Xn ) for all c ∈ R. Therefore, T is a function of D = (D1 , ..., Dn−1 ), Di = Xi − Xn , i = 1, ..., n − 1. From (4.25), the joint Lebesgue p.d.f. of D is R ∞ d1 dn−1 1 f + t, ..., + t, t dt, (4.27) n−1 σ σ σ −∞

which is of the form (4.22) with n replaced by n−1 and xi ’s replaced by di ’s. It follows from Theorem 4.8 that if T0 (D) is any finite risk scale invariant estimator of σ h based on D, then an MRIE of σ h is T∗ (D) = T0 (D)/u∗ (W ),

(4.28)

where W = (W1 , ..., Wn−1 ), Wi = Di /Dn−1 , i = 1, ..., n − 2, Wn−1 = ˜1 [L(T0 (D)/u(w))|W = w] Dn−1 /|Dn−1 |, u∗ (w) is any number minimizing E ˜ over all positive Borel functions u, and E1 is the conditional expectation calculated under the assumption that D has p.d.f. (4.27) with σ = 1. Consider next the estimation of µ. Under the location-scale transformation gc,r , it can be shown (exercise) that a loss function is invariant if and only if it is of the form L a−µ . (4.29) σ An estimator T of µ is location-scale invariant if and only if T (rX1 + c, ..., rXn + c) = rT (X1 , ..., Xn ) + c. Again, by Theorem 4.7, the risk of an invariant T is a constant. The following result is an analogue of Proposition 4.3 or 4.5. Proposition 4.6. Let T0 be any estimator of µ invariant under locationscale transformation and let T1 be any estimator of σ satisfying (4.26) with h = 1 and T1 > 0. Then an estimator T of µ is location-scale invariant if and only if there is a Borel function u on Rn−1 such that T (X) = T0 (X) − u(W )T1 (X), where W is given in (4.28).

259

4.2. Invariance

The proofs of Proposition 4.6 and the next result, an analogue of Theorem 4.5 or 4.8, are left as exercises. Theorem 4.9. Let X be distributed with p.d.f. given by (4.25) and consider the estimation of µ under the loss function given by (4.29). Suppose that there is a location-scale invariant estimator T0 of µ with finite risk. Let T1 be given in Proposition 4.6. Then an MRIE of µ is T∗ (X) = T0 (X) − u∗ (W )T1 (X), where W is given in (4.28), u∗ (w) is any number minimizing E0,1 [L(T0 (X) − u(w)T1 (X))|W = w] over all Borel functions u, and E0,1 is computed under the assumption that X has the p.d.f. (4.25) with µ = 0 and σ = 1. Corollary 4.2. Under the conditions of Theorem 4.9 and the loss function (a − µ)2 /σ 2 , u∗ (w) in Theorem 4.9 is equal to u∗ (w) =

E0,1 [T0 (X)T1 (X)|W = w] . E0,1 {[T1 (X)]2 |W = w}

Example 4.16. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ), where µ ∈ R and σ 2 > 0 are unknown. Consider first the estimation of σ 2 under loss function (4.23). The sample variance S 2 is location-scale invariant and is independent of W in (4.28) (Basu’s theorem). Thus, by (4.28), S 2 /u∗ is an MRIE, ˜1 [L(S 2 /u)] over all u > 0. If the loss where u∗ is a constant minimizing E function is given by (4.24) with p = 2, then by Corollary 4.1, the MRIE of σ 2 is n

T∗ (X) =

˜1 (S 2 ) S 2E S2 1 X ¯ 2, = 2 = (Xi − X) 2 2 2 ˜ (n − 1)/(n − 1) n + 1 i=1 E1 (S )

since (n − 1)S 2 has a chi-square distribution χ2n−1 when σ = 1. Next, consider the estimation of µ under the loss function (4.29). Since ¯ is a location-scale invariant estimator of µ and is independent of W in X (4.28) (Basu’s theorem), by Theorem 4.9, an MRIE of µ is ¯ − u∗ S 2 , T∗ (X) = X where u∗ is a constant. If L in (4.29) is convex and even, then u∗ = 0 (see ¯ is an MRIE of µ. Theorem 4.5(iii)) and, hence, X Example 4.17. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (µ − 12 σ, µ + 12 σ), where µ ∈ R and σ > 0 are unknown. Consider first the

260

4. Estimation in Parametric Models

estimation of σ under the loss function (4.24) with p = 2. The sample range X(n) − X(1) is a location-scale invariant estimator of σ and is independent of W in (4.28) (Basu’s theorem). By (4.28) and Corollary 4.1, the MRIE of σ is T∗ (X) =

˜1 (X(n) − X(1) ) (X(n) − X(1) )E (n + 2)(X(n) − X(1) ) = . 2 ˜ n E1 (X(n) − X(1) )

Consider now the estimation of µ under the loss function (4.29). Since (X(1) + X(n) )/2 is a location-scale invariant estimator of µ and is independent of W in (4.28) (Basu’s theorem), by Theorem 4.9, an MRIE of µ is X(1) + X(n) T∗ (X) = − u∗ (X(n) − X(1) ), 2 where u∗ is a constant. If L in (4.29) is convex and even, then u∗ = 0 (see Theorem 4.5(iii)) and, hence, (X(1) + X(n) )/2 is an MRIE of µ. Finding MRIE’s in various location-scale families under transformations AX +c, where A ∈ T and c ∈ C with given T and C, can be done in a similar way. We only provide some brief discussions for two important cases. The first case is the two-sample location-scale problem in which two samples, X = (X1 , ..., Xm ) and Y = (Y1 , ..., Yn ), are taken from a distribution with Lebesgue p.d.f. yn −µy x1 −µx xm −µx y1 −µy 1 , (4.30) σm σn f σx , ..., σx , σy , ..., σy x

y

where f is known, µx ∈ R and µy ∈ R are unknown location parameters, and σx > 0 and σy > 0 are unknown scale parameters. The family of distributions is invariant under the transformations g(X, Y ) = (rX1 + c, ..., rXm + c, r′ Y1 + c′ , ..., r′ Yn + c′ ),

(4.31)

where r > 0, r′ > 0, c ∈ R, and c′ ∈ R. The parameters to be estimated in this problem are usually ∆ = µy − µx and η = (σy /σx )h with a fixed h 6= 0. If X and Y are from two populations, ∆ and η are measures of the difference between the two populations. For estimating η, results similar to those in this section can be established. For estimating ∆, MRIE’s can be obtained under some conditions. See Exercises 63-65. The second case is the general linear model (3.25) under the assumption that εi ’s are i.i.d. with the p.d.f. σ −1 f (x/σ), where f is a known Lebesgue p.d.f. The family of populations is invariant under the transformations g(X) = rX + Zc,

r ∈ (0, ∞), c ∈ Rp

(4.32)

261

4.3. Minimaxity and Admissibility

(exercise). The estimation of lτ β with l ∈ R(Z) is invariant under the a−lτ β and the LSE lτ βˆ is an invariant estimator of lτ β loss function L σ (exercise). When f is normal, the following result can be established using an argument similar to that in Example 4.16. Theorem 4.10. Consider model (3.25) with assumptionA1. τ (i) Under transformations (4.32) and the loss function L a−lσ β , where L is convex and even, the LSE lτ βˆ is an MRIE of lτ β for any l ∈ R(Z). (ii) Under transformations (4.32) and the loss function (a − σ 2 )2 /σ 4 , the MRIE of σ 2 is SSR/(n − r + 2), where SSR is given by (3.35) and r is the rank of Z. MRIE’s in a parametric family with a multi-dimensional θ are often inadmissible. See Lehmann (1983, p. 285) for more discussions.

4.3 Minimaxity and Admissibility Consider the estimation of a real-valued ϑ = g(θ) based on a sample X from Pθ , θ ∈ Θ, under a given loss function. A minimax estimator minimizes the maximum risk supθ∈Θ RT (θ) over all estimators T (see §2.3.2). A unique minimax estimator is admissible, since any estimator better than a minimax estimator is also minimax. This indicates that we should consider minimaxity and admissibility together. The situation is different for a UMVUE (or an MRIE), since if a UMVUE (or an MRIE) is inadmissible, it is dominated by an estimator that is not unbiased (or invariant).

4.3.1 Estimators with constant risks By minimizing the maximum risk, a minimax estimator tries to do as well as possible in the worst case. Such an estimator can be very unsatisfactory. However, if a minimax estimator has some other good properties (e.g., it is a Bayes estimator), then it is often a reasonable estimator. Here we study when estimators having constant risks (e.g., MRIE’s) are minimax. Theorem 4.11. Let Π be a proper prior on Θ and δ be a Bayes estimator of ϑ w.r.t. Π. Let ΘΠ = {θ : Rδ (θ) = supθ∈Θ Rδ (θ)}. If Π(ΘΠ ) = 1, then δ is minimax. If, in addition, δ is the unique Bayes estimator w.r.t. Π, then it is the unique minimax estimator. Proof. Let T be any other estimator of ϑ. Then Z Z sup RT (θ) ≥ RT (θ)dΠ ≥ Rδ (θ)dΠ = sup Rδ (θ). θ∈Θ

ΘΠ

ΘΠ

θ∈Θ

262

4. Estimation in Parametric Models

If δ is the unique Bayes estimator, then the second inequality in the previous expression should be replaced by > and, therefore, δ is the unique minimax estimator. The condition of Theorem 4.11 essentially means that δ has a constant risk. Thus, a Bayes estimator having constant risk is minimax. Example 4.18. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). Consider the estimation of p under the squared er¯ has risk p(1 − p)/n which is not constant. In fact, ror loss. The UMVUE X ¯ is not minimax (Exercise 67). To find a minimax estimator by applying X Theorem 4.11, we consider the Bayes estimator w.r.t. the beta distribution B(α, β) with known α and β (Exercise 1): ¯ (α + β + n). δ(X) = (α + nX)

A straightforward calculation shows that

Rδ (p) = [np(1 − p) + (α − αp − βp)2 ] (α + β + n)2 .

To apply Theorem 4.11, we need to find values of α > 0 and β > 0 such that Rδ (p) √ is constant. It can be shown that Rδ (p) is constant if and only if α = β = n/2, which leads to the unique minimax estimator √ √ ¯ + n/2) (n + n). T (X) = (nX √ The risk of T is RT = 1/[4(1 + n)2 ]. Note that T is a Bayes estimator and has some good properties. Com¯ we find that T has smaller risk if and paring the risk of T with that of X, only if q q 1 1 n 1 1 n p ∈ 2 − 2 1 − (1+√n)2 , 2 + 2 1 − (1+√n)2 . (4.33)

¯ for most Thus, for a small n, T is better (and can be much better) than X of the range of p (Figure 4.1). When n → ∞, the interval in (4.33) shrinks ¯ is better than T toward 12 . Hence, for a large (and even moderate) n, X for most of the range of p (Figure 4.1). The limit of the asymptotic relative ¯ is 4p(1 − p), which is always smaller than 1 when efficiency of T w.r.t. X 1 p 6= 2 and equals 1 when p = 12 . The minimax estimator depends strongly on the loss function. To see this, let us consider the loss function L(p, a) = (a−p)2 /[p(1−p)]. Under this ¯ has constant risk and is the unique Bayes estimator w.r.t. loss function, X ¯ is the unique minimax the uniform prior on (0, 1). By Theorem 4.11, X √ estimator. On the other hand, the risk of T is equal to 1/[4(1+ n)2 p(1−p)], which is unbounded.

263

4.3. Minimaxity and Admissibility n=4

n=9

n=16

0.0

0.10

mse

0.20

0.30

0.0

0.10

mse

0.20

0.30

n=1

0.0

0.25

0.5 p

0.75

1.0

0.0

0.25

0.5

0.75

1.0

p

¯ (curve) and T (X) (straight line) in Example 4.18 Figure 4.1: mse’s of X In many cases a constant risk estimator is not a Bayes estimator (e.g., an unbiased estimator under the squared error loss), but a limit of Bayes estimators w.r.t. a sequence of priors. Then the following result may be used to find a minimax estimator. Theorem 4.12. Let Πj , j = 1, 2, ..., be a sequence of priors and rj be the Bayes risk of a Bayes estimator of ϑ w.r.t. Πj . Let T be a constant risk estimator of ϑ. If lim inf j rj ≥ RT , then T is minimax. The proof of this theorem is similar to that of Theorem 4.11. Although Theorem 4.12 is more general than Theorem 4.11 in finding minimax estimators, it does not provide uniqueness of the minimax estimator even when there is a unique Bayes estimator w.r.t. each Πj . In Example 2.25, we actually applied the result in Theorem 4.12 to show ¯ as an estimator of µ = EX1 when X1 , ..., Xn are i.i.d. the minimaxity of X from a normal distribution with a known σ 2 = Var(X1 ), under the squared ¯ in the case where σ 2 is unknown, error loss. To discuss the minimaxity of X we need the following lemma.

264

4. Estimation in Parametric Models

Lemma 4.3. Let Θ0 be a subset of Θ and T be a minimax estimator of ϑ when Θ0 is the parameter space. Then T is a minimax estimator if sup RT (θ) = sup RT (θ). θ∈Θ

θ∈Θ0

Proof. If there is an estimator T0 with supθ∈Θ RT0 (θ) < supθ∈Θ RT (θ), then sup RT0 (θ) ≤ sup RT0 (θ) < sup RT (θ) = sup RT (θ), θ∈Θ0

θ∈Θ

θ∈Θ

θ∈Θ0

which contradicts the minimaxity of T when Θ0 is the parameter space. Hence, T is minimax when Θ is the parameter space. Example 4.19. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown θ = (µ, σ 2 ). Consider the estimation of µ under the squared error loss. Suppose first that Θ = R × (0, c] with a constant c > 0. Let Θ0 = R × {c}. From ¯ is a minimax estimator of µ when the parameter space Example 2.25, X ¯ is also minimax when is Θ0 . An application of Lemma 4.3 shows that X 2 the parameter space is Θ. Although σ is assumed to be bounded by c, the ¯ does not depend on c. minimax estimator X Consider next the case where Θ = R × (0, ∞), i.e., σ2 is unbounded. Let T be any estimator of µ. For any fixed σ 2 , σ2 ≤ sup RT (θ), n µ∈R ¯ that is minimax when σ 2 is known (Example since σ 2 /n is the risk of X 2 2.25). Letting σ → ∞, we obtain that supθ RT (θ) = ∞ for any estimator T . Thus, minimaxity is meaningless (any estimator is minimax). Theorem 4.13. Suppose that T as an estimator of ϑ has constant risk and is admissible. Then T is minimax. If the loss function is strictly convex, then T is the unique minimax estimator. Proof. By the admissibility of T , if there is another estimator T0 with supθ RT0 (θ) ≤ RT , then RT0 (θ) = RT for all θ. This proves that T is minimax. If the loss function is strictly convex and T0 is another minimax estimator, then R(T +T0 )/2 (θ) < (RT0 + RT )/2 = RT for all θ and, therefore, T is inadmissible. This shows that T is unique if the loss is strictly convex. Combined with Theorem 4.7, Theorem 4.13 tells us that if an MRIE is admissible, then it is minimax. From the discussion at the end of §4.2.1, MRIE’s in one-parameter location families (such as Pitman’s estimators) are usually minimax.

265

4.3. Minimaxity and Admissibility

4.3.2 Results in one-parameter exponential families The following result provides a sufficient condition for the admissibility of a class of estimators when the population Pθ is in a one-parameter exponential family. Using this result and Theorem 4.13, we can obtain a class of minimax estimators. The proof of this result is an application of the information inequality introduced in §3.1.3. Theorem 4.14. Suppose that X has the p.d.f. c(θ)eθT (x) w.r.t. a σ-finite measure ν, where T (x) is real-valued and θ ∈ (θ− , θ+ ) ⊂ R. Consider the estimation of ϑ = E[T (X)] under the squared error loss. Let λ ≥ 0 and γ be known constants and let Tλ,γ (X) = (T + γλ)/(1 + λ). Then a sufficient condition for the admissibility of Tλ,γ is that Z

θ+

θ0

e−γλθ dθ = [c(θ)]λ

Z

θ0

θ−

e−γλθ dθ = ∞, [c(θ)]λ

(4.34)

where θ0 ∈ (θ− , θ+ ). Proof. From Theorem 2.1, ϑ = E[T (X)] = −c′ (θ)/c(θ) and dϑ dθ = Var(T ) = I(θ), the Fisher information defined in (3.5). Suppose that there is an estimator δ of ϑ such that for all θ, Rδ (θ) ≤ RTλ,γ (θ) = [I(θ) + λ2 (ϑ − γ)2 ]/(1 + λ)2 . Let bδ (θ) be the bias of δ. From the information inequality (3.6), Rδ (θ) ≥ [bδ (θ)]2 + [I(θ) + b′δ (θ)]2 /I(θ). Let h(θ) = bδ (θ) − λ(γ − ϑ)/(1 + λ). Then [h(θ)]2 −

2λh(θ)(ϑ − γ) + 2h′ (θ) [h′ (θ)]2 + ≤ 0, 1+λ I(θ)

which implies [h(θ)]2 −

2λh(θ)(ϑ − γ) + 2h′ (θ) ≤ 0. 1+λ

(4.35)

Let a(θ) = h(θ)[c(θ)]λ eγλθ . Differentiation of a(θ) reduces (4.35) to [a(θ)]2 e−γλθ 2a′ (θ) ≤ 0. + [c(θ)]λ 1+λ

(4.36)

Suppose that a(θ0 ) < 0 for some θ0 ∈ (θ− , θ+ ). From (4.36), a′ (θ) ≤ 0 for all θ. Hence a(θ) < 0 for all θ ≥ θ0 and, for θ > θ0 , (4.36) can be written as d 1 (1 + λ)e−γλθ ≥ . dθ a(θ) 2[c(θ)]λ

266

4. Estimation in Parametric Models

Integrating both sides from θ0 to θ gives 1+λ 2

Z

θ

θ0

e−γλθ 1 1 1 − ≤− . dθ ≤ λ [c(θ)] a(θ) a(θ0 ) a(θ0 )

Letting θ → θ+ , the left-hand side of the previous expression diverges to ∞ by condition (4.34), which is impossible. This shows that a(θ) ≥ 0 for all θ. Similarly, we can show that a(θ) ≤ 0 for all θ. Thus, a(θ) = 0 for all θ. This means that h(θ) = 0 for all θ and b′δ (θ) = −λϑ′ /(1 + λ) = −λI(θ)/(1 + λ), which implies Rδ (θ) ≡ RTλ,γ (θ). This proves the admissibility of Tλ,γ . The reason why Tλ,γ is considered is that it is often a Bayes estimator w.r.t. some prior; see, for example, Examples 2.25, 4.1, 4.7, and 4.8. To find minimax estimators, we may use the following result. Corollary 4.3. Assume that X has the p.d.f. as described in Theorem 4.14 with θ− = −∞ and θ+ = ∞. (i) As an estimator of ϑ = E(T ), T (X) is admissible under the squared error loss and the loss (a − ϑ)2 /Var(T ). (ii) T is the unique minimax estimator of ϑ under the loss (a − ϑ)2 /Var(T ). Proof. (i) With λ = 0, condition (4.34) is clearly satisfied. Hence, Theorem 4.14 applies under the squared error loss. The admissibility of T under the loss (a − ϑ)2 /Var(T ) follows from the fact that T is admissible under the squared error loss and Var(T ) 6= 0. (ii) This is a consequence of part (i) and Theorem 4.13. Example 4.20. P Let X1 , ..., Xn be i.i.d. from N (0, σ 2 ) with an unknown 2 σ > 0. Let Y = ni=1 Xi2 . From Example 4.14, Y /(n+2) is the MRIE of σ 2 and has constant risk under the loss (a − σ 2 )2 /σ 4 . We now apply Theorem 4.14 to show that Y /(n + 2) is admissible. Note that the joint p.d.f. of Xi ’s is of the form c(θ)eθT (x) with θ = −n/(4σ 2 ), c(θ) = (−2θ/n)n/2 , T (X) = 2Y /n, θ− = −∞, and θ+ = 0. By Theorem 4.14, Tλ,γ = (T + γλ)/(1 + λ) is admissible under the squared error loss if Z

−c

−∞

e

−γλθ

−2θ n

−nλ/2

dθ =

Z

0

c

eγλθ θ−nλ/2 dθ = ∞

for some c > 0. This means that Tλ,γ is admissible if γ = 0 and λ = 2/n, or if γ > 0 and λ ≥ 2/n. In particular, 2Y /(n + 2) is admissible for estimating E(T ) = 2E(Y )/n = 2σ 2 , under the squared error loss. It is easy to see that Y /(n + 2) is then an admissible estimator of σ 2 under the squared error loss and the loss (a − σ 2 )2 /σ 4 . Hence Y /(n + 2) is minimax under the loss (a − σ 2 )2 /σ 4 . Note that we cannot apply Corollary 4.3 directly since θ+ = 0.

267

4.3. Minimaxity and Admissibility

Example 4.21. Let X1 , ..., Xn be i.i.d. from the Poisson distribution P (θ) with an unknown θ > 0. The joint p.d.f. of Xi ’s w.r.t. the counting measure is (x1 ! · · · xn !)−1 e−nθ en¯x log θ . For η = n log θ, the conditions of Corollary ¯ Since E(T ) = θ and Var(T ) = θ/n, 4.3 are satisfied with T (X) = X. ¯ by Corollary 4.3, X is the unique minimax estimator of θ under the loss function (a − θ)2 /θ.

4.3.3 Simultaneous estimation and shrinkage estimators In this chapter (and most of Chapter 3) we have focused on the estimation of a real-valued ϑ. The problem of estimating a vector-valued ϑ under the decision theory approach is called simultaneous estimation. Many results for the case of a real-valued ϑ can be extended to simultaneous estimation in a straightforward manner. ˜ A Let ϑ be a p-vector of parameters (functions of θ) with range Θ. vector-valued estimator T (X) can be viewed as a decision rule taking values ˜ Let L(θ, a) be a given nonnegative loss function in the action space A = Θ. on Θ × A. A natural generalization of the squared error loss is L(θ, a) = ka − ϑk2 =

p X i=1

(ai − ϑi )2 ,

(4.37)

where ai and ϑi are the ith components of a and ϑ, respectively. A vector-valued estimator T is called unbiased if and only if E(T ) = ϑ for all θ ∈ Θ. If there is an unbiased estimator of ϑ, then ϑ is called estimable. It can be seen that the result in Theorem 3.1 extends to the case of vector-valued ϑ with any L strictly convex in a. If the loss function is given by (4.37) and Ti is a UMVUE of ϑi for each i, then T = (T1 , ..., Tp ) is a UMVUE of ϑ. If there is a sufficient and complete statistic U (X) for θ, then by Theorem 2.5 (Rao-Blackwell theorem), T must be a function of U (X) and is the unique best unbiased estimator of ϑ. Example 4.22. Consider the general linear model (3.25) with assumption A1 and a full rank Z. Let ϑ = β. An unbiased estimator of β is then the ˆ From the proof of Theorem 3.7, βˆ is a function of the sufficient and LSE β. complete statistic for θ = (β, σ 2 ). Hence, βˆ is the unique best unbiased estimator of ϑ under any strictly convex loss function. In particular, βˆ is the UMVUE of β under the loss function (4.37). Next, we consider Bayes estimators of ϑ, which is still defined to be Bayes actions considered as functions of X. Under the loss function (4.37), the Bayes estimator is still given by (4.4) with vector-valued g(θ) = ϑ.

268

4. Estimation in Parametric Models

Example 4.23. Let X = (X0 , X1 , ..., Xk ) have the multinomial distribution given in Example 2.7. Consider the estimation of the vector θ = (p0 , p1 , ..., pk ) under the loss function (4.37), and the Dirichlet prior for θ that has the Lebesgue p.d.f. Γ(α0 + · · · + αk ) α0 −1 k −1 p · · · pα IA (θ), k Γ(α0 ) · · · Γ(αk ) 0

(4.38)

Pk where αj ’s are known positive constants and A = {θ : 0 ≤ pj , j=0 pj = 1}. It turns out that the Dirichlet prior is conjugate so that the posterior of θ given X = x is also a Dirichlet distribution having the p.d.f. given by (4.38) with αj replaced by αj + xj , j = 0, 1, ..., k. Thus, the Bayes estimator of θ is δ = (δ0 , δ1 , ..., δk ) with δj (X) =

αj + Xj , α0 + α1 + · · · + αk + n

j = 0, 1, ..., k.

After a suitable class of transformations is defined, the results in §4.2 for invariant estimators and MRIE’s are still valid. This is illustrated by the following example. Example 4.24. Let X be a sample with the Lebesgue p.d.f. f (x − θ), where f is a known Lebesgue p.d.f. on Rp with a finite second moment and θ ∈ Rp is an unknown parameter. Consider the estimation of θ under the loss function (4.37). This problem is invariant under the location transformations g(X) = X + c, where c ∈ Rp . Invariant estimators of θ are of the form X + l, l ∈ Rp . It is easy to show that any invariant estimator has constant bias and risk (a generalization of Proposition 4.4) and the MRIE of θ is the unbiased invariant estimator. In particular, if f is the p.d.f. of Np (0, Ip ), then the MRIE is X. The definition of minimax estimators applies without changes. Example 4.25. Let X be a sample from Np (θ, Ip ) with an unknown θ ∈ Rp . Consider the estimation of θ under the loss function (4.37). A modification of the proof of Theorem 4.12 with independent priors for θi ’s shows that X is a minimax estimator of θ (exercise). Example 4.26. Consider Example 4.23. If we choose α0 = · · · = αk = √ n/(k + 1), then the Bayes estimator of θ in Example 4.23 has constant risk. Using the same argument in the proof of Theorem 4.11, we can show that this Bayes estimator is minimax. The previous results for simultaneous estimation are fairly straightforward generalizations of those for the case of a real-valued ϑ. Results for

4.3. Minimaxity and Admissibility

269

admissibility in simultaneous estimation, however, are quite different. A surprising result, due to Stein (1956), is that in estimating the vector mean θ = EX of a normally distributed p-vector X (Example 4.25), X is inadmissible under the loss function (4.37) when p ≥ 3, although X is the UMVUE, MRIE (Example 4.24), and minimax estimator (Example 4.25). Since any estimator better than a minimax estimator is also minimax, there exist many (in fact, infinitely many) minimax estimators in Example 4.25 when p ≥ 3, which is different from the case of p = 1 in which X is the unique admissible minimax estimator (Example 4.6 and Theorem 4.13). We start with the simple case where X is from Np (θ, Ip ) with an unknown θ ∈ Rp . James and Stein (1961) proposed the following class of estimators of ϑ = θ having smaller risks than X when the loss is given by (4.37) and p ≥ 3: p−2 δc = X − (X − c), (4.39) kX − ck2 where c ∈ Rp is fixed. The choice of c is discussed next and at the end of this section. Before we prove that δc in (4.39) is better than X, we try to motivate δc from two viewpoints. First, suppose that it were thought a priori likely, though not certain, that θ = c. Then we might first test a hypothesis H0 : θ = c and estimate θ by c if H0 is accepted and by X otherwise. The best rejection region has the form kX − ck2 > t for some constant t > 0 (see Chapter 6) so that we might estimate θ by I(t,∞) (kX − ck2 )X + [1 − I(t,∞) (kX − ck2 )]c. It can be seen that δc in (4.39) is a smoothed version of this estimator, since δc = ψ(kX − ck2 )X + [1 − ψ(kX − ck2 )]c (4.40) for some function ψ. Any estimator having the form of the right-hand side of (4.40) shrinks the observations toward a given point c and, therefore, is called a shrinkage estimator. Next, δc in (4.40) can be viewed as an empirical Bayes estimator (§4.1.2). In view of (2.25) in Example 2.25, a Bayes estimator of θ is of the form δ = (1 − B)X + Bc, where c is the prior mean of θ and B involves prior variances. If 1 − B is “estimated” by ψ(kX − ck2 ), then δc is an empirical Bayes estimator. Theorem 4.15. Suppose that X is from Np (θ, Ip ) with p ≥ 3. Then, under the loss function (4.37), the risks of the following estimators of θ, δc,r = X −

r(p − 2) (X − c), kX − ck2

(4.41)

270

4. Estimation in Parametric Models

are given by Rδc,r (θ) = p − (2r − r2 )(p − 2)2 E(kX − ck−2 ),

(4.42)

where c ∈ Rp and r ∈ R are known. Proof. Let Z = X − c. Then

2

r(p − 2)

. Rδc,r (θ) = Ekδc,r − E(X)k2 = E Z − E(Z) 1 −

2 kZk

Hence, we only need to show the case of c = 0. Let h(θ) = Rδ0,r (θ), g(θ) be 2 the right-hand side of (4.42) with c = 0, and πα (θ) = (2πα)−p/2 e−kθk /(2α) , which is the p.d.f. of Np (0, αIp ). Note that the distribution of X can be viewed as the conditional distribution of X given θ = θ, where θ has the Lebesgue p.d.f. πα (θ). Then Z g(θ)πα (θ)dθ = p − (2r − r2 )(p − 2)2 E[E(kXk−2 |θ)] Rp

= p − (2r − r2 )(p − 2)2 E(kXk−2 ) = p − (2r − r2 )(p − 2)/(α + 1),

where the expectation in the second line of the previous expression is w.r.t. the joint distribution of (X, θ) and the last equality follows from the fact that the marginal distribution of X is Np (0, (α+1)Ip ), kXk2/(α+1) has the chi-square distribution χ2p and, therefore, E(kXk−2) = 1/[(p − 2)(α + 1)]. ˆ = r(p − 2)/kXk2. Then Let B = 1/(α + 1) and B Z ˆ h(θ)πα (θ)dθ = Ek(1 − B)X − θk2 Rp

ˆ = E{E[k(1 − B)X − θk2 |X]}

= E{E[kθ − E(θ|X)k2 |X] 2 ˆ } + kE(θ|X) − (1 − B)Xk ˆ − B)2 kXk2 } = E{p(1 − B) + (B

= E{p(1 − B) + B 2 kXk2 − 2Br(p − 2) + r2 (p − 2)2 kXk−2 }

= p − (2r − r2 )(p − 2)B,

where the fourth equality follows from the fact that the conditional distribution of θ given X is Np (1 − B)X, (1 − B)Ip and the last equality follows from EkXk−2 = B/(p − 2) and EkXk2 = p/B. This proves Z Z g(θ)πα (θ)dθ = h(θ)πα (θ)dθ, α > 0. (4.43) Rp

Rp

4.3. Minimaxity and Admissibility

271

Note that h(θ) and g(θ) are expectations of functions of kXk2 , θτ X, and kθk2 . Make an orthogonal transformation from X to Y such that Y1 = θτ X/kθk, EYj = 0 for j > 1, and P Var(Y ) = Ip . Then h(θ) and g(θ) are expectations of functions of Y1 , pj=2 Yj2 , and kθk2 . Thus, both h and g are functions of kθk2 . For the family of p.d.f.’s {πα (θ) : α > 0}, kθk2 is a complete and sufficient “statistic”. Hence, (4.43) and the fact that h and g are functions of kθk2 imply that h(θ) = g(θ) a.e. w.r.t. the Lebesgue measure. From Theorem 2.1, both h and g are continuous functions of kθk2 and, therefore, h(θ) = g(θ) for all θ ∈ Rp . This completes the proof. It follows from Theorem 4.15 that the risk of δc,r is smaller than that of X (for every value of θ) when p ≥ 3 and 0 < r < 2, since the risk of X is p under the loss function (4.37). From Example 4.6, X is admissible when p = 1. When p = 2, X is still admissible (Stein, 1956). But we have just shown that X is inadmissible when p ≥ 3. The James-Stein estimator δc in (4.39), which is a special case of (4.41) with r = 1, is better than any δc,r in (4.41) with r 6= 1, since the factor 2r − r2 takes on its maximum value 1 if and only if r = 1. To see that δc may have a substantial improvement over X in terms of risks, consider the special case where θ = c. Since kX − ck2 has the chi-square distribution χ2p when θ = c, EkX −ck−2 = (p−2)−1 and the right-hand side of (4.42) equals 2. Thus, the ratio RX (θ)/Rδc (θ) equals p/2 when θ = c and, therefore, can be substantially larger than 1 near θ = c when p is large. Since X is minimax (Example 4.25), any shrinkage estimator of the form (4.41) is minimax provided that p ≥ 3 and 0 < r < 2. Unfortunately, the James-Stein estimator with any c is also inadmissible. It is dominated by p−2 δc+ = X − min 1, (X − c); (4.44) kX − ck2 see, for example, Lehmann (1983, Theorem 4.6.2). This estimator, however, is still inadmissible. An example of an admissible estimator of the form (4.40) is provided by Strawderman (1971); see also Lehmann (1983, p. 304). Although neither the James-Stein estimator δc nor δc+ in (4.44) is admissible, it is found that no substantial improvements over δc+ are possible (Efron and Morris, 1973). To extend Theorem 4.15 to general Var(X), we consider the case where Var(X) = σ 2 D with an unknown σ 2 > 0 and a known positive definite matrix D. If σ 2 is known, then an extended James-Stein estimator is δ˜c,r = X −

r(p − 2)σ 2 D−1 (X − c). kD−1 (X − c)k2

(4.45)

272

4. Estimation in Parametric Models

One can show (exercise) that under the loss (4.37), the risk of δ˜c,r is σ 2 tr(D) − (2r − r2 )(p − 2)2 σ 2 E(kD−1 (X − c)k−2 ) . (4.46)

When σ 2 is unknown, we assume that there exists a statistic S02 such that S02 is independent of X and S02 /σ 2 has the chi-square distribution χ2m (see Example 4.27). Replacing rσ 2 in (4.45) by σ ˆ 2 = tS02 with a constant t > 0 leads to the following extended James-Stein estimator: δ˜c = X −

(p − 2)ˆ σ2 D−1 (X − c). kD−1 (X − c)k2

(4.47)

By (4.46) and the independence of σ ˆ 2 and X, the risk of δ˜c (as an estimator of ϑ = EX) is h i Rδ˜c (θ) = E E(kδ˜c − ϑk2 |ˆ σ2 ) h i = E E(kδ˜c,(ˆσ2 /σ2 ) − ϑk2 |ˆ σ2 ) = σ 2 E tr(D) − [2(ˆ σ 2 /σ 2 ) − (ˆ σ 2 /σ 2 )2 ](p − 2)2 σ 2 κ(θ) σ 2 /σ 2 )2 ](p − 2)2 σ 2 κ(θ) σ 2 /σ 2 ) − E(ˆ = σ 2 tr(D) − [2E(ˆ = σ 2 tr(D) − [2tm − t2 m(m + 2)](p − 2)2 σ 2 κ(θ) ,

where θ = (ϑ, σ 2 ) and κ(θ) = E(kD−1 (X − c)k−2 ). Since 2tm − t2 m(m + 2) is maximized at t = 1/(m + 2), replacing t by 1/(m + 2) leads to Rδ˜c (θ) = σ 2 tr(D) − m(m + 2)−1 (p − 2)2 σ 2 E(kD−1 (X − c)k−2 ) .

Hence, the risk of the extended James-Stein estimator in (4.47) is smaller than that of X for any fixed θ, when p ≥ 3. Example 4.27. Consider the general linear model (3.25) with assumption A1, p ≥ 3, and a full rank Z, and the estimation of ϑ = β under the loss function (4.37). From Theorem 3.8, the LSE βˆ is from N (β, σ 2 D) with a ˆ and S 2 /σ 2 known matrix D = (Z τ Z)−1 ; S02 = SSR is independent of β; 0 2 has the chi-square distribution χn−p . Hence, from the previous discussion, the risk of the shrinkage estimator βˆ −

(p − 2)ˆ σ2 Z τ Z(βˆ − c) kZ τ Z(βˆ − c)k2

is smaller than that of βˆ for any β and σ 2 , where c ∈ Rp is fixed and σ ˆ 2 = SSR/(n − p + 2). From the previous discussion, the James-Stein estimators improve X substantially when we shrink the observations toward a vector c that is near

273

4.4. The Method of Maximum Likelihood

ϑ = EX. Of course, this cannot be done since ϑ is unknown. One may consider shrinking the observations toward the mean of the observations rather than a given point; that is, one may obtain a shrinkage estimator by ¯ p , where X ¯ = p−1 Pp Xi and Jp is replacing c in (4.39) or (4.47) by XJ i=1 the p-vector of ones. However, we have to replace the factor p − 2 in (4.39) or (4.47) by p − 3. This leads to shrinkage estimators p−3 ¯ ¯ p k2 (X − XJp ) kX − XJ

(4.48)

(p − 3)ˆ σ2 −1 ¯ ¯ p )k2 D (X − XJp ). kD−1 (X − XJ

(4.49)

X− and X−

These estimators are better than X (and, hence, are minimax) when p ≥ 4, under the loss function (4.37) (exercise). The results discussed in this section for the simultaneous estimation of a vector of normal means can be extended to a wide variety of cases where the loss functions are not given by (4.37) (Brown, 1966). The results have also been extended to exponential families and to general location parameter families. For example, Berger (1976) studied the inadmissibility of generalized Bayes estimators of a location vector; Berger (1980) considered simultaneous estimation of gamma scale parameters; and Tsui (1981) investigated simultaneous estimation of several Poisson parameters. See Lehmann (1983, pp. 320-330) for some further references.

4.4 The Method of Maximum Likelihood So far we have studied estimation methods in parametric families using the decision theory approach. The maximum likelihood method introduced next is the most popular method for deriving estimators in statistical inference that does not use any loss function.

4.4.1 The likelihood function and MLE’s To introduce the idea, let us consider an example. Example 4.28. Let X be a single observation taking values from {0, 1, 2} according to Pθ , where θ = θ0 or θ1 and the values of Pθj ({i}) are given by the following table: θ = θ0 θ = θ1

x=0 0.8 0.2

x=1 0.1 0.3

x=2 0.1 0.5

274

4. Estimation in Parametric Models

If X = 0 is observed, it is more plausible that it came from Pθ0 , since Pθ0 ({0}) is much larger than Pθ1 ({0}). We then estimate θ by θ0 . On the other hand, if X = 1 or 2, it is more plausible that it came from Pθ1 , although in this case the difference between the probabilities is not as large as that in the case of X = 0. This suggests the following estimator of θ: θ0 X=0 T (X) = X 6= 0. θ1 The idea in Example 4.28 can be easily extended to the case where Pθ is a discrete distribution and θ ∈ Θ ⊂ Rk . If X = x is observed, θ1 is more plausible than θ2 if and only if Pθ1 ({x}) > Pθ2 ({x}). We then estimate θ by a θˆ that maximizes Pθ ({x}) over θ ∈ Θ, if such a θˆ exists. The word plausible rather than probable is used because θ is considered to be nonrandom and Pθ is not a distribution of θ. Under the Bayesian approach with a prior that is the discrete uniform distribution on {θ1 , ..., θm }, Pθ ({x}) is proportional to the posterior probability and we can say that θ1 is more probable than θ2 if Pθ1 ({x}) > Pθ2 ({x}). Note that Pθ ({x}) in the previous discussion is the p.d.f. w.r.t. the counting measure. Hence, it is natural to extend the idea to the case of continuous (or arbitrary) X by using the p.d.f. of X w.r.t. some σ-finite measure on the range X of X. This leads to the following definition. Definition 4.3. Let X ∈ X be a sample with a p.d.f. fθ w.r.t. a σ-finite measure ν, where θ ∈ Θ ⊂ Rk . (i) For each x ∈ X, fθ (x) considered as a function of θ is called the likelihood function and denoted by ℓ(θ). ˆ = max ¯ ℓ(θ) is ¯ be the closure of Θ. A θˆ ∈ Θ ¯ satisfying ℓ(θ) (ii) Let Θ θ∈Θ ˆ called a maximum likelihood estimate (MLE) of θ. If θ is a Borel function of X a.e. ν, then θˆ is called a maximum likelihood estimator (MLE) of θ. (iii) Let g be a Borel function from Θ to Rp , p ≤ k. If θˆ is an MLE of θ, ˆ is defined to be an MLE of ϑ = g(θ). then ϑˆ = g(θ) ¯ instead of Θ is used in the definition of an MLE. This is Note that Θ because a maximum of ℓ(θ) may not exist when Θ is an open set (Examples 4.29 and 4.30). As an estimator, an MLE is defined a.e. ν. Part (iii) of Definition 4.3 is motivated by a fact given in Exercise 95 of §4.6. ¯ = Θ If the parameter space Θ contains finitely many points, then Θ and an MLE can always be obtained by comparing finitely many values ℓ(θ), θ ∈ Θ. If ℓ(θ) is differentiable on Θ◦ , the interior of Θ, then possible candidates for MLE’s are the values of θ ∈ Θ◦ satisfying ∂ℓ(θ) = 0, ∂θ

(4.50)

4.4. The Method of Maximum Likelihood

275

which is called the likelihood equation. Note that θ’s satisfying (4.50) may be local or global minima, local or global maxima, or simply stationary points. Also, extrema may occur at the boundary of Θ or when kθk → ∞. Furthermore, if ℓ(θ) is not always differentiable, then extrema may occur at nondifferentiable or discontinuity points of ℓ(θ). Hence, it is important to analyze the entire likelihood function to find its maxima. Since log x is a strictly increasing function and ℓ(θ) can be assumed to be positive without loss of generality, θˆ is an MLE if and only if it maximizes the log-likelihood function log ℓ(θ). It is often more convenient to work with log ℓ(θ) and the following analogue of (4.50) (which is called the log-likelihood equation or likelihood equation for simplicity): ∂ log ℓ(θ) = 0. ∂θ

(4.51)

Example 4.29. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ Θ = (0, 1). When (X1 , ..., Xn ) = (x1 , ..., xn ) is observed, the likelihood function is ℓ(p) =

n Y

i=1

Pn −1

pxi (1 − p)1−xi = pn¯x (1 − p)n(1−¯x) ,

◦ ¯ where x ¯=n i=1 xi . Note that Θ = [0, 1] and Θ = Θ. The likelihood equation (4.51) reduces to

n¯ x n(1 − x¯) − = 0. p 1−p If 0 < x ¯ < 1, then this equation has a unique solution x ¯. The second-order derivative of log ℓ(p) is n¯ x n(1 − x ¯) − 2 − , p (1 − p)2 which is always negative. Also, when p tends to 0 or 1 (the boundary of Θ), ℓ(p) → 0. Thus, x ¯ is the unique MLE of p. When x ¯ = 0, ℓ(p) = (1 − p)n is a strictly decreasing function of p and, therefore, its unique maximum is 0. Similarly, the MLE is 1 when x¯ = 1. Combining these results with the previous result, we conclude that the MLE of p is x ¯. When x¯ = 0 or 1, a maximum of ℓ(p) does not exist on Θ = (0, 1), although supp∈(0,1) ℓ(p) = 1; the MLE takes a value outside of Θ and, hence, is not a reasonable estimator. However, if p ∈ (0, 1), the probability that x ¯ = 0 or 1 tends to 0 quickly as n → ∞.

276

4. Estimation in Parametric Models

Example 4.29 indicates that, for small n, a maximum of ℓ(θ) may not exist on Θ and an MLE may be an unreasonable estimator; however, this is unlikely to occur when n is large. A rigorous result of this sort is given in §4.5.2, where we study asymptotic properties of MLE’s. Example 4.30. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with an unknown θ = (µ, σ 2 ), where n ≥ 2. Consider first the case where Θ = R × (0, ∞). When (X1 , ..., Xn ) = (x1 , ..., xn ) is observed, the log-likelihood function is log ℓ(θ) = −

n 1 X n n (xi − µ)2 − log σ 2 − log(2π). 2σ 2 i=1 2 2

The likelihood equation (4.51) becomes n 1 X (xi − µ) = 0 σ 2 i=1

and

n 1 X n (xi − µ)2 − 2 = 0. 4 σ i=1 σ

(4.52)

Solving ¯= Pn the first equation in (4.52) for µ, we obtain a unique solution x n−1 i=1 xi , and substituting x ¯ for µ in the second equation in (4.52), Pn we obtain a unique solution σ ˆ 2 = n−1 i=1 (xi − x ¯)2 . To show that θˆ = 2 (¯ x, σ ˆ ) is an MLE, first note that Θ is an open set and ℓ(θ) is differentiable everywhere; as θ tends to the boundary of Θ or kθk → ∞, ℓ(θ) tends to 0; and ! n 1 Pn ∂ 2 log ℓ(θ) i=1 (xi − µ) σ2 σ4 =− Pn 1 Pn ∂θ∂θτ (xi − µ) 16 (xi − µ)2 − n4 4 σ

i=1

σ

i=1

2σ

is negative definite when µ = x ¯ and σ = σ ˆ . Hence θˆ is the unique MLE. Sometimes we can avoid the calculation of the second-order derivatives. For instance, in this example we know that ℓ(θ) is bounded and ℓ(θ) → 0 as kθk → ∞ or θ tends to the boundary of Θ; hence the unique solution to (4.52) must be the MLE. Another way to show that θˆ is the MLE is indicated by the following discussion. Consider next the case where Θ = (0, ∞) × (0, ∞), i.e., µ is known to be positive. The likelihood function is differentiable on Θ◦ = Θ and ¯ = [0, ∞) × [0, ∞). If x Θ ¯ > 0, then the same argument for the previous ¯ ≤ 0, then the first case can be used to show that (¯ x, σ ˆ 2 ) is the MLE. If x equation in (4.52) does not have a solution in Θ. However, the function log ℓ(θ) = log ℓ(µ, σ 2 ) is strictly decreasing in µ for any fixed σ 2 . Hence, a maximum of log ℓ(µ, σ 2 ) is µ = 0, which does not depend on σ 2 . Then, the MLE is (0, σ ˜ 2 ), where σ ˜ 2 is the value maximizing log ℓ(0, σ 2 ) overPσ 2 ≥ 0. n Applying (4.51) to the function log ℓ(0, σ 2 ) leads to σ ˜ 2 = n−1 i=1 x2i . Thus, the MLE is x ¯>0 (¯ x, σ ˆ2 ) θˆ = x ¯ ≤ 0. (0, σ ˜2 ) 2

2

277

4.4. The Method of Maximum Likelihood

Again, the MLE in this case is not in Θ if x¯ ≤ 0. One can show that a maximum of ℓ(θ) does not exist on Θ when x ¯ ≤ 0. Example 4.31. Let X1 , ..., Xn be i.i.d. from the uniform distribution on an interval Iθ with an unknown θ. First, consider the case where Iθ = (0, θ) and θ > 0. The likelihood function is ℓ(θ) = θ−n I(x(n) ,∞) (θ), which is not always differentiable. In this case Θ◦ = (0, x(n) ) ∪ (x(n) , ∞). But, on (0, x(n) ), ℓ ≡ 0 and on (x(n) , ∞), ℓ′ (θ) = −nθn−1 < 0 for all θ. Hence, the method of using the likelihood equation is not applicable to this problem. Since ℓ(θ) is strictly decreasing on (x(n) , ∞) and is 0 on (0, x(n) ), a unique maximum of ℓ(θ) is x(n) , which is a discontinuity point of ℓ(θ). This shows that the MLE of θ is the largest order statistic X(n) . Next, consider the case where Iθ = (θ − 12 , θ + 12 ) with θ ∈ R. The likelihood function is ℓ(θ) = I(x(n) − 21 ,x(1) + 12 ) (θ). Again, the method of using the likelihood equation is not applicable. However, it follows from Definition 4.3 that any statistic T (X) satisfying x(n) − 21 ≤ T (x) ≤ x(1) + 12 is an MLE of θ. This example indicates that MLE’s may not be unique and can be unreasonable. Example 4.32. Let X be an observation from the hypergeometric distribution HG(r, n, θ − n) (Table 1.1, page 18) with known r, n, and an unknown θ = n + 1, n + 2, .... In this case, the likelihood function is defined on integers and the method of using the likelihood equation is certainly not applicable. Note that ℓ(θ) (θ − r)(θ − n) = , ℓ(θ − 1) θ(θ − n − r + x)

which is larger than 1 if and only if θ < rn/x and is smaller than 1 if and only if θ > rn/x. Thus, ℓ(θ) has a maximum θ = the integer part of rn/x, which is the MLE of θ. Example 4.33. Let X1 , ..., Xn be i.i.d. from the gamma distribution Γ(α, γ) with unknown α > 0 and γ > 0. The log-likelihood function is log ℓ(θ) = −nα log γ − n log Γ(α) + (α − 1) and the likelihood equation (4.51) becomes n

−n log γ − and −

n X i=1

log xi −

nΓ′ (α) X + log xi = 0 Γ(α) i=1

n 1 X nα + 2 xi = 0. γ γ i=1

n 1X xi γ i=1

278

4. Estimation in Parametric Models

The second equation yields γ = x ¯/α. Substituting γ = x ¯/α into the first equation we obtain that n

log α −

1X Γ′ (α) + log xi − log x ¯ = 0. Γ(α) n i=1

In this case, the likelihood equation does not have an explicit solution, although it can be shown (exercise) that a solution exists almost surely and it is the unique MLE. A numerical method has to be applied to compute the MLE for any given observations x1 , ..., xn . These examples indicate that we need to use various methods to derive MLE’s. In applications, MLE’s typically do not have analytic forms and some numerical methods have to be used to compute MLE’s. A commonly used numerical method is the Newton-Raphson iteration method, which repeatedly computes −1 2 ∂ log ℓ(θ) ∂ log ℓ(θ) (t+1) (t) ˆ ˆ θ =θ − ˆ(t) , ∂θ∂θτ θ=θˆ(t) ∂θ θ=θ

(4.53)

t = 0, 1, ..., where θˆ(0) is an initial value and ∂ 2 log ℓ(θ)/∂θ∂θτ is assumed of full rank for every θ ∈ Θ. If, at each iteration, we replace ∂ 2 log ℓ(θ)/∂θ∂θτ in (4.53) by its expected value E[∂ 2 log ℓ(θ)/∂θ∂θτ ], where the expectation is taken under Pθ , then the method is known as the Fisher-scoring method. If the iteration converges, then θˆ(∞) or θˆ(t) with a sufficiently large t is a numerical approximation to a solution of the likelihood equation (4.51). The following example shows that the MCMC methods discussed in §4.1.4 can also be useful in computing MLE’s. Example 4.34. Let X be a random k-vector from Pθ with the following p.d.f. w.r.t. a σ-finite measure ν: Z fθ (x) = fθ (x, y)dν(y), where fθ (x, y) is a joint p.d.f. w.r.t. ν × ν. This type of distribution is called a mixture distribution. Thus, the likelihood ℓ(θ) = fθ (x) involves a k-dimensional integral. In many cases this integral has to be computed in order to compute an MLE of θ. Let ℓ˜m (θ) be the MCMC approximation to ℓ(θ) based on one of the MCMC methods described in §4.1.4 and a Markov chain of length m. Under the conditions of Theorem 4.4, ℓ˜m (θ) →a.s. ℓ(θ) for every fixed θ and x. Suppose that, for each m, there exists θ˜m that maximizes ℓ˜m (θ) over θ ∈ Θ. Geyer (1994) studies the convergence of θ˜m to an MLE.

4.4. The Method of Maximum Likelihood

279

In terms of their mse’s, MLE’s are not necessarily better than UMVUE’s or Bayes estimators. Also, MLE’s are frequently inadmissible. This is not surprising, since MLE’s are not derived under any given loss function. The main theoretical justification for MLE’s is provided in the theory of asymptotic efficiency considered in §4.5.

4.4.2 MLE’s in generalized linear models Suppose that X has a distribution from a natural exponential family so that the likelihood function is ℓ(η) = exp{η τ T (x) − ζ(η)}h(x), where η ∈ Ξ is a vector of unknown parameters. The likelihood equation (4.51) is then ∂ζ(η) ∂ log ℓ(η) = T (x) − = 0, ∂η ∂η which has a unique solution T (x) = ∂ζ(η)/∂η, assuming that T (x) is in the range of ∂ζ(η)/∂η. Note that ∂ 2 log ℓ(η) ∂ 2 ζ(η) = − = −Var(T ) ∂η∂η τ ∂η∂η τ

(4.54)

(see the proof of Proposition 3.2). Since Var(T ) is positive definite, − log ℓ(η) is convex in η and T (x) is the unique MLE of the parameter µ(η) = ∂ζ(η)/∂η. By (4.54) again, the function µ(η) is one-to-one so that µ−1 exists. By Definition 4.3, the MLE of η is ηˆ = µ−1 (T (x)). If the distribution of X is in a general exponential family and the likelihood function is ℓ(θ) = exp{[η(θ)]τ T (x) − ξ(θ)}h(x), η ), if η −1 exists and ηˆ is in the range of η(θ). then the MLE of θ is θˆ = η −1 (ˆ Of course, θˆ is also the solution of the likelihood equation ∂η(θ) ∂ξ(θ) ∂ log ℓ(θ) = T (x) − = 0. ∂θ ∂θ ∂θ The results for exponential families lead to an estimation method in a class of models that have very wide applications. These models are generalizations of the normal linear model (model (3.25) with assumption A1) discussed in §3.3.1-§3.3.2 and, therefore, are named generalized linear models (GLM).

280

4. Estimation in Parametric Models

A GLM has the following structure. The sample X = (X1 , ..., Xn ) ∈ Rn has independent components and Xi has the p.d.f. o n i) h(xi , φi ), i = 1, ..., n, (4.55) exp ηi xi −ζ(η φi

w.r.t. a σ-finite measure ν, where ηi and φi are unknown, φi > 0, R ηi ∈ Ξ = η : 0 < h(x, φ)eηx/φ dν(x) < ∞ ⊂ R

for all i, ζ and h are known functions, and ζ ′′ (η) > 0 is assumed for all η ∈ Ξ◦ , the interior of Ξ. Note that the p.d.f. in (4.55) belongs to an exponential family if φi is known. As a consequence, E(Xi ) = ζ ′ (ηi ) and Var(Xi ) = φi ζ ′′ (ηi ),

i = 1, ..., n.

(4.56)

Define µ(η) = ζ ′ (η). It is assumed that ηi is related to Zi , the ith value of a p-vector of covariates (see (3.24)), through g(µ(ηi )) = β τ Zi ,

i = 1, ..., n,

(4.57)

where β is a p-vector of unknown parameters and g, called a link function, is a known one-to-one, third-order continuously differentiable function on {µ(η) : η ∈ Ξ◦ }. If µ = g −1 , then ηi = β τ Zi and g is called the canonical or d natural link function. If g is not canonical, we assume that dη (g ◦ µ)(η) 6= 0 for all η. In a GLM, the parameter of interest is β. We assume that the range of β is B = {β : (g ◦ µ)−1 (β τ z) ∈ Ξ◦ for all z ∈ Z}, where Z is the range of Zi ’s. φi ’s are called dispersion parameters and are considered to be nuisance parameters. It is often assumed that φi = φ/ti ,

i = 1, ..., n,

(4.58)

with an unknown φ > 0 and known positive ti ’s. As we discussed earlier, the linear model (3.24) with εi = N (0, φ) is a special GLM. One can verify this by taking g(µ) ≡ µ and ζ(η) = η 2 /2. The usefulness of the GLM is that it covers situations where the relationship between E(Xi ) and Zi is nonlinear and/or Xi ’s are discrete (in which case the linear model (3.24) is clearly not appropriate). The following is an example. Example 4.35. Let Xi ’s be independent discrete random variables taking values in {0, 1, ..., m}, where m is a known positive integer. First, suppose that Xi has the binomial distribution Bi(pi , m) with an unknown pi ∈ pi (0, 1), i = 1, ..., n. Let ηi = log 1−p and ζ(ηi ) = m log(1 + eηi ). Then the i p.d.f. of Xi (w.r.t. the counting measure) is given by (4.55) with φi = 1,

281

4.4. The Method of Maximum Likelihood

h(xi , φi ) = xmi , and Ξ = R. Under (4.57) and the logit link (canonical t link) g(t) = log m−t , τ

meηi meβ Zi E(Xi ) = mpi = = . 1 + eηi 1 + eβ τ Zi Another popular link in this problem is the probit link g(t) = Φ−1 (t/m), where Φ is the c.d.f. of the standard normal. Under the probit link, E(Xi ) = mΦ(β τ Zi ). The variance of Xi is mpi (1 − pi ) under the binomial distribution assumption. This assumption is often violated in applications, which results in an over-dispersion, i.e., the variance of Xi exceeds the nominal variance mpi (1 − pi ). Over-dispersion can arise in a number of ways, but the most common one is clustering in the population. Families, households, and litters Pm are common instances of clustering. For example, suppose that Xi = j=1 Xij , where Xij are binary random variables having a common distribution. If Xij ’s are independent, then Xi has a binomial distribution. However, if Xij ’s are from the same cluster (family or household), then they are often positively correlated. Suppose that the correlation coefficient (§1.3.2) between Xij and Xil , j 6= l, is ρi > 0. Then Var(Xi ) = mpi (1 − pi ) + m(m − 1)ρi pi (1 − pi ) = φi mpi (1 − pi ), where φi = 1 + (m − 1)ρi is the dispersion parameter. Of course, overdispersion can occur only if m > 1 in this case. This motivates the consideration of GLM (4.55)-(4.57) with dispersion parameters φi . If Xi has the p.d.f. (4.55) with ζ(ηi ) = m log(1 + eηi ), then E(Xi ) =

meηi 1 + eηi

and

Var(Xi ) = φi

meηi , (1 + eηi )2

which is exactly (4.56). Of course, the distribution of Xi is not binomial unless φi = 1. We now derive an MLE of β in a GLM under assumption (4.58). Let θ = (β, φ) and ψ = (g ◦ µ)−1 . Then the log-likelihood function is log ℓ(θ) =

n X i=1

ψ(β τ Zi )xi − ζ(ψ(β τ Zi )) log h(xi , φ/ti ) + φ/ti

and the likelihood equation is n

1X ∂ log ℓ(θ) = {[xi − µ(ψ(β τ Zi ))]ψ ′ (β τ Zi )ti Zi } = 0 ∂β φ i=1

(4.59)

282

4. Estimation in Parametric Models

and n

∂ log ℓ(θ) X = ∂φ i=1

∂ log h(xi , φ/ti ) ti [ψ(β τ Zi )xi − ζ(ψ(β τ Zi ))] − ∂φ φ2

= 0.

From the first equation, an MLE of β, if it exists, can be obtained without estimating φ. The second equation, however, is usually difficult to solve. Some other estimators of φ are suggested by various researchers; see, for example, McCullagh and Nelder (1989). Suppose that there is a solution βˆ ∈ B to equation (4.59). (The existence of βˆ is studied in §4.5.2.) We now study whether βˆ is an MLE of β. Let n X Mn (β) = [ψ ′ (β τ Zi )]2 ζ ′′ (ψ(β τ Zi ))ti Zi Ziτ (4.60) i=1

and

Rn (β) =

n X

[xi − µ(ψ(β τ Zi ))]ψ ′′ (β τ Zi )ti Zi Ziτ .

Var

i=1

Then

and

∂ log ℓ(θ) ∂β

= Mn (β)/φ

∂ 2 log ℓ(θ) = [Rn (β) − Mn (β)]/φ. ∂β∂β τ

(4.61)

(4.62)

(4.63)

Consider first the simple case of canonical g. Then ψ ′′ ≡ 0 and Rn ≡ 0. If Mn (β) is positive definite for all β, then − log ℓ(θ) is strictly convex in β for any fixed φ and, therefore, βˆ is the unique MLE of β. For the case of noncanonical g, Rn (β) 6= 0 and βˆ is not necessarily an MLE. If Rn (β) is dominated by Mn (β) (i.e., [Mn (β)]−1/2 Rn (β)[Mn (β)]−1/2 → 0 in some sense), then − log ℓ(θ) is convex and βˆ is an MLE for large n; see more details in the proof of Theorem 4.18 in §4.5.2. Example 4.36. Consider the GLM (4.55) with ζ(η) = η 2 /2, η ∈ R. If g in (4.57) is the canonical link, then the model is the same as (3.24) with independent εi ’s distributed as N (0, φi ). If (4.58) holds with ti ≡ 1, then (4.59) is exactly the same as equation (3.27). If Z is of full rank, then Mn (β) = Z τ Z is positive definite. Thus, we have shown that the LSE βˆ given by (3.28) is actually the unique MLE of β. Suppose now that g is noncanonical but (4.58) still holds with ti ≡ 1. Then the model reduces to the one with independent Xi ’s and Xi = N g −1 (β τ Zi ), φ , i = 1, ..., n. (4.64)

283

4.4. The Method of Maximum Likelihood

This type of model is called a nonlinear regression model (with normal errors) and an MLE of β under this model is also called a nonlinear LSE, since maximizing the log-likelihood is equivalent to minimizing the sum of P squares ni=1 [Xi −g −1 (β τ Zi )]2 . Under certain conditions the matrix Rn (β) is dominated by Mn (β) and an MLE of β exists. More details can be found in §4.5.2. Example 4.37 (The Poisson model). Consider the GLM (4.55) with ζ(η) = eη , η ∈ R. If φi ≡ 1, then Xi has the Poisson distribution with mean eηi . Assume that (4.58) holds. Under the canonical link g(t) = log t, Mn (β) =

n X

eβ

τ

Zi

ti Zi Ziτ ,

i=1

√ √ τ which is positive definite if inf i eβ Zi > 0 and the matrix ( t1 Z1 , ..., tn Zn ) is of full rank. There is one noncanonical link that deserves attention. Suppose that ′ 2 ′′ we Pn choose aτ link function so that [ψ (t)] ζ (ψ(t)) ≡ 1. Then Mn (β) ≡ i=1 ti Zi Zi does not depend on β. In §4.5.2 it is shown that the asymptotic variance of the MLE βˆ is φ[Mn (β)]−1 . The fact that Mn (β) does not depend on β makes the estimation of the asymptotic variance (and, thus, statistical inference) easy. Under the Poisson model, ζ ′′ (t) = et and, therefore, we need to solve the differential equation [ψ ′ (t)]2 eψ(t) = 1. A solution √ is ψ(t) = 2 log(t/2), which gives the link function g(µ) = 2 µ. In a GLM, an MLE βˆ usually does not have an analytic form. A numerical method such as the Newton-Raphson or the Fisher-scoring method has to be applied. Using the Newton-Raphson method, we have the following iteration procedure: βˆ(t+1) = βˆ(t) − [Rn (βˆ(t) ) − Mn (βˆ(t) )]−1 sn (βˆ(t) ),

t = 0, 1, ...,

where sn (β) = φ∂ log ℓ(θ)/∂β. Note that E[Rn (β)] = 0 if β is the true parameter value and xi is replaced by Xi . This means that the Fisherscoring method uses the following iteration procedure: βˆ(t+1) = βˆ(t) + [Mn (βˆ(t) )]−1 sn (βˆ(t) ),

t = 0, 1, ....

If the canonical link is used, then the two methods are identical.

4.4.3 Quasi-likelihoods and conditional likelihoods We now introduce two variations of the method of using likelihoods.

284

4. Estimation in Parametric Models

Consider a GLM (4.55)-(4.57). Assumption (4.58) is often unrealistic in applications. If there is no restriction on φi ’s, however, there are too many parameters and an MLE of β may not exist. (Note that assumption (4.58) reduces n nuisance parameters to one.) One way to solve this problem is to assume that φi = h ¯ (Zi , ξ) for some known function h ¯ and unknown parameter vector ξ (which may include β as a subvector). Let θ = (β, ξ). Then we can try to solve the likelihood equation ∂ log ℓ(θ)/∂θ = 0 to obtain an MLE of β and/or ξ. We omit the details, which can be found, for example, in Smyth (1989). Suppose that we do not impose any assumptions on φi ’s but still estimate β by solving s˜n (β) =

n X i=1

{[xi − µ(ψ(β τ Zi ))]ψ ′ (β τ Zi )ti Zi } = 0.

(4.65)

Note that (4.65) is not a likelihood equation unless (4.58) holds. In the special case of Example 4.36 where Xi = N (β τ Zi , φi ), i = 1, ..., n, a solution to (4.65) is simply an LSE of β whose properties are discussed at the end of §3.3.3. Estimating β by solving equation (4.65) is motivated by the following facts. First, if (4.58) does hold, then our estimate is an MLE. Second, if (4.58) is slightly violated, the performance of our estimate is still nearly the same as that of an MLE under assumption (4.58) (see the discussion of robustness at the end of §3.3.3). Finally, estimators obtained by solving (4.65) usually have good asymptotic properties. As a special case of a general result in §5.4, a solution to (4.65) is asymptotically normal under some regularity conditions. In general, an equation such as (4.65) is called a quasi-likelihood equation if and only if it is a likelihood equation when certain assumptions hold. The “likelihood” corresponding to a quasi-likelihood equation is called quasilikelihood and a maximum of the quasi-likelihood is then called a maximum quasi-likelihood estimate (MQLE). Thus, a solution to (4.65) is an MQLE. Note that (4.65) is a likelihood equation if and only if both (4.55) and (4.58) hold. The LSE (§3.3) without normality assumption on Xi ’s is a simple example of an MQLE without (4.55). Without assumption (4.55), the model under consideration is usually nonparametric and, therefore, the MQLE’s are studied in §5.4. While the quasi-likelihoods are used to relax some assumptions in our models, the conditional likelihoods discussed next are used mainly in cases where MLE’s are difficult to compute. We consider two cases. In the first case, θ = (θ1 , θ2 ), θ1 is the main parameter vector of interest, and θ2 is a nuisance parameter vector. Suppose that there is a statistic T2 (X) that is sufficient for θ2 for each fixed θ1 . By the sufficiency, the conditional distribution of X given T2 does not depend on θ2 . The likelihood function

285

4.4. The Method of Maximum Likelihood

corresponding to the conditional p.d.f. of X given T2 is called the conditional likelihood function. A conditional MLE of θ1 can then be obtained by maximizing the conditional likelihood function. This method can be applied to the case where the dimension of θ is considerably larger than the dimension of θ1 so that computing the unconditional MLE of θ is much more difficult than computing the conditional MLE of θ1 . Note that the conditional MLE’s are usually different from the unconditional MLE’s. As a more specific example, suppose that X has a p.d.f. in an exponential family: fθ (x) = exp{θ1τ T1 (x) + θ2τ T2 (x) − ζ(θ)}h(x). Then T2 is sufficient for θ2 for any given θ1 . Problems of this type are from comparisons of two binomial distributions or two Poisson distributions (Exercises 119-120). The second case is when our sample X = (X1 , ..., Xn ) follows a firstorder autoregressive time series model: Xt − µ = ρ(Xt−1 − µ) + εt ,

t = 2, ..., n,

where µ ∈ R and ρ ∈ (−1, 1) are unknown and εi ’s are i.i.d. from N (0, σ 2 ) with an unknown σ 2 > 0. This model is often a satisfactory representation of the error time series in economic models, and is one of the simplest and most heavily used models in time series analysis (Fuller, 1996). Let θ = (µ, ρ, σ2 ). The log-likelihood function is n n 1 log(2π) − log σ 2 + log(1 − ρ2 ) 2 ( 2 2 ) n X 1 2 2 2 − 2 (x1 − µ) (1 − ρ ) + [xt − µ − ρ(xt−1 − µ)] . 2σ t=2

log ℓ(θ) = −

The computation of the MLE is greatly simplified if we consider the conditional likelihood given X1 = x1 : log ℓ(θ|x1 ) = −

n n−1 n−1 1 X log(2π)− log σ 2 − 2 [xt −µ−ρ(xt−1 −µ)]2 . 2 2 2σ t=2

Let (¯ x−1 , x ¯0 ) = (n − 1)−1 ρˆ =

n X t=2

Pn

t=2 (xt−1 , xt ).

If

X n (xt − x¯0 )(xt−1 − x ¯−1 ) (xt−1 − x ¯−1 )2 t=2

is between −1 and 1, then it is the conditional MLE of ρ and the conditional MLE’s of µ and σ 2 are, respectively, ¯−1 )/(1 − ρˆ) µ ˆ = (¯ x0 − ρˆx

286

4. Estimation in Parametric Models

and

n

σ ˆ2 =

1 X [xt − x ¯0 − ρˆ(xt−1 − x ¯−1 )]2 . n − 1 t=2

Obviously, the result can be extended to the case where X follows a pth-order autoregressive time series model: Xt − µ = ρ1 (Xt−1 − µ) + · · · + ρp (Xt−p − µ) + εt ,

t = p + 1, ..., n, (4.66)

where ρj ’s are unknown parameters satisfying the constraint that the roots (which may be complex) of the polynomial xp − ρ1 xp−1 − · · · − ρp = 0 are less than one in absolute value (exercise). Some other likelihood based methods are introduced in §5.1.4. Although they can also be applied to parametric models, the methods in §5.1.4 are more useful in nonparametric models.

4.5 Asymptotically Efficient Estimation In this section, we consider asymptotic optimality of point estimators in parametric models. We use the asymptotic mean squared error (amse, see §2.5.2) or its multivariate generalization to assess the performance of an estimator. Reasons for considering asymptotics have been discussed in §2.5. We focus on estimators that are asymptotically normal, since this covers the majority of cases. Some cases of asymptotically nonnormal estimators are studied in Exercises 111-114 in §4.6.

4.5.1 Asymptotic optimality Let {θˆn } be a sequence of estimators of θ based on a sequence of samples {X = (X1 , ..., Xn ) : n = 1, 2, ...} whose distributions are in a parametric family indexed by θ. Suppose that as n → ∞, [Vn (θ)]−1/2 (θˆn − θ) →d Nk (0, Ik ),

(4.67)

where, for each n, Vn (θ) is a k × k positive definite matrix depending on θ. If θ is one-dimensional (k = 1), then Vn (θ) is the asymptotic variance as well as the amse of θˆn (§2.5.2). When k > 1, Vn (θ) is called the asymptotic covariance matrix of θˆn and can be used as a measure of asymptotic performance of estimators. If θˆjn satisfies (4.67) with asymptotic covariance matrix Vjn (θ), j = 1, 2, and V1n (θ) ≤ V2n (θ) (in the sense that V2n (θ) − V1n (θ) is nonnegative definite) for all θ ∈ Θ, then θˆ1n is said to be asymptotically more efficient than θˆ2n . Of course, some sequences of estimators are

4.5. Asymptotically Efficient Estimation

287

not comparable under this criterion. Also, since the asymptotic covariance matrices are unique only in the limiting sense, we have to make our comparison based on their limits. When Xi ’s are i.i.d., Vn (θ) is usually of the form n−δ V (θ) for some δ > 0 (= 1 in the majority of cases) and a positive definite matrix V (θ) that does not depend on n. Note that (4.67) implies that θˆn is an asymptotically unbiased estimator of θ. If Vn (θ) = Var(θˆn ), then, under some regularity conditions, it follows from Theorem 3.3 that Vn (θ) ≥ [In (θ)]−1 , (4.68) where, for every n, In (θ) is the Fisher information matrix (see (3.5)) for X of size n. (Note that (4.68) holds if and only if lτ Vn (θ)l ≥ lτ [In (θ)]−1 l for every l ∈ Rk .) Unfortunately, when Vn (θ) is an asymptotic covariance matrix, (4.68) may not hold (even in the limiting sense), even if the regularity conditions in Theorem 3.3 are satisfied.

Example 4.38 (Hodges). Let X1 , ..., Xn be i.i.d. from N (θ, 1), θ ∈ R. Then In (θ) = n. Define ¯ ¯ ≥ n−1/4 X |X| θˆn = ¯ ¯ tX |X| < n−1/4 , where t is a fixed constant. By Proposition 3.2, all conditions in Theorem 3.3 are satisfied. It can be shown (exercise) that (4.67) holds with Vn (θ) = V (θ)/n, where V (θ) = 1 if θ 6= 0 and V (θ) = t2 if θ = 0. If t2 < 1, (4.68) does not hold when θ = 0. However, the following result, due to Le Cam (1953), shows that (4.68) holds for i.i.d. Xi ’s except for θ in a set of Lebesgue measure 0. Theorem 4.16. Let X1 , ..., Xn be i.i.d. from a p.d.f. fθ w.r.t. a σ-finite measure ν on (R, B), where θ ∈ Θ and Θ is an open set in Rk . Suppose that for every x in the range of X1 , fθ (x) is twice continuously differentiable in θ and satisfies Z Z ∂ ∂ ψθ (x)dν = ψθ (x)dν ∂θ ∂θ for ψθ (x) = fθ (x) and = ∂fθ (x)/∂θ; the Fisher information matrix τ ∂ ∂ I1 (θ) = E log fθ (X1 ) log fθ (X1 ) ∂θ ∂θ is positive definite; and for any given θ ∈ Θ, there exists a positive number cθ and a positive function hθ such that E[hθ (X1 )] < ∞ and

2

∂ log fγ (x)

sup (4.69)

∂γ∂γ τ ≤ hθ (x) γ:kγ−θk [I1 (θ)]1/2 , Z Pθn (Kn (X, θ) ≤ t) = ℓ(θn )dν × · · · × dν Kn (x,θ)≤t

=

Z

Kn (x,θ)≤t

= e−I1 (θ)/2 =e

Z

ℓ(θn ) dPθ (x) ℓ(θ)

e[I1 (θ)]

Kn (x,θ)≤t Z t −I1 (θ)/2 [I1 (θ)]1/2 z

e

−∞ Z t

1/2

1/2

Kn (x,θ)

dPθ (x)

dHn (z)

e[I1 (θ)] z dΦ(z) + o(1) = e−I1 (θ)/2 −∞ = Φ t − [I1 (θ)]1/2 + o(1),

4.5. Asymptotically Efficient Estimation

289

where Hn denotes the distribution of Kn (X, θ) and the next to last equality follows from (4.70) and the dominated convergence theorem. This result and result (4.71) imply that there is a sequence {nj } such that for j = 1, 2, ..., Pθnj (θˆnj ≤ θnj ) < Pθnj (Knj (X, θ) ≤ t). (4.72) By the Neyman-Pearson lemma (Theorem 6.1 in §6.1.1), we conclude that (4.72) implies that for j = 1, 2, ..., Pθ (θˆnj ≤ θnj ) < Pθ (Knj (X, θ) ≤ t).

(4.73)

(The reader should come back to this after reading §6.1.1.) From (4.70) and (4.67) with Vn (θ) = V (θ)/n, (4.73) implies Φ [V (θ)]−1/2 ≤ Φ(t). Hence [V (θ)]−1/2 ≤ t. Since In (θ) = nI1 (θ) (Proposition 3.1(i)) and t is arbitrary but > [I1 (θ)]1/2 , we conclude that (4.68) holds.

Points at which (4.68) does not hold are called points of superefficiency. Motivated by the fact that the set of superefficiency points is of Lebesgue measure 0 under some regularity conditions, we have the following definition. Definition 4.4. Assume that the Fisher information matrix In (θ) is well defined and positive definite for every n. A sequence of estimators {θˆn } satisfying (4.67) is said to be asymptotically efficient or asymptotically optimal if and only if Vn (θ) = [In (θ)]−1 . Suppose that we are interested in estimating ϑ = g(θ), where g is a differentiable function from Θ to Rp , 1 ≤ p ≤ k. If θˆn satisfies (4.67), then, by Theorem 1.12(i), ϑˆn = g(θˆn ) is asymptotically distributed as Np (ϑ, [∇g(θ)]τ Vn (θ)∇g(θ)). Thus, inequality (4.68) becomes [∇g(θ)]τ Vn (θ)∇g(θ) ≥ [I˜n (ϑ)]−1 , where I˜n (ϑ) is the Fisher information matrix about ϑ contained in X. If p = k and g is one-to-one, then [I˜n (ϑ)]−1 = [∇g(θ)]τ [In (θ)]−1 ∇g(θ) and, therefore, ϑˆn is asymptotically efficient if and only if θˆn is asymptotically efficient. For this reason, in the case of p < k, ϑˆn is considered to be asymptotically efficient if and only if θˆn is asymptotically efficient, and we can focus on the estimation of θ only.

290

4. Estimation in Parametric Models

4.5.2 Asymptotic efficiency of MLE’s and RLE’s We now show that under some regularity conditions, a root of the likelihood equation (RLE), which is a candidate for an MLE, is asymptotically efficient. Theorem 4.17. Assume the conditions of Theorem 4.16. (i) There is a sequence of estimators {θˆn } such that and θˆn →p θ, P sn (θˆn ) = 0 → 1

(4.74)

where sn (γ) = ∂ log ℓ(γ)/∂γ. (ii) Any consistent sequence θ˜n of RLE’s is asymptotically efficient. Proof. (i) Let Bn (c) = {γ : k[In (θ)]1/2 (γ − θ)k ≤ c} for c > 0. Since Θ is open, for each c > 0, Bn (c) ⊂ Θ for sufficiently large n. Since Bn (c) shrinks to {θ} as n → ∞, the existence of θˆn satisfying (4.74) is implied by the fact that for any ǫ > 0, there exists c > 0 and n0 > 1 such that P log ℓ(γ) − log ℓ(θ) < 0 for all γ ∈ ∂Bn (c) ≥ 1 − ǫ, n ≥ n0 , (4.75)

where ∂Bn (c) is the boundary of Bn (c). (For a proof of the measurability of θˆn , see Serfling (1980, pp. 147-148).) For γ ∈ ∂Bn (c), the Taylor expansion gives log ℓ(γ) − log ℓ(θ) = cλτ [In (θ)]−1/2 sn (θ)

(4.76)

+ (c2 /2)λτ [In (θ)]−1/2 ∇sn (γ ∗ )[In (θ)]−1/2 λ,

where λ = [In (θ)]1/2 (γ − θ)/c satisfying kλk = 1, ∇sn (γ) = ∂sn (γ)/∂γ, and γ ∗ lies between γ and θ. Note that E

k∇sn (γ) − ∇sn (θ)k k∇sn (γ ∗ ) − ∇sn (θ)k ≤ E max n γ∈Bn (c) n

2

∂ log fγ (X1 ) ∂ 2 log fθ (X1 )

≤ E max −

γ∈Bn (c) ∂γ∂γ τ ∂θ∂θτ → 0, (4.77)

which follows from (a) ∂ 2 log fγ (x)/∂γ∂γ τ is continuous in a neighborhood of θ for any fixed x; (b) Bn (c) shrinks to {θ}; and (c) for sufficiently large n,

2

∂ log fγ (X1 ) ∂ 2 log fθ (X1 )

≤ 2hθ (X1 ) max −

γ∈Bn (c) ∂γ∂γ τ ∂θ∂θτ under condition (4.69). By the SLLN (Theorem 1.13) and Proposition 3.1, n−1 ∇sn (θ) →a.s. −I1 (θ) (i.e., kn−1 ∇sn (θ) + I1 (θ)k →a.s. 0). These results, together with (4.76), imply that log ℓ(γ) − log ℓ(θ) = cλτ [In (θ)]−1/2 sn (θ) − [1 + op (1)]c2 /2.

(4.78)

4.5. Asymptotically Efficient Estimation

291

Note that maxλ {λτ [In (θ)]−1/2 sn (θ)} = k[In (θ)]−1/2 sn (θ)k. Hence, (4.75) follows from (4.78) and P k[In (θ)]−1/2 sn (θ)k < c/4 ≥ 1 − (4/c)2 Ek[In (θ)]−1/2 sn (θ)k2 = 1 − k(4/c)2 ≥ 1−ǫ

by choosing c sufficiently large. This completes the proof of (i). (ii) Let Aǫ = {γ : kγ − θk ≤ ǫ} for ǫ > 0. Since Θ is open, Aǫ ⊂ Θ for sufficiently small ǫ. Let {θ˜n } be a sequence of consistent RLE’s, i.e., P (sn (θ˜n ) = 0 and θ˜n ∈ Aǫ ) → 1 for any ǫ > 0. Hence, we can focus on the set on which sn (θ˜n ) = 0 and θ˜n ∈ Aǫ . Using the mean-value theorem for vector-valued functions, we obtain that Z 1 −sn (θ) = ∇sn θ + t(θ˜n − θ) dt (θ˜n − θ). 0

Note that

Z 1

1 k∇sn (γ) − ∇sn (θ)k ˜

. ∇sn θ + t(θn − θ) dt − ∇sn (θ) ≤ max

γ∈A n 0 n ǫ

Using the argument in proving (4.77) and the fact that P (θ˜n ∈ Aǫ ) → 1 for arbitrary ǫ > 0, we obtain that

Z 1

1 ˜

→p 0. ∇s − θ) dt − ∇s (θ) θ + t( θ n n n

n 0 Since n−1 ∇sn (θ) →a.s. −I1 (θ) and In (θ) = nI1 (θ),

−sn (θ) = −In (θ)(θ˜n − θ) + op kIn (θ)(θ˜n − θ)k .

√ This and Slutsky’s theorem (Theorem 1.11) imply that n(θ˜n − θ) has the same asymptotic distribution as √ n[In (θ)]−1 sn (θ) = n−1/2 [I1 (θ)]−1 sn (θ) →d Nk 0, [I1 (θ)]−1

by the CLT (Corollary 1.2), since Var(sn (θ)) = In (θ).

Theorem 4.17(i) shows the asymptotic existence of a sequence of consistent RLE’s, and Theorem 4.17(ii) shows the asymptotic efficiency of any sequence of consistent RLE’s. However, for a given sequence of RLE’s, its consistency has to be checked unless the RLE’s are unique for sufficiently large n, in which case the consistency of the RLE’s is guaranteed by Theorem 4.17(i).

292

4. Estimation in Parametric Models

RLE’s are not necessarily MLE’s. We still have to use the techniques discussed in §4.4 to check whether an RLE is an MLE. However, according to Theorem 4.17, when a sequence of RLE’s is consistent, then it is asymptotically efficient and, therefore, we may not need to search for MLE’s, if asymptotic efficiency is the only criterion to select estimators. The method of estimating θ by solving sn (γ) = 0 over γ ∈ Θ is called scoring and the function sn (γ) is called the score function. Example 4.39. Suppose that Xi has a distribution in a natural exponential family, i.e., the p.d.f. of Xi is fη (xi ) = exp{η τ T (xi ) − ζ(η)}h(xi ).

(4.79)

Since ∂ 2 log fη (xi )/∂η∂η τ = −∂ 2 ζ(η)/∂η∂η τ , condition (4.69) is satisfied. From Proposition 3.2, other conditions in Theorem 4.16 are also satisfied. For i.i.d. Xi ’s, n X ∂ζ(η) . T (Xi ) − sn (η) = ∂η i=1

Pn If θˆn = n−1 i=1 T (Xi ) ∈ Θ, the range of θ = g(η) = ∂ζ(η)/∂η, then θˆn is a unique RLE of θ, which is also a unique MLE of θ since ∂ 2 ζ(η)/∂η∂η τ = Var(T (Xi )) is positive definite. Also, η = g −1 (θ) exists and a unique RLE (MLE) of η is ηˆn = g −1 (θˆn ). However, θˆn may not be in Θ and the previous argument fails (e.g., Example 4.29). What Theorem 4.17 tells us in this case is that as n → ∞, P (θˆn ∈ Θ) → 1 and, therefore, θˆn (or ηˆn ) is the unique asymptotically efficient RLE (MLE) of θ (or η) in the limiting sense. In an example like this we can directly show that P (θˆn ∈ Θ) → 1, using the fact that θˆn →a.s. E[T (X1 )] = g(η) (the SLLN).

The next theorem provides a similar result for the MLE or RLE in the GLM (§4.4.2). Theorem 4.18. Consider the GLM (4.55)-(4.58) with ti ’s in a fixed interval (t0 , t∞ ), 0 < t0 ≤ t∞ < ∞. Assume that the range of the unknown parameter β in (4.57) is an open subset of Rp ; at the true parameter value β, 0 < inf i ϕ(β τ Zi ) ≤ supi ϕ(β τ Zi ) < ∞, where ϕ(t) = [ψ ′ (t)]2 ζ ′′ (ψ(t)); as n → ∞, maxi≤n Ziτ (Z τ Z)−1 Zi → 0 and λ− [Z τ Z] → ∞, where Z is the n × p matrix whose ith row is the vector Zi and λ− [A] is the smallest eigenvalue of the matrix A. (i) There is a unique sequence of estimators {βˆn } such that P sn (βˆn ) = 0 → 1

and

βˆn →p β,

(4.80)

293

4.5. Asymptotically Efficient Estimation

where sn (γ) is the score function defined to be the left-hand side of (4.59) with γ = β. (ii) Let In (β) = Var(sn (β)). Then [In (β)]1/2 (βˆn − β) →d Np (0, Ip ).

(4.81)

(iii) If φ in (4.58) is known or the p.d.f. in (4.55) indexed by θ = (β, φ) satisfies the conditions for fθ in Theorem 4.16, then βˆn is asymptotically efficient. Proof. (i) The proof of the existence of βˆn satisfying (4.80) is the same as that of Theorem 4.17(i) with θ = β, except that we need to show

max [In (β)]−1/2 ∇sn (γ)[In (β)]−1/2 + Ip →p 0, γ∈Bn (c)

where Bn (c) = {γ : k[In (β)]1/2 (γ − β)k ≤ c}. From (4.62) and (4.63), In (β) = Mn (β)/φ and ∇sn (γ) = [Rn (γ) − Mn (γ)]/φ, where Mn (γ) and Rn (γ) are defined by (4.60)-(4.61) with γ = β. Hence, it suffices to show that for any c > 0,

max [Mn (β)]−1/2 [Mn (γ) − Mn (β)][Mn (β)]−1/2 → 0 (4.82) γ∈Bn (c)

and

max [Mn (β)]−1/2 Rn (γ)[Mn (β)]−1/2 →p 0.

γ∈Bn (c)

The left-hand side of (4.82) is bounded by √ 1 − ϕ(γ τ Zi )/ϕ(β τ Zi ) , p max γ∈Bn (c),i≤n

which converges to 0 since ϕ is continuous and, for γ ∈ Bn (c), |γ τ Zi − β τ Zi |2 = |(γ − β)τ [In (β)]1/2 [In (β)]−1/2 Zi |2 ≤ k[In (β)]1/2 (γ − β)k2 k[In (β)]−1/2 Zi k2 ≤ c2 max Ziτ [In (β)]−1 Zi i≤n

−1 ≤ c φ t0 inf ϕ(β τ Zi ) max Ziτ (Z τ Z)−1 Zi 2

i

i≤n

→0

under the assumed conditions. This proves (4.82). Let ei = Xi − µ(ψ(β τ Zi )), Un (γ) =

n X i=1

[µ(ψ(β τ Zi )) − µ(ψ(γ τ Zi ))]ψ ′′ (γ τ Zi )ti Zi Ziτ ,

(4.83)

294

4. Estimation in Parametric Models

Vn (γ) =

n X i=1

and

ei [ψ ′′ (γ τ Zi ) − ψ ′′ (β τ Zi )]ti Zi Ziτ ,

Wn (β) =

n X

ei ψ ′′ (β τ Zi )ti Zi Ziτ .

i=1

Then Rn (γ) = Un (γ) + Vn (γ) + Wn (β). Using the same argument as that in proving (4.82), we can show that

max [Mn (β)]−1/2 Un (γ)[Mn (β)]−1/2 → 0. γ∈Bn (c)

Note that [Mn (β)]−1/2 Vn (γ)[Mn (β)]−1/2 is bounded by the product of [Mn (β)]−1/2

n X i=1

|ei |ti Zi Ziτ [Mn (β)]−1/2 = Op (1)

and max

γ∈Bn (c),i≤n

′′ τ ψ (γ Zi ) − ψ ′′ (β τ Zi ) ,

which can be shown to be o(1) using the same argument as that in proving (4.82). Hence,

max [Mn (β)]−1/2 Vn (γ)[Mn (β)]−1/2 →p 0 γ∈Bn (c)

and (4.83) follows from

[Mn (β)]−1/2 Wn (β)[Mn (β)]−1/2 →p 0.

To show this result, we apply Theorem 1.14(ii). Since E(ei ) = 0 and ei ’s are independent, it suffices to show that n X 1+δ E ei ψ ′′ (β τ Zi )ti Ziτ [Mn (β)]−1 Zi →0

(4.84)

i=1

for some δ ∈ (0, 1). Note that supi E|ei |1+δ < ∞. Hence, there is a constant C > 0 such that the left-hand side of (4.84) is bounded by C

n X τ τ −1 1+δ Zi (Z Z) Zi ≤ pC max |Ziτ (Z τ Z)−1 Zi |δ → 0. i=1

i≤n

Hence, (4.84) follows from Theorem 1.14(ii). This proves (4.80). The uniqueness of βˆn follows from (4.83) and the fact that Mn (γ) is positive definite in a neighborhood of β. This completes the proof of (i).

4.5. Asymptotically Efficient Estimation

295

(ii) The proof of (ii) is very similar to that of Theorem 4.17(ii). Using the results in the proof of (i) and Taylor’s expansion, we can establish (exercise) that [In (β)]1/2 (βˆn − β) = [In (β)]−1/2 sn (β) + op (1). (4.85) Using the CLT (e.g., Corollary 1.3) and Theorem 1.9(iii), we can show (exercise) that [In (β)]−1/2 sn (β) →d Np (0, Ip ). (4.86) Result (4.81) follows from (4.85)-(4.86) and Slutsky’s theorem. (iii) The result is obvious if φ is known. When φ is unknown, it follows from (4.59) that ∂ ∂ log ℓ(θ) sn (β) =− . ∂φ ∂β φ Since E[sn (β)] = 0, the Fisher information about θ = (β, φ) is 2 ∂ log ℓ(θ) 0 In (β) , In (β, φ) = −E = 0 I˜n (φ) ∂θ∂θτ where I˜n (φ) is the Fisher information about φ. The result then follows from (4.81) and the discussion in the end of §4.5.1.

4.5.3 Other asymptotically efficient estimators To study other asymptotically efficient estimators, we start with MRIE’s in location-scale families. Since MLE’s and RLE’s are invariant (see Exercise 109 in §4.6), MRIE’s are often asymptotically efficient; see, for example, Stone (1974). Assume the conditions in Theorem 4.16 and let sn (γ) be the score func(0) tion. Let θˆn be an estimator of θ that may not be asymptotically efficient. The estimator θˆn(1) = θˆn(0) − [∇sn (θˆn(0) )]−1 sn (θˆn(0) ) (4.87) is the first iteration in computing an MLE (or RLE) using the Newton(0) Raphson iteration method with θˆn as the initial value (see (4.53)) and, (1) therefore, is called the one-step MLE. Without any further iteration, θˆn (1) can be used as a numerical approximation to an MLE or RLE; and θˆn is asymptotically efficient under some conditions, as the following result shows. Theorem 4.19. Assume that the conditions in Theorem 4.16 hold and √ (0) that θˆn is n-consistent for θ (Definition 2.10). (1) (i) The one-step MLE θˆn is asymptotically efficient. (ii) The one-step MLE obtained by replacing ∇sn (γ) in (4.87) with its

296

4. Estimation in Parametric Models

expected value, −In (γ) (the Fisher-scoring method), is asymptotically efficient. √ (0) (0) Proof. Since θˆn is n-consistent, we can focus on the event θˆn ∈ Aǫ = {γ : kγ − θk ≤ ǫ} for a sufficiently small ǫ such that Aǫ ⊂ Θ. From the mean-value theorem, Z 1 sn (θˆn(0) ) = sn (θ) + ∇sn θ + t(θˆn(0) − θ) dt (θˆn(0) − θ). 0

Substituting this into (4.87) we obtain that

θˆn(1) − θ = −[∇sn (θˆn(0) )]−1 sn (θ) + [Ik − Gn (θˆn(0) )](θˆn(0) − θ), where Gn (θˆn(0) ) = [∇sn (θˆn(0) )]−1

Z

1

0

∇sn θ + t(θˆn(0) − θ) dt.

(0) From (4.77), k[In (θ)] [∇sn (θˆn )]−1 [In (θ)]1/2 + Ik k →p 0. Using an argument similar to those in the proofs of (4.77) and (4.82), we can show that √ (0) (0) kGn (θˆn ) − Ik k →p 0. These results and the fact that n(θˆn − θ) = Op (1) imply √ √ (1) n(θˆn − θ) = n[In (θ)]−1 sn (θ) + op (1). 1/2

This proves (i). The proof for (ii) is similar. Example 4.40. Let X1 , ..., Xn be i.i.d. from the Weibull distribution W (θ, 1), where θ > 0 is unknown. Note that n

sn (θ) =

n

X n X + log Xi − Xiθ log Xi θ i=1 i=1

and

n

∇sn (θ) = −

X n − Xiθ (log Xi )2 . 2 θ i=1

Hence, the one-step MLE of θ is " # (0) ˆn(0) (Pn log Xi − Pn X θˆn log Xi ) n + θ i=1 i=1 i . θˆn(1) = θˆn(0) 1 + Pn ˆ(0) (0) n + (θˆn )2 i=1 Xiθn (log Xi )2

Usually one can use a moment estimator (§3.5.2) as the initial estimator (0) ¯ = θˆn . In this example, a moment estimator of θ is the solution of X −1 Γ(θ + 1). Results similar to that in Theorem 4.19 can be obtained in non-i.i.d. cases, for example, the GLM discussed in §4.4.2 (exercise); see also §5.4.

297

4.5. Asymptotically Efficient Estimation

As we discussed in §4.1.3, Bayes estimators are usually consistent. The next result, due to Bickel and Yahav (1969) and Ibragimov and Has’minskii (1981), states that Bayes estimators are asymptotically efficient when Xi ’s are i.i.d. Theorem 4.20. Assume the conditions of Theorem 4.16. Let π(γ) be a prior p.d.f. (which may be improper) w.r.t. the Lebesgue measure on Θ and pn (γ) be the posterior p.d.f., given X1 , ..., Xn , n = 1, 2, .... Assume that Rthere exists an n0 such R that pn0 (γ) is continuous and positive for all γ ∈ Θ, pn0 (γ)dγ = 1 and kγkpn0 (γ)dγ < ∞. Suppose further that, for any ǫ > 0, there exists a δ > 0 such that ! log ℓ(γ) − log ℓ(θ) lim P > −δ = 0 (4.88) sup n→∞ n kγ−θk≥ǫ and lim P

n→∞

k∇sn (γ) − ∇sn (θ)k ≥ǫ sup n kγ−θk≤δ

!

= 0,

(4.89)

where ℓ(γ) is the likelihood function and s√ n (γ) is the score function. (i) Let p∗n (γ) be the posterior p.d.f. of n(γ − Tn ), where Tn = θ + [In (θ)]−1 sn (θ) and θ is the true parameter value, and let ψ(γ) be the p.d.f. of Nk (0, [I1 (θ)]−1 ). Then Z (1 + kγk) p∗n (γ) − ψ(γ) dγ →p 0. (4.90)

(ii) The Bayes estimator of θ under the squared error loss is asymptotically efficient.

The proof of Theorem 4.20 is lengthy and is omitted; see Lehmann (1983, §6.7) for a proof of the case of univariate θ. A number of conclusions can be drawn from Theorem 4.20. First, result (4.90) shows that the posterior p.d.f. is approximately normal with mean θ + [In (θ)]−1 sn (θ) and covariance matrix [In (θ)]−1 . This result is useful in Bayesian computation; see Berger (1985, §4.9.3). Second, (4.90) shows that the posterior distribution and its first-order moments converge to the degenerate distribution at θ and its first-order moments, which implies the consistency and asymptotic unbiasedness of Bayes estimators such as the posterior means. Third, the Bayes estimator under the squared error loss is asymptotically efficient, which provides an additional support for the early suggestion that the Bayesian approach is a useful method for generating estimators. Finally, the results hold regardless of the prior being used, indicating that the effect of the prior declines as n increases.

298

4. Estimation in Parametric Models

In addition to the regularity conditions in Theorem 4.16, Theorem 4.20 requires two more nontrivial regularity conditions, (4.88) and (4.89). Let us verify these conditions for natural exponential families (Example 4.39), i.e., Xi ’s are i.i.d. with p.d.f. (4.79). Since ∇sn (η) = −n∂ 2 ζ(η)/∂η∂η τ , (4.89) follows from the continuity of the second-order derivatives of ζ. To show (4.88), consider first the case of univariate η. Without loss of generality, we assume that γ > η. Note that log ℓ(γ) − log ℓ(η) ζ(γ) − ζ(η) ′ ′ ¯ = T − ζ (η) + ζ (η) − (γ − η), (4.91) n γ−η

where T¯ is the average of T (Xi )’s. Since ζ(γ) is strictly convex, γ > η implies ζ ′ (η) < [ζ(γ) − ζ(η)]/(γ − η). Also, T¯ →a.s. ζ ′ (η). Hence, with probability tending to 1, the factor in front of (γ − η) on the right-hand side of (4.91) is negative. Then (4.88) holds with ǫ ζ(γ) − ζ(η) δ= inf − ζ ′ (η) . 2 γ≥η+ǫ γ −η To show how to extend this to multivariate η, consider the case of bivariate η. Let ηj , γj , and ξj be the jth components of η, γ, and T¯ − ∇ζ(η), respectively. Assume γ1 > η1 and γ2 > η2 . Let ζj′ be the derivative of ζ w.r.t. the jth component of η. Then the left-hand side of (4.91) is the sum of (γ1 − η1 )ξ1 − [ζ(η1 , γ2 ) − ζ(η1 , η2 ) − (γ2 − η2 )ζ2′ (η1 , η2 )] and

(γ2 − η2 )ξ2 − [ζ(γ1 , γ2 ) − ζ(η1 , γ2 ) − (γ1 − η1 )ζ1′ (η1 , η2 )],

where the last quantity is bounded by

(γ2 − η2 )ξ2 − [ζ(γ1 , γ2 ) − ζ(η1 , γ2 ) − (γ1 − η1 )ζ1′ (η1 , γ2 )], since ζ1′ (η1 , η2 ) ≤ ζ1′ (η1 , γ2 ). The rest of the proof is the same as the case of univariate η. When Bayes estimators have explicit forms under a specific prior, it is usually easy to prove the asymptotic efficiency of the Bayes estimators directly. For instance, in Example 4.7, the Bayes estimator of θ is −1 ¯ + γ −1 ¯ nX ¯ + γ − (α − 1)X = X ¯ + O 1 a.s., =X n+α−1 n+α−1 n

¯ is the MLE of θ. Hence the Bayes estimator is asymptotically where X efficient by Slutsky’s theorem. A similar result can be obtained for the Bayes estimator δt (X) in Example 4.7. Theorem 4.20, however, is useful in cases where Bayes estimators do not have explicit forms and/or the prior is not specified clearly. One such example is the problem in Example 4.40 (Exercises 153 and 154).

299

4.6. Exercises

4.6 Exercises 1. Show that the priors in the following cases are conjugate priors: (a) X1 , ..., Xn are i.i.d. from Nk (θ, Ik ), θ ∈ Rk , and Π = Nk (µ0 , Σ0 ) (Normal family); (b) X1 , ..., Xn are i.i.d. from the binomial distribution Bi(θ, k), θ ∈ (0, 1), and Π = B(α, β) (Beta family); (c) X1 , ..., Xn are i.i.d. from the uniform distribution U (0, θ), θ > 0, and Π = P a(a, b) (Pareto family); (d) X1 , ..., Xn are i.i.d. from the exponential distribution E(0, θ), θ > 0, Π = the inverse gamma distribution Γ−1 (α, γ) (a random variable Y has the inverse gamma distribution Γ−1 (α, γ) if and only if Y −1 has the gamma distribution Γ(α, γ)). (e) X1 , ..., Xn are i.i.d. from the exponential distribution E(θ, 1), θ ∈ R, and Π has a Lebesgue p.d.f. b−1 e−a/b eθ/b I(−∞,a) (θ), a ∈ R, b > 0. 2. In Exercise 1, find the posterior mean and variance for each case. 3. Let X1 , ..., Xn be i.i.d. from the N (θ, 1) distribution and let the prior be the double exponential distribution DE(0, 1). Obtain the posterior mean. 4. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (0, θ), where θ > 0 is unknown. Let the prior of θ be the log-normal distribution LN (µ0 , σ02 ), where µ0 ∈ R and σ0 > 0 are known constants. (a) Find the posterior p.d.f. of ϑ = log θ. (b) Find the rth posterior moment of θ. (c) Find a value that maximizes the posterior p.d.f. of θ. 5. Show that if T (X) is a sufficient statistic for θ ∈ Θ, then the Bayes action δ(x) in (4.3) is a function of T (x). ¯ be the sample mean of n i.i.d. observations from N (θ, σ 2 ) with 6. Let X a known σ > 0 and an unknown θ ∈ R. Let π(θ) be a prior p.d.f. w.r.t. a σ-finite measure on R. ¯ = x, is of the form (a) Show that the posterior mean of θ, given X σ 2 d log(p(x)) , n dx ¯ unconditional on θ. where p(x) is the marginal p.d.f. of X, ¯ = x) as a function (b) Express the posterior variance of θ (given X of the first two derivatives of log(p(x)) w.r.t. x. (c) Find explicit expressions for p(x) and δ(x) in (a) when the prior is N (µ0 , σ02 ) with probability 1 − ǫ and a point mass at µ1 with probability ǫ, where µ0 , µ1 , and σ02 are known constants. δ(x) = x +

300

4. Estimation in Parametric Models

7. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). Find the Bayes action w.r.t. the uniform prior on [0, 1] in the problem of estimating p under the loss L(p, a) = (p−a)2 /[p(1−p)]. 8. Consider the estimation of θ in Exercise 41 of §2.6 under the squared error loss. Suppose that the prior of θ is the uniform distribution U (0, 1), the prior of j is P (j = 1) = P (j = 2) = 21 , and the joint prior of (θ, j) is the product probability of the two marginal priors. Show that the Bayes action is δ(x) =

H(x)B(t + 1) + G(t + 1) , H(x)B(t) + G(t)

where x = (x1 , ..., xn ) is the vector of observations, t = x1 + · · · + xn , R1 R1 B(t) = 0 θt (1 − θ)n−t dθ, G(t) = 0 θt e−nθ dθ, and H(x) is a function of x with range {0, 1}. 9. Consider the estimation problem in Example 4.1 R with the loss function L(θ, a) = w(θ)[g(θ) − a]2 , where w(θ) ≥ 0 and Θ w(θ)[g(θ)]2 dΠ < ∞. Show that the Bayes action is R w(θ)g(θ)fθ (x)dΠ . δ(x) = ΘR Θ w(θ)fθ (x)dΠ

10. Let X be a sample from Pθ , θ ∈ Θ ⊂ R. Consider the estimation of θ under the loss L(|θ − a|), where L is an increasing function on [0, ∞). Let π(θ|x) be the posterior p.d.f. of θ given X = x. Suppose that π(θ|x) is symmetric about δ(x) ∈ Θ and that π(θ|x) is nondecreasing for θ ≤ δ(x) and nonincreasing for θ ≥ δ(x). Show that δ(x) is a Bayes action, assuming that all integrals involved are finite. 11. Let X be a sample of size 1 from the geometric distribution G(p) with an unknown p ∈ (0, 1]. Consider the estimation of p with A = [0, 1] and the loss function L(p, a) = (p − a)2 /p. (a) RShow that δ is aR Bayes action w.r.t. Π if and only if δ(x) = 1 − (1 − p)x dΠ(p)/ (1 − p)x−1 dΠ(p), x = 1, 2, .... (b) Let δ0 be a rule such that δ0 (1) = 1/2 and δ0 (x) = 0 for all x > 1. Show that δ0 is a limit of Bayes actions. (c) Let δ0 be a rule such that δ0 (x) = 0 for all x > 1 and δ0 (1) is arbitrary. Show that δ0 is a generalized Bayes action. 12. Let X be a single observation from N (µ, σ 2 ) with a known σ 2 and an unknown µ > 0. Consider the estimation of µ under the squared error loss and the noninformative prior Π = the Lebesgue measure on (0, ∞). Show that the generalized Bayes action when X = x is δ(x) = x + σΦ′ (x/σ)/[1 − Φ(−x/σ)], where Φ is the c.d.f. of the standard normal distribution and Φ′ is its derivative.

301

4.6. Exercises

13. Let X be a sample from Pθ having the p.d.f. h(x) exp{θτ x − ζ(θ)} w.r.t. ν. Let Π be the Lebesgue measure on Θ = Rp . Show that the generalized Bayes action under the loss L(θ, a) = kE(X) − ak2 is δ(x) = x when X = x. 14. Let , ..., Xn be i.i.d. random variables with the Lebesgue p.d.f. p X1−(x−θ) 2 /2 2/πe I(θ,∞) (x), where θ ∈ R is unknown. Find the generalized Bayes action for estimating θ under the squared error loss, when the (improper) prior of θ is the Lebesgue measure on R.

15. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) and π(µ, σ 2 ) = σ −2 I(0,∞) (σ 2 ) be an improper prior for (µ, σ 2 ) w.r.t. the Lebesgue measure on R2 . (a) Show that the posterior p.d.f. of (µ, σ2 ) given x = (x1 , ..., xn ) is π(µ, σ 2 |x) = π1 (µ|σ 2 , x)π2 (σ 2 |x), where π1 (µ|σ 2 , x) is the p.d.f. of 2 N (¯ x, σ 2 /n) and π2 (σP |x) is the p.d.f. of the inverse gamma distribun −1 tion Γ ((n − 1)/2, [ i=1 (xi − x ¯)2 /2]−1 ) (see Exercise 1(d)). x (b) Show that the marginal posterior p.d.f. of µ given x is f µ−¯ , τ P n where τ 2 = ¯)2 /[n(n − 1)] and f is the p.d.f. of the ti=1 (xi − x distribution tn−1 . (c) Obtain the generalized Bayes action for estimating µ/σ under the squared error loss. 16. Consider Example 3.13. Under the squared error loss and the prior 2 with the improper Lebesgue density π(µ1 , ..., µm , σP ) = σ −2 , obtain m −2 the generalized P Bayes action for estimating θ = σ ¯)2 , i=1 ni (µi − µ m −1 where µ ¯=n i=1 ni µi . 17. Let X be a single observation from the Lebesgue p.d.f. e−x+θ I(θ,∞) (x), where θ > 0 is an unknown parameter. Consider the estimation of j θ ∈ (j − 1, j], j = 1, 2, 3, ϑ= 4 θ>3 under the loss L(i, j), 1 ≤ i, j 0 1 1 3

≤ 4, given by the following matrix: 1 1 2 0 2 2 . 2 0 2 3

3 0

When X = 4, find the Bayes action w.r.t. the prior with the Lebesgue p.d.f. e−θ I(0,∞) (θ).

18. (Bayesian hypothesis testing). Let X be a sample from Pθ , where θ ∈ Θ. Let Θ0 ⊂ Θ and Θ1 = Θc0 , the complement of Θ0 . Consider the problem of testing H0 : θ ∈ Θ0 versus H1 : θ ∈ Θ1 under the loss 0 θ ∈ Θi L(θ, ai ) = Ci θ 6∈ Θi ,

302

4. Estimation in Parametric Models

where Ci > 0 are known constants and {a0 , a1 } is the action space. Let Πθ|x be the posterior distribution of θ w.r.t. a prior distribution Π, given X = x. Show that the Bayes action δ(x) = a1 if and only if Πθ|x (Θ1 ) ≥ C1 /(C0 + C1 ). 19. In (b)-(d) of Exercise 1, assume that the parameters in priors are unknown. Using the method of moments, find empirical Bayes actions under the squared error loss. 20. In Example 4.5, assume that both µ0 and σ02 in the prior for µ are unknown. Let the second-stage joint prior for (µ0 , σ02 ) be the product of N (a, v 2 ) and the Lebesgue measure on (0, ∞), where a and v are known. Under the squared error loss, obtain a formula for the hierarchical Bayes action in terms of a one-dimensional integral. 21. Let X1 , ..., Xn be i.i.d. random variables from the uniform distribution U (0, θ), where θ > 0 is unknown. Let π(θ) = bab θ−(b+1) I(a,∞) (θ) be a prior p.d.f. w.r.t. the Lebesgue measure, where b > 1 is known but a > 0 is an unknown hyperparameter. Consider the estimation of θ under the squared error loss. (a) Show that the empirical Bayes method using the method of moments produces the empirical Bayes action δ(ˆ a), where δ(a) = 2(b−1) Pn b+n ˆ = bn i=1 Xi , and X(n) is the largest orb+n−1 max{a, X(n) }, a der statistic. (b) Let h(a) = a−1 I(0,∞) (a) be an improper Lebesgue prior density for a. Obtain explicitly the hierarchical generalized Bayes action. 22. Let X be a sample and δ(X) with any fixed X R = x ∈ A be a Bayes action, where δ is a measurable function and Θ Pθ (A)dΠ = 1. Show that δ(X) is a Bayes rule as defined in §2.3.2. 23. Let X1 , ..., random variables with the Lebesgue p.d.f. p Xn be i.i.d. 2 fθ (x) = 2θ/πe−θx /2 I[0,∞) (x), where θ > 0 is unknown. Let the prior of θ be the gamma distribution Γ(α, γ) with known α and γ. Find the Bayes estimator of fθ (0) and its Bayes risk under the loss function L(θ, a) = (a − θ)2 /θ. 24. Let X be a single observation from N (θ, θ2 ) and consider a prior p.d.f. −1 2 2 πξ (θ) = c(α, µ, τ )|θ|−α e−(θ −µ) /(2τ ) w.r.t. the Lebesgue measure, where ξ = (α, µ, τ ) is a vector of hyperparameters and c(α, µ, τ ) ensures that πξ (θ) is a p.d.f. (a) Identify the constraints on the hyperparameters for πξ (θ) to be a proper prior. (b) Show that the posterior p.d.f. is πξ∗ (θ) for given X = x and identify ξ∗ .

303

4.6. Exercises

(c) Express the Bayes estimator of |θ| and its Bayes risk in terms of the function c and ξ∗ and state any additional constraints needed on the hyperparameters. 25. Let X1 , X2 , ... be i.i.d. from the exponential distribution E(0, 1). Suppose that we observe T = X1 + · · · + Xθ , where θ is an unknown integer ≥ 1. Consider the estimation of θ under the loss function L(θ, a) = (θ − a)2 /θ and the geometric distribution G(p) as the prior for θ, where p ∈ (0, 1) is known. (a) Show that the posterior expected loss is E[L(θ, a)|T = t] = 1 + ξ − 2a + (1 − e−ξ )a2 /ξ, where ξ = (1 − p)t. (b) Find theP Bayes estimator of θ and show that its posterior expected −mξ loss is 1 − ξ ∞ . m=1 e (c) Find the marginal distribution of (1 − p)T , unconditional on θ. (d) Obtain an explicit expression for the Bayes risk of the Bayes estimator in part (b). 26. Prove (ii) and (iii) of Theorem 4.2. 27. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). ¯ is an admissible estimator of p under the loss function (a) Show that X 2 (a − p) /[p(1 − p)]. ¯ is an admissible estimator of p under the squared (b) Show that X error loss. 28. Let X be a sample (of size 1) from N (µ, 1). Consider the estimation of µ under the loss function L(µ, a) = |µ − a|. Show that X is an admissible estimator. 29. In Exercise 1, consider the posterior mean to be the Bayes estimator of the corresponding parameter in each case. (a) Show that the bias of the Bayes estimator converges to 0 if n → ∞. (b) Show that the Bayes estimator is consistent. (c) Discuss whether the Bayes estimator is admissible. 30. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). (a) Obtain the Bayes estimator of p(1 − p) w.r.t. Π = the beta distribution B(α, β) with known α and β, under the squared error loss. (b) Compare the Bayes estimator in part (a) with the UMVUE of p(1 − p). (c) Discuss the bias, consistency, and admissibility of the Bayes estimator in (a).

304

4. Estimation in Parametric Models (d) Let π(p) = [p(1 − p)]−1 I(0,1) (p) be an improper Lebesgue prior density for p. Show that the posterior of p given Xi ’s is a p.d.f. pro¯ ∈ (0, 1). vided that the sample mean X (e) Under the squared error loss, find the generalized Bayes estimator of p(1 − p) w.r.t. the improper prior in (d).

31. Let X be an observation from the negative binomial distribution N B(p, r) with a known r and an unknown p ∈ (0, 1). (a) Under the squared error loss, find Bayes estimators of p and p−1 w.r.t. Π = the beta distribution B(α, β) with known α and β. (b) Show that the Bayes estimators in (a) are consistent as r → ∞. 32. In Example 4.7, show that ¯ is the generalized Bayes estimator of θ w.r.t. the improper (a) X prior dΠ dω = I(0,∞) (ω) and is a limit of Bayes estimators (as α → 1 and γ → ∞); (b) under the squared error loss for estimating θ, the Bayes estimator ¯ +γ −1 )/(n+α−1) is admissible, but the limit of Bayes estimators, (nX ¯ nX/(n + α − 1) with an α 6= 2, is inadmissible. ¯ is a generalized 33. Consider Example 4.8. Show that the sample mean X ¯ is admissible Bayes estimator of µ under the squared error loss and X using (a) Theorem 4.3 and (b) the result in Example 4.6. 34. Let X be an observation from the gamma distribution Γ(α, θ) with a known α and an unknown θ > 0. Show that X/(α+1) is an admissible estimator of θ under the squared error loss, using Theorem 4.3. 35. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (θ, θ + 1), θ ∈ R. Consider the estimation of θ under the squared error loss. (a) Let π(θ) be a continuous and positive Lebesgue p.d.f. on R. Derive the Bayes estimator w.r.t. the prior π and show that it is a consistent estimator of θ. (b) Show that (X(1) + X(n) − 1)/2 is an admissible estimator of θ and obtain its risk, where X(j) is the jth order statistic. 36. Consider the normal linear model X = Nn (Zβ, σ 2 In ), where Z is an n × p known matrix of full rank, p < n, β ∈ Rp , and σ 2 > 0. (a) Assume that σ 2 is known. Derive the posterior distribution of β when the prior distribution for β is Np (β0 , σ 2 V ), where β0 ∈ Rp is known and V is a known positive definite matrix, and find the Bayes estimator of lτ β under the squared error loss, where l ∈ Rp is known. (b) Show that the Bayes estimator in (a) is admissible and consistent as n → ∞, assuming that the minimum eigenvalue of Z τ Z → ∞. (c) Repeat (a) and (b) when σ 2 is unknown and has the inverse gamma distribution Γ−1 (α, γ) (see Exercise 1(d)), where α and γ are known.

305

4.6. Exercises

(d) In part (c), obtain Bayes estimators of σ 2 and lτ β/σ under the squared error loss and show that they are consistent under the condition in (b). 37. In Example 4.9, suppose that εij has the Lebesgue p.d.f. o n κ(δ)σi−1 exp −c(δ)|x/σi |2/(1+δ) , where

c(δ) =

Γ

3(1+δ) 2 Γ 1+δ 2

(

)

1 1+δ

,

κ(δ) =

Γ

3(1+δ) 2

1/2

3/2

(1+δ)[Γ( 1+δ 2 )]

,

−1 < δ ≤ 1 and σi > 0. −2/(1+δ) (a) Assume that δ is known. Let ωi = c(δ)σi . Under the squared error loss and the same prior in Example 4.9, show that the Bayes estimator of σi2 is 1+δ Z ni X 1 qi (δ) + |xij − β τ Zi |2/(1+δ) f (β|x, δ)dβ, γ j=1

where qi (δ) = [c(δ)]1+δ Γ

1+δ 2 ni

+α−δ

Γ

1+δ 2 ni

+ α + 1 and

−(α+1+ 1+δ 2 ni ) ni k Y X 1 + f (β|x, δ) ∝ π(β) |xij − β τ Zi |2/(1+δ) . γ i=1 j=1

(b) Assume that δ has a prior p.d.f. f (δ) and that given δ, ωi still has the same prior in (a). Derive a formula (similar to that in (a)) for the Bayes estimator of σi2 . 38. Suppose that we have observations Xij = µi + εij ,

i = 1, ..., k, j = 1, ..., m,

where εij ’s are i.i.d. from N (0, σε2 ), µi ’s are i.i.d. from N (µ, σµ2 ), and εij ’s and µi ’s are independent. Suppose that the distribution for σε2 is the inverse gamma distribution Γ−1 (α1 , β1 ) (see Exercise 1(d)); the distribution for σµ2 is the inverse gamma distribution Γ−1 (α2 , β2 ); the distribution for µ is N (µ0 , σ02 ); and σε , σµ , and µ are independent. Describe a Gibbs sampler and obtain explicit forms of (a) the distribution of µ, given Xij ’s, µi ’s, σε2 , and σµ2 ; (b) the distribution of µi , given Xij ’s, µ, σε2 , and σµ2 ; (c) the distribution of σε2 , given Xij ’s, µi ’s, µ, and σµ2 ; (d) the distribution of σµ2 , given Xij ’s, µi ’s, µ, and σε2 . 39. Prove (4.16).

306

4. Estimation in Parametric Models

40. Consider a Lebesgue p.d.f. p(y) ∝ (2+y)125 (1−y)38 y 34 I(0,1) (y). Generate Markov chains of length 10,000 and compute approximations to R yp(y)dy, using the Metropolis kernel with q(y, z) being the p.d.f. of N (y, r2 ), given y, where (a) r = 0.001; (b) r = 0.05; (c) r = 0.12. 41. Prove Proposition 4.4 for the cases of variance and risk. 42. In the proof of Theorem 4.5, show that if L is (strictly) convex and not monotone, then E[L(T0 (x) − a)|D = d] is (strictly) convex and not monotone in a. 43. Prove part (iii) of Theorem 4.5. 44. Under the conditions of Theorem 4.5 and the loss function L(µ, a) = |µ − a|, show that u∗ (d) in Theorem 4.5 is any median (Exercise 92 in §2.6) of T0 (X) under the conditional distribution of X given D = d when µ = 0. 45. Show that if there is a location invariant estimator T0 of µ with finite mean, then E0 [T (X)|D = d] is finite a.s. P for any location invariant estimator T . 46. Show (4.21) under the squared error loss. 47. In Exercise 14, find the MRIE of θ under the squared error loss. 48. In Example 4.12, (a) show that X(1) − θ log 2/n is an MRIE of µ under the absolute error loss L(µ − a) = |µ − a|; (b) show that X(1) − t is an MRIE under the loss function L(µ − a) = I(t,∞) (|µ − a|). 49. In Example 4.13, show that T∗ is also an MRIE of µ if the loss function is convex and even. (Hint: the distribution of T∗ (X) given D depends only on X(n) − X(1) and is symmetric about 0 when µ = 0.) 50. Let X1 , ..., Xn be i.i.d. from the double exponential distribution DE(µ, 1) with an unknown µ ∈ R. Under the squared error loss, find the MRIE of µ. (Hint: for x1 < · · · < xn and xk < t < xk+1 , Pn Pk Pn i=1 |xi − t| = i=k+1 xi − i=1 xi + (2k − n)t.) 51. In Example 4.11, find the MRIE of µ under the loss function −α(µ − a) µ 0 satisfying Z c Z ∞ xdPx|z = xdPx|z , 0

c

where Px|z is the conditional distribution of X given Z = z when σ = 1. −1

56. In Example 4.15, show that the MRIE is 2(n+1) X(n) when the loss is given by (4.24) with p = 1. 57. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(0, θ) with an unknown θ > 0. (a) Find the MRIE of θ under the loss (4.24) with p = 2. (b) Find the MRIE of θ under the loss (4.24) with p = 1. (c) Find the MRIE of θ2 under the loss (4.24) with p = 2. 58. Let X1 , ..., Xn be i.i.d. with a Lebesgue p.d.f. (2/σ)[1−(x/σ)]I(0,σ) (x), where σ > 0 is an unknown scale parameter. Find Pitman’s estimator of σ h for n = 2, 3, and 4. 59. Let X1 , ..., Xn be i.i.d. from the Pareto distribution P a(σ, α), where σ > 0 is an unknown parameter and α > 2 is known. Find the MRIE of σ under the loss function (4.24) with p = 2. 60. Assume that the sample X has a joint Lebesgue p.d.f. given by (4.25). Show that a loss function for the estimation of µ is invariant under the location-scale transformations gc,r (X) = (rX1 + c, ..., rXn + c), r > 0, c ∈ R, if and only if it is of the form L a−µ . σ 61. Prove Proposition 4.6, Theorem 4.9, and Corollary 4.2.

62. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(µ, σ), where µ ∈ R and σ > 0 are unknown. (a) Find the MRIE of σ under the loss (4.24) with p = 1 or 2. (b) Under the loss function (a − µ)2 /σ 2 , find the MRIE of µ. (c) Compute the bias of the MRIE of µ in (b).

308

4. Estimation in Parametric Models

63. Suppose that X and Y are two samples with p.d.f. given by (4.30). (a) Suppose that µx = µy = 0 and consider the estimation of η = (σy /σx )h with a fixed h 6= 0 under the loss L(a/η). Show that the problem is invariant under the transformations g(X, Y ) = (rX, r′ Y ), r > 0, r′ > 0. Generalize Proposition 4.5, Theorem 4.8, and Corollary 4.1 to the present problem. (b) Generalize the result in (a) to the case of unknown µx and µy under the transformations in (4.31). 64. Under the conditions of part (a) of the previous exercise and the loss function (a − η)2 /η 2 , determine the MRIE of η in the following cases: (a) m = n = 1, X and Y are independent, X has the gamma distribution Γ(αx , γ) with a known αx and an unknown γ = σx > 0, and Y has the gamma distribution Γ(αy , γ) with a known αy and an unknown γ = σy > 0; (b) X is Nm (0, σx2 Im ), Y is Nn (0, σy2 In ), and X and Y are independent; (c) X and Y are independent, the components of X are i.i.d. from the uniform distribution U (0, σx ), and the components of Y are i.i.d. from the uniform distribution U (0, σy ). samples, where Xi ’s 65. Let X1 , ..., Xm and Y1 , ..., Yn betwo independent x−µx −1 are i.i.d. having the p.d.f. σx f with µx ∈ R and σx > 0, and σx x−µy with µy ∈ R and σy > 0. Yi ’s are i.i.d. having the p.d.f. σy−1 f σy

Under the loss function (a − η)2 /η 2 and the transformations in (4.31), obtain the MRIE of η = σy /σx when (a) f is the p.d.f. of N (0, 1); (b) f is the p.d.f. of the exponential distribution E(0, 1); (c) f is the p.d.f. of the uniform distribution U − 12 , 12 ; (d) In (a)-(c), find the MRIE of ∆ = µy − µx under the assumption that σx = σy = σ and under the loss function (a − ∆)2 /σ 2 .

66. Consider the general linear model (3.25) under the assumption that εi ’s are i.i.d. with the p.d.f. σ −1 f (x/σ), where f is a known Lebesgue p.d.f. (a) Show that the family of populations is invariant under the transformations in (4.32). τ (b) Show that the estimation τ of l β with l ∈ R(Z) is invariant under the loss function L a−lσ β . (c) Show that the LSE lτ βˆ is an invariant estimator of lτ β, l ∈ R(Z). (d) Prove Theorem 4.10. 67. In Example 4.18, let T be a randomized estimator of p with probabil¯ and probability 1/(n + 1) being 1 . Show that ity n/(n + 1) being X 2

309

4.6. Exercises

¯ T has a constant risk that is smaller than the maximum risk of X. 68. Let X be a single sample from the geometric distribution G(p) with an unknown p ∈ (0, 1). Show that I{1} (X) is a minimax estimator of p under the loss function (a − p)2 /[p(1 − p)]. ¯ is a minimax estimator of µ under the 69. In Example 4.19, show that X 2 2 loss function (a − µ) /σ when Θ = R × (0, ∞). 70. Let T be a minimax (or admissible) estimator of ϑ under the squared error loss. Show that c1 T + c0 is a minimax (or admissible) estimator of c1 ϑ+c0 under the squared error loss, where c1 and c0 are constants. 71. Let X be a sample from Pθ with an unknown θ = (θ1 , θ2 ), where θj ∈ Θj , j = 1, 2, and let Π2 be a probability R measure on Θ2 . Suppose that an estimator T0 minimizesRsupθ1 ∈Θ1 RT (θ)dΠ2 (θ2 ) over all estimators T and that supθ1 ∈Θ1 RT0 (θ)dΠ2 (θ2 ) = supθ1 ∈Θ1 ,θ2 ∈Θ2 RT0 (θ). Show that T0 is a minimax estimator. 72. Let X1 , ..., Xm be i.i.d. from N (µx , σx2 ) and Y1 , ..., Yn be i.i.d. from N (µy , σy2 ). Assume that Xi ’s and Yj ’s are independent. Consider the estimation of ∆ = µy − µx under the squared error loss. ¯ is a minimax estimator of ∆ when σx and σy (a) Show that Y¯ − X ¯ are known, where X and Y¯ are the sample means based on Xi ’s and Yi ’s, respectively. ¯ is a minimax estimator of ∆ when σx ∈ (0, cx ] (b) Show that Y¯ − X and σy ∈ (0, cy ], where cx and cy are constants. 73. Consider the general linear model (3.25) with assumption A1 and the estimation of lτ β under the squared error loss, where l ∈ R(Z). Show that the LSE lτ βˆ is minimax if σ 2 ∈ (0, c] with a constant c. 74. Let X be a random variable having the hypergeometric distribution HG(r, θ, N − θ) (Table 1.1, page 18) with known N and r but an unknown θ. Consider the estimation of θ/N under the squared error loss. (a) Show that the p risk function of T (X) = αX/r + β is constant, where α = {1 + (N − r)/[r(N − 1)]}−1 and β = (1 − α)/2. (b) Show that T in (a) is the minimax estimator of θ/N and the Bayes estimator w.r.t. the prior Π({θ}) =

Γ(2c) [Γ(c)]2

Z

0

1

N θ+c−1 t (1 − t)N −θ+c−1 dt, θ = 1, ..., N, θ

where c = β/(α/r − 1/N ).

310

4. Estimation in Parametric Models

75. Let X be a single observation from N (µ, 1) and let µ have the improper Lebesgue prior density π(µ) = eµ . Under the squared error loss, show that the generalized Bayes estimator of µ is X + 1, which is neither minimax nor admissible. 76. Let X be a random variable having the Poisson distribution P (θ) with an unknown θ > 0. Consider the estimation of θ under the squared error loss. (a) Show that supθ RT (θ) = ∞ for any estimator T = T (X). (b) Let ℑ = {aX + b : a ∈ R, b ∈ R}. Show that 0 is a ℑ-admissible estimator of θ. 77. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(a, θ) with a known θ and an unknown a ∈ R. Under the squared error loss, show that X(1) − θ/n is the unique minimax estimator of a. 78. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (µ − 12 , µ + 12 ) with an unknown µ ∈ R. Under the squared error loss, show that (X(1) + X(n) )/2 is the unique minimax estimator of µ. 79. Let X1 , ..., Xn be i.i.d. from the double exponential distribution DE(µ, 1) with an unknown µ ∈ R. Under the squared error loss, find a minimax estimator of µ. ¯ + b)/(n + 1) is an admissi80. Consider Example 4.7. Show that (nX ble estimator of θ under the squared error loss for any b ≥ 0 and ¯ that nX/(n + 1) is a minimax estimator of θ under the loss function L(θ, a) = (a − θ)2 /θ2 . 81. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p ∈ (0, 1). Consider the estimation of p under the squared error loss. ¯ and (X ¯ + γλ)/(1 + λ) with λ > 0 Using Theorem 4.14, show that X and 0 ≤ γ ≤ 1 are admissible. 82. Let X be a single observation. Using Theorem 4.14, find values of α and β such that αX + β are admissible for estimating EX under the squared error loss when (a) X has the Poisson distribution P (θ) with an unknown θ > 0; (b) X has the negative binomial distribution N B(p, r) with a known r and an unknown p ∈ (0, 1). 83. Let X be a single observation having the Lebesgue p.d.f. 21 c(θ)eθx−|x| , |θ| < 1. (a) Show that c(θ) = 1 − θ2 . (b) Show that if 0 ≤ α ≤ 21 , then αX + β is admissible for estimating E(X) under the squared error loss.

311

4.6. Exercises

84. Let X be a single observation from the discrete p.d.f. fθ (x) = [x!(1 − e−θ )]−1 θx e−θ I{1,2,...} (x), where θ > 0 is unknown. Consider the estimation of ϑ = θ/(1 − e−θ ) under the squared error loss. (a) Show that the estimator X is admissible. (b) Show that X is not minimax unless supθ RT (θ) = ∞ for any estimator T = T (X). (c) Find a loss function under which X is minimax and admissible. 85. In Example 4.23, find the UMVUE of θ = (p1 , ..., pk ) under the loss function (4.37). 86. Let X be a sample from Pθ , θ ∈ Θ ⊂ Rp . Consider the estimation of θ under the loss (θ − a)τ Q(θ − a), where a ∈ A = Θ and Q is a known positive definite matrix. Show that the Bayes action is the posterior mean E(θ|X = x), assuming that all integrals involved are finite. 87. In Example 4.24, show that X is the MRIE of θ under the loss function (4.37), if Qp (a) f (x − θ) = j=1 fj (xj − θj ), where each fj is a known Lebesgue p.d.f. with mean 0; R (b) f (x − θ) = f (kx − θk) with xf (kxk)dx = 0.

88. Prove that X in Example 4.25 is a minimax estimator of θ under the loss function (4.37). 89. Let X1 , ..., Xk be independent random variables, where Xi has the binomial distribution Bi(pi , ni ) with an unknown pi ∈ (0, 1) and a known ni . For estimating θ = (p1 , ..., pk ) under the loss (4.37), find a minimax estimator of θ and determine whether it is admissible. 90. Show that the risk function in (4.42) tends to p as kθk → ∞.

91. Suppose that X is Np (θ, Ip ). Consider the estimation of θ under the loss (a − θ)τ Q(a − θ) with a positive definite p × p matrix Q. Show that the risk of the estimator Q δc,r =X−

r(p − 2) Q−1 (X − c) kQ−1/2 (X − c)k2

is equal to tr(Q) − (2r − r2 )(p − 2)2 E(kQ−1/2 (X − c)k−2 ). 92. Show that under the loss (4.37), the risk of δ˜c,r in (4.45) is given by (4.46).

312

4. Estimation in Parametric Models

93. Suppose that X is Np (θ, V ) with p ≥ 4. Consider the estimation of θ under the loss function (4.37). (a) When V = Ip , show that the risk of the estimator in (4.48) is ¯ p k−2 ). p − (p − 3)2 E(kX − XJ 2 (b) When V = σ D with an unknown σ 2 > 0 and a known matrix D, show that the risk function of the estimator in (4.49) is smaller than that of X for any θ and σ 2 . 94. Let X be a sample from a p.d.f. fθ and T (X) be a sufficient statistic for θ. Show that if an MLE exists, it is a function of T but it may not be sufficient for θ. 95. Let {fθ : θ ∈ Θ} be a family of p.d.f.’s w.r.t. a σ-finite measure, where Θ ⊂ Rk ; h be a Borel function from Θ onto Λ ⊂ Rp , 1 ≤ p ≤ k; and ˜ let ℓ(λ) = supθ:h(θ)=λ ℓ(θ) be the induced likelihood function for the transformed parameter λ. Show that if θˆ ∈ Θ is an MLE of θ, then ˆ = h(θ) ˆ maximizes ˜l(λ). λ 96. Let X1 , ..., Xn be i.i.d. with a p.d.f. fθ . Find an MLE of θ in each of the following cases. (a) fθ (x) = θ−1 I{1,...,θ}(x), θ is an integer between 1 and θ0 . (b) fθ (x) = e−(x−θ)I(θ,∞) (x), θ > 0. (c) fθ (x) = θ(1 − x)θ−1 I(0,1) (x), θ > 1. θ x(2θ−1)/(1−θ) I(0,1) (x), θ ∈ ( 12 , 1). (d) fθ (x) = 1−θ −1 −|x−θ| (e) fθ (x) = 2 e , θ ∈ R. (f) fθ (x) = θx−2 I(θ,∞) (x), θ > 0. (g) fθ (x) = θx (1 − θ)1−x I{0,1} (x), θ ∈ [ 21 , 34 ]. (h) fθ (x) is the p.d.f. of N (θ, θ2 ), θ ∈ R, θ 6= 0. (i) fθ (x) is the p.d.f. of the exponential distribution E(µ, σ), θ = (µ, σ) ∈ R × (0, ∞). (j) fθ (x) is the p.d.f. of the log-normal distribution LN (µ, σ 2 ), θ = (µ, σ 2 ) ∈ R × (0, ∞). √ (k) fθ (x) = I(0,1) (x) if θ = 0 and fθ (x) = (2 x)−1 I(0,1) (x) if θ = 1. (l) fθ (x) = β −α αxα−1 I(0,β) (x), θ = (α, β) ∈ (0, ∞) × (0, ∞). (m) fθ (x) = xθ px (1 − p)θ−xI{0,1,...,θ} (x), θ = 1, 2, ..., where p ∈ (0, 1) is known. (n) fθ (x) = 12 (1 − θ2 )eθx−|x|, θ ∈ (−1, 1). 97. In Exercise 14, obtain an MLE of θ when (a) θ ∈ R and (b) θ ≤ 0. 98. Suppose that n observations are taken from N (µ, 1) with an unknown µ. Instead of recording all the observations, one records only whether the observation is less than 0. Find an MLE of µ. 99. Find an MLE of θ in Exercise 43 of §2.6.

4.6. Exercises

313

100. Let (Y1 , Z1 ), ..., (Yn , Zn ) be i.i.d. random 2-vectors such that Y1 and Z1 are independently distributed as the exponential distributions E(0, λ) and E(0, µ), respectively, where λ > 0 and µ > 0. (a) Find the MLE of (λ, µ). (b) Suppose that we only observe Xi = min{Yi , Zi } and ∆i = 1 if Xi = Yi and ∆i = 0 if Xi = Zi . Find the MLE of (λ, µ). 101. In Example 4.33, show that almost surely the likelihood equation has a unique solution that is the MLE of θ = (α, γ). Obtain iteration equation (4.53) for this example. Discuss how to apply the Fisherscoring method in this example. 102. Let X1 , ..., Xn be i.i.d. from the discrete p.d.f. in Exercise 84 with an unknown θ > 0. Show that the likelihood equation has a unique root when the sample mean > 1. Show whether this root is an MLE of θ. 103. Let X1 , ..., Xn be i.i.d. from the logistic distribution LG(µ, σ) (Table 1.2, page 20). (a) Show how to find an MLE of µ when µ ∈ R and σ is known. (b) Show how to find an MLE of σ when σ > 0 and µ is known. 104. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. from a two-dimensional normal distribution with E(X1 ) = E(Y1 ) = 0, Var(X1 ) = Var(Y1 ) = 1, and an unknown correlation coefficient ρ ∈ (−1, 1). Show that the likelihood equation is a cubic in ρ and the probability that it has a unique root tends to 1 as n → ∞. 105. Let X1 , ..., Xn be i.i.d. from the Weibull distribution W (α, θ) (Table 1.2, page 20) with unknown α > 0 and θ > P 0. Show that n −1 the likelihood equation is equivalent to h(α) = n i=1 log xi and Pn Pn Pn −1 α α −1 α −1 θ=n x , where h(α) = ( x ) x , i=1 i i=1 i i=1 i log xi − α and that the likelihood equation has a unique solution. 106. Consider the random effects model in Example 3.17. Assume that µ = 0 and ni = n0 for all i. Provide a condition on Xij ’s under which a unique MLE of (σa2 , σ 2 ) exists and find this MLE. 107. Let X1 , ..., Xn be i.i.d. with the p.d.f. θf (θx), where f is a Lebesgue p.d.f. on (0, ∞) or symmetric about 0, and θ > 0 is an unknown parameter. Show that the likelihood equation has a unique root if xf ′ (x)/f (x) is continuous in x and strictly decreasing for x > 0. Verify that this condition is satisfied if f is the p.d.f. of the Cauchy distribution C(0, 1). 108. Let X1 , ..., Xn be i.i.d. with the Lebesgue p.d.f. fθ (x) = θf1 (x) + (1 − θ)f2 (x), where fj ’s are two different known Lebesgue p.d.f.’s and

314

4. Estimation in Parametric Models

θ ∈ (0, 1) is unknown. (a) Provide a necessary and sufficient condition for the likelihood equation to have a unique solution and show that if there is a solution, it is the MLE of θ. (b) Derive the MLE of θ when the likelihood equation has no solution. 109. Consider the location family in §4.2.1 and the scale family in §4.2.2. In each case, show that an MLE or an RLE (root of the likelihood equation) of the parameter, if it exists, is invariant. 110. Let X be a sample from Pθ , θ ∈ R. Suppose that Pθ ’s have p.d.f.’s fθ w.r.t. a common σ-finite measure and that {x : fθ (x) > 0} does not depend on θ. Assume further that an estimator θˆ of θ attains the Cram´er-Rao lower bound and that the conditions in Theorem 3.3 ˆ Show that θˆ is a unique MLE of θ. hold for θ. 111. Let Xij , j = 1, ..., r > 1, i = 1, ..., n, be independently distributed as N (µi , σ 2 ). Find the MLE of (µ1 , ..., µn , σ 2 ). Show that the MLE of σ 2 is not a consistent estimator (as n → ∞). 112. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (0, θ), where θ > 0 is unknown. Let θˆ be the MLE of θ and T be the UMVUE. (a) Obtain the ratio mseT (θ)/mseθˆ(θ) and show that the MLE is inadmissible when n ≥ 2. (b) Let Za,θ be a random variable having the exponential distribution ˆ →d Z0,θ and n(θ − T ) →d Z−θ,θ . Obtain the E(a, θ). Prove n(θ − θ) asymptotic relative efficiency of θˆ w.r.t. T . 113. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(a, θ) with unknown a and θ. Obtain the asymptotic relative efficiency of the MLE of a (or θ) w.r.t. the UMVUE of a (or θ). 114. Let X1 , ..., Xn be i.i.d. from the Pareto distribution P a(a, θ) with unknown a and θ. (a) Find the MLE of (a, θ). (b) Find the asymptotic relative efficiency of the MLE of a w.r.t. the UMVUE of a. 115. In Exercises 40 and 41 of §2.6, (a) obtain an MLE of (θ, j); (b) show whether the MLE of j in part (a) is consistent; (c) show that the MLE of θ is consistent and derive its nondegenerated asymptotic distribution. 116. In Example 4.36, obtain the MLE of β under the canonical link and assumption (4.58) but ti 6≡ 1.

4.6. Exercises

315

117. Consider the GLM P in Example 4.35 with φi ≡ 1 and the canonical n link. Assume that i=1 Zi Ziτ is positive definite for n ≥ n0 . Show that the likelihood equation has at most one solution when n ≥ n0 and a solution exists with probability tending to 1. 118. Consider the linear model (3.25) with ε = Nn (0, V ), where V is an unknown positive definite matrix. Show that the LSE βˆ defined by (3.29) is an MQLE and that βˆ is an MLE if and only if one of (a)-(e) in Theorem 3.10 holds. 119. Let Xj be a random variable having the binomial distribution Bi(pj , nj ) with a known nj and an unknown pj ∈ (0, 1), j = 1, 2. Assume that Xj ’s are independent. Obtain a conditional likelihood p1 p2 function of the odds ratio θ = 1−p 1−p2 , given X1 + X2 . 1 120. Let X1 and X2 be independent from Poisson distributions P (µ1 ) and P (µ2 ), respectively. Suppose that we are interested in θ1 = µ1 /µ2 . Derive a conditional likelihood function of θ1 , using (a) θ2 = µ1 ; (b) θ2 = µ1 + µ2 ; and (c) θ2 = µ1 µ2 . 121. Assume model (4.66) with p = 2 and normally distributed i.i.d. εt ’s. Obtain the conditional likelihood given (X1 , X2 ) = (x1 , x2 ). 122. Prove the claim in Example 4.38. 123. Prove (4.70). (Hint: Show, using the argument in proving (4.77), that ∂2 ∂2 n−1 | ∂θ 2 log ℓ(ξn ) − ∂θ 2 log ℓ(θ)| = op (1) for any random variable ξn satisfying |ξn − θ| ≤ |θ − θn |.) 124. Let X1 , ..., Xn be i.i.d. from N (µ, 1) truncated at two known points α < β, i.e., the Lebesgue p.d.f. of Xi is √ 2 { 2π[Φ(β − µ) − Φ(α − µ)]}−1 e−(x−µ) /2 I(α,β) (x). ¯ is asymptotically efficient for esti(a) Show that the sample mean X mating θ = EX1 . ¯ is the unique MLE of θ. (b) Show that X 125. Let X1 , ..., Xn be i.i.d. from the discrete p.d.f. x m−x fθ (x) = [1 − (1 − θ)m ]−1 m I{1,2,...,m} (x), x θ (1 − θ)

where θ ∈ (0, 1) is unknown and m ≥ 2 is a known integer. ¯ = m, show that X/m ¯ (a) When the sample mean X is an MLE of θ. ¯ (b) When 1 < X < m, show that the likelihood equation has at least one solution. (c) Show that the regularity conditions of Theorem 4.16 are satisfied and find the asymptotic variance of a consistent RLE of θ.

316

4. Estimation in Parametric Models

126. In Exercise 96, check whether the regularity conditions of Theorem 4.16 are satisfied for cases (b), (c), (d), (e), (g), (h), (j) and (n). Obtain nondegenerated asymptotic distributions of RLE’s for cases in which Theorem 4.17 can be applied. 127. Let X1 , ..., Xn be i.i.d. random variables such that log Xi is N (θ, θ) with an unknown θ > 0. (a) Obtain the likelihood equation and show that one of the solutions of the likelihood equation is the unique MLE of θ. (b) Using Theorem 4.17, obtain the asymptotic distribution of the MLE of θ. 128. In Exercise 107 of §3.6, find the MLE’s of α and β and obtain their nondegenerated asymptotic joint distribution. 129. In Example 4.30, show that the MLE (or RLE) of θ is asymptotically efficient by (a) applying Theorem 4.17 and (b) directly deriving the asymptotic distribution of the MLE. 130. In Example 4.23, show that there is a unique asymptotically efficient RLE of θ = (p1 , ..., pk ). Discuss whether this RLE is the MLE. 131. Let X1 , ..., Xn be i.i.d. with P (X1 = 0) = 6θ2 − 4θ + 1, P (X1 = 1) = θ − 2θ2 , and P (X1 = 2) = 3θ − 4θ2 , where θ ∈ (0, 21 ) is unknown. Apply Theorem 4.17 to obtain the asymptotic distribution of an RLE of θ. 132. Let X1 , ..., Xn be i.i.d. random variables from N (µ, 1), where µ ∈ R is unknown. Let θ = P (X1 ≤ c), where c is a known constant. Find the asymptotic relative efficiency ofP the MLE of θ w.r.t. (a) the UMVUE of θ and (b) the estimator n−1 ni=1 I(−∞,c] (Xi ).

133. In Exercise 19 of §3.6, find the MLE’s of θ and ϑ = P (Y1 > 1) and find the asymptotic relative efficiency of the MLE of ϑ w.r.t. the UMVUE of ϑ in part (b). 134. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors. Suppose that both X1 and Y1 are binary, P (X1 = 1) = 21 , P (Y1 = 1|X1 = 0) = e−aθ , and P (Y1 = 1|X1 = 0) = e−bθ , where θ > 0 is unknown and a > 0 and b > 0 are known constants. (a) Suppose that (Xi , Yi ), i = 1, ..., n, are observed. Find the MLE of θ and its nondegenerated asymptotic distribution. (b) Suppose that only Y1 , ..., Yn are observed. Find the MLE of θ and its nondegenerated asymptotic distribution. (c) Calculate the asymptotic relative efficiency of the MLE in (a) w.r.t. the MLE in (b). How much efficiency is lost in the special case of a = b?

317

4.6. Exercises

135. In Exercise 110 of §3.6, derive (a) the MLE of (θ1 , θ2 ); (b) a nondegenerated asymptotic distribution of the MLE of (θ1 , θ2 ); (c) the asymptotic relative efficiencies of the MLE’s w.r.t. the moment estimators in Exercise 110 of §3.6. 136. In Exercise 104, show that the RLE of ρ is asymptotically distributed as N ρ, (1 − ρ2 )2 /[n(1 + ρ2 )] .

137. In Exercise 107, obtain a nondegenerated asymptotic distribution of the RLE of θ when f is the p.d.f. of the Cauchy distribution C(0, 1). 138. Let X1 , ..., Xn be i.i.d. from the logistic distribution LG(µ, σ) with unknown µ ∈ R and σ > 0. Obtain a nondegenerated asymptotic distribution of the RLE of (µ, σ). 139. In Exercise 105, show that the conditions of Theorem 4.16 are satisfied. 140. Let X1 , ..., Xn be i.i.d. binary random variables with P (X1 = 1) = p, where p ∈ (0, 1) is unknown. Let ϑˆn be the MLE of ϑ = p(1 − p). (a) Show that ϑˆn is asymptotically normal when p 6= 21 . (b) When p = 12 , derive a nondegenerated asymptotic distribution of ϑˆn with an appropriate normalization. 141. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors satisfying 0 ≤ X1 ≤ 1, 0 ≤ Y1 ≤ 1, and P (X1 > x, Y1 > y) = (1 − x)(1 − y)(1 − max{x, y})θ for 0 ≤ x ≤ 1, 0 ≤ y ≤ 1, where θ ≥ 0 is unknown. (a) Obtain the likelihood function and the likelihood equation. (b) Show that an RLE of θ is asymptotically normal and derive its amse. 142. Assume the conditions in Theorem 4.16. Suppose that θ = (θ1 , ..., θk ) and there is a positive integer p < k such that ∂ log ℓ(θ)/∂θi and ∂ log ℓ(θ)/∂θj are uncorrelated whenever i ≤ p < j. Show that the asymptotic distribution of the RLE of (θ1 , ..., θp ) is unaffected by whether θp+1 , ..., θk are known. 143. Let X1 , ..., Xn be i.i.d. random p-vectors from Np (µ, Σ) with unknown µ and Σ. Find the MLE’s of µ and Σ and derive their nondegenerated asymptotic distributions. 144. Let X1 , ..., Xn be i.i.d. bivariate normal random vectors with mean 0 and an unknown covariance matrix whose diagonal elements are

318

4. Estimation in Parametric Models σ12 and σ22 and off-diagonal element is σ1 σ2 ρ. Let θ = (σ12 , σ22 , ρ). Obtain In (θ) and [In (θ)]−1 and derive a nondegenerated asymptotic distribution of the MLE of θ.

145. Let X1 , ..., Xn be i.i.d. each with probability p as N (µ, σ 2 ) and probability 1 − p as N (η, τ 2 ), where θ = (µ, η, σ 2 , τ 2 , p) is unknown. (a) Show that the conditions in Theorem 4.16 are satisfied. (b) Show that the likelihood function is unbounded. (c) Show that an MLE may be inconsistent. 146. Let X1 , ..., Xn and Y1 , ..., Yn be independently distributed as N (µ, σ 2 ) and N (µ, τ 2 ), respectively, with unknown θ = (µ, σ 2 , τ 2 ). Find the MLE of θ and show that it is asymptotically efficient. 147. Find a nondegenerated asymptotic distribution of the MLE of (σa2 , σ 2 ) in Exercise 106. 148. Under the conditions in Theorem 4.18, prove (4.85) and (4.86). 149. Assume linear model (3.25) with ε = Nn (0, σ 2 In ) and a full rank Z. Apply Theorem 4.18 to show that the LSE βˆ is asymptotically efficient. Compare this result with that in Theorem 3.12. 150. Apply Theorem 4.18 to obtain the asymptotic distribution of the RLE of β in (a) Example 4.35 and (b) Example 4.37. 151. Let X1 , ..., Xn be i.i.d. from the logistic distribution LG(µ, σ), µ ∈ R, σ > 0. Using Newton-Raphson and Fisher-scoring methods, find (a) one-step MLE’s of µ when σ is known; (b) one-step MLE’s of σ when µ is known; (c) one-step MLE’s of (µ, σ); √ (d) n-consistent initial estimators in (a)-(c). 152. Under the GLM (4.55)-(4.58), (a) show how to obtain a one-step MLE of β, if an initial estimator (0) βˆn is available; (b) show that under the conditions in Theorem 4.18, the one-step (0) MLE satisfies (4.81) if k[In (β)]1/2 (βˆn − β)k = Op (1). 153. In Example 4.40, show that the conditions in Theorem 4.20 concerning the likelihood function are satisfied. 154. Let X1 , ..., Xn be i.i.d. from the logistic distribution LG(µ, σ) with unknown µ ∈ R and σ > 0. Show that the conditions in Theorem 4.20 concerning the likelihood function are satisfied.

Chapter 5

Estimation in Nonparametric Models Estimation methods studied in this chapter are useful for nonparametric models as well as for parametric models in which the parametric model assumptions might be violated (so that robust estimators are required) or the number of unknown parameters is exceptionally large. Some such methods have been introduced in Chapter 3; for example, the methods that produce UMVUE’s in nonparametric models, the U- and V-statistics, the LSE’s and BLUE’s, the Horvitz-Thompson estimators, and the sample (central) moments. The theoretical justification for estimators in nonparametric models, however, relies more on asymptotics than that in parametric models. This means that applications of nonparametric methods usually require large sample sizes. Also, estimators derived using parametric methods are asymptotically more efficient than those based on nonparametric methods when the parametric models are correct. Thus, to choose between a parametric method and a nonparametric method, we need to balance the advantage of requiring weaker model assumptions (robustness) against the drawback of losing efficiency, which results in requiring a larger sample size. It is assumed in this chapter that a sample X = (X1 , ..., Xn ) is from a population in a nonparametric family, where Xi ’s are random vectors.

5.1 Distribution Estimators In many applications the c.d.f.’s of Xi ’s are determined by a single c.d.f. F on Rd ; for example, Xi ’s are i.i.d. random d-vectors. In this section, we 319

320

5. Estimation in Nonparametric Models

consider the estimation of F or F (t) for several t’s, under a nonparametric model in which very little is assumed about F .

5.1.1 Empirical c.d.f.’s in i.i.d. cases For i.i.d. random variables X1 , ..., Xn , the empirical c.d.f. Fn is defined in (2.28). The definition of the empirical c.d.f. based on X = (X1 , ..., Xn ) in the case of Xi ∈ Rd is analogously given by n

Fn (t) =

1X I(−∞,t] (Xi ), n i=1

t ∈ Rd ,

(5.1)

where (−∞, a] denotes the set (−∞, a1 ] × · · · × (−∞, ad ] for any a = (a1 , ..., ad ) ∈ Rd . Similar to the case of d = 1 (Example 2.26), Fn (t) as an estimator of F (t) has the following properties. For any t ∈ Rd , nFn (t) has the binomial distribution Bi(F (t), n); Fn (t) is unbiased with variance F (t)[1 − F (t)]/n; √Fn (t) is the UMVUE under some nonparametric models; and Fn (t) is n-consistent for F (t). For any m fixed distinct points t1 , ..., tm in Rd , it follows from the multivariate CLT (Corollary 1.2) and (5.1) that as n → ∞, √ n Fn (t1 ), ..., Fn (tm ) − F (t1 ), ..., F (tm ) →d Nm (0, Σ), (5.2) where Σ is the m × m matrix whose (i, j)th element is P X1 ∈ (−∞, ti ] ∩ (−∞, tj ] − F (ti )F (tj ).

Note that these results hold without any assumption on F . Considered as a function of t, Fn is a random element taking values in √ F, the collection of all c.d.f.’s on Rd . As n → ∞, n(Fn − F ) converges in some sense to a random element defined on some probability space. A detailed discussion of such a result is beyond our scope and can be found, for example, in Shorack and Wellner (1986). To discuss some global properties of Fn as an estimator of F ∈ F, we need to define a closeness measure between the elements (c.d.f.’s) in F. Definition 5.1. Let F0 be a collection of c.d.f.’s on Rd . (i) A function ̺ from F0 × F0 to [0, ∞) is called a distance or metric on F0 if and only if for any Gj in F0 , (a) ̺(G1 , G2 ) = 0 if and only if G1 = G2 ; (b) ̺(G1 , G2 ) = ̺(G2 , G1 ); and (c) ̺(G1 , G2 ) ≤ ̺(G1 , G3 ) + ̺(G3 , G2 ). (ii) Let D = {c(G1 − G2 ) : c ∈ R, Gj ∈ F0 , j = 1, 2}. A function k · k from D to [0, ∞) is called a norm on D if and only if (a) k∆k = 0 if and only if ∆ = 0; (b) kc∆k = |c|k∆k for any ∆ ∈ D and c ∈ R; and (c) k∆1 + ∆2 k ≤ k∆1 k + k∆2 k for any ∆j ∈ D, j = 1, 2.

321

5.1. Distribution Estimators

Any norm k·k on D induces a distance given by ̺(G1 , G2 ) = kG1 −G2 k. The most commonly used distance is the sup-norm distance ̺∞ , i.e., the distance induced by the sup-norm kG1 − G2 k∞ = sup |G1 (t) − G2 (t)|, t∈Rd

Gj ∈ F.

(5.3)

The following result concerning the sup-norm distance between Fn and F is due to Dvoretzky, Kiefer, and Wolfowitz (1956). Lemma 5.1. (DKW’s inequality). Let Fn be the empirical c.d.f. based on i.i.d. X1 , ..., Xn from a c.d.f. F on Rd . (i) When d = 1, there exists a positive constant C (not depending on F ) such that 2 P ̺∞ (Fn , F ) > z ≤ Ce−2nz , z > 0, n = 1, 2, ....

(ii) When d ≥ 2, for any ǫ > 0, there exists a positive constant Cǫ,d (not depending on F ) such that 2 P ̺∞ (Fn , F ) > z ≤ Cǫ,d e−(2−ǫ)nz , z > 0, n = 1, 2, .... The proof of this lemma is omitted. The following results useful in statistics are direct consequences of Lemma 5.1. Theorem 5.1. Let Fn be the empirical c.d.f. based on i.i.d. X1 , ..., Xn from a c.d.f. F on Rd . Then (i) ̺∞ √ (Fn , F ) →a.s. 0 as n → ∞; (ii) E[ n̺∞ (Fn , F )]s = O(1) for any s > 0. Proof. (i) From DKW’s inequality, ∞ X

n=1

P ̺∞ (Fn , F ) > z < ∞.

Hence, the result follows from Theorem 1.8(v). √ (ii) Using DKW’s inequality with z = y 1/s / n and the result in Exercise 55 of §1.6, we obtain that Z ∞ √ √ E[ n̺∞ (Fn , F )]s = P n̺∞ (Fn , F ) > y 1/s dy 0 Z ∞ 2/s ≤ Cǫ,d e−(2−ǫ)y dy 0

= O(1)

as long as 2 − ǫ > 0.

322

5. Estimation in Nonparametric Models

Theorem 5.1(i) means that Fn (t) →a.s. F (t) uniformly in t ∈ Rd , a result stronger than√the strong consistency of Fn (t) for every t. Theorem √ 5.1(ii) implies that n̺∞ (Fn , F ) = Op (1), a result stronger than the nconsistency of Fn (t). These results hold without any condition on F . R Let p ≥ 1 and Fp = {G ∈ F : ktkp dG < ∞}, which is the subset of c.d.f.’s in F having finite pth moments. Mallows’ distance between G1 and G2 in Fp is defined to be ̺Mp (G1 , G2 ) = inf(EkY1 − Y2 kp )1/p ,

(5.4)

where the infimum is taken over all pairs of Y1 and Y2 having c.d.f.’s G1 and G2 , respectively. Let {Gj : j = 0, 1, 2, ...} ⊂ Fp . Then ̺Mp (Gj , G0 ) → 0 as R R j → ∞ if and only if ktkp dGj → ktkp dG0 and Gj (t) → G0 (t) for every t ∈ Rd at which G0 is continuous. It follows from Theorem 5.1 and the SLLN (Theorem 1.13) that ̺Mp (Fn , F ) →a.s. 0 if F ∈ Fp . When d = 1, another useful distance for measuring the closeness between Fn and F is the Lp distance ̺Lp induced by the Lp -norm (p ≥ 1) Z 1/p p |G1 (t) − G2 (t)| dt , Gj ∈ F1 . (5.5) kG1 − G2 kLp = A result similar to Theorem 5.1 is given as follows. Theorem 5.2. Let Fn be the empirical c.d.f. based on i.i.d. random variables X1 , ..., Xn from a c.d.f. F ∈ F1 . Then (i) ̺Lp (Fn , F ) →a.s. 0; R √ (ii) E[ n̺Lp (Fn , F )] = O(1) if 1 ≤ p < 2 and {F (t)[1 − F (t)]}p/2 dt < ∞, or p ≥ 2. Proof. (i) Since [̺Lp (Fn , F )]p ≤ [̺∞ (Fn , F )]p−1 [̺L1 (Fn , F )] and, by Theorem 5.1, ̺∞ (Fn , F ) →a.s. 0, it suffices to show the result for p = 1. Let R0 Yi = −∞ [I(−∞,t] (Xi ) − F (t)]dt. Then Y1 , ..., Yn are i.i.d. and Z Z E|Yi | ≤ E|I(−∞,t] (Xi ) − F (t)|dt = 2 F (t)[1 − F (t)]dt,

which is finite under the condition that F ∈ F1 . By the SLLN, Z 0 n 1X [Fn (t) − F (t)]dt = Yi →a.s. E(Y1 ) = 0. (5.6) n i=1 −∞ R0 Since [Fn (t) − F (t)]− ≤ F (t) and −∞ F (t)dt < ∞ (Exercise 55 in §1.6), it R 0follows from Theorem 5.1 and the dominated convergence theorem that −∞ [Fn (t) − F (t)]− dt →a.s. 0, which with (5.6) implies Z 0 |Fn (t) − F (t)|dt →a.s. 0. (5.7) −∞

323

5.1. Distribution Estimators

The result follows since we can similarly show that (5.7) holds with R∞ replaced by 0 . (ii) When 1 ≤ p < 2, the result follows from E[̺Lp (Fn , F )] ≤ ≤

Z

Z

=n

p

E|Fn (t) − F (t)| dt

Z

−∞

1/p

[E|Fn (t) − F (t)|2 ]p/2 dt

−1/2

R0

{F (t)[1 − F (t)]}

1/p

p/2

dt

= O(n−1/2 ),

1/p

where the two inequalities follow from Jensen’s inequality. When p ≥ 2, o n E[̺Lp (Fn , F )] ≤ E [̺∞ (Fn , F )]1−2/p [̺L2 (Fn , F )]2/p n o1/q 1/p ≤ E[̺∞ (Fn , F )](1−2/p)q E[̺L2 (Fn , F )]2 1/p n o1/q Z −(1−2/p)q/2 2 E |Fn (t) − F (t)| dt = O(n ) = O(n

−(1−2/p)/2

= O(n−1/2 ),

)

1 n

Z

F (t)[1 − F (t)]dt

1/p

where 1q + 1p = 1, the second inequality follows from H¨older’s inequality (see (1.40) in §1.3.2), and the first equality follows from Theorem 5.1(ii).

5.1.2 Empirical likelihoods In §4.4 and §4.5, we have shown that the method of using likelihoods provides some asymptotically efficient estimators. We now introduce some likelihoods in nonparametric models. This not only provides another justification for the use of the empirical c.d.f. in (5.1), but also leads to a useful method of deriving estimators in various (possibly non-i.i.d.) cases, some of which are discussed later in this chapter. Let X1 , ..., Xn be i.i.d. with F ∈ F and PG be the probability measure corresponding to G ∈ F. Given X1 = x1 , ..., Xn = xn , the nonparametric likelihood function is defined to be the following functional from F to [0, ∞): ℓ(G) =

n Y

i=1

PG ({xi }),

G ∈ F.

(5.8)

324

5. Estimation in Nonparametric Models

Apparently, ℓ(G) = 0 if PG ({xi }) = 0 for at least one i. The following result, due to Kiefer and Wolfowitz (1956), shows that the empirical c.d.f. Fn is a nonparametric maximum likelihood estimator of F . Theorem 5.3. Let X1 , ..., Xn be i.i.d. with F ∈ F and ℓ(G) be defined by (5.8). Then Fn maximizes ℓ(G) over G ∈ F. Proof. We only need to consider G ∈ F such that ℓ(G) > 0. Let c ∈ (0, 1] and F(c) be the subset Pn of F containing G’s satisfying pi = PG ({xi }) > 0, i = 1, ..., n, and i=1 pi = c. We now apply the Lagrange multiplier method to solve the problem of maximizing ℓ(G) over G ∈ F(c). Define ! n n Y X H(p1 , ..., pn , λ) = pi + λ pi − c , i=1

i=1

where λ is the Lagrange multiplier. Set n

X ∂H = pi − c = 0, ∂λ i=1

n Y ∂H = p−1 pi + λ = 0, j ∂pj i=1

j = 1, ..., n.

The solution is pi = c/n, i = 1, ..., n, λ = −(c/n)n−1 . It can be shown (exercise) that Pn this solution is a maximum of H(p1 , ..., pn , λ) over pi > 0, i = 1, ..., n, i=1 pi = c. This shows that max ℓ(G) = (c/n)n ,

G∈F(c)

which is maximized at c = 1 for any fixed n. The result follows from PFn ({xi }) = n−1 for given Xi = xi , i = 1, ..., n. From the proof of Theorem 5.3, P Fn maximizes the likelihood ℓ(G) in (5.8) over pi > 0, i = 1, ..., n, and ni=1 pi = 1, where pi = PG ({xi }). This method of deriving an estimator of F can be extended to various situations with some modifications of (5.8) and/or constraints on pi ’s. Modifications of the likelihood in (5.8) are called empirical likelihoods (Owen, 1988, 2001; Qin and Lawless, 1994). An estimator obtained by maximizing an empirical likelihood is then called a maximum empirical likelihood estimator (MELE). We now discuss several applications of the method of empirical likelihoods. Consider first the estimation of F with auxiliary information about F (and i.i.d. X1 , ..., Xn ). For instance, suppose that there is a known Borel function u from Rd to Rs such that Z u(x)dF = 0 (5.9) (e.g., some components of the mean of F are 0). It is thenR reasonable to expect that any estimate Fˆ of F has property (5.9), i.e., u(x)dFˆ = 0,

325

5.1. Distribution Estimators

which is not true for the empirical c.d.f. Fn in (5.1), since Z n 1X u(x)dFn = u(Xi ) 6= 0 n i=1

even if E[u(X1 )] = 0. Using the method of empirical likelihoods, a natural solution is to put another constraint in the process of maximizing the likelihood. That is, we maximize ℓ(G) in (5.8) subject to pi > 0,

i = 1, ..., n,

n X

pi = 1,

and

i=1

n X

pi u(xi ) = 0,

(5.10)

i=1

where pi = PG ({xi }). Using the Lagrange multiplier method and an argument similar to the proof of Theorem 5.3, it can be shown (exercise) that an MELE of F is n X Fˆ (t) = pˆi I(−∞,t] (Xi ), (5.11) i=1

where the notation (−∞, t] is the same as that in (5.1), pˆi = n−1 [1 + λτn u(Xi )]−1 ,

i = 1, ..., n,

(5.12)

and λn ∈ Rs is the Lagrange multiplier satisfying n X i=1

n

pˆi u(Xi ) =

u(Xi ) 1X = 0. n i=1 1 + λτn u(Xi )

(5.13)

Note that Fˆ reduces to Fn if u ≡ 0. To see that (5.13) has a solution asymptotically, note that " n # n ∂ 1X u(Xi ) 1X τ log 1 + λ u(Xi ) = ∂λ n i=1 n i=1 1 + λτ u(Xi )

and

∂2 ∂λ∂λτ

"

# n n 1X 1 X u(Xi )[u(Xi )]τ τ log 1 + λ u(Xi ) = − , n i=1 n i=1 [1 + λτ u(Xi )]2

which is negative definite if Var(u(X1 )) is positive definite. Also, " n ) # ( ∂ 1X = E[u(X1 )] = 0. log 1 + λτ u(Xi ) E ∂λ n i=1 λ=0

Hence, using the same argument as in the proof of Theorem 4.18, we can show that there exists a unique sequence {λn (X)} such that as n → ∞, ! n u(Xi ) 1X =0 →1 and λn →p 0. (5.14) P n i=1 1 + λτn u(Xi )

326

5. Estimation in Nonparametric Models

Theorem 5.4. Let X1 , ..., Xn be i.i.d. with F ∈ F, u be a Borel function on Rd satisfying (5.9), and Fˆ be given by (5.11)-(5.13). Suppose that U = Var(u(X1 )) is positive definite. Then, for any m fixed distinct t1 , ..., tm in Rd , √ n[ Fˆ (t1 ), ..., Fˆ (tm ) − F (t1 ), ..., F (tm ) ] →d Nm (0, Σu ), (5.15)

where

Σu = Σ − W τ U −1 W, Σ is given in (5.2), W = W (t1 ), ..., W (tm ) , W (tj ) = E[u(X1 )I(−∞,tj ] (X1 )], and the notation (−∞, t] is the same as that in (5.1). Proof. We prove the case Pn of m = 1. The case of m ≥ 2 is left as an exercise. Let u ¯ = n−1 i=1 u(Xi ). It follows from (5.13), (5.14), and Taylor’s expansion that n

u ¯=

1X u(Xi )[u(Xi )]τ λn [1 + op (1)]. n i=1

By the SLLN and CLT,

U −1 u ¯ = λn + op (n−1/2 ). Using Taylor’s expansion and the SLLN again, we have n n 1X 1X 1 − 1 I(−∞,t] (Xi )(nˆ pi − 1) = I(−∞,t] (Xi ) n i=1 n i=1 1 + λτn u(Xi ) n

=−

1X I(−∞,t] (Xi )λτn u(Xi ) + op (n−1/2 ) n i=1

= −λτn W (t) + op (n−1/2 )

= −¯ uτ U −1 W (t) + op (n−1/2 ).

Thus, n

1X Fˆ (t) − F (t) = Fn (t) − F (t) + I(−∞,t] (Xi )(nˆ pi − 1) n i=1

¯τ U −1 W (t) + op (n−1/2 ) = Fn (t) − F (t) − u n 1 X I(−∞,t] (Xi )−F (t)−[u(Xi )]τ U −1 W (t) + op (n−1/2 ). = n i=1

The result follows from the CLT and the fact that Var [W (t)]τ U −1 u(Xi ) = [W (t)]τ U −1 U U −1 W (t) = [W (t)]τ U −1 W (t)

= E{[W (t)]τ U −1 u(Xi )I(−∞,t] (Xi )}

= Cov I(−∞,t] (Xi ), [W (t)]τ U −1 u(Xi ) .

327

5.1. Distribution Estimators

Comparing (5.15) with (5.2), we conclude that Fˆ is asymptotically more efficient than Fn . Example 5.1 (Survey problems). An example of situations in which we have auxiliary information expressed as (5.9) is a survey problem (Example 2.3) where the population P = {y1 , ..., yN } consists of two-dimensional yj ’s, PN yj = (y1j , y2j ), and the population mean Y¯2 = N −1 j=1 y2j is known. For example, suppose that y1j is the current year’s income of unit j in the population and y2j is the last year’s income. In many applications the population total or mean of y2j ’s is known, for example, from tax return records. Let X1 , ..., Xn be a simple random sample (see Example 2.3) selected from P with replacement. Then Xi ’s are i.i.d. bivariate random vectors whose c.d.f. is F (t) =

N 1 X I(−∞,t] (yj ), N j=1

(5.16)

where the notation (−∞, t] is the same as that in (5.1). If Y¯2 is known, then it can be expressed as (5.9) with u(x1 , x2 ) = x2 − Y¯2 . In survey problems Xi ’s are usually sampled without replacement so that X1 , ..., Xn are not i.i.d. However, for a simple random sample without replacement, (5.8) can still be treated as an empirical likelihood, given Xi ’s. Note that F in (5.16) is the c.d.f. of Xi , regardless of whether Xi ’s are sampled with replacement. If X = (X1 , ..., Xn ) is not a simple random sample, then the likelihood (5.8) has to be modified. Suppose that πi is the probability that the ith unit is selected (see Theorem 3.15). Given X = {yi , i ∈ s}, an empirical likelihood is Y Y 1/π ℓ(G) = (5.17) [PG ({yi })]1/πi = pi i , i∈s i∈s where pi = PG ({yi }). With the auxiliary information (5.9), an MELE of F in (5.16) can be obtained by maximizing ℓ(G) in (5.17) subject to (5.10). In this case F may not be the c.d.f. of Xi , but the c.d.f.’s of Xi ’s are determined by F and πi ’s. It can be shown (exercise) that an MELE is given by (5.11) with X 1 1 pˆi = (5.18) πi [1 + λτn u(yi )] i∈s πi and

X

u(yi ) = 0. π [1 + λτn u(yi )] i∈s i

(5.19)

If πi = a constant, then the MELE reduces to that in (5.11)-(5.13). If

328

5. Estimation in Nonparametric Models

u(x) = 0 (no auxiliary information), then the MELE is X X 1 1 ˆ I(−∞,t] (yi ) , F (t) = π π i∈s i i∈s i which is a ratio of two Horvitz-Thompson estimators (§3.4.2). Some asymptotic properties of the MELE Fˆ can be found in Chen and Qin (1993). The second part of Example 5.1 shows how to use empirical likelihoods in a non-i.i.d. problem. Applications of empirical likelihoods in non-i.i.d. problems are usually straightforward extensions of those in i.i.d. cases. The following is another example. Example 5.2 (Biased sampling). Biased sampling is often used in applications. Suppose that n = n1 + · · · + nk , k ≥ 2; Xi ’s are independent random variables; X1 , ..., Xn1 are i.i.d. with F ; and Xn1 +···+nj +1 , ..., Xn1 +···+nj+1 are i.i.d. with the c.d.f. Z ∞ Z t wj+1 (s)dF (s) wj+1 (s)dF (s), −∞

−∞

j = 1, ..., k − 1, where wj ’s are some nonnegative Borel functions. A simple example is that X1 , ..., Xn1 are sampled from F and Xn1 +1 , ..., Xn1 +n2 are sampled from F but conditional on the fact that each sampled value exceeds a given value x0 (i.e., w2 (s) = I(x0 ,∞) (s)). For instance, Xi ’s are blood pressure measurements; X1 , ..., Xn1 are sampled from ordinary people and Xn1 +1 , ..., Xn1 +n2 are sampled from patients whose blood pressures are higher than x0 . The name biased sampling comes from the fact that there is a bias in the selection of samples. For simplicity we consider the case of k = 2, since the extension to k ≥ 3 is straightforward. Denote w2 by w. An empirical likelihood is ℓ(G) =

n1 Y

i=1

=

"

PG ({xi })

n X i=1

n Y

i=n1

w(xi )PG ({xi }) R w(s)dG(s) +1

#−n2

pi w(xi )

n Y

i=1

pi

n Y

w(xi ),

(5.20)

i=n1 +1

the where pi = PG ({xi }). An MELE of F can be obtained by maximizing Pn empirical likelihood (5.20) subject to pi > 0, i = 1, ..., n, and i=1 pi = 1. Using the Lagrange multiplier method we can show (exercise) that an MELE Fˆ is given by (5.11) with pˆi = [n1 + n2 w(Xi )/w] ˆ −1 ,

i = 1, ..., n,

(5.21)

329

5.1. Distribution Estimators

where w ˆ satisfies w ˆ=

n X i=1

w(Xi ) . n1 + n2 w(Xi )/w ˆ

An asymptotic result similar to that in Theorem 5.4 can be established (Vardi, 1985; Qin, 1993). If the function w depends on an unknown parameter vector θ, then the method of profile empirical likelihood (see §5.1.4) can be applied. Our last example concerns an important application in survival analysis. Example 5.3 (Censored data). Let T1 , ..., Tn be survival times that are i.i.d. nonnegative random variables from a c.d.f. F , and C1 , ..., Cn be i.i.d. nonnegative random variables independent of Ti ’s. In a variety of applications in biostatistics and life-time testing, we are only able to observe the smaller of Ti and Ci and an indicator of which variable is smaller: Xi = min{Ti , Ci },

δi = I(0,Ci ) (Ti ),

i = 1, ..., n.

This is called a random censorship model and Ci ’s are called censoring times. We consider the estimation of the survival distribution F ; see Kalbfleisch and Prentice (1980) for other problems involving censored data. An MELE of F can be derived as follows. Let x(1) ≤ · · · ≤ x(n) be ordered values of Xi ’s and δ(i) be the δ-value associated with x(i) . Consider a c.d.f. G that assigns its mass to the points x(1) , ..., x(n) and the interval (x(n) , ∞). Let pi = PG ({x(i) }), i = 1, ..., n, and pn+1 = 1 − G(x(n) ). An MELE of F is then obtained by maximizing 1−δ(i) n n+1 Y X δ(i) pi pj (5.22) ℓ(G) = i=1

j=i+1

subject to

pi ≥ 0,

i = 1, ..., n + 1,

n+1 X

pi = 1.

(5.23)

i=1

It can be shown (exercise) that an MELE is Fˆ (t) =

n+1 X

pˆi I(0,t] (X(i) ),

(5.24)

i=1

where X(0) = 0, X(n+1) = ∞, X(1) ≤ · · · ≤ X(n) are order statistics, and pˆi =

δ(i) n−i+1

i−1 Y

j=1

1−

δ(j) n−j+1

,

i = 1, ..., n,

pˆn+1 = 1 −

n X j=1

pˆj .

330

5. Estimation in Nonparametric Models

The Fˆ in (5.24) can also be written as (exercise) Y δ(i) 1 − n−i+1 , Fˆ (t) = 1 −

(5.25)

X(i) ≤t

which is the well-known Kaplan-Meier (1958) product-limit estimator. Some asymptotic results for Fˆ in (5.25) can be found, for example, in Shorack and Wellner (1986).

5.1.3 Density estimation Suppose that X1 , ..., Xn are i.i.d. random variables from F and that F is unknown but has a Lebesgue p.d.f. f . Estimation of F can be done by estimating f , which is called density estimation. Note that estimators of F derived in §5.1.1 and §5.1.2 do not have Lebesgue p.d.f.’s. Since f (t) = F ′ (t) a.e., a simple estimator of f (t) is the difference quotient Fn (t + λn ) − Fn (t − λn ) fn (t) = , t ∈ R, (5.26) 2λn where Fn is the empirical c.d.f. given by (2.28) or (5.1) with d = 1, and {λn } is a sequence of positive constants. Since 2nλn fn (t) has the binomial distribution Bi(F (t + λn ) − F (t − λn ), n), E[fn (t)] → f (t) and

Var fn (t) → 0

if λn → 0 as n → ∞ if λn → 0 and nλn → ∞.

Thus, we should choose λn converging to 0 slower than n−1 . If we assume that λn → 0, nλn → ∞, and f is continuously differentiable at t, then it can be shown (exercise) that f (t) 1 + O(λ2n ) msefn (t) (F ) = +o (5.27) 2nλn nλn and, under the additional condition that nλ3n → 0, p nλn [fn (t) − f (t)] →d N 0, 12 f (t) .

(5.28)

A useful class of estimators is the class of kernel density estimators of the form n 1 X t−Xi fˆ(t) = (5.29) w λn , nλn i=1

331

5.1. Distribution Estimators

where w is a known Lebesgue p.d.f. on R and is called the kernel. If we ˆ in (5.29) is essentially the same as the choose w(t) = 12 I[−1,1] (t), then f(t) so-called histogram. The bias of fˆ(t) in (5.29) is Z 1 E[fˆ(t)] − f (t) = f (z)dz − f (t) w t−z λ n λn Z = w(y)[f (t − λn y) − f (t)]dy. If f is bounded and continuous at t, then, by the dominated convergence theorem (Theorem 1.1(iii)), the bias of fˆ(t) converges to 0 as λn → 0; if f ′ R ˆ is bounded and continuous at t and |t|w(t)dt < ∞, then the bias of f(t) ˆ is O(λn ). The variance of f(t) is 1 t−X1 Var fˆ(t) = Var w λn nλ2n Z h i2 1 t−z = w f (z)dz λn nλ2n 2 Z 1 1 f (z)dz − w t−z λn n λn Z 1 1 [w(y)]2 f (t − λn y)dy + O = nλn n 1 w0 f (t) +o = nλn nλn R if f is bounded and continuous at t and w0 = [w(t)]2 dt < ∞. Hence, if λn → 0, nλn → ∞, and f ′ is bounded and continuous at t, then msefˆ(t) (F ) =

w0 f (t) + O(λ2n ). nλn

Using the CLT (Theorem 1.15), one can show (exercise) that if λn → 0, nλn → ∞, and f is bounded and continuous at t, then p ˆ − E[fˆ(t)]} →d N 0, w0 f (t) . nλn {f(t) (5.30) R Furthermore, if f ′ is bounded and continuous at t, |t|w(t)dt < ∞, and nλ3n → 0, then p p ˆ nλn {E[f(t)] − f (t)} = O nλn λn → 0

and, therefore, (5.30) holds with E[fˆ(t)] replaced by f (t). Similar to the estimation of a c.d.f., we can also study global properties of fn or fˆ as an estimator of the density curve f , using a suitably defined

332

5. Estimation in Nonparametric Models

0.0

0.1

0.2

f(t)

0.3

0.4

0.5

True p.d.f. Estimator (5.26) Estimator (5.29)

-2

-1

0

1

2

t

Figure 5.1: Density estimates in Example 5.4 distance between f and its density estimator. R For example, we may study the convergence of supt∈R |fˆ(t) − f (t)| or |fˆ(t) − f (t)|2 dt. More details can be found, for example, in Silverman (1986). Example 5.4. An i.i.d. sample of size n = 200 was generated from N (0, 1). Density curve estimates (5.26) and (5.29) are plotted in Figure 5.1 with the curve of the true p.d.f. For the kernel density estimator (5.29), w(t) = 21 e−|t| is used and λn = 0.4. From Figure 5.1, it seems that the kernel estimate (5.29) is much better than the estimate (5.26). There are many other density estimation methods, for example, the nearest neighbor method (Stone, 1977), the smoothing splines (Wahba, 1990), and the method of empirical likelihoods described in §5.1.2 (see, e.g., Jones (1991)), which produces estimators of the form n 1 X i fˆ(t) = . pˆi w t−X λn λn i=1

333

5.1. Distribution Estimators

5.1.4 Semi-parametric methods Suppose that the sample X is from a population in a family indexed by (θ, ξ), where θ is a parameter vector, i.e., θ ∈ Θ ⊂ Rk with a fixed positive integer k, but ξ is not vector-valued, e.g., ξ is a c.d.f. Such a model is often called a semi-parametric model, although it is nonparametric according to our definition in §2.1.2. A semi-parametric method refers to a statistical inference method that combines a parametric method and a nonparametric method in making an inference about the parametric component θ and the nonparametric component ξ. In the following, we consider two important examples of semi-parametric methods. Partial likelihoods and proportional hazards models The idea of partial likelihood (Cox, 1972) is similar to that of conditional likelihood introduced in §4.4.3. To illustrate this idea, we assume that X has a p.d.f. fθ,ξ and ξ is also a vector-valued parameter. Suppose that X can be transformed into a sequence of pairs (V1 , U1 ), ..., (Vm , Um ) such that "m #" m # Y Y fθ,ξ (x) = gθ (ui |v1 , u1 , ..., ui−1 , vi ) hθ,ξ (vi |v1 , u1 , ..., vi−1 , ui−1 ) , i=1

i=1

where gθ (·|v1 , u1 , ..., ui−1 , vi ) is the conditional p.d.f. of Ui given V1 = v1 , U1 = u1 , ..., Ui−1 = ui−1 , Vi = vi , which does not depend on ξ, and hθ,ξ (·|v1 , u1 , ..., vi−1 , ui−1 ) is the conditional p.d.f. of Vi given V1 = v1 , U1 = u1 , ..., Vi−1 = vi−1 , Ui−1 = ui−1 . The first product in the previous expression for fθ,ξ (x) is called the partial likelihood for θ. When ξ is a nonparametric component, the partial likelihood for θ can be similarly defined, in which case the full likelihood fθ,ξ (x) should be replaced by a nonparametric likelihood or an empirical likelihood. As long as the conditional distributions of Ui given V1 , U1 , ..., Ui−1 , Vi , i = 1, ..., m, are in a parametric family (indexed by θ), the partial likelihood is parametric. A semi-parametric estimation method consists of a parametric method (typically the maximum likelihood method in §4.4) for estimating θ and a nonparametric method for estimating ξ. To illustrate the application of the method of partial likelihoods, we consider the estimation of the c.d.f. of survival data in the random censorship model described in Example 5.3. Following the notation in Example 5.3, we assume that {T1 , ..., Tn } (survival times) and {C1 , ..., Cn } (censoring times) are two sets of independent nonnegative random variables and that Xi = min{Ti , Ci } and δi = I(0,Ci ) (Ti ), i = 1, ..., n, are independent observations. In addition, we assume that there is a p-vector Zi of covariate values associated with Xi and δi . The situation considered in Example 5.3

334

5. Estimation in Nonparametric Models

can be viewed as a special homogeneous case with Zi ≡ a constant. The survival function when the covariate vector is equal to z is defined to be Sz (t) = 1 − Fz (t), where Fz is the c.d.f. of the survival time T having the same distribution as Ti . Assume that fz (t) = Fz′ (t) exists for all t > 0. The function λz (t) = fz (t)/Sz (t) is called the hazard function and the Rt function Λz (t) = 0 λz (s)ds is called the cumulative hazard function, when the covariate vector is equal to z. A commonly adopted model for λz is the following proportional hazards model: λz (t) = λ0 (t)φ(β τ z),

(5.31)

where φ is a known function (typically φ(x) = ex ), z is a value of the pvector of covariates, β ∈ Rp is an unknown parameter vector, and λ0 (t) is the unknown hazard function when the covariate vector is 0 and is referred to as the baseline hazard function. Under model (5.31), 1 − Fz (t) = exp{−Λz (t)} = exp{−φ(β τ z)Λ0 (t)}. Thus, the estimation of the c.d.f. Fz or the survival function Sz can be done through the estimation of β, the parametric component of model (5.31), and Λ0 , the nonparametric component of model (5.31). Consider first the estimation of β using the method of partial likelihoods. Suppose that there are l observed failures at times T(1) < · · · < T(l) , where (i) is the label for the ith failure ordered according to the time to failure. (Note that a failure occurs when δi = 1.) Suppose that there are mi items censored at or after T(i) but before T(i+1) at times T(i,1) , ..., T(i,mi ) (setting T(0) = 0). Let Ui = (i) and Vi = (T(i) , T(i−1,1) , ..., T(i−1,mi−1 ) ), i = 1, ..., l. Then the partial likelihood is l Y P (Ui = (i)|V1 , U1 , ..., Ui−1 , Vi ). i=1

Since λz (t) = lim∆>0,∆→0 ∆−1 Pz (t ≤ T < t + ∆|T > t), where Pz denotes the probability measure of T when the covariate is equal to z, P (Ui = (i)|V1 , U1 , ..., Ui−1 , Vi ) = P

λZ(i) (ti ) φ(β τ Z(i) ) = P , τ j∈Ri λZj (ti ) j∈Ri φ(β Zj )

where ti is the observed value of T(i) , Ri = {j : Xj ≥ ti } is called the risk set, and the last equality follows from assumption (5.31). This leads to the partial likelihood #δi " l n Y Y φ(β τ Z(i) ) φ(β τ Zi ) P P ℓ(β) = = , τ τ j∈Ri φ(β Zj ) j∈Ri φ(β Zj ) i=1 i=1 which is a function of the parameter β, given the observed data. The maximum likelihood method introduced for parametric models in §4.4 can

335

5.1. Distribution Estimators

be applied to obtain a maximum partial likelihood estimator βˆ of β. It is shown in Tsiatis (1981) that βˆ is consistent for β and is asymptotically normal under some regularity conditions. We now consider the estimation of Λ0 . First, assume that the covariate vector Zi is random, (Ti , Ci , Zi ) are i.i.d., and Ti and Ci are conditionally independent given Zi . Let (T, C, Z) be the random vector having the same distribution as (Ti , Ci , Zi ), X = min{T, C}, and δ = I(0,C) (T ). Under assumption (5.31), it can be shown (exercise) that ZZ ∞ λ0 (s)φ(β τ z)H(s|z)dsdG(z), (5.32) Q(t) = P (X > t, δ = 1) = t

where H(s|z) = P (X > s|Z = z) and G is the c.d.f. of Z. Then Z dQ(t) = −λ0 (t) φ(β τ z)H(t|z)dG(z) dt and λ0 (t) = −

dQ(t) 1 , dt K(t)

(5.33)

(5.34)

where K(t) = E[φ(β τ Z)I(t,∞) (X)] (exercise). Consequently, Z t Z t dQ(s) . Λ0 (t) = λ0 (s)ds = − 0 0 K(s) An estimator of Λ0 can then be obtained by substituting Q and K in the previous expression by their estimators n

X ˆ = 1 Q(t) I{Xi >t,δi =1} n i=1 and

n

1 X ˆτ ˆ φ(β Zi )I(t,∞) (Xi ). K(t) = n i=1

(5.35)

This estimator is known as Breslow’s estimator. When Z1 , ..., Zn are nonrandom, we can still use Breslow’s estimator. Its asymptotic properties can be found, for example, in Fleming and Harrington (1991). Profile likelihoods Let ℓ(θ, ξ) be a likelihood (or empirical likelihood), where θ and ξ are not necessarily vector-valued. It may be difficult to maximize the likelihood ℓ(θ, ξ) simultaneously over θ and ξ. For each fixed θ, let ξ(θ) satisfy ℓ(θ, ξ(θ)) = sup ℓ(θ, ξ). ξ

336

5. Estimation in Nonparametric Models

The function ℓP (θ) = ℓ(θ, ξ(θ)) is called a profile likelihood function for θ. Suppose that θˆP maximizes ℓP (θ). Then θˆP is called a maximum profile likelihood estimator of θ. Note that θˆP may be different from an MLE of θ. Although this idea can be applied to parametric models, it is more useful in nonparametric models, especially when θ is a parametric component. For example, consider the empirical likelihood in (5.8) subject to the constraints in (5.10). Sometimes it is more convenient to allow the function u in (5.10) to depend on an unknown parameter vector θ ∈ Rk , where k ≤ s. This leads to the empirical likelihood ℓ(G) in (5.8) subject to (5.10) with u(x) replaced by ψ(x, θ), where ψ is a known function from Rd × Rk to Rs . Maximizing this empirical likelihood is equivalent to maximizing ! n n n Y X X ℓ(p1 , ..., pn , ω, λ, θ) = pi + ω 1 − pi + pi λτ ψ(xi , θ), i=1

i=1

i=1

where ω and λ are Lagrange multipliers. It follows from (5.12) and (5.13) that ω = n, p˜i (θ) = n−1 {1 + [λn (θ)]τ ψ(xi , θ)}−1 with a λn (θ) satisfying n

ψ(xi , θ) 1X =0 n i=1 1 + [λn (θ)]τ ψ(xi , θ)

P maximize ℓ(p1 , ...pn , ω, λ, θ) for any fixed θ. Substituting p˜i with ni=1 p˜i = 1 into ℓ(p1 , ...pn , ω, λ, θ) leads to the following profile empirical likelihood for θ: n Y 1 ℓP (θ) = . (5.36) τ ψ(x , θ)} n{1 + [λ (θ)] n i i=1

If θˆ is a maximum of ℓP (θ) in (5.36), then θˆ is a maximum profile empirical ˆ A likelihood estimator of θ and the corresponding estimator of pi is p˜i (θ). result similar to Theorem 5.4 and a result on asymptotic normality of θˆ are established in Qin and Lawless (1994), under some conditions on ψ. Another example is the empirical likelihood (5.20) in the problem of biased sampling with a function w(x) = wθ (x) depending on an unknown θ ∈ Rk . The profile empirical likelihood for θ is then ℓP (θ) = w ˆθ−n2

n Y

1 n + n w ˆθ 2 θ (xi )/w i=1 1

where w ˆθ satisfies w ˆθ =

n X i=1

n Y

i=n1 +1

wθ (xi ) . n1 + n2 wθ (xi )/w ˆθ

wθ (xi ),

337

5.1. Distribution Estimators

Finally, we consider the problem of missing data. Assume that X1, ...,Xn are i.i.d. random variables from an unknown c.d.f. F and some Xi ’s are missing. Let δi = 1 if Xi is observed and δi = 0 if Xi is missing. Suppose that (Xi , δi ) are i.i.d. Let π(x) = P (δi = 1|Xi = x). If Xi and δi are independent, i.e., π(x) ≡ π does not depend on x, then the empirical c.d.f. based on observed data, i.e., the c.d.f. putting mass r−1 to each observed Xi , where r is the number of observed Xi ’s, is an unbiased and consistent estimator of F , provided that π > 0. On the other hand, if π(x) depends on x, then the empirical c.d.f. based on observed data is a biased and inconsistent estimator of F . In fact, it can be shown (exercise) that the empirical c.d.f. based on observed data is an unbiased estimator of P (Xi ≤ x|δi = 1), which is generally different from the unconditional probability F (x) = P (Xi ≤ x). If both π and F are in parametric models, then we can apply the method of maximum likelihood. For example, if π(x) = πθ (x) and F (x) = Fϑ (x) has a p.d.f. fϑ , where θ and ϑ are vectors of unknown parameters, then a parametric likelihood of (θ, ϑ) is ℓ(θ, ϑ) =

n Y

[πθ (xi )fϑ (xi )]δi (1 − π)1−δi ,

i=1

R

where π = πθ (x)dF (x). Suppose now that π(x) = πθ (x) is the parametric component and F is the nonparametric component. Then an empirical likelihood can be defined as ℓ(θ, G) =

n Y

[πθ (xi )PG ({xi })]δi (1 − π)1−δi

i=1

Pn

Pn subject to pi ≥ 0, i=1 δi pi = 1, i=1 δi pi [πθ (xi ) − π] = 0, where pi = PG ({xi }), i = 1, ..., n. It can be shown (exercise) that the logarithm of the profile empirical likelihood for (θ, π) (with a Lagrange multiplier) is n X i=1

δi log πθ (xi ) +(1−δi ) log(1−π)−δi log 1+λ[πθ (xi )−π] . (5.37)

Under some regularity conditions, Qin, Leung, and Shao (2002) show that ˆ π ˆ obtained by maximizing the likelihood in (5.37) the estimators θ, ˆ , and λ are consistent and asymptotically normal and that the empirical c.d.f. ˆ ˆ(Xi ) − π putting mass pˆi = r−1 {1 + λ[π ˆ ]}−1 to each observed Xi is conθ sistent for F . The results are also extended to the case where a covariate vector Zi associated with Xi is observed for all i.

338

5. Estimation in Nonparametric Models

5.2 Statistical Functionals In many nonparametric problems, we are interested in estimating some characteristics (parameters) of the unknown population, not the entire population. We assume in this section that Xi ’s are i.i.d. from an unknown c.d.f. F on Rd . Most characteristics of F can be written as T(F ), where T is a functional from F to Rs . If we estimate F by the empirical c.d.f. Fn in (5.1), then a natural estimator of T(F ) is T(Fn ), which is called a statistical functional. Many commonly used statistics can be written as T(F R n ) for some T. Two simple examples are given as follows. Let T(F ) = ψ(x)dF R P (x) with an integrable function ψ, and T(Fn ) = ψ(x)dFn (x) = n−1 ni=1 ψ(Xi ). The sample moments discussed in §3.5.2 are particular examples of this kind of statistical functional. For d = 1, let T(F ) = F −1 (p) = inf{x : F (x) ≥ p}, where p ∈ (0, 1) is a fixed constant. F −1 (p) is called the pth quantile of F . The statistical functional T(Fn ) = Fn−1 (p) is called the pth sample quantile. More examples of statistical functionals are provided in §5.2.1 and §5.2.2. In this section, we study asymptotic distributions of T(Fn ). We focus on the case of real-valued T (s = 1), since the extension to the case of s ≥ 2 is straightforward.

5.2.1 Differentiability and asymptotic normality Note that T(Fn ) is a function of the “statistic” Fn . In Theorem 1.12 (and §3.5.1) we have studied how to use Taylor’s expansion to establish asymptotic normality of differentiable functions of statistics that are asymptotically normal. This leads to the approach of establishing asymptotic normality of T(Fn ) by using some generalized Taylor expansions for functionals and using asymptotic properties of Fn given in §5.1.1. First, we need a suitably defined differential of T. Several versions of differentials are given in the following definition. Definition 5.2. Let T be a functional on F0 , a collection of c.d.f.’s on Rd , and let D = {c(G1 − G2 ) : c ∈ R, Gj ∈ F0 , j = 1, 2}. (i) A functional T on F0 is Gˆ ateaux differentiable at G ∈ F0 if and only if there is a linear functional LG on D (i.e., LG (c1 ∆1 + c2 ∆2 ) = c1 LG (∆1 ) + c2 LG (∆2 ) for any ∆j ∈ D and cj ∈ R) such that ∆ ∈ D and G + t∆ ∈ F0 imply T(G + t∆) − T(G) − LG (∆) = 0. lim t→0 t (ii) Let ̺ be a distance on F0 induced by a norm k · k on D. A functional T on F0 is ̺-Hadamard differentiable at G ∈ F0 if and only if there is a

339

5.2. Statistical Functionals

linear functional LG on D such that for any sequence of numbers tj → 0 and {∆, ∆j , j = 1, 2, ...} ⊂ D satisfying k∆j − ∆k → 0 and G + tj ∆j ∈ F0 , T(G + tj ∆j ) − T(G) − LG (∆j ) = 0. lim j→∞ tj (iii) Let ̺ be a distance on F0 . A functional T on F0 is ̺-Fr´echet differentiable at G ∈ F0 if and only if there is a linear functional LG on D such that for any sequence {Gj } satisfying Gj ∈ F0 and ̺(Gj , G) → 0, lim

j→∞

T(Gj ) − T(G) − LG (Gj − G) = 0. ̺(Gj , G)

The functional LG is called the differential of T at G. If we define h(t) = T(G + t∆), then the Gˆ ateaux differentiability is equivalent to the differentiability of the function h(t) at t = 0, and LG (∆) is simply h′ (0). Let διx denote the d-dimensional c.d.f. degenerated at the point x and φG (x) = LG (διx − G). Then φF (x) is called the influence function of T at F , which is an important tool in robust statistics (see Hampel (1974)). If T is Gˆateaux differentiable √ at F , then we have the following expansion (taking t = n−1/2 and ∆ = n(Fn − F )): √ √ n[T(Fn ) − T(F )] = LF n(Fn − F ) + Rn . (5.38) Since LF is linear, LF

n √ 1 X n(Fn − F ) = √ φF (Xi ) →d N (0, σF2 ) n i=1

(5.39)

by the CLT, provided that E[φF (X1 )] = 0

and

σF2 = E[φF (X1 )]2 < ∞

(5.40)

(which is usually true when φF is bounded or when F has some finite moments). By Slutsky’s theorem and (5.39), √ n[T(Fn ) − T(F )] →d N (0, σF2 ) (5.41) if Rn in (5.38) is op (1). Unfortunately, Gˆ ateaux differentiability is too weak to be useful in establishing Rn = op (1) (or (5.41)). This is why we need other types of differentiability. Hadamard differentiability, which is also referred to as compact differentiability, is clearly stronger than Gˆateaux differentiability but weaker than Fr´echet differentiability (exercise). For a given functional

340

5. Estimation in Nonparametric Models

T, we can first find LG by differentiating h(t) = T(G + t∆) at t = 0 and then check whether T is ̺-Hadamard (or ̺-Fr´echet) differentiable with a given ̺. The most commonly used distances on F0 are the sup-norm distance ̺∞ and the Lp distance ̺Lp . Their corresponding norms are given by (5.3) and (5.5), respectively. Theorem 5.5. Let X1 , ..., Xn be i.i.d. from a c.d.f. F on Rd . (i) If T is ̺∞ -Hadamard differentiable at F , then Rn in (5.38) is op (1). (ii) If T is ̺-Fr´echet differentiable at F with a distance ̺ satisfying √ n̺(Fn , F ) = Op (1), (5.42) then Rn in (5.38) is op (1). (iii) In either (i) or (ii), if (5.40) is also satisfied, then (5.41) holds. Proof. Part (iii) follows directly from (i) or (ii). The proof of (i) involves some high-level mathematics and is omitted; see, for example, Fernholz (1983). We now prove (ii).√From Definition 5.2(iii), for any ǫ > 0, there is a δ > 0 such that |Rn | < ǫ n̺(Fn , F ) whenever ̺(Fn , F ) < δ. Then √ n̺(Fn , F ) > η/ǫ + P (̺(Fn , F ) ≥ δ) P (|Rn | > η) ≤ P for any η > 0, which implies

lim sup P (|Rn | > η) ≤ lim sup P n

n

√ n̺(Fn , F ) > η/ǫ .

The result follows from (5.42) and the fact that ǫ can be made arbitrarily small. Since ̺-Fr´echet differentiability implies ̺-Hadamard differentiability, Theorem 5.5(ii) is useful when ̺ is not the sup-norm distance. There are functionals that are not ̺∞ -Hadamard differentiable (and hence R not ̺∞ -Fr´echet differentiable). For example, if d = 1 and T(G) = g( xdG) with a differentiable function g, then T is not necessarily ̺∞ -Hadamard differentiable, but is ̺L1 -Fr´echet differentiable (exercise). From Theorem 5.2, condition (5.42) holds for ̺Lp under the moment conditions on F given in Theorem 5.2. Note that if ̺ and ̺˜ are two distances on F0 satisfying ̺˜(G1 , G2 ) ≤ c̺(G1 , G2 ) for a constant c and all Gj ∈ F0 , then ̺˜-Hadamard (Fr´echet) differentiability implies ̺-Hadamard (Fr´echet) differentiability. This suggests the use of the distance ̺∞+p = ̺∞ + ̺Lp , which also satisfies (5.42) under the moment conditions in Theorem 5.2. The distance ̺∞+p is useful in some cases (Theorem 5.6). A ̺∞ -Hadamard differentiable T having a bounded and continuous influence function φF is robust in Hampel’s sense (see, e.g., Huber (1981)).

5.2. Statistical Functionals

341

This is motivated by the fact that the asymptotic behavior of T(Fn ) is determined by that of LF (Fn − F ), and a small change in the sample, i.e., small changes in all xi ’s (rounding, grouping) or large changes in a few xi ’s (gross errors, blunders), will result in a small change of T(Fn ) if and only if φF is bounded and continuous. We now consider some examples. For the sample moments related to R functionals of the form T(G) = ψ(x)dG(x), it is clear that T is a linear functional. Any linear functional is trivially ̺-Fr´echet differentiable for any ̺. Next, if F is one-dimensional and F ′ (x) > 0 for all x, then the quantile functional T(G) = G−1 (p) is ̺∞ -Hadamard differentiable at F (Fernholz, 1983). Hence, Theorem 5.5 applies to these functionals. But the asymptotic normality of sample quantiles can be established under weaker conditions, which are studied in §5.3.1. Example 5.5 (Convolution functionals). Suppose that F is on R and for a fixed z ∈ R, Z T(G) = G(z − y)dG(y), G ∈ F. If X1 and X2 are i.i.d. with c.d.f. G, then T(G) is the c.d.f. of X1 + X2 (Exercise 47 in §1.6), and is also called the convolution of G evaluated at z. For tj → 0 and k∆j − ∆k∞ → 0, Z Z T(G + tj ∆j ) − T(G) = 2tj ∆j (z − y)dG(y) + t2j ∆j (z − y)d∆j (y) (for ∆ = c1 G1 + c2 G2 , Gj ∈ F0 , and cj ∈ R, d∆ denotes c1 dG1 + c2 dG2 ). Using Lemma 5.2, one can show (exercise) that Z ∆j (z − y)d∆j (y) = O(1). (5.43) Hence T is ̺∞ -Hadamard differentiable at any G R∈ F with LG (∆) = R 2 ∆(z−y)dG(y). The influence function, φF (x) = 2 (διx −F )(z−y)dF (y), is a bounded function and clearly satisfies (5.40). Thus, (5.41) holds. If F is continuous, then T is robust in Hampel’s sense (exercise). Three important classes of statistical functionals, i.e., L-estimators, Mestimators, and rank statistics and R-estimators, are considered in §5.2.2. Lemma 5.2. Let ∆ ∈ D and h be a continuous function on R such that R h(x)d∆(x) is finite. Then Z h(x)d∆(x) ≤ khkV k∆k∞ ,

342

5. Estimation in Nonparametric Models

where khkV is the variation norm defined by m X sup khkV = lim |h(xj ) − h(xj−1 )| a→−∞,b→∞

j=1

with the supremum being taken over all partitions a = x0 < · · · < xm = b of the interval [a, b]. The proof of Lemma 5.2 can be found in Natanson (1961, p. 232). The differentials in Definition 5.2 are first-order differentials. For some functionals, we can also consider their second-order differentials, which provides a way of defining the order of the asymptotic biases via expansion (2.37). Definition 5.3. Let T be a functional on F0 and ̺ be a distance on F0 . (i) T is second-order ̺-Hadamard differentiable at G ∈ F0 if and only if there is a functional QG on D such that for any sequence of numbers tj → 0 and {∆, ∆j , j = 1, 2, ...} ⊂ D satisfying k∆j − ∆k → 0 and G + tj ∆j ∈ F0 , lim

j→∞

T(G + tj ∆j ) − T(G) − QG (tj ∆j ) = 0, t2j

RR ψG (x, y)d(GR + where QG (∆) = R ∆)(x)d(G + ∆)(y) for a function ψG satisfying ψG (x, y) = ψG (y, x), ψG (x, y)dG(x)dG(y) = 0, and D and k · k are the same as those in Definition 5.2(ii). (ii) T is second-order ̺-Fr´echet differentiable at G ∈ F0 if and only if, for any sequence {Gj } satisfying Gj ∈ F0 and ̺(Gj , G) → 0, lim

j→∞

T (Gj ) − T (G) − QG (Gj − G) = 0, [̺(Gj , G)]2

where QG is the same as that in (i). For a second-order differentiable T, we have the following expansion: n[T(Fn ) − T(F )] = nVn + Rn ,

(5.44)

where Vn = QF (Fn − F ) =

Z Z

ψF (x, y)dFn (x)dFn (y) =

n n 1 XX ψF (Xi , Xj ) n2 j=1 i=1

is a “V-statistic” (§3.5.3) whose asymptotic properties are given by Theorem 3.16. If Rn in (5.44) is op (1), then the asymptotic behavior of T(Fn ) − T(F ) is the same as that of Vn .

343

5.2. Statistical Functionals

Proposition 5.1. Let X1 , ..., Xn be i.i.d. from F . (i) If T is second-order ̺∞ -Hadamard differentiable at F , then Rn in (5.44) is op (1). (ii) If T is second-order ̺-Fr´echet differentiable at F with a distance ̺ satisfying (5.42), then Rn in (5.44) is op (1). Combining Proposition 5.1 with Theorem 3.16, we conclude that if Z ψF (X1 , y)dF (y) > 0, ζ1 = Var then (5.41) holds with σF2 = 4ζ1 and amseT (Fn ) (P ) = σF2 /n; if ζ1 = 0, then n[T(Fn ) − T(F )] →d

∞ X

λj χ21j

j=1

and amseT (Fn ) (P ) = {2Var(ψF (X1 , X2 )) + [EψF (X1 , X1 )]2 }/n2 . In any case, expansion (2.37) holds and the n−1 order asymptotic bias of T (Fn ) is EψF (X1 , X1 )/n. If T is also first-order differentiable, then it can be shown (exercise) that Z φF (x) = 2 ψF (x, y)dF (y). (5.45)

Then ζ1 = 4−1 Var(φF (X1 )) and ζ1 = 0 corresponds to the case of φF (x) ≡ 0. However, second-order ̺-Hadamard (Fr´echet) differentiability does not imply first-order ̺-Hadamard (Fr´echet) differentiability (exercise). The technique in this section can be applied to non-i.i.d. Xi ’s when the c.d.f.’s of Xi ’s are determined by an unknown c.d.f. F , provided that results similar to (5.39) and (5.42) (with Fn replaced by some other estimator Fˆ ) can be established.

5.2.2 L-, M-, and R-estimators and rank statistics Three large classes of statistical functionals based on i.i.d. Xi ’s are studied in this section. L-estimators Let J(t) be a Borel function on [0, 1]. An L-functional is defined as Z (5.46) T(G) = xJ(G(x))dG(x), G ∈ F0 , where F0 contains all c.d.f.’s on R for which T is well defined. For X1 , ..., Xn i.i.d. from F ∈ F0 , T(Fn ) is called an L-estimator of T(F ).

344

5. Estimation in Nonparametric Models

Example 5.6. The following are some examples of commonly used Lestimators. ¯ the sample mean. (i) When J ≡ 1, T(Fn ) = X, (ii) When J(t) = 4t − 2, T(Fn ) is proportional to Gini’s mean difference. (iii) When J(t) = (β − α)−1 I(α,β) (t) for some constants α < β, T(Fn ) is called the trimmed sample mean. For an L-functional T, it can be shown (exercise) that Z T(G) − T(F ) = φF (x)d(G − F )(x) + R(G, F ), where φF (x) = − R(G, F ) = − and WG (x) =

(

[G(x) − F (x)]−1 0

Z

Z

(διx − F )(y)J(F (y))dy,

(5.47)

(5.48)

WG (x)[G(x) − F (x)]dx,

R G(x) F (x)

J(t)dt − J(F (x))

G(x) 6= F (x) G(x) = F (x).

A sufficient condition for (5.40) in this case is that J is bounded and F has a finite variance (exercise). However, (5.40) is also satisfied if φF is bounded. The differentiability of T can be verified under some conditions on J. Theorem 5.6. Let T be an L-functional defined by (5.46). (i) Suppose that J is bounded, J(t) = 0 when t ∈ [0, α] ∪ [β, 1] for some constants α < β, and that the set D = {x : J is discontinuous at F (x)} has Lebesgue measure 0. Then T is ̺∞ -Fr´echet differentiable at F with the influence function φF given by (5.48), and φF is bounded and continuous and satisfies (5.40). (ii) Suppose that J is bounded, the set D in (i) has Lebesgue measure 0, and J is continuous on [0, α] ∪ [β, 1] for some constants α < β. Then T is ̺∞+1 -Fr´echet differentiable at F . (iii) Suppose that |J(t) − J(s)| ≤ C|t − s|p−1 , where C > 0 and p > 1 are some constants. Then T is ̺Lp -Fr´echet differentiable at F . (iv) If, in addition to the conditions in part (i), J ′ is continuous on [α, β], then T is second-order ̺∞ -Fr´echet differentiable at F with Z ψF (x, y) = φF (x) + φF (y) − (διx − F )(z)(διy − F )(z)J ′ (F (z))dz. (v) Suppose that J ′ is continuous on [0, 1]. Then T is second-order ̺L2 Fr´echet differentiable at F with the same ψF given in (iv).

345

5.2. Statistical Functionals

Proof. We prove (i)-(iii). The proofs for (iv) and (v) are similar and are left to the reader. (i) Let Gj ∈ F and ̺∞ (Gj , F ) → 0. Let c and d be two constants such that F (c) > β and F (d) < α. Then, for sufficiently large j, Gj (x) ∈ [0, α] ∪ [β, 1] if x > c or x < d. Hence, for sufficiently large j, Z c |R(Gj , F )| = WGj (x)(Gj − F )(x)dx d Z c ≤ ̺∞ (Gj , F ) |WGj (x)|dx. d

Since J is continuous at F (x) when x 6∈ D and D has Lebesgue measure R0,c WGj (x) → 0 a.e. Lebesgue. By the dominated convergence theorem, |WGj (x)|dx → 0. This proves that T is ̺∞ -Fr´echet differentiable. The d assertions on φF can be proved by noting that Z c (διx − F )(y)J(F (y))dy. φF (x) = − d

(ii) From the proof of (i), we only need to show that Z ̺∞+1 (Gj , F ) → 0, W (x)(G − F )(x)dx Gj j

(5.49)

A

where A = {x : F (x) ≤ α or F (x) > β}. The quantity on the left-hand side of (5.49) is bounded by supx∈A |WGj (x)|, which converges to 0 under the continuity assumption of J on [0, α] ∪ [β, 1]. Hence (5.49) follows. (iii) The result follows from Z |R(G, F )| ≤ C |G(x) − F (x)|p dx = O [̺Lp (G, F )]p and the fact that p > 1.

An L-estimator with J(t) = 0 when t ∈ [0, α] ∪ [β, 1] is called a trimmed L-estimator. Theorem 5.6(i) shows that trimmed L-estimators satisfy (5.41) and are robust in Hampel’s sense. In cases (ii) and (iii) of Theorem 5.6, (5.41) holds if Var(X1 ) < ∞, but T(Fn ) may not be robust in Hampel’s sense. It can be shown (exercise) that one or several of (i)-(v) of Theorem 5.6 can be applied to each of the L-estimators in Example 5.6. M-estimators Let ρ(x, t) be a Borel function on Rd × R and Θ be an open subset of R. An M-functional is defined to be a solution of Z Z ρ(x, T(G))dG(x) = min ρ(x, t)dG(x), G ∈ F0 , (5.50) t∈Θ

346

5. Estimation in Nonparametric Models

where F0 contains all c.d.f.’s on Rd for which the integrals in (5.50) are well defined. For X1 , ..., Xn i.i.d. from F ∈ F0 , T(Fn ) is called an M-estimator of T(F ). Assume that ψ(x, t) = ∂ρ(x, t)/∂t exists a.e. and Z Z ∂ λG (t) = ψ(x, t)dG(x) = ρ(x, t)dG(x). (5.51) ∂t Then λG (T(G)) = 0. Example 5.7. The following are some examples of M-estimators. R (i) If ρ(x, t) = (x − t)2 /2, then ψ(x, t) = t − x; T(G) = xdG(x) is the ¯ is the sample mean. mean functional; and T(Fn ) = X (ii) If ρ(x, t) = |x − t|p /p, where p ∈ [1, 2), then x≤t |x − t|p−1 ψ(x, t) = x > t. −|x − t|p−1 When p = 1, T(Fn ) is the sample median. When 1 < p < 2, T (Fn ) is called the pth least absolute deviations estimator or the minimum Lp distance estimator. (iii) Let F0 = {fθ : θ ∈ Θ} be a parametric family of p.d.f.’s with Θ ⊂ R and ρ(x, t) = − log ft (x). Then T(Fn ) is an MLE. This indicates that Mestimators are extensions of MLE’s in parametric models. (iv) Let C > 0 be a constant. Huber (1964) considers ( 1 2 |x − t| ≤ C 2 (x − t) ρ(x, t) = 1 2 |x − t| > C 2C with

t−x |x − t| ≤ C 0 |x − t| > C. The corresponding T(Fn ) is a type of trimmed sample mean. (v) Let C > 0 be a constant. Huber (1964) considers 1 2 |x − t| ≤ C 2 (x − t) ρ(x, t) = |x − t| > C C|x − t| − 12 C 2 ψ(x, t) =

t−x >C C ψ(x, t) = t−x |x − t| ≤ C −C t − x < −C. The corresponding T(Fn ) is a type of Winsorized sample mean. (vi) Hampel (1974) considers ψ(x, t) = ψ0 (t − x) with ψ0 (s) = −ψ0 (−s) and s 0≤s≤a a a<s≤b ψ0 (s) = a(c−s) b <s≤c c−b 0 s > c,

with

347

5.2. Statistical Functionals

where 0 < a < b < c are constants. A smoothed version of ψ0 is sin(as) 0 ≤ s < π/a ψ1 (s) = 0 s > π/a. For bounded and continuous ψ, the following result shows that T is ̺∞ Hadamard differentiable with a bounded and continuous influence function and, hence, T(Fn ) satisfies (5.41) and is robust in Hampel’s sense. Theorem 5.7. Let T be an M-functional defined by (5.50). Assume that ψ is a bounded and continuous function on Rd × R and that λF (t) is continuously differentiable at T(F ) and λ′F (T(F )) 6= 0. Then T is ̺∞ Hadamard differentiable at F with φF (x) = −ψ(x, T(F ))/λ′F (T(F )). Proof. Let tj → 0, ∆j ∈ D, k∆j − ∆k∞ → 0, and Gj = F + tj ∆j ∈ F. Since λG (T(G)) = 0, Z |λF (T(Gj )) − λF (T(F ))| = tj ψ(x, T(Gj ))d∆j (x) → 0

by k∆j − ∆k∞ → 0 and the boundedness of ψ. Note that λ′F (T(F )) 6= 0. Hence, the inverse of λF (t) exists and is continuous in a neighborhood of 0 = λF (T(F )). Therefore, T(Gj ) − T(F ) → 0.

(5.52)

Let hF (T(F )) = λ′F (T(F )), hF (t) = [λF (t) − λF (T(F ))]/[t − T(F )] if t 6= T(F ), Z 1 1 − , R1j = ψ(x, T(F ))d∆j (x) ′ λF (T(F )) hF (T(Gj )) Z 1 [ψ(x, T(Gj )) − ψ(x, T(F ))]d∆j (x), R2j = hF (T(Gj )) and LF (∆) = −

1 ′ λF (T(F ))

Z

ψ(x, T(F ))d∆(x),

∆ ∈ D.

Then T(Gj ) − T(F ) = −LF (tj ∆j ) + tj (R1j − R2j ). By (5.52), k∆j − ∆k∞ → 0, and the boundedness of ψ, Rj1 → 0. The result then follows from R2j → 0, which follows from k∆j − ∆k∞ → 0 and the boundedness and continuity of ψ (exercise).

348

5. Estimation in Nonparametric Models

Some ψ functions in Example 5.7 satisfy the conditions in Theorem 5.7 (exercise). Under more conditions on ψ, it can be shown that an Mfunctional is ̺∞ -Fr´echet differentiable at F (Clarke, 1986; Shao, 1993). Some M-estimators that satisfy (5.41) but are not differentiable functionals are studied in §5.4. Rank statistics and R-estimators Assume that X1 , ..., Xn are i.i.d. from a c.d.f. F on R. The rank of Xi among X1 , ..., Xn , denoted by Ri , is defined to be the number of Xj ’s satisfying Xj ≤ Xi , i = 1, ..., n. The rank of |Xi | among |X1 |, ..., |Xn | is ˜ i . A statistic that is a function of Ri ’s similarly defined and denoted by R ˜ i ’s is called a rank statistic. For G ∈ F, let or R ˜ G(x) = G(x) − G (−x)− , x > 0,

where g(x−) denotes the left limit of the function g at x. Define a functional T by Z ∞ ˜ T(G) = J(G(x))dG(x), G ∈ F, (5.53) 0

where J is a function on [0, 1] with a bounded derivative J ′ . Then Z ∞ n 1 X R˜ i T(Fn ) = J(F˜n (x))dFn (x) = J n I(0,∞) (Xi ) n i=1 0

is a (one-sample) signed rank statistic. If J(t) = t, then T(Fn ) is the wellknown Wilcoxon signed rank test statistic (§6.5.1). Statistics based on ranks (or signed ranks) are robust against changes in values of xi ’s, but may not provide efficient inference procedures, since the values of xi ’s are discarded after ranks (or signed ranks) are determined. It can be shown (exercise) that T in (5.53) is ̺∞ -Hadamard differentiable at F with the differential Z ∞ Z ∞ ˜ LF (∆) = J ′ (F˜ (x))∆(x)dF (x) + J(F˜ (x))d∆(x), (5.54) 0

0

˜ where ∆ ∈ D and ∆(x) = ∆(x) − ∆((−x)−). These results can be extended to the case where X1 , ..., Xn are i.i.d. from a c.d.f. F on R2 . For any c.d.f. G on R2 , let J be a function on [0, 1] with J(1 − t) = −J(t) and a bounded J ′ , ¯ G(y) = [G(y, ∞) + G(∞, y)]/2,

and T(G) =

Z

¯ J(G(y))dG(y, ∞).

y ∈ R, (5.55)

349

5.2. Statistical Functionals

Let Xi = (Yi , Zi ), Ri be the rank of Yi , and Ui be the number of Zj ’s satisfying Zj ≤ Yi , i = 1, ..., n. Then Z n 1X +Ui ¯ T(Fn ) = J(Fn (y))dFn (y, ∞) = J Ri2n n i=1 is called a two-sample linear rank statistic. It can be shown (exercise) that T in (5.55) is ̺∞ -Hadamard differentiable at F with the differential Z Z ¯ LF (∆) = J ′ (F¯ (y))∆(y)dF (y, ∞) + J(F¯ (y))d∆(y, ∞), (5.56)

¯ where ∆(y) = [∆(y, ∞) + ∆(∞, y)]/2. Rank statistics (one-sample or two-sample) are asymptotically normal and robust in Hampel’s sense (exercise). These results are useful in testing hypotheses (§6.5). Let F be a continuous c.d.f. on R symmetric about an unknown parameter θ ∈ R. An estimator of θ closely related to a rank statistic can be derived as follows. Let Xi be i.i.d. from F and Wi = (Xi , 2t − Xi ) with a fixed t ∈ R. The functional T in (5.55) evaluated at the c.d.f. of Wi is equal to Z (2t−x) λF (t) = J F (x)+1−F dF (x). (5.57) 2

If J is strictly increasing and F is strictly increasing in a neighborhood of θ, then λF (t) = 0 if and only if t = θ (exercise). For G ∈ F, define T(G) to be a solution of Z J G(x)+1−G(2T(G)−x) dG(x) = 0. (5.58) 2 T(Fn ) is called an R-estimator of T(F ) = θ. When J(t) = t − 12 (which is related to the Wilcoxon signed rank test), T(Fn ) is the well-known HodgesLehmann estimator and is equal to any value between the two middle points of the values (Xi + Xj )/2, i = 1, ..., n, j = 1, ..., n.

Theorem 5.8. Let T be the functional defined by (5.58). Suppose that F is continuous and symmetric about θ, the derivatives F ′ and J ′ exist, and J ′ is bounded. Then T is ̺∞ -Hadamard differentiable at F with the influence function φF (x) = R

J(F (x)) J ′ (F (x))F ′ (x)dF (x)

.

Proof. Since F is symmetric about θ, F (x) + RF (2θ − x) = 1. Under the assumed conditions, λF (t) is continuous and J ′ (F (x))F ′ (x)dF (x) = −λ′F (θ) 6= 0 (exercise). Hence, the inverse of λF exists and is continuous

350

5. Estimation in Nonparametric Models

at 0 = λF (θ). Suppose that tj → 0, ∆j ∈ D, k∆j − ∆k∞ → 0, and Gj = F + tj ∆j ∈ F. Then Z [J(Gj (x, t)) − J(F (x, t))]dGj (x) → 0 uniformly in t, where G(x, t) = [G(x) + 1 − G(2t − x)]/2, and Z Z J(F (x, t))d(Gj − F )(x) = (F − Gj )(x)J ′ (F (x, t))dF (x, t) → 0 uniformly in t. Let λG (t) be defined by (5.57) with F replaced by G. Then λGj (t) − λF (t) → 0 uniformly in t. Thus, λF (T(Gj )) → 0, which implies T(Gj ) → T(F ) = θ.

R

(5.59)

Let ξG (t) = J(F (x, t))dG(x), hF (t) = [λ R F (t) − λF (θ)]/(t − θ) if t 6= θ, and hF (θ) = λ′F (θ). Then T(Gj ) − T(F ) − φF (x)d(Gj − F )(x) is equal to λF (T(Gj )) − ξGj (θ) 1 1 − + . (5.60) ξGj (θ) ′ λF (θ) hF (T(Gj )) hF (T(Gj )) Note that ξGj (θ) =

Z

J(F (x))dGj (x) = tj

Z

J(F (x))d∆j (x).

By (5.59), Lemma 5.2, and k∆j − ∆k∞ → 0, the first term in (5.60) is o(tj ). The second term in (5.60) is the sum of Z tj − [J(F (x, T(Gj ))) − J(F (x))]d∆j (x) (5.61) hF (T(Gj )) and 1 hF (T(Gj ))

Z

[J(F (x, T(Gj ))) − J(Gj (x, T(Gj )))]dGj (x).

(5.62)

From the continuity of J and F , the quantity in (5.61) is o(tj ). Similarly, the quantity in (5.62) is equal to Z 1 [J(F (x, T(Gj ))) − J(Gj (x, T(Gj )))]dF (x) + o(tj ). (5.63) hF (T(Gj )) From Taylor’s expansion, (5.59), and k∆j − ∆k∞ → 0, the quantity in (5.63) is equal to Z tj J ′ (F (x))∆(x, θ)dF (x) + o(tj ). (5.64) hF (T(Gj ))

351

5.3. Linear Functions of Order Statistics

Since J(1 − t) = −J(t), the integral in (5.64) is 0. This proves that the second term in (5.60) is o(tj ) and thus the result. It is clear that the influence function φF for an R-estimator is bounded and continuous if J and F are continuous. Thus, R-estimators satisfy (5.41) and are robust in Hampel’s sense. Example 5.8. Let J(t) = t − 12 . Then T(Fn ) is the Hodges-Lehmann estiR mator. From Theorem 5.8, φF (x) = [F (x)− 21 ]/γ, where γ = F ′ (x)dF (x). Since F (X1 ) has a uniform √ distribution on [0, 1], φF (X1 ) has mean 0 and variance (12γ 2 )−1 . Thus, n[T(Fn ) − T(F )] →d N (0, (12γ 2)−1 ).

5.3 Linear Functions of Order Statistics In this section, we study statistics that are linear functions of order statistics X(1) ≤ · · · ≤ X(n) based on independent random variables X1 , ..., Xn (in §5.3.1 and §5.3.2, X1 , ..., Xn are assumed i.i.d.). Order statistics, first introduced in Example 2.9, are usually sufficient and often complete (or minimal sufficient) for nonparametric families (Examples 2.12 and 2.14). L-estimators defined in §5.2.2 are in fact linear functions of order statistics. If T is given by (5.46), then T(Fn ) =

Z

n

xJ(Fn (x))dFn (x) =

1X J n i=1

i n

X(i) ,

(5.65)

since Fn (X(i) ) = i/n, i = 1, ..., n. If J is a smooth function, such as those given in Example 5.6 or those satisfying the conditions in Theorem 5.6, the corresponding L-estimator is often called a smooth L-estimator. Asymptotic properties of smooth L-estimators can be obtained using Theorem 5.6 and the results in §5.2.1. Results on L-estimators that are slightly different from that in (5.65) can be found in Serfling (1980, Chapter 8). In §5.3.1, we consider another useful class of linear functions of order statistics, the sample quantiles described in the beginning of §5.2. In §5.3.2, we study robust linear functions of order statistics (in Hampel’s sense) ¯ an efficient but and their relative efficiencies w.r.t. the sample mean X, nonrobust estimator. In §5.3.3, extensions to linear models are discussed.

5.3.1 Sample quantiles Recall that G−1 (p) is defined to be inf{x : G(x) ≥ p} for any c.d.f. G on R, where p ∈ (0, 1) is a fixed constant. For i.i.d. X1 , ..., Xn from F , let θp = F −1 (p) and θˆp = Fn−1 (p) denote the pth quantile of F and the pth

352

5. Estimation in Nonparametric Models

sample quantile, respectively. Then θˆp = cnp X(mp ) + (1 − cnp )X(mp +1) ,

(5.66)

where mp is the integer part of np, cnp = 1 if np is an integer, and cnp = 0 if np is not an integer. Thus, θˆp is a linear function of order statistics. Note that F (θp −) ≤ p ≤ F (θp ) and F is not flat in a neighborhood of θp if and only if p < F (θp + ǫ) for any ǫ > 0. Theorem 5.9. Let X1 , ..., Xn be i.i.d. random variables from a c.d.f. F satisfying p < F (θp +ǫ) for any ǫ > 0. Then, for every ǫ > 0 and n = 1, 2,..., 2 (5.67) P |θˆp − θp | > ǫ ≤ 2Ce−2nδǫ ,

where δǫ is the smaller of F (θp + ǫ) − p and p − F (θp − ǫ) and C is the same constant in Lemma 5.1(i). Proof. Let ǫ > 0 be fixed. Note that G(x) ≥ t if and only if x ≥ G−1 (t) for any c.d.f. G on R (exercise). Hence P θˆp > θp + ǫ = P p > Fn (θp + ǫ) = P F (θp + ǫ) − Fn (θp + ǫ) > F (θp + ǫ) − p ≤ P ̺∞ (Fn , F ) > δǫ 2

≤ Ce−2nδǫ ,

where the last inequality follows from DKW’s inequality (Lemma 5.1(i)). Similarly, 2 P θˆp < θp − ǫ ≤ Ce−2nδǫ . This proves (5.67).

Result (5.67) implies that θˆp is strongly consistent for θp (exercise) and √ that θˆp is n-consistent for θp if F ′ (θp −) and F ′ (θp +) (the left and right derivatives of F at θp ) exist (exercise). The exact distribution of θˆp can be obtained as follows. Since nFn (t) has the binomial distribution Bi(F (t), n) for any t ∈ R, P θˆp ≤ t = P Fn (t) ≥ p n X n [F (t)]i [1 − F (t)]n−i , (5.68) = i i=lp

where lp = np if np is an integer and lp = 1+ the integer part of np if np is not an integer. If F has a Lebesgue p.d.f. f , then θˆp has the Lebesgue p.d.f. n−1 ϕn (t) = n [F (t)]lp −1 [1 − F (t)]n−lp f (t). (5.69) lp − 1

353

5.3. Linear Functions of Order Statistics

The following result provides an asymptotic distribution for

√ ˆ n(θp −θp ).

Theorem 5.10. Let X1 , ..., √ Xn be i.i.d. random variables from F . (i) If F (θp ) = p, then P ( n(θˆp − θp ) ≤ 0) → Φ(0) = 12 , where Φ is the c.d.f. of the standard normal. (ii) If F is continuous at θp and there exists F ′ (θp −) > 0, then P

√ n(θˆp − θp ) ≤ t → Φ(t/σ − ), F

t < 0,

p where σF− = p(1 − p)/F ′ (θp −). (iii) If F is continuous at θp and there exists F ′ (θp +) > 0, then P

√ n(θˆp − θp ) ≤ t → Φ(t/σF+ ),

t > 0,

p where σF+ = p(1 − p)/F ′ (θp +). (iv) If F ′ (θp ) exists and is positive, then

√ n(θˆp − θp ) →d N (0, σF2 ),

(5.70) p where σF = p(1 − p)/F ′ (θp ). Proof. The proof of (i) is left as an exercise. Part (iv) is a direct consequence of (i)-(iii) and the proofs of (ii) and (iii) are similar. Thus, we only give a proof for (iii). p √ Let t > 0, pnt = F (θp +p tσF+ n−1/2 ), cnt = n(pnt − p)/ pnt (1 − pnt ), and Znt = [Bn (pnt )−npnt ]/ npnt (1 − pnt ), where Bn (q) denotes a random variable having the binomial distribution Bi(q, n). Then P θˆp ≤ θp + tσF+ n−1/2 = P p ≤ Fn (θp + tσF+ n−1/2 ) = P Znt ≥ −cnt .

Under the assumed conditions on F , pnt → p and cnt → t. Hence, the result follows from P Znt < −cnt − Φ(−cnt ) → 0.

But this follows from the CLT (Example 1.33) and P´olya’s theorem (Proposition 1.16). If both F ′ (θp −) and F ′ (θp +) exist and are positive, but F ′ (θp −) 6= √ F (θp +), then the asymptotic distribution of n(θˆp − θp ) has the c.d.f. Φ(t/σF− )I(−∞,0) (t) + Φ(t/σF+ )I[0,∞) (t), a mixture of two normal distributions. An example of such a case when p = 12 is ′

F (x) = xI[0, 21 ) (x) + (2x − 12 )I[ 12 , 34 ) (x) + I[ 34 ,∞) (x).

354

5. Estimation in Nonparametric Models

When F ′ (θp −) = F ′ (θp +) = F ′ (θp ) > 0, (5.70) shows that the asymptotic √ √ distribution of n(θˆp −θp ) is the same as that of n[Fn (θp )−F (θp )]/F ′ (θp ) (see (5.2)). The following result reveals a stronger relationship between sample quantiles and the empirical c.d.f. Theorem 5.11 (Bahadur’s representation). Let X1 , ..., Xn be i.i.d. random variables from F . Suppose that F ′ (θp ) exists and is positive. Then F (θp ) − Fn (θp ) √1 + o θˆp = θp + . p n F ′ (θp )

(5.71)

Zn (t) − Zn (0) = op (1).

(5.72)

√ Proof. Let t ∈ R, θnt = θp +tn−1/2 , Zn (t) = n[F (θnt )−Fn (θnt )]/F ′ (θp ), √ and Un (t) = n[F (θnt ) − Fn (θˆp )]/F ′ (θp ). It can be shown (exercise) that

Note that |p − Fn (θˆp )| ≤ n−1 . Then Un (t) = =

√

√

→ t.

n[F (θnt ) − p + p − Fn (θˆp )]/F ′ (θp ) n[F (θnt ) − p]/F ′ (θp ) + O(n−1/2 )

(5.73)

√ ˆ n(θp − θp ). Then, for any t ∈ R and ǫ > 0, P ξn ≤ t, Zn (0) ≥ t + ǫ = P Zn (t) ≤ Un (t), Zn (0) ≥ t + ǫ (5.74) ≤ P |Zn (t) − Zn (0)| ≥ ǫ/2 + P |Un (t) − t| ≥ ǫ/2

Let ξn =

→0

by (5.72) and (5.73). Similarly, P ξn ≥ t + ǫ, Zn (0) ≤ t → 0.

(5.75)

It follows from the result in Exercise 128 of §1.6 that ξn − Zn (0) = op (1), which is the same as (5.71). If F has a positive Lebesgue p.d.f., then θˆp viewed as a statistical functional (§5.2) is ̺∞ -Hadamard differentiable at F (Fernholz, 1983) with the influence function φF (x) = [F (θp ) − I(−∞,θp ] (x)]/F ′ (θp ).

355

5.3. Linear Functions of Order Statistics

This implies result (5.71). Note that φF is bounded and is continuous except when x = θp . Corollary 5.1. Let X1 , ..., Xn be i.i.d. random variables from F having positive derivatives at θpj , where 0 < p1 < · · · < pm < 1 are fixed constants. Then √ n[(θˆp1 , ..., θˆpm ) − (θp1 , ..., θpm )] →d Nm (0, D), where D is the m × m symmetric matrix whose (i, j)th element is pi (1 − pj )/[F ′ (θpi )F ′ (θpj )],

i ≤ j.

The proof of this corollary is left to the reader. Example 5.9 (Interquartile range). One application of Corollary 5.1 is the derivation of the asymptotic distribution of the interquartile range θˆ0.75 − θˆ0.25 . The interquartile range is used as a measure of the variability among Xi ’s. It can be shown (exercise) that √ n[(θˆ0.75 − θˆ0.25 ) − (θ0.75 − θ0.25 )] →d N (0, σF2 ) with σF2 =

3 3 1 . + − 16[F ′ (θ0.75 )]2 16[F ′ (θ0.25 )]2 8F ′ (θ0.75 )F ′ (θ0.25 )

There are some applications of using extreme order statistics such as X(1) and X(n) . One example is given in Example 2.34. Some other examples and references can be found in Serfling (1980, pp. 89-91).

5.3.2 Robustness and efficiency Let F be a c.d.f. on R symmetric about θ ∈ R with F ′ (θ) > 0. Then θ = θ0.5 and is called the median of F . If F has a finite mean, then θ is also equal to the mean. In this section, we consider the estimation of θ based on i.i.d. Xi ’s from F . If F is normal, it has been shown in previous chapters that the sample ¯ is the UMVUE, MRIE, and MLE of θ, and is asymptotically mean X efficient. On the other hand, if F is the c.d.f. of the Cauchy distribution ¯ has the same distribution C(θ, 1), it follows from Exercise 78 in §1.6 that X ¯ as X1 , i.e., X is as variable as X1 , and is inconsistent as an estimator of θ. ¯ perform so differently? An important difference between Why does X the normal and Cauchy p.d.f.’s is that the former tends to 0 at the rate

356

5. Estimation in Nonparametric Models 2

the latter tends to 0 at the much slower rate e−x /2 as |x| → ∞, whereas R ¯ in the x−2 , which results in |x|dF (x) = ∞. The poor performance of X Cauchy case is due to the high probability of getting extreme observations ¯ is sensitive to large changes in a few of the Xi ’s. (Note and the fact that X R ¯ that X is not robust in Hampel’s sense, since the functional xdG(x) has an unbounded influence function at F .) This suggests the use of a robust estimator that discards some extreme observations. The sample median, which is defined to be the 50%th sample quantile θˆ0.5 described in §5.3.1, is insensitive to the behavior of F as |x| → ∞. Since both the sample mean and the sample median can be used to estimate θ, a natural question is when is one better than the other, using a criterion such as the amse. Unfortunately, a general answer does not exist, since the asymptotic relative efficiency between these two estimators depends on the unknown distribution F . If F does not have a finite vari¯ = ∞ and X ¯ may be inconsistent. In such a case the ance, then Var(X) sample median is certainly preferred, since θˆ0.5 is consistent and asymptotically normal as long as F ′ (θ) > 0, and may have a finite variance (Exercise 60). The following example, which compares the sample mean and median in some cases, shows that the sample median can be better even if Var(X1 ) < ∞. Example 5.10. Suppose that Var(X1 ) < ∞. Then, by the CLT, √ ¯ − θ) →d N (0, Var(X1 )). n(X By Theorem 5.10(iv), √ ˆ n(θ0.5 − θ) →d N (0, [2F ′ (θ)]−2 ). ¯ is Hence, the asymptotic relative efficiency of θˆ0.5 w.r.t. X e(F ) = 4[F ′ (θ)]2 Var(X1 ). √ (i) If F is the c.d.f. of N (θ, σ 2 ), then Var(X1 ) = σ 2 , F ′ (θ) = ( 2πσ)−1 , and e(F ) = 2/π = 0.637. (ii) If F is the c.d.f. of the logistic distribution LG(θ, σ), then Var(X1 ) = σ 2 π 2 /3, F ′ (θ) = (4σ)−1 , and e(F ) = π 2 /12 = 0.822. (iii) If F (x) = F0 (x − θ) and F0 is the c.d.f. of the√t-distribution tν with ν ν ≥ 3, then Var(X1 ) = ν/(ν − 2), F ′ (θ) = Γ( ν+1 2 )/[ νπΓ( 2 )], e(F ) = 1.62 when ν = 3, e(F ) = 1.12 when ν = 4, and e(F ) = 0.96 when ν = 5. (iv) If F is the c.d.f. of the double exponential distribution DE(θ, σ), then F ′ (θ) = (2σ)−1 and e(F ) = 2. (v) Consider the Tukey model F (x) = (1 − ǫ)Φ x−θ + ǫΦ x−θ (5.76) σ τσ ,

357

5.3. Linear Functions of Order Statistics

2 2 2 where σ > 0, τ > 0, and √ 0 < ǫ < 1. Then Var(X1 ) = 2(1 − ǫ)σ + ǫτ 2 σ , ′ F (θ) = (1 − ǫ + ǫ/τ )/( 2πσ), and e(F ) = 2(1 − ǫ + ǫτ )(1 − ǫ + ǫ/τ ) /π. Note that limǫ→0 e(F ) = 2/π and limτ →∞ e(F ) = ∞.

Since the sample median uses at most two actual values of xi ’s, it may go too far in discarding observations, which results in a possible loss of efficiency. The trimmed sample mean introduced in Example 5.6(iii) is a natural compromise between the sample mean and median. Since F is symmetric, we consider β = 1 − α in the trimmed mean, which results in the following L-estimator: ¯α = X

n−m Xα 1 X(j) , (1 − 2α)n j=m +1

(5.77)

α

where mα is the integer part of nα and α ∈ (0, 12 ). The estimator in (5.77) is called the α-trimmed sample mean. It discards the mα smallest and mα largest observations. The sample mean and median can be viewed as two ¯ α as α → 0 and 1 , respectively. extreme cases of X 2 It follows from Theorem 5.6 that if F (x) = F0 (x − θ), where F0 is symmetric about 0 and has a Lebesgue p.d.f. positive in the range of X1 , then √ ¯ α − θ) →d N (0, σ 2 ), n(X (5.78) α where

σα2

2 = (1 − 2α)2

(Z

0

F0−1 (1−α)

2

x dF0 (x) +

α[F0−1 (1

2

− α)]

)

.

Lehmann (1983, §5.4) provides various values of the asymptotic relative efficiency eX¯ α ,X¯ (F ) = Var(X1 )/σα2 . For instance, when F (x) = F0 (x − θ) and F0 is the c.d.f. of the t-distribution t3 , eX¯ α ,X¯ (F ) = 1.70, 1.91, and 1.97 for α = 0.05, 0.125, and 0.25, respectively; when F is given by (5.76) with τ = 3 and ǫ = 0.05, eX¯ α ,X¯ (F ) = 1.20, 1.19, and 1.09 for α = 0.05, 0.125, and 0.25, respectively; when F is given by (5.76) with τ = 3 and ǫ = 0.01, eX¯ α ,X¯ (F ) = 1.04, 0.98, and 0.89 for α = 0.05, 0.125, and 0.25, respectively. Robustness and efficiency of other L-estimators can be discussed similarly. For an L-estimator T(Fn ) with T given by (5.46), if the conditions in one of (i)-(iii) of Theorem 5.6 are satisfied, then (5.41) holds with Z ∞Z ∞ σF2 = J(F (x))J(F (y))[F (min{x, y}) − F (x)F (y)]dxdy, (5.79) −∞

−∞

provided that σF2 < ∞ (exercise). If F is symmetric about θ, J is symmetric R1 about 21 , and 0 J(t)dt = 1, then T(F ) = θ (exercise) and, therefore, the ¯ is Var(X1 )/σ 2 . asymptotic relative efficiency of T(Fn ) w.r.t. X F

358

5. Estimation in Nonparametric Models

5.3.3 L-estimators in linear models In this section, we extend L-estimators to the following linear model: Xi = β τ Z i + ε i ,

i = 1, ..., n,

(5.80)

with i.i.d. εi ’s having an unknown c.d.f. F0 and a full rank Z whose ith row is the vector Zi . Note that the c.d.f. of Xi is F0 (x − β τ Zi ). Instead of assuming E(εi ) = 0 (as we did in Chapter 3), we assume that Z (5.81) xJ(F0 (x))dF0 (x) = 0, where J is a Borel function on [0, 1] (the same as that in (5.46)). Note that (5.81) may hold without any assumption on the existence of E(εi ). For instance, (5.81) holds if F0 is symmetric about 0, J is symmetric about 21 , R1 and 0 J(t)dt = 1 (Exercise 69). Since Xi ’s are not identically distributed, the use of the order statistics and the empirical c.d.f. based on X1 , ..., Xn may not be appropriate. Inˆ i = 1, ..., n, stead, we consider the ordered values of residuals ri = Xi −Ziτ β, τ ˆ and some empirical c.d.f.’s based on residuals, where β = (Z Z)−1 Z τ X is the LSE of β (§3.3.1). To illustrate the idea, let us start with the case where β and Zi are univariate. First, assume thatP Zi ≥ 0 for all i (or Zi ≤ 0 for all i). Let Fˆ0 be the c.d.f. putting mass Zi / ni=1 Zi at ri , i = 1, ..., n. An L-estimator of β is defined to be X Z n n X ˆ ˆ ˆ ˆ βL = β + xJ(F0 (x))dF0 (x) Zi Zi2 . i=1

i=1

When J(t) = (1 − 2α)−1 I(α,1−α) (t) with an α ∈ (0, 21 ), βˆL is similar to the α-trimmed sample mean in the i.i.d. case. If not all Zi ’s have the same sign, we can define L-estimators as follows. Let Zi+ = P max{Zi , 0} and Zi− = Zi+ − Zi . Let Fˆ0± be the c.d.f. putting ± mass Zi / ni=1 Zi± at ri , i = 1, ..., n. An L-estimator of β is defined to be βˆL = βˆ + −

Z

Z

xJ(Fˆ0+ (x))dFˆ0+ (x)

xJ(Fˆ0− (x))dFˆ0− (x)

n X

Zi+

i=1

n X i=1

Zi−

X n

Zi2

i=1

X n

Zi2 .

i=1

For a general p-vector Zi , let zij be the jth component of Zi , j = 1, ..., p. + − + ± = max{zij , 0}, zij = zij − zij , and Fˆ0j be the c.d.f. putting mass Let zij

359

5.4. Generalized Estimating Equations

± Pn ± / i=1 zij at ri , i = 1, ..., n. For any j, if zij ≥ 0 for all i (or zij ≤ 0 for zij + − all i), then we set Fˆ0j ≡ 0 (or Fˆ0j ≡ 0). An L-estimator of β is defined to be βˆL = βˆ + (Z τ Z)−1 (A+ − A− ), (5.82)

where

A± =

Z

± ± xJ(Fˆ01 (x))dFˆ01 (x)

n X i=1

± zi1 , ...,

Z

± ± xJ(Fˆ0p (x))dFˆ0p (x)

n X i=1

± zip

!

.

Obviously, βˆL in (5.82) reduces to the previously defined βˆL when β and Zi are univariate. Theorem 5.12. Assume model (5.80) with i.i.d. εi ’s from a c.d.f. F0 satisfying (5.81) for a given J. Suppose that F0 has a uniformly continuous, positive, and bounded derivative on the range of ε1 . Suppose further that the conditions on Zi ’s in Theorem 3.12 are satisfied. (i) If the function J is continuous on (α1 , α2 ) and equals 0 on [0, α1 ]∪[α2 , 1], where 0 < α1 < α2 < 1 are constants, then σF−1 (Z τ Z)1/2 (βˆL − β) →d Np (0, Ip ), 0

(5.83)

where σF2 0 is given by (5.79) with F = F0 . (ii) Result (5.83) also holds if J ′ is bounded on [0, 1], E|ε1 | < ∞, and σF2 0 is finite. The proof of this theorem can be found in Bickel (1973). Robustness and efficiency comparisons between the LSE βˆ and L-estimators βˆL can be made in a way similar to those in §5.3.2.

5.4 Generalized Estimating Equations The method of generalized estimating equations (GEE) is a powerful and general method of deriving point estimators, which includes many previously described methods as special cases. In §5.4.1, we begin with a description of this method and, to motivate the idea, we discuss its relationship with other methods that have been studied. Consistency and asymptotic normality of estimators derived from generalized estimating equations are studied in §5.4.2 and §5.4.3. Throughout this section, we assume that X1 , ..., Xn are independent (not necessarily identically distributed) random vectors, where the dimension of Xi is di , i = 1, ..., n (supi di < ∞), and that we are interested in estimating θ, a k-vector of unknown parameters related to the unknown population.

360

5. Estimation in Nonparametric Models

5.4.1 The GEE method and its relationship with others The sample mean and, more generally, the LSE in linear models are solutions of equations of the form n X i=1

(Xi − γ τ Zi )Zi = 0.

Also, MLE’s (or RLE’s) in §4.4 and, more generally, M-estimators in §5.2.2 are solutions to equations of the form n X

ψ(Xi , γ) = 0.

n X

ψi (Xi , γ),

i=1

This leads to the following general estimation method. Let Θ ⊂ Rk be the range of θ, ψi be a Borel function from Rdi × Θ to Rk , i = 1, ..., n, and sn (γ) =

i=1

γ ∈ Θ.

(5.84)

ˆ = 0, then θˆ is called a GEE If θ is estimated by θˆ ∈ Θ satisfying sn (θ) estimator. The equation sn (γ) = 0 is called a GEE. Apparently, the LSE’s, RLE’s, MQLE’s, and M-estimators are special cases of GEE estimators. Usually GEE’s are chosen so that E[sn (θ)] =

n X

E[ψi (Xi , θ)] = 0,

(5.85)

i=1

where the expectation E may be replaced by an asymptotic expectation defined in §2.5.2 if the exact expectation does not exist. If this is true, ˆ = 0 is a sample analogue of then θˆ is motivated by the fact that sn (θ) E[sn (θ)] = 0. To motivate the idea, let us study the relationship between the GEE method and other methods that have been introduced. M-estimators The M-estimators defined in §5.2.2 for univariate θ = T(F ) in the i.i.d. case are special cases of GEE estimators. Huber (1981) also considers regression M-estimators in the linear model (5.80). A regression M-estimator of β is defined as a solution to the GEE n X ψ(Xi − γ τ Zi )Zi = 0, i=1

where ψ is one of the functions given in Example 5.7.

361

5.4. Generalized Estimating Equations

LSE’s in linear and nonlinear regression models Suppose that Xi = f (Zi , θ) + εi ,

i = 1, ..., n,

(5.86)

where Zi ’s are the same as those in (5.80), θ is an unknown k-vector of parameters, f is a known function, and εi ’s are independent random variables. Model (5.86) is the same as model (5.80) if f is linear in θ and is called a nonlinear regression model otherwise. Note that model (4.64) is a special case of P model (5.86). The LSE under model (5.86) is any point in n Θ minimizing i=1 [Xi − f (Zi , γ)]2 over γ ∈ Θ. If f is differentiable, then the LSE is a solution to the GEE n X i=1

[Xi − f (Zi , γ)]

∂f (Zi , γ) = 0. ∂γ

Quasi-likelihoods This is a continuation of the discussion of the quasi-likelihoods introduced in §4.4.3. Assume first that Xi ’s are univariate (di ≡ 1). If Xi ’s follow a GLM, i.e., Xi has the p.d.f. in (4.55) and (4.57) holds, and if (4.58) holds, then the likelihood equation (4.59) can be written as n X xi − µi (γ) i=1

vi (γ)

Gi (γ) = 0,

(5.87)

where µi (γ) = µ(ψ(γ τ Zi )), Gi (γ) = ∂µi (γ)/∂γ, vi (γ) = Var(Xi )/φ, and we have used the following fact: ψ ′ (t) = (µ−1 )′ (g −1 (t))(g −1 )′ (t) = (g −1 )′ (t)/ζ ′′ (ψ(t)). Equation (5.87) is a quasi-likelihood equation if either Xi does not have the p.d.f. in (4.55) or (4.58) does not hold. Note that this generalizes the discussion in §4.4.3. If Xi does not have the p.d.f. in (4.55), then the problem is often nonparametric. Let sn (γ) be the left-hand side of (5.87). Then sn (γ) = 0 is a GEE and E[sn (β)] = 0 is satisfied as long as the first condition in (4.56), E(Xi ) = µi (β), is satisfied. For general di ’s, let Xi = (Xi1 , ..., Xidi ), i = 1, ..., n, where each Xit satisfies (4.56) and (4.57), i.e., E(Xit ) = µ(ηit ) = g −1 (β τ Zit ) and Var(Xit ) = φi µ′ (ηit ), and Zit ’s are k-vector values of covariates. In biostatistics and life-time testing problems, components of Xi are repeated measurements at different times from subject i and are called longitudinal data. Although Xi ’s are

362

5. Estimation in Nonparametric Models

assumed independent, Xit ’s are likely to be dependent for each i. Let Ri be the di × di correlation matrix whose (t, l)th element is the correlation coefficient between Xit and Xil . Then Var(Xi ) = φi [Di (β)]1/2 Ri [Di (β)]1/2 ,

(5.88)

where Di (γ) is the di × di diagonal matrix with the tth diagonal element (g −1 )′ (γ τ Zit ). If Ri ’s in (5.88) are known, then an extension of (5.87) to the multivariate xi ’s is n X i=1

Gi (γ){[Di (γ)]1/2 Ri [Di (γ)]1/2 }−1 [xi − µi (γ)] = 0,

(5.89)

where µi (γ) = (µ(ψ(γ τ Zi1 )), ..., µ(ψ(γ τ Zidi ))) and Gi (γ) = ∂µi (γ)/∂γ. In ˜ i be a most applications, Ri is unknown and its form is hard to model. Let R known correlation matrix (called a working correlation matrix). Replacing ˜ i leads to the quasi-likelihood equation Ri in (5.89) by R n X i=1

˜ i [Di (γ)]1/2 }−1 [xi − µi (γ)] = 0. Gi (γ){[Di (γ)]1/2 R

(5.90)

For example, we may assume that the components of Xi are independent ˜ i = Idi . Although the working correlation matrix R ˜ i may not be and take R the same as the true unknown correlation matrix Ri , an MQLE obtained from (5.90) is still consistent and asymptotically normal (§5.4.2 and §5.4.3). ˜ i is closer to Ri . Of course, MQLE’s are asymptotically more efficient if R ˜ Even if Ri = Ri and φi ≡ φ, (5.90) is still a quasi-likelihood equation, since the covariance matrix of Xi cannot determine the distribution of Xi unless Xi is normal. ˜ i closer to Ri results in a better MQLE, sometimes it is Since an R ˜ i in (5.90) by R ˆ i , an estimator of Ri (Liang and suggested to replace R Zeger, 1986). The resulting equation is called a pseudo-likelihood p equation. ˆ i − Ui k →p 0 as n → ∞, where kAk = tr(Aτ A) for As long as maxi≤n kR a matrix A and Ui is a correlation matrix (not necessarily the same as Ri ), i = 1, ..., n, MQLE’s are consistent and asymptotically normal. Empirical likelihoods The previous discussion shows that the GEE method coincides with the method of deriving M-estimators, LSE’s, MLE’s, or MQLE’s. The following discussion indicates that the GEE method is also closely related to the method of empirical likelihoods introduced in §5.1.4. Assume that Xi ’s are i.i.d. from a c.d.f. F on Rd and ψi = ψ for all i. Then condition (5.85) reduces to E[ψ(X1 , θ)] = 0. Hence, we can consider

363

5.4. Generalized Estimating Equations

the empirical likelihood ℓ(G) =

n Y

i=1

subject to pi ≥ 0,

n X

PG ({xi }),

pi = 1,

i=1

and

G∈F n X

pi ψ(xi , θ) = 0,

(5.91)

i=1

where pi = PG ({xi }). However, in this case the dimension of the function ψ is the same as the dimension of the parameter θ and, hence, the last equation in (5.91) does not impose any restriction on pi ’s. Then, it follows from Theorem 5.3 that (p1 , ..., pn ) = (n−1 , ..., n−1 ) maximizes ℓ(G) for any fixed θ. Substituting pi = n−1 into the last equation in (5.91) leads to n

1X ψ(xi , θ) = 0. n i=1

That is, any MELE θˆ of θ is a GEE estimator.

5.4.2 Consistency of GEE estimators We now study under what conditions (besides (5.85)) GEE estimators are consistent. For each n, let θˆn be a GEE estimator, i.e., sn (θˆn ) = 0, where sn (γ) is defined by (5.84). First, Theorem 5.7 and its proof can be extended to multivariate T in a straightforward manner. Hence, we have the following result. Proposition 5.2. Suppose that X1 , ..., Xn are i.i.d. from F and ψi ≡ d k Rψ, a bounded and continuous function from R × Θ to R . Let Ψ(t) = ψ(x, t)dF (x). Suppose that Ψ(θ) = 0 and ∂Ψ(t)/∂t exists and is of full rank at t = θ. Then θˆn →p θ. For unbounded ψ in the i.i.d. case, the following result and its proof can be found in Qin and Lawless (1994). Proposition 5.3. Suppose that X1 , ..., Xn are i.i.d. from F and ψi ≡ ψ. Assume that ϕ(x, γ) = ∂ψ(x, γ)/∂γ exists in Nθ , a neighborhood of θ, and is continuous at θ; there is a function h(x) such that supγ∈Nθ kϕ(x, γ)k ≤ h(x), supγ∈Nθ kψ(x, γ)k3 ≤ h(x), and E[h(X1 )] < ∞; E[ϕ(X1 , θ)] is of full rank; E{ψ(X1 , θ)[ψ(X1 , θ)]τ } is positive definite; and (5.85) holds. Then, there exists a sequence of random vectors {θˆn } such that and θˆn →p θ. (5.92) P sn (θˆn ) = 0 → 1

364

5. Estimation in Nonparametric Models

Next, we consider non-i.i.d. Xi ’s. Proposition 5.4. Suppose that X1 , ..., Xn are independent and θ is univariate. Assume that ψi (x, γ) is real-valued and nonincreasing in γ for all i; there is a δ > 0 such that supi E|ψi (Xi , γ)|1+δ < ∞ for any γ in Nθ , a neighborhood of θ (this condition can be replaced by E|ψ(X1 , γ)| < ∞ for any γ in Nθ when Xi ’s are i.i.d. and ψi ≡ ψ); ψi (x, γ) are continuous in Nθ ; (5.85) holds; and lim sup E[Ψn (θ + ǫ)] < 0 < lim inf E[Ψn (θ − ǫ)] n

n

(5.93)

for any ǫ > 0, where Ψn (γ) = n−1 sn (γ). Then, there exists a sequence of random variables {θˆn } such that (5.92) holds. Furthermore, any sequence {θˆn } satisfying sn (θˆn ) = 0 satisfies (5.92). Proof. Since ψi ’s are nonincreasing, the functions Ψn (γ) and E[Ψn (γ)] are nonincreasing. Let ǫ > 0 be fixed so that θ ± ǫ ∈ Nθ . Under the assumed conditions, Ψn (θ ± ǫ) − E[Ψn (θ ± ǫ)] →p 0 (Theorem 1.14(ii)). By condition (5.93), P Ψn (θ + ǫ) < 0 < Ψn (θ − ǫ) → 1.

The rest of the proof is left as an exercise.

To establish the next result, we need the following lemma. First, we need the following concept. A sequence of functions {gi } from Rk to Rk is called equicontinuous on an open set O ⊂ Rk if and only if, for any ǫ > 0, there is a δǫ > 0 such that supi kgi (t) − gi (s)k < ǫ whenever t ∈ O, s ∈ O, and kt − sk < δǫ . Since a continuous function on a compact set is uniformly continuous, functions such as gi (γ) = g(ti , γ) form an equicontinuous sequence on O if ti ’s vary in a compact set containing O and g(t, γ) is a continuous function in (t, γ). Lemma 5.3. Suppose that Θ is a compact subset of Rk . Let hi (Xi ) = supγ∈Θ kψi (Xi , γ)k, i = 1, 2,.... Suppose that supi E|hi (Xi )|1+δ < ∞ and supi EkXi kδ < ∞ for some δ > 0 (this condition can be replaced by E|h(X1 )| < ∞ when Xi ’s are i.i.d. and ψi ≡ ψ). Suppose further that for any c > 0 and sequence {xi } satisfying kxi k ≤ c, the sequence of functions {gi (γ) = ψi (xi , γ)} is equicontinuous on any open subset of Θ. Then

n

1 X

{ψi (Xi , γ) − E[ψi (Xi , γ)]} sup

→p 0. n γ∈Θ i=1

365

5.4. Generalized Estimating Equations

Proof. Since we only need to consider components of ψi ’s, without loss of generality we can assume that ψi ’s are functions from Rdi × Θ to R. For any c > 0, " n # 1X sup E hi (Xi )I(c,∞) (kXi k) ≤ sup E[hi (Xi )I(c,∞) (kXi k)]. n i=1 n i Let c0 = supi E|hi (Xi )|1+δ and c1 = supi EkXi kδ . By H¨older’s inequality, 1/(1+δ) δ/(1+δ) [P (kXi k > c)] E[hi (Xi )I(c,∞) (kXi k)] ≤ E|hi (Xi )|1+δ 1/(1+δ) δ/(1+δ) −δ 2 /(1+δ) c1 c

≤ c0

1/(1+δ) δ/(1+δ)

2

for all i. For ǫ > 0 and ǫ˜ > 0, choose a c such that c0 c1 c−δ /(1+δ) < ǫ˜ ǫ/4. Then, for any O ⊂ Θ, the probability ! ( ) n ǫ 1X (5.94) sup ψi (Xi , γ) − inf ψi (Xi , γ) I(c,∞) (kXi k) > P γ∈O n i=1 γ∈O 2 is bounded by ǫ˜ (exercise). From the equicontinuity of {ψi (xi , γ)}, there is a δǫ > 0 such that ( ) n 1X ǫ sup ψi (Xi , γ) − inf ψi (Xi , γ) I[0,c] (kXi k) < γ∈Oǫ n i=1 γ∈Oǫ 2 for sufficiently large n, where Oǫ denotes any open ball in Rk with radius less than δǫ . These results, together with Theorem 1.14(ii) and the fact that kψi (Xi , γ)k ≤ hi (Xi ), imply that ( ! ) n 1X sup ψi (Xi , γ) − E inf ψi (Xi , γ) > ǫ → 0. (5.95) P γ∈Oǫ n i=1 γ∈Oǫ Let Hn (γ) = n−1

Pn

i=1 {ψi (Xi , γ) n

1X sup Hn (γ) ≤ n i=1 γ∈Oǫ

(

− E[ψi (Xi , γ)]}. Then

sup ψi (Xi , γ) − E

γ∈Oǫ

) inf ψi (Xi , γ) ,

γ∈Oǫ

which with (5.95) implies that

P Hn (γ) > ǫ for all γ ∈ Oǫ = P

sup Hn (γ) > ǫ γ∈Oǫ

Similarly we can show that P Hn (γ) < −ǫ for all γ ∈ Oǫ → 0.

!

→ 0.

366

5. Estimation in Nonparametric Models

Since Θ is compact, there exists mǫ open balls Oǫ,j such that Θ ⊂ ∪Oǫ,j . Then, the result follows from ! X mǫ P sup |Hn (γ)| > ǫ ≤ P sup |Hn (γ)| > ǫ → 0. γ∈Θ

j=1

γ∈Oǫ,j

˜i} Example 5.11. Consider the quasi-likelihood equation (5.90). Let {R be a sequence of working correlation matrices and ˜ i [Di (γ)]1/2 }−1 [xi − µi (γ)]. ψi (xi , γ) = Gi (γ){[Di (γ)]1/2 R

(5.96)

It can be shown (exercise) that ψi ’s satisfy the conditions of Lemma 5.3 if Θ is compact and supi kZi k < ∞. Proposition 5.5. Assume (5.85) and the conditions in Lemma 5.3 (with Θ replaced by any compact subset of the parameter space). Suppose that the functions ∆n (γ) = E[n−1 sn (γ)] have the property that limn→∞ ∆n (γ) = 0 if and only if γ = θ. (If ∆n converges to a function ∆, then this condition and (5.85) imply that ∆ has a unique 0 at θ.) Suppose that {θˆn } is a sequence of GEE estimators and that θˆn = Op (1). Then θˆn →p θ. Proof. First, assume that Θ is a compact subset of Rk . From Lemma 5.3 and sn (θˆn ) = 0, ∆n (θˆn ) →p 0. By Theorem 1.8(vi), there is a subsequence {ni } such that (5.97) ∆ni (θˆni ) →a.s. 0.

Let x1 , x2 , ... be a fixed sequence such that (5.97) holds and let θ0 be a limit point of {θˆn }. Since Θ is compact, θ0 ∈ Θ and there is a subsequence {mj } ⊂ {ni } such that θˆmj → θ0 . Using the argument in the proof of Lemma 5.3, it can be shown (exercise) that {∆n (γ)} is equicontinuous on any open subset of Θ. Then ∆mj (θˆmj ) − ∆mj (θ0 ) → 0, which with (5.97) implies ∆mj (θ0 ) → 0. Under the assumed condition, θ0 = θ. Since this is true for any limit point of {θˆn }, θˆn →p θ. Next, consider a general Θ. For any ǫ > 0, there is an Mǫ > 0 such that P (kθˆn k ≤ Mǫ ) > 1 − ǫ. The result follows from the previous proof by considering the closure of Θ ∩ {γ : kγk ≤ Mǫ } as the parameter space.

Condition θˆn = Op (1) in Proposition 5.5 is obviously necessary for the consistency of θˆn . It has to be checked in any particular problem. If a GEE is a likelihood equation under some conditions, then we can often show, using an argument similar to the proof of Theorem 4.17 or 4.18, that there exists a consistent sequence of GEE estimators.

5.4. Generalized Estimating Equations

367

Proposition 5.6. Suppose that sn (γ) = ∂ log ℓn (γ)/∂γ for some function ℓn ; Dn (θ) = Var(sn (θ)) → 0; ϕi (x, γ) = ∂ψi (x, γ)/∂γ exists and the sequence of functions {ϕij , i = 1, 2, ...} satisfies the conditions in Lemma 5.3 with Θ replaced by a compact neighborhood of θ, where ϕij is the jth row of ϕi , j = 1, ..., k; − lim inf n [Dn (θ)]1/2 E[∇sn (θ)][Dn (θ)]1/2 is positive definite, where ∇sn (γ) = ∂sn (γ)/∂γ; and (5.85) holds. Then, there exists a sequence of estimators {θˆn } satisfying (5.92). The proof of Proposition 5.6 is similar to that of Theorem 4.17 or Theorem 4.18 and is left as an exercise. ˜i = Example 5.12. Consider the quasi-likelihood equation (5.90) with R Idi for all i. Then the GEE is a likelihood equation under a GLM (§4.4.2) assumption. It can be shown (exercise) that the conditions of Proposition 5.6 are satisfied if supi kZi k < ∞.

5.4.3 Asymptotic normality of GEE estimators Asymptotic normality of a consistent sequence of GEE estimators can be established under some conditions. We first consider the special case where θ is univariate and X1 , ..., Xn are i.i.d. Theorem 5.13. LetR X1 , ..., Xn be i.i.d. from F , ψi ≡ ψ, and θ ∈ R. Suppose that Ψ(γ) = ψ(x, γ)dF (x) = 0 if and only if γ = θ, Ψ′ (θ) exists and Ψ′ (θ) 6= 0. R (i) Assume that ψ(x, γ) is nonincreasing in γ and that [ψ(x, γ)]2 dF (x) is finite for γ in a neighborhood of θ and is continuous at θ. Then, any sequence of GEE estimators (M-estimators) {θˆn } satisfies √ n(θˆn − θ) →d N (0, σF2 ), (5.98) where σF2

=

Z

[ψ(x, θ)]2 dF (x)/[Ψ′ (θ)]2 .

R (ii) Assume that [ψ(x, θ)]2 dF (x) < ∞, ψ(x, γ) is continuous in x, and limγ→θ kψ(·, γ) − ψ(·, θ)kV = 0, where k · kV is the variation norm defined in Lemma 5.2. Then, any consistent sequence of GEE estimators {θˆn } satisfies (5.98). Proof. (i) Let Ψn (γ) = n−1 sn (γ). Since Ψn is nonincreasing, P (Ψn (t) < 0) ≤ P (θˆn ≤ t) ≤ P (Ψn (t) ≤ 0) for any t ∈ R. Then, (5.98) follows from lim P Ψn (tn ) < 0 = lim P Ψn (tn ) ≤ 0 = Φ(t) n→∞

n→∞

368

5. Estimation in Nonparametric Models

for all t ∈ R, where tn = θ + tσF n−1/2 . Let s2t,n = Var(ψ(X1 , tn )) and Yni = [ψ(Xi , tn ) − Ψ(tn )]/st,n . Then, it suffices to show that ! √ n nΨ(tn ) 1 X = Φ(t) lim P √ Yni ≤ − n→∞ st,n n i=1

√ for all t. Under the assumed conditions, nΨ(tn ) → Ψ′ (θ)tσF and st,n → −Ψ′ (θ)σF . Hence, it suffices to show that n

1 X √ Yni →d N (0, 1). n i=1 Note that Yn1 , ..., Ynn are i.i.d. random variables. Hence we can apply Lindeberg’s CLT (Theorem 1.15). In this case, Lindeberg’s condition (1.92) is implied by Z lim [ψ(x, tn )]2 dF (x) = 0 √ n→∞

|ψ(x,tn )|> nǫ

for any ǫ > 0. For any η > 0, ψ(x, θ + η) ≤ ψ(x, tn ) ≤ ψ(x, θ − η) for all x and sufficiently large n. Let u(x) = max{|ψ(x, θ − η)|, |ψ(x, θ + η)|}. Then Z Z 2 [ψ(x, t )] dF (x) ≤ [u(x)]2 dF (x), n √ √ |ψ(x,tn )|> nǫ

u(x)> nǫ

R

which converges to 0 since [ψ(x, γ)]2 dF (x) is finite for γ in a neighborhood of θ. This proves (i). (ii) Let φF (x) = −ψ(x, θ)/Ψ′ (θ). Following the proof of Theorem 5.7, we have n √ 1 X n(θˆn − θ) = √ φF (Xi ) + R1n − R2n , n i=1 where

" # n 1 1 X 1 − ψ(Xi , θ) , R1n = √ n i=1 Ψ′ (θ) hF (θˆn ) √ Z n [ψ(x, θˆn ) − ψ(x, θ)]d(Fn − F )(x), R2n = hF (θˆn )

and hF is defined in the proof of Theorem 5.7 with Ψ = λF . By the CLT and the consistency of θˆn , R1n = op (1). Hence, the result follows if we can show that R2n = op (1). By Lemma 5.2, √ |R2n | ≤ n|hF (θˆn )|−1 ̺∞ (Fn , F )kψ(·, θˆn ) − ψ(·, θ)kV . The result follows from the assumed condition on ψ and the fact that √ n̺∞ (Fn , F ) = Op (1) (Theorem 5.1).

369

5.4. Generalized Estimating Equations

Note that the result in Theorem 5.13 coincides with the result in Theorem 5.7 and (5.41). Example 5.13. Consider the M-estimators given in Example 5.7 based on i.i.d. random variables X1 , ..., Xn . If ψ is bounded and continuous, then Theorem 5.7 applies and (5.98) holds. For case (ii), ψ(x, γ) is not bounded but is nondecreasing in γ (−ψ(x, γ) is nonincreasing in γ). Hence Theorem 5.13 can be applied to this case. Consider Huber’s ψ given in Example 5.7(v). Assume that F is continuous at θ − C and θ + C. Then Z γ+C Ψ(γ) = (γ − x)dF (x) + CF (γ − C) − C[1 − F (γ + C)] γ−C

is differentiable at θ (exercise); Ψ(θ) = 0 if F is symmetric about θ (exercise); and Z

2

[ψ(x, γ)] dF (x) =

Z

γ+C

γ−C

(γ − x)2 dF (x) + C 2 F (γ − C) + C 2 [1 − F (γ + C)]

is continuous at θ (exercise). Therefore, (5.98) holds with σF2

=

R θ+C θ−C

(θ − x)2 dF (x) + C 2 F (θ − C) + C 2 [1 − F (θ + C)] [F (θ + C) − F (θ − C)]2

(exercise). Note that Huber’s M-estimator is robust in Hampel’s sense. ¯ can be obAsymptotic relative efficiency of θˆn w.r.t. the sample mean X tained (exercise). The next result is for general θ and independent Xi ’s. Theorem 5.14. Suppose that ϕi (x, γ) = ∂ψi (x, γ)/∂γ exists and the sequence of functions {ϕij , i = 1, 2, ...} satisfies the conditions in Lemma 5.3 with Θ replaced by a compact neighborhood of θ, where ϕij is the jth row of ϕi ; supi Ekψi (Xi , θ)k2+δ < ∞ for some δ > 0 (this condition can be replaced by Ekψ(X1 , θ)k2 < ∞ if Xi ’s are i.i.d. and ψi ≡ ψ); E[ψi (Xi , θ)] = 0; lim inf n λ− [n−1 Var(sn (θ))] > 0 and lim inf n λ− [n−1 Mn (θ)] > 0, where Mn (θ) = −E[∇sn (θ)] and λ− [A] is the smallest eigenvalue of the matrix A. If {θˆn } is a consistent sequence of GEE estimators, then Vn−1/2 (θˆn − θ) →d Nk (0, Ik ),

(5.99)

Vn = [Mn (θ)]−1 Var(sn (θ))[Mn (θ)]−1 .

(5.100)

where

370

5. Estimation in Nonparametric Models

Proof. The proof is similar to that of Theorem 4.17. By the consistency of θˆn , we can focus on the event {θˆn ∈ Aǫ }, where Aǫ = {γ : kγ − θk ≤ ǫ} with a given ǫ > 0. For sufficiently small ǫ, it can be shown (exercise) that max

γ∈Aǫ

k∇sn (γ) − ∇sn (θ)k = op (1), n

(5.101)

using an argument similar to the proof of Lemma 5.3. From the mean-value theorem and sn (θˆn ) = 0, Z 1 ∇sn θ + t(θˆn − θ) dt (θˆn − θ). −sn (θ) = 0

It follows from (5.101) that

Z 1

1 ˆ

= op (1). θ + t( θ ∇s − θ) dt − ∇s (θ) n n n

n 0 Also, by Theorem 1.14(ii),

n−1 k∇sn (θ) + Mn (θ)k = op (1). This and lim inf n λ− [n−1 Mn (θ)] > 0 imply [Mn (θ)]−1 sn (θ) = [1 + op (1)](θˆn − θ). The result follows if we can show that Vn−1/2 [Mn (θ)]−1 sn (θ) →d Nk (0, Ik ).

(5.102)

For any nonzero l ∈ Rk , n X 1 E|lτ [Mn (θ)]−1 ψi (Xi , θ)|2+δ → 0, (lτ Vn l)1+δ/2 i=1

(5.103)

since lim inf n λ− [n−1 Var(sn (θ))] > 0 and supi Ekψi (Xi , θ)k2+δ < ∞ (exercise). Applying the CLT (Theorem 1.15) with Liapounov’s condition (5.103), we obtain that p lτ [Mn (θ)]−1 sn (θ)/ lτ Vn l →d N (0, 1) (5.104) for any l, which implies (5.102) (exercise).

Asymptotic normality of GEE estimators can be established under various other conditions; see, for example, Serfling (1980, Chapter 7) and He and Shao (1996).

371

5.5. Variance Estimation

If Xi ’s are i.i.d. and ψi ≡ ψ, the asymptotic covariance matrix in (5.100) reduces to Vn = n−1 {E[ϕ(X1 , θ)]}−1 E{ψ(X1 , θ)[ψ(X1 , θ)]τ }{E[ϕ(X1 , θ)]}−1 , where ϕ(x, γ) = ∂ψ(x, γ)/∂γ. When θ is univariate, Vn further reduces to Vn = n−1 E[ψ(X1 , θ)]2 /{E[ϕ(X1 , θ)]}2 . Under the conditions of Theorem 5.14, Z Z ∂ ∂ψ(x, θ) dF (x) = ψ(x, θ)dF (x). E[ϕ(X1 , θ)] = ∂θ ∂θ Hence, the result in Theorem 5.14 coincides with that in Theorem 5.13. Example 5.14. Consider the quasi-likelihood equation in (5.90) and ψi in (5.96). If supi kZi k < ∞, then ψi satisfies the conditions in Theorem 5.14 ˜ i [Di (γ)]1/2 . Then (exercise). Let V˜n (γ) = [Di (γ)]1/2 R Var(sn (θ)) =

n X

Gi (θ)[V˜n (θ)]−1 Var(Xi )[V˜n (θ)]−1 [Gi (θ)]τ

i=1

and Mn (θ) =

n X

Gi (θ)[V˜n (θ)]−1 [Gi (θ)]τ .

i=1

˜ i = Ri (the true correlation matrix) for all i, then If R Var(sn (θ)) =

n X

φi Gi (θ)[V˜n (θ)]−1 [Gi (θ)]τ .

i=1

If, in addition, φi ≡ φ, then Vn = [Mn (θ)]−1 Var(sn (θ))[Mn (θ)]−1 = φ[Mn (θ)]−1 .

5.5 Variance Estimation In statistical inference the accuracy of a point estimator is usually assessed by its mse or amse. If the bias or asymptotic bias of an estimator is (asymptotically) negligible w.r.t. its mse or amse, then assessing the mse or amse is equivalent to assessing variance or asymptotic variance. Since variances and asymptotic variances usually depend on the unknown population, we have to estimate them in order to report accuracies of point estimators. Variance estimation is an important part of statistical inference, not only for

372

5. Estimation in Nonparametric Models

assessing accuracy, but also for constructing inference procedures studied in Chapters 6 and 7. See also the discussion at the end of §2.5.1. Let θ be a parameter of interest and θˆn be its estimator. Suppose that, as the sample size n → ∞, Vn−1/2 (θˆn − θ) →d Nk (0, Ik ),

(5.105)

where Vn is the covariance matrix or an asymptotic covariance matrix of θˆn . An essential asymptotic requirement in variance estimation is the consistency of variance estimators according to the following definition. See also (3.60) and Exercise 116 in §3.6. Definition 5.4. Let {Vn } be a sequence of k × k positive definite matrices and Vˆn be a positive definite matrix estimator of Vn for each n. Then {Vˆn } or Vˆn is said to be consistent for Vn (or strongly consistent for Vn ) if and only if kVn−1/2 Vˆn Vn−1/2 − Ik k →p 0 (5.106) (or (5.106) holds with →p replaced by →a.s. ). Note that (5.106) is different from kVˆn −Vn k →p 0, because kVn k → 0 in most applications. It can be shown (Exercise 93) that (5.106) holds if and only if lnτ Vˆn ln /lnτ Vn ln →p 1 for any sequence of nonzero vectors {ln } ⊂ Rk . If (5.105) and (5.106) hold, then Vˆn−1/2 (θˆn − θ) →d Nk (0, Ik ) (exercise), a result useful for asymptotic inference discussed in Chapters 6 and 7. If the unknown population is in a parametric family indexed by θ, then Vn is a function of θ, say Vn = Vn (θ), and it is natural to estimate Vn (θ) by Vn (θˆn ). Consistency of Vn (θˆn ) according to Definition 5.4 can usually be directly established. Thus, variance estimation in parametric problems is usually simple. In a nonparametric problem, Vn may depend on unknown quantities other than θ and, thus, variance estimation is much more complex. We introduce three commonly used variance estimation methods in this section, the substitution method, the jackknife, and the bootstrap.

5.5.1 The substitution method Suppose that we can obtain a formula for the covariance or asymptotic covariance matrix Vn in (5.105). Then a direct method of variance estimation is to substitute unknown quantities in the variance formula by some

373

5.5. Variance Estimation

estimators. To illustrate, consider the simplest case where X1 , ..., Xn are ¯ i.i.d. random d-vectors with EkX1 k2 < ∞, θ = g(µ), µ = EX1 , θˆn = g(X), d k and g is a function from R to R . Suppose that g is differentiable at µ. Then, by the CLT and Theorem 1.12(i), (5.105) holds with Vn = [∇g(µ)]τ Var(X1 )∇g(µ)/n,

(5.107)

which depends on unknown quantities µ and Var(X1 ). A substitution estimator of Vn is ¯ τ S 2 ∇g(X)/n, ¯ Vˆn = [∇g(X)] (5.108) where

n

S2 =

1 X ¯ ¯ τ (Xi − X)(X i − X) n − 1 i=1

is the sample covariance matrix, an extension of the sample variance to the multivariate Xi ’s. ¯ →a.s. µ and S 2 →a.s. Var(X1 ). Hence, Vˆn in (5.108) By the SLLN, X is strongly consistent for Vn in (5.107), provided that ∇g(µ) 6= 0 and ∇g is continuous at µ. Example 5.15. Let Y1 , ..., Yn be i.i.d. random variables with finite µy = EY1 , σy2 = Var(Y1 ), γy = EY13 , and κy = EY14 . Consider the estimation Pn ¯ σ of θ = (µy , σy2 ). Let θˆn = (X, ˆy2 ), where σ ˆy2 = n−1 i=1 (Yi − Y¯ )2 . If ¯ with g(x) = (x1 , x2 − x2 ). Hence, (5.105) Xi = (Yi , Yi2 ), then θˆn = g(X) 1 holds with γy − µy (σy2 + µ2y ) σy2 Var(X1 ) = γy − µy (σy2 + µ2y ) κy − (σy2 + µ2y )2 and ∇g(x) =

1 −2x1

0 1

.

The estimator Vˆn in (5.108) is strongly consistent, since ∇g(x) is obviously a continuous function. Similar results can be obtained for problems in Examples 3.21 and 3.23 and Exercises 100 and 101 in §3.6. A key step in the previous discussion is the derivation of formula (5.107) ¯ via Taylor’s expansion for the asymptotic covariance matrix of θˆn = g(X) (Theorem 1.12) and the CLT. Thus, the idea can be applied to the case where θˆn = T(Fn ), a differentiable statistical functional. We still consider i.i.d. random d-vectors X1 , ..., Xn from F . Suppose that T is a vector-valued functional whose components are ̺-Hadamard

374

5. Estimation in Nonparametric Models

differentiable at F , where ̺ is either ̺∞ or a distance satisfying (5.42). Let φF be the vector of influence functions of components of T. If the components of φF satisfy (5.40), then (5.105) holds with θ = T(F ), θˆn = T(Fn ), Fn = the empirical c.d.f. in (5.1), and Z 1 Var(φF (X1 )) = φF (x)[φF (x)]τ dF (x). Vn = (5.109) n n Formula (5.109) leads to a natural substitution variance estimator 1 Vˆn = n

Z

φFn (x)[φFn (x)]τ dFn (x) =

n 1 X φF (Xi )[φFn (Xi )]τ , (5.110) n2 i=1 n

provided that φFn (x) is well defined, i.e., the components of T are Gˆateaux differentiable at Fn for sufficiently large n. Under some more conditions on φFn we can establish the consistency of Vˆn in (5.110). Theorem 5.15. Let X1 , ..., Xn be i.i.d. random d-vectors from F , T be a vector-valued functional whose components are Gˆateaux differentiable at F and Fn , and φF be the vector of influence functions of components of T. Suppose that supkxk≤c kφFn (x) − φF (x)k = op (1) for any c > 0 and that there exist a constant c0 > 0 and a function h(x) ≥ 0 such that E[h(X1 )] < ∞ and P kφFn (x)k2 ≤ h(x) for all kxk ≥ c0 → 1. Then Vˆn in (5.110) is consistent for Vn in (5.109). Proof. Let ζ(x) = φF (x)[φF (x)]τ and ζn (x) = φFn (x)[φFn (x)]τ . By the SLLN, Z n 1X ζ(Xi ) →a.s. ζ(x)dF (x). n i=1

Hence the result follows from

n

1 X

[ζn (Xi ) − ζ(Xi )] = op (1).

n i=1

Using the assumed conditions and the argument in the proof of Lemma 5.3, we can show that for any ǫ > 0, there is a c > 0 such that ! n 1X ǫ P ≤ǫ kζn (Xi ) − ζ(Xi )kI(c,∞) (kXi k) > n i=1 2

and

n

P

1X ǫ kζn (Xi ) − ζ(Xi )kI[0,c] (kXi k) > n i=1 2

for sufficiently large n. This completes the proof.

!

≤ǫ

375

5.5. Variance Estimation

Example 5.16. Consider the L-functional defined in (5.46) and the Lestimator θˆn = T(Fn ). Theorem 5.6 shows that T is Hadamard differentiable at F under some conditions on J. It can be shown (exercise) that T is Gˆateaux differentiable at Fn with φFn (x) given by (5.48) (with F replaced by Fn ). Then the difference φFn (x) − φF (x) is equal to Z Z (Fn − F )(y)J(Fn (y))dy + (F − διx )(y)[J(Fn (y)) − J(F (y))]dy. One can show (exercise) that the conditions in Theorem 5.15 are satisfied if the conditions in Theorem 5.6(i) or (ii) (with E|X1 | < ∞) hold. Substitution variance estimators for M-estimators and U-statistics can also be derived (exercises). The substitution method can clearly be applied to non-i.i.d. cases. For example, the LSE βˆ in linear model (3.25) with a full rank Z and i.i.d. εi ’s ˆ = σ 2 (Z τ Z)−1 , where σ 2 = Var(ε1 ). A consistent substitution has Var(β) ˆ ˆ can be obtained by replacing σ 2 in the formula of Var(β) estimator of Var(β) by a consistent estimator of σ 2 such as SSR/(n − p) (see (3.35)). We now consider variance estimation for the GEE estimators described in §5.4.1. By Theorem 5.14, the asymptotic covariance matrix of the GEE estimator θˆn is given by (5.100), where Var(sn (θ)) =

n X i=1

Mn (θ) =

E{ψi (Xi , θ)[ψi (Xi , θ)]τ }, n X

E[ϕi (Xi , θ)],

i=1

and ϕi (x, γ) = ∂ψi (x, γ)/∂γ. Substituting θ by θˆn and the expectations by their empirical analogues, we obtain the substitution estimator Vˆn = ˆ −1 Var(s ˆ −1 , where d n )M M n n and

d n) = Var(s

n X

ψi (Xi , θˆn )[ψi (Xi , θˆn )]τ

i=1

ˆn = M

n X

ϕi (Xi , θˆn ).

i=1

The proof of the following result is left as an exercise. Theorem 5.16. Let X1 , ..., Xn be independent and {θˆn } be a consistent sequence of GEE estimators. Assume the conditions in Theorem 5.14. Suppose further that the sequence of functions {hij , i = 1, 2, ...} satisfies the

376

5. Estimation in Nonparametric Models

conditions in Lemma 5.3 with Θ replaced by a compact neighborhood of θ, where hij (x, γ) is the jth row of ψi (x, γ)[ψi (x, γ)]τ , j = 1, ..., k. Let Vn be ˆ −1 Var(s ˆ −1 is consistent for Vn . d n )M given by (5.100). Then Vˆn = M n n

5.5.2 The jackknife

Applying the substitution method requires the derivation of a formula for the covariance matrix or asymptotic covariance matrix of a point estimator. There are variance estimation methods that can be used without actually deriving such a formula (only the existence of the covariance matrix or asymptotic covariance matrix is assumed), at the expense of requiring a large number of computations. These methods are called resampling methods, replication methods, or data reuse methods. The jackknife method introduced here and the bootstrap method in §5.5.3 are the most popular resampling methods. The jackknife method was proposed by Quenouille (1949) and Tukey (1958). Let θˆn be a vector-valued estimator based on independent Xi ’s, where each Xi is a random di -vector and supi di < ∞. Let θˆ−i be the same estimator but based on X1 , ..., Xi−1 , Xi+1 , ..., Xn , i = 1, ..., n. Note that θˆ−i also depends on n but the subscript n is omitted for simplicity. Since θˆn and θˆ−1 , ..., θˆ−n are estimators of the same quantity, the “sample covariance matrix” τ 1 X ˆ θ−i − θ¯n θˆ−i − θ¯n n − 1 i=1 n

(5.111)

can be used as a measure of the variation of θˆn , where θ¯n is the average of θˆ−i ’s. There are two major differences between the quantity in (5.111) and the sample covariance matrix S 2 previously discussed. First, θˆ−i ’s are not independent. Second, θˆ−i − θˆ−j usually converges to 0 at a fast rate (such as n−1 ). Hence, to estimate the asymptotic covariance matrix of θˆn , the ¯ quantity in (5.111) should be multiplied by a correction factor cn . If θˆn = X −1 ¯ ˆ ¯ (di ≡ d), then θ−i − θn = (n − 1) (X − Xi ) and the quantity in (5.111) reduces to n X 1 1 ¯ Xi − X ¯ τ = S2, Xi − X (n − 1)3 i=1 (n − 1)2

where S 2 is the sample covariance matrix. Thus, the correction factor cn ¯ since, by the SLLN, S 2 /n is strongly is (n − 1)2 /n for the case of θˆn = X ¯ consistent for Var(X).

377

5.5. Variance Estimation

It turns out that the same correction factor works for many other estimators. This leads to the following jackknife variance estimator for θˆn : τ n − 1 X ˆ θ−i − θ¯n θˆ−i − θ¯n . VˆJ = n i=1 n

(5.112)

Theorem 5.17. Let X1 , ..., Xn be i.i.d. random d-vectors from F with ¯ Suppose that ∇g is finite µ = E(X1 ) and Var(X1 ), and let θˆn = g(X). continuous at µ and ∇g(µ) 6= 0. Then the jackknife variance estimator VˆJ in (5.112) is strongly consistent for Vn in (5.107). Proof. We prove the case where g is real-valued. The proof of the gen¯ −i be the sample mean based on eral case is left to the reader. Let X X1 , ..., Xi−1 , Xi+1 , ..., Xn . From the mean-value theorem, we have ¯ −i ) − g(X) ¯ θˆ−i − θˆn = g(X τ ¯ ¯ = [∇g(ξn,i )] (X−i − X) ¯ −i − X) ¯ + Rn,i , ¯ τ (X = [∇g(X)]

¯ τ (X ¯ −i − X) ¯ and ξn,i is a point on the where Rn,i = ∇g(ξn,i ) − ∇g(X) ¯ ¯ ¯ ¯ = (n − 1)−1 (X ¯ − Xi ), it line segmentP between X−i and X. From X−i − X n ¯ ¯ follows that i=1 (X−i − X) = 0 and n

n

1X ˆ 1X ¯n. (θ−i − θˆn ) = Rn,i = R n i=1 n i=1

From the definition of the jackknife estimator in (5.112), VˆJ = An + Bn + 2Cn , where n

An =

X n−1 ¯ τ ¯ −i − X)( ¯ X ¯ −i − X) ¯ τ ∇g(X), ¯ [∇g(X)] (X n i=1 n

Bn = and

n−1 X ¯ n )2 , (Rn,i − R n i=1

n

Cn =

n−1X ¯ n )[∇g(X)] ¯ τ (X ¯ −i − X). ¯ (Rn,i − R n i=1

¯ −i − X ¯ = (n − 1)−1 (X ¯ − Xi ), the SLLN, and the continuity of ∇g at By X µ, An /Vn →a.s. 1.

378

5. Estimation in Nonparametric Models

Also, (n − 1) Hence

n X i=1

n

¯ −i − Xk ¯ 2= kX

1 X ¯ 2 = O(1) a.s. kXi − Xk n − 1 i=1

(5.113)

¯ −i − Xk ¯ 2 →a.s. 0, max kX i≤n

¯ ≤ kX ¯ −i − Xk, ¯ which, together with the continuity of ∇g at µ and kξn,i − Xk implies that ¯ →a.s. 0. un = max k∇g(ξn,i ) − ∇g(X)k i≤n P ¯ −i − Xk ¯ 2 /Vn = O(1) a.s. Hence From (5.107) and (5.113), ni=1 kX n n Bn n−1X 2 un X ¯ ¯ 2 →a.s. 0. ≤ Rn,i ≤ kX−i − Xk Vn Vn n i=1 Vn i=1

By the Cauchy-Schwarz inequality, (Cn /Vn )2 ≤ (An /Vn )(Bn /Vn ) →a.s. 0. This proves the result. A key step in the proof of Theorem 5.17 is that θˆ−i − θˆn can be approx¯ τ (X ¯ −i − X) ¯ and the contributions of the remainders, imated by [∇g(X)] Rn,1 , ..., Rn,n , are sufficiently small, i.e., Bn /Vn →a.s. 0. This indicates that the jackknife estimator (5.112) is consistent for θˆn that can be well approximated by some linear statistic. In fact, the jackknife estimator (5.112) has been shown to be consistent when θˆn is a U-statistic (Arvesen, 1969) or a statistical functional that is Hadamard differentiable and continuously Gˆateaux differentiable at F (which includes certain types of L-estimators and M-estimators). More details can be found in Shao and Tu (1995, Chapter 2). The jackknife method can be applied to non-i.i.d. problems. A detailed discussion of the use of the jackknife method in survey problems can be found in Shao and Tu (1995, Chapter 6). We now consider the jackknife variance estimator for the LSE βˆ in linear model (3.25). For simplicity, assume that Z is of full rank. Assume also that εi ’s are independent with E(εi ) = 0 and Var(εi ) = σi2 . Then ˆ = (Z τ Z)−1 Var(β)

n X

σi2 Zi Ziτ (Z τ Z)−1 .

i=1

Let βˆ−i be the LSE of β based on the data with the ith pair (Xi , Zi ) deleted. Using the fact that (A + ccτ )−1 = A−1 − A−1 ccτ A−1 /(1 + cτ A−1 c) for a matrix A and a vector c, we can show that (exercise) βˆ−i = βˆ − ri Zi /(1 − hi ),

(5.114)

379

5.5. Variance Estimation

where ri = Xi − Ziτ βˆ is the ith residual and hi = Ziτ (Z τ Z)−1 Zi . Hence " n # n n τ X r 2 Zi Z τ X X r Z r Z n − 1 1 i i i i i i (Z τ Z)−1 (Z τ Z)−1 . − VˆJ = 2 n (1 − h ) n 1 − h 1 − h i i i i=1 i=1 i=1 Wu (1986) proposed the following weighted jackknife variance estimator that improves VˆJ : VˆW J =

n τ X ri2 Zi Ziτ τ −1 (1 − hi ) βˆ−i − βˆ βˆ−i − βˆ = (Z τ Z)−1 (Z Z) . 1 − hi i=1 i=1

n X

Theorem 5.18. Assume the conditions in Theorem 3.12 and that εi ’s are ˆ independent. Then both VˆJ and VˆW J are consistent for Var(β). p Proof. Let ln ∈ R , n = 1, 2, ..., be nonzero vectors and li = lnτ (Z τ Z)−1 Zi . Since maxi≤n hi → 0, the result for VˆW J follows from X n n X 2 2 li ri li2 σi2 →p 1 (5.115) i=1

i=1

(see Exercise 93). By the WLLN (Theorem 1.14(ii)) and maxi≤n hi → 0, X n n X 2 2 li εi li2 σi2 →p 1. i=1

i=1

ˆ and Note that ri = εi + Ziτ (β − β) ˆ 2 ≤ kZ(β − β)k ˆ 2 max hi = op (1). max[Ziτ (β − β)] i≤n

i≤n

Hence (5.115) holds. The consistency of VˆJ follows from (5.115) and !2 n n X n − 1 X li ri li2 σi2 = op (1). n2 1 − h i i=1 i=1

(5.116)

The proof of (5.116) is left as an exercise. Finally, let us consider the jackknife estimators for GEE estimators in §5.4.1. Under the conditions of Proposition 5.5 or 5.6, it can be shown that ˆ = op (1), max kθˆ−i − θk i≤n

where θˆ−i is a root of sni (γ) = 0 and X ψj (Xj , γ). sni (γ) = j6=i,j≤n

(5.117)

380

5. Estimation in Nonparametric Models

Assume that ψi (x, γ) is continuously differentiable w.r.t. γ in a neighborhood of θ. Using Taylor’s expansion and the fact that sni (θˆ−i ) = 0 and sn (θˆn ) = 0, we obtain that Z 1 ∇sn θˆn + t(θˆ−i − θˆn ) dt (θˆ−i − θˆn ). ψi (Xi , θˆ−i ) = 0

Following the proof of Theorem 5.14, we obtain that VˆJ = [Mn (θ)]−1

n X i=1

ψi (Xi , θˆ−i )[ψi (Xi , θˆ−i )]τ [Mn (θ)]−1 + Rn ,

−1/2

−1/2

where Rn satisfies kVn Rn Vn k = op (1) for Vn in (5.100). Under the conditions of Theorem 5.16, it follows from (5.117) that VˆJ is consistent. If θˆn is computed using an iteration method, then the computation of ˆ VJ requires n additional iteration processes. We may use the idea of a one-step MLE to reduce the amount of computation. For each i, let θˆ−i = θˆn − [∇sni (θˆn )]−1 sni (θˆn ),

(5.118)

which is the result from the first iteration when the Newton-Raphson method is applied in computing a root of sni (γ) = 0 and θˆn is used as the initial point. Note that θˆ−i ’s in (5.118) satisfy (5.117) (exercise). If the jackknife variance estimator is based on θˆ−i ’s in (5.118), then VˆJ = [Mn (θ)]−1

n X

˜n, ψi (Xi , θˆn )[ψi (Xi , θˆn )]τ [Mn (θ)]−1 + R

i=1

˜ n satisfies kVn−1/2 R ˜ n Vn−1/2 k = op (1). These results are summarized where R in the following theorem. Theorem 5.19. Assume the conditions in Theorems 5.14 and 5.16. Assume further that θˆ−i ’s are given by (5.118) or GEE estimators satisfying (5.117). Then the jackknife variance estimator VˆJ is consistent for Vn given in (5.100).

5.5.3 The bootstrap The basic idea of the bootstrap method can be described as follows. Suppose that P is a population or model that generates the sample X and that ˆ where θˆ = θ(X) ˆ we need to estimate Var(θ), is an estimator, a statistic based on X. Suppose further that the unknown population P is estimated by Pˆ , based on the sample X. Let X ∗ be a sample (called a bootstrap

381

5.5. Variance Estimation

sample) taken from the estimated population Pˆ using the same or a similar ˆ ∗ ), which is the sampling procedure used to obtain X, and let θˆ∗ = θ(X ∗ ˆ same as θ but with X replaced by X . If we believe that P = Pˆ (i.e., ˆ = Var∗ (θˆ∗ ), we have a perfect estimate of the population), then Var(θ) where Var∗ is the conditional variance w.r.t. the randomness in generating ˆ 6= Var∗ (θˆ∗ ). But X ∗ , given X. In general, P 6= Pˆ and, therefore, Var(θ) ˆ and can be used as an VˆB = Var∗ (θˆ∗ ) is an empirical analogue of Var(θ) ˆ estimate of Var(θ). In a few cases, an explicit form of VˆB = Var∗ (θˆ∗ ) can be obtained. First, consider i.i.d. X1 , ..., Xn from a c.d.f. F on Rd . The population is determined by F . Suppose that we estimate F by the empirical c.d.f. Fn ¯ its bootstrap in (5.1) and that X1∗ , ..., Xn∗ are i.i.d. from Fn . For θˆ = X, ∗ ∗ ∗ ˆ ¯ analogue is θ = X , the average of Xi ’s. Then n X n−1 2 ¯ ∗) = 1 ¯ ¯ τ (Xi − X)(X S , VˆB = Var∗ (X i − X) = n2 i=1 n2

¯ ∗ ) is where S 2 is the sample covariance matrix. In this case VˆB = Var∗ (X ¯ a strongly consistent estimator for Var(X). Next, consider i.i.d. random variables X1 , ..., Xn from a c.d.f. F on R and θˆ = Fn−1 ( 12 ), the sample median. Suppose that n = 2l − 1 for an integer l. Let X1∗ , ..., Xn∗ be i.i.d. from Fn and θˆ∗ be the sample median based on X1∗ , ..., Xn∗ . Then !2 n n X X ∗ ˆ pj X(j) − pi X(i) , VˆB = Var∗ (θ ) = j=1

i=1

where X(1) ≤ · · · ≤ X(n) are order statistics and pj = P (θˆ∗ = X(j) |X). It can be shown (exercise) that l−1 X n (j − 1)t (n − j + 1)n−t − j t (n − j)n−t pj = . (5.119) nn t t=0 However, in most cases VˆB does not have a simple explicit form. When P is known, the Monte Carlo method described in §4.1.4 can be used to ˆ That is, we draw repeatedly new data sets from P and approximate Var(θ). then use the sample covariance matrix based on the values of θˆ computed ˆ This idea from new data sets as a numerical approximation to Var(θ). can be used to approximate VˆB , since Pˆ is a known population. That is, we can draw m bootstrap data sets X ∗1 , ..., X ∗m independently from Pˆ ˆ ∗j ), j = 1, ..., m, and approximate (conditioned on X), compute θˆ∗j = θ(X VˆB by m 1 X ˆ∗j ¯∗ ˆ∗j ¯∗ τ VˆBm = θ −θ θ −θ , m j=1

382

5. Estimation in Nonparametric Models

where θ¯∗ is the average of θˆ∗j ’s. Since each X ∗j is a data set generated from Pˆ , VˆBm is a resampling estimator. From the SLLN, as m → ∞, VˆBm →a.s. VˆB , conditioned on X. Both VˆB and its Monte Carlo approximation VˆBm are ˆ Vˆ m is more useful in practical called bootstrap variance estimators for θ. B applications, whereas in theoretical studies, we usually focus on VˆB . The consistency of the bootstrap variance estimator VˆB is a much more complicated problem than that of the jackknife variance estimator in §5.5.2. Some examples can be found in Shao and Tu (1995, §3.2.2). The bootstrap method can also be applied to estimate quantities other ˆ For example, let K(t) = P (θˆ ≤ t) be the c.d.f. of a real-valued than Var(θ). ˆ From the previous discussion, a bootstrap estimator of K(t) estimator θ. is the conditional probability P (θˆ∗ ≤ t|X), which can be approximated Pm by the Monte Carlo approximation m−1 j=1 I(−∞,t] (θˆ∗j ). An important application of bootstrap distribution estimators in problems of constructing confidence sets is studied in §7.4. Here, we study the use of a bootstrap distribution estimator to form a consistent estimator of the asymptotic ˆ variance of a real-valued estimator θ. Suppose that √ n(θˆ − θ) →d N (0, v), (5.120) √ ˆ where v is unknown. Let Hn (t) be the c.d.f. of n(θ − θ) and √ ˆ ≤ t|X) ˆ B (t) = P ( n(θˆ∗ − θ) H

(5.121)

be a bootstrap estimator of Hn (t). If ˆ B (t) − Hn (t) →p 0 H for any t, then, by (5.120), √ ˆ B (t) − Φ t/ v →p 0, H

which implies (Exercise 112) that

ˆ −1 (α) →p H B

√ vzα

for any α ∈ (0, 1), where zα = Φ−1 (α). Then, for α 6= 12 , ˆ −1 (α) →p ˆ −1 (1 − α) − H H B B

√ v(z1−α − zα ).

ˆ is Therefore, a consistent estimator of v/n, the asymptotic variance of θ, " #2 ˆ −1 (α) ˆ −1 (1 − α) − H H 1 B B . V˜B = n z1−α − zα

383

5.6. Exercises

ˆ B (t)−Hn (t) →p 0. The following result gives some conditions under which H The proof of part (i) is omitted. The proof of part (ii) is given in Exercises 113-115 in §5.6. Theorem 5.20. Suppose that X1 , ..., Xn are i.i.d. from a c.d.f. F on Rd . Let θˆ = T(Fn ), where T is a real-valued functional, θˆ∗ = T(Fn∗ ), where Fn∗ is the empirical c.d.f. based on a bootstrap sample X1∗ , ..., Xn∗ i.i.d. from Fn , ˆ B be given by (5.121). and let H (i) If T is ̺∞ -Hadamard differentiable at F and (5.40) holds, then ˆ B , Hn ) →p 0. ̺∞ (H

(5.122) R (ii) If d = 1 and T is ̺Lp -Fr´echet differentiable at F ( {F (t)[1 − F (t)]}p/2 dt < ∞ if 1 ≤ p < 2) and (5.40) holds, then (5.122) holds. Applications of the bootstrap method to non-i.i.d. cases can be found, for example, in Efron and Tibshirani (1993), Hall (1992), and Shao and Tu (1995).

5.6 Exercises 1. Let ̺∞ be the sup-norm distance. Find an example of a sequence {Gn } of c.d.f.’s satisfying Gn →w G for a c.d.f. G, but ̺∞ (Gn , G) does not converge to 0. 2. Let X1 , ..., Xn be i.i.d. random d-vectors with c.d.f. F and Fn be the empirical c.d.f. defined by (5.1). Show that for any t > 0 and ǫ > 0, there is a Cǫ,d such that for all n = 1, 2, ..., P

sup ̺∞ (Fm , F ) > t

m≥n

2

≤

Cǫ,d e−(2−ǫ)t n . 1 − e−(2−ǫ)t2

3. Show that ̺Mp defined by (5.4) is a distance on Fp , p ≥ 1. 4. Show that k · kLp in (5.5) is a norm for any p ≥ 1. 5. Let F1 be the collection of c.d.f.’s finite means. R 1 on R with −1 −1 (a) Show that ̺M1 (G1 , G2 ) = 0 |G−1 (z) 1 (z) − G2 (z)|dz, where G = inf{t : G(t) ≥ z} for any G ∈ F. (b) Show that ̺M1 (G1 , G2 ) = ̺L1 (G1 , G2 ). 6. Find an example of a sequence {Gj } ⊂ F for which (a) limj→∞ ̺∞ (Gj , G0 ) = 0 but ̺M2 (Gj , G0 ) does not converge to 0; (b) limj→∞ ̺M2 (Gj , G0 ) = 0 but ̺∞ (Gj , G0 ) does not converge to 0.

384

5. Estimation in Nonparametric Models

7. Repeat the previous exercise with ̺M2 replaced by ̺L2 . 8. Let X be a random variable having c.d.f. F . Show that R (a) E|X|2 < ∞ implies {F (t)[1 − F (t)]}p/2Rdt < ∞ for p ∈ (1, 2); (b) E|X|2+δ < ∞ with some δ > 0 implies {F (t)[1 − F (t)]}1/2 dt < ∞. 9. For Gj ∈ F1 , j = 1, 2, show that ̺L1 (G1 , G2 ) ≥ R any one-dimensional R | xdG1 − xdG2 |.

10. In the proof of Theorem 5.3, show that pi = c/n, i = 1, ..., n, λ = −(c/n)n−1 is Pan maximum of the function H(p1 , ..., pn , λ) over pi > 0, i = 1, ..., n, i=1 pi = c.

11. Show that (5.11)-(5.13) is a solution to the problem of maximizing ℓ(G) in (5.8) subject to (5.10). 12. In the proof of Theorem 5.4, prove the case of m ≥ 2. 13. Show that a maximum of ℓ(G) in (5.17) subject to (5.10) is given by (5.11) with pˆi defined by (5.18) and (5.19). 14. In Example 5.2, show that an MELE is given by (5.11) with pˆi ’s given by (5.21). 15. In Example 5.3, show that (a) maximizing (5.22) subject to (5.23) is equivalent to maximizing n Y

i=1

Pn+1

δ

qi (i) (1 − qi )n−i+1−δ(i) ,

where qi = pi / j=i pj , i = 1, ..., n; (b) Fˆ given by (5.24) maximizes (5.22) subject to (5.23); (Hint: use Qi−1 part (a) and the fact that pi = qi j=1 (1 − qj ).) (c) Fˆ given by (5.25) is the same as that in (5.24); (d) if δi = 1 for all i (no censoring), then Fˆ in (5.25) is the same as the empirical c.d.f. in (5.1). 16. Let fn be given by (5.26). (a) Show that fn is a Lebesgue p.d.f. on R. (b) Suppose that f is continuously differentiable at t, λn → 0, and nλn → ∞. Show that (5.27) holds. (c) Under nλ3n → 0 and the conditions of (b), show that (5.28) holds. (d) Suppose that f is continuous on [a, b], −∞ < a < b < ∞, λn → 0, Rb Rb and nλn → ∞. Show that a fn (t)dt →p a f (t)dt.

5.6. Exercises

385

17. Let fˆ be given by (5.29). (a) Show that fˆ is a Lebesgue p.d.f. on R. (b) Prove (5.30) under the condition that λn → 0, nλn → ∞, and R f is bounded and continuous at t and [w(t)]2 dt < ∞. (Hint: check Lindeberg’s condition and apply Theorem 1.15.) (c) Assume that λn → 0, nλn → ∞, w is bounded, and f is bounded Rb ˆ and continuous on [a, b], −∞ < a < b < ∞. Show that a f(t)dt →p Rb a f (t)dt.

18. Prove (5.32)-(5.34) under the conditions described in §5.1.4.

ˆ 19. Show that K(t) in (5.35) is a consistent estimator of K(t) in (5.34), ˆ assuming that β →p β, φ is a continuous function on R, (Xi , Zi )’s are i.i.d., and kZi k ≤ c for a constant c > 0. 20. Let ℓ(θ, ξ) be a likelihood. Show that a maximum profile likelihood estimator θˆ of θ is an MLE if ξ(θ), the maximum of supξ ℓ(θ, ξ) for a fixed θ, does not depend on θ. 21. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ). Derive the profile likelihood function for µ or σ 2 . Discuss in each case whether the maximum profile likelihood estimator is the same as the MLE.

22. Derive the profile empirical likelihoods in (5.36) and (5.37). 23. Let X1 , ..., Xn be i.i.d. random variables from a c.d.f. F and let π(x) = P (δi = 1|Xi = x), where δi = 1R if Xi is observed and δi = 0 if Xi is missing. Assume that 0 < π = π(x)dF (x) < 1. (a) Let F1 (x) = P (Xi ≤ x|δi = 1). Show that F and F1 are the same if and only if π(x) ≡ π. (b) Let Fˆ be the c.d.f. putting mass r−1 to each observed Xi , where r is the number of observed Xi ’s. Show that Fˆ (x) is unbiased and consistent for F1 (x), x ∈ R. (c) When π(x) ≡ π, show that Fˆ (x) in part (b) is unbiased and consistent for F (x), x ∈ R. When π(x) is not constant, show that Fˆ (x) is biased and inconsistent for F (x) for some x ∈ R. 24. Show that ̺-Fr´echet differentiability implies ̺-Hadamard differentiability. 25. Suppose that a functional T is Gˆ ateaux differentiable at F with a continuous differential LF in the sense that ̺∞ (∆j , ∆) → 0 implies LF (∆j ) → LF (∆). Show that φF is bounded. 26. Suppose that a functional T is Gˆ ateaux differentiable at F with a bounded and continuous influence function φF . Show that the differential LF is continuous in the sense described in the previous exercise.

386

5. Estimation in Nonparametric Models

R 27. Let T(G) = g( xdG) be a functional defined on F1 , the collection of one-dimensional c.d.f.’s with finite means. (a) Find a differentiable function g for which the functional T is not ̺∞ -Hadamard differentiable at F . (b) Show that if g is a differentiable function, then T is ̺L1 -Fr´echet differentiable at F . (Hint: use the result in Exercise 9.) 28. In Example 5.5, show that (5.43) holds. (Hint: for ∆ = c(G1 − G2 ), show that k∆kV ≤ |c|(kG1 kV + kG2 kV ) = 2|c|.) 29. In Example 5.5, show that φF is continuous if F is continuous. 30. In Example 5.5, show that T is not ̺∞ -Fr´echet differentiable at F . 31. Prove Proposition 5.1(ii). 32. Suppose that T is first-order and second-order ̺-Hadamard differentiable at F . Prove (5.45). 33. Find an example of a second-order ̺-Fr´echet differentiable functional T that is not first-order ̺-Hadamard differentiable. 34. Prove (5.47) and that (5.40) is satisfied for an L-functional if J is bounded and F has a finite variance. 35. Prove (iv) and (v) of Theorem 5.6. 36. Discuss which of (i)-(v) in Theorem 5.6 can be applied to each of the L-estimators in Example 5.6. 37. Obtain explicit forms of the influence functions for L-estimators in Example 5.6. Discuss which of them are bounded and continuous. 38. Provide an example in which the L-functional T given by (5.46) is not ̺∞ -Hadamard differentiable at F . (Hint: consider an untrimmed J.) 39. Discuss which M-functionals defined in (i)-(vi) of Example 5.7 satisfy the conditions of Theorem 5.7. 40. In the proof of Theorem 5.7, show that R2j → 0. 41. Show that the second equality in (5.51) holds when ψ is Borel and bounded. 42. Show that the functional T in (5.53) is ̺∞ -Hadamard differentiable at F with the differential given by (5.54). Obtain the influence function φF and show that it is bounded and continuous if F is continuous.

387

5.6. Exercises

43. Show that the functional T in (5.55) is ̺∞ -Hadamard differentiable at F with the differential given by (5.56). Obtain the influence function φF and show that it is bounded and continuous if F (y, ∞) and F (∞, z) are continuous. 44. Let F be a continuous c.d.f. on R. Suppose that F is symmetric about θ and is strictly increasing in a neighborhood of θ. Show that λF (t) = 0 if and only if t = θ, where λF (t) is defined by (5.57) with a strictly increasing J satisfying J(1 − t) = −J(t). 45. Show in (5.57) is differentiable at θ and λ′F (θ) is equal to R ′that λF (t) ′ − J (F (x))F (x)dF (x).

46. Let T(Fn ) be an R-estimator satisfying the conditions in Theorem 5.8. Show that (5.41) holds with σF2 =

Z

0

1

[J(t)]2 dt

Z

∞

−∞

2 J ′ (F (x))F ′ (x)dF (x) .

47. Calculate the asymptotic relative efficiency of the Hodges-Lehmann estimator in Example 5.8 w.r.t. the sample mean based on an i.i.d. sample from F when (a) F is the c.d.f. of N (µ, σ 2 ); (b) F is the c.d.f. of the logistic distribution LG(µ, σ); (c) F is the c.d.f. of the double exponential distribution DE(µ, σ); (d) F (x) = F0 (x − θ), where F0 (x) is the c.d.f. of the t-distribution tν with ν ≥ 3. 48. Let G be a c.d.f. on R. Show that G(x) ≥ t if and only if x ≥ G−1 (t). 49. Show that (5.67) implies that θˆp is strongly consistent for θp and is √ n-consistent for θp if F ′ (θp −) and F ′ (θp +) exist and are positive. 2

50. Under the condition of Theorem 5.9, show that, for ρǫ = e−2δǫ , 2Cρnǫ , n = 1, 2, .... P sup |θˆp − θp | > ǫ ≤ 1 − ρǫ m≥n 51. Prove that ϕn (t) in (5.69) is the Lebesgue p.d.f. of the pth sample quantile θˆp when F has the Lebesgue p.d.f. f by (a) differentiating the c.d.f. of θˆp in (5.68); (b) using result (5.66) and the result in Example 2.9. 52. Let X1 , ..., Xn be i.i.d. random variables from F with a finite mean. Show that θˆp has a finite jth moment for sufficiently large n, j = 1, 2,....

388

5. Estimation in Nonparametric Models

53. Prove Theorem 5.10(i). 54. Suppose that a c.d.f. F has a Lebesgue p.d.f. f that is continuous at the pth quantile of F , p ∈ (0, 1). Using the p.d.f. in (5.69) and Scheff´e’s theorem (Proposition 1.18), prove part (iv) of Theorem 5.10. 55. Let {kn } be a sequence of integers satisfying kn /n = p + o(n−1/2 ) with p ∈ (0, 1), and let X1 , ..., Xn be i.i.d. random variables from a c.d.f. F with F ′ (θp ) > 0. Show that √ n(X(kn ) − θp ) →d N (0, p(1 − p)/[F ′ (θp )]2 ). 56. In the proof of Theorem 5.11, prove (5.72), (5.75), and inequality (5.74). 57. Prove Corollary 5.1. 58. Prove the claim in Example 5.9. 59. Let T (G) = G−1 (p) be the pth quantile functional. Suppose that F has a positive derivative F ′ in a neighborhood of θ = F −1 (p). Show that T is Gˆateaux differentiable at F and obtain the influence function. 60. Let X1 , ..., Xn be i.i.d. from the Cauchy distribution C(0, 1). (a) Show that E(X(j) )2 < ∞ if and only if 3 ≤ j ≤ n − 2. (b) Show that E(θˆ0.5 )2 < ∞ for n ≥ 5. 61. Suppose that F is the c.d.f. of the uniform distribution U (θ− 21 , θ+ 12 ), θ ∈ R. Obtain the asymptotic relative efficiency of the sample median w.r.t. the sample mean, based on an i.i.d. sample of size n from F . 62. Suppose that F (x) = F0 (x − θ) and F0 is the c.d.f. of the Cauchy distribution C(0, 1) truncated R c at c and −c, i.e., F0 has the Lebesgue p.d.f. (1 + x2 )−1 I(−c,c) (x)/ −c (1 + x2 )−1 dt. Obtain the asymptotic relative efficiency of the sample median w.r.t. the sample mean, based on an i.i.d. sample of size n from F . 63. Let X1 , ..., Xn be i.i.d. with the c.d.f. (1−ǫ)Φ x−µ +ǫD x−µ , where σ σ ǫ ∈ (0, 1) is a known constant, Φ is the c.d.f. of the standard normal distribution, D is the c.d.f. of the double exponential distribution D(0, 1), and µ ∈ R and σ > 0 are unknown parameters. Consider the estimation of µ. Obtain the asymptotic relative efficiency of the sample mean w.r.t. the sample median. 64. Let X1 , ..., Xn be i.i.d. with the Lebesgue p.d.f. 2−1 (1 − θ2 )eθx−|x|, where θ ∈ (−1, 1) is unknown. (a) Show that the median of the distribution of X1 is given by m(θ) =

5.6. Exercises

389

(1 − θ)−1 log(1 + θ) when θ ≥ 0 and m(θ) = −m(−θ) when θ < 0. (b) Show that the mean of the distribution of X1 is µ(θ) = 2θ/(1−θ2 ). (c) Show that the inverse functions of m(θ) and µ(θ) exist. Obtain ¯ where m the asymptotic relative efficiency of m−1 (m) ˆ w.r.t. µ−1 (X), ˆ ¯ is the sample median and X is the sample mean. ¯ in (d) asymptotically efficient in estimating θ? (e) Is µ−1 (X) ¯ α in (5.77) is the L-estimator corresponding to the J 65. Show that X function given in Example 5.6(iii) with β = 1 − α. 66. Let X1 , ..., Xn be i.i.d. random variables from F , where F is symmetric about θ. (a) Show that P X(j) − θ and θ − X(n−j+1) have the same distribution. n (b) Show that j=1 wj X(j) has a c.d.f. symmetric about θ, if wi ’s are Pn constants satisfying i=1 wi = 1 and wj = wn−j+1 for all j. ¯ α has a c.d.f. symmetric (c) Show that the trimmed sample mean X about θ. 67. Under the conditions in one of (i)-(iii) of Theorem 5.6, show that (5.41) holds for T(Fn ) with σF2 given by (5.79), if σF2 < ∞. 68. Prove (5.78) under the assumed conditions. 69. For the functional T given by (5.46), show that T(F ) = θ if F is R1 symmetric about θ, J is symmetric about 12 , and 0 J(t)dt = 1.

70. Obtain the asymptotic relative efficiency of the trimmed sample mean ¯ α w.r.t. the sample mean, based on an i.i.d. sample of size n from the X double exponential distribution DE(θ, 1), where θ ∈ R is unknown. 71. Obtain the asymptotic relative efficiency of the trimmed sample mean ¯ α w.r.t. the sample median, based on an i.i.d. sample of size n from X the Cauchy distribution C(θ, 1), where θ ∈ R is unknown. 72. Consider the α-trimmed sample mean defined in (5.77). Show that σα2 in (5.78) is the same as σF2 in (5.79) with J(t) = (1−2α)−1 I(α,1−α) (t), when F (x) = F0 (x − θ) and F0 is symmetric about 0. 73. For σα2 in (5.78), show that (a) if F0′ (0) exists and is positive, then limα→ 12 σα2 = 1/[2F0′ (0)]2 ; R (b) if σ 2 = x2 dF0 (x) < ∞, then limα→0 σα2 = σ 2 .

74. Show that if J ≡ 1, then σF2 in (5.79) is equal to the variance of the c.d.f. F .

75. Calculate σF2 in (5.79) with J(t) = 4t − 2 and F being the double exponential distribution DE(θ, 1), θ ∈ R.

390

5. Estimation in Nonparametric Models

76. Consider the simple linear model in Example 3.12 with positive ti ’s. Derive the L-estimator of β defined by (5.82) with a J symmetric about 12 and compare it with the LSE of β. 77. Consider the one-way ANOVA model in Example 3.13. Derive the L-estimator of β defined by (5.82) when (a) J is symmetric about 12 and (b) J(t) = (1 − 2α)−1 I(α,1−α) (t). Compare these L-estimators with the LSE of β. 78. Show that the method of moments in §3.5.2 is a special case of the GEE method. 79. Complete the proof of Proposition 5.4. 80. In the proof of Lemma 5.3, show that the probability in (5.94) is bounded by ǫ. 81. In Example 5.11, show that ψi ’s satisfy the conditions of Lemma 5.3 if Θ is compact and supi kZi k < ∞. 82. In the proof of Proposition 5.5, show that {∆n (γ)} is equicontinuous on any open subset of Θ. 83. Prove Proposition 5.6. 84. Prove the claim in Example 5.12. 85. Prove the claims in Example 5.13. 86. For Huber’s M-estimator discussed in Example 5.13, obtain a formula ¯ when F is for e(F ), the asymptotic relative efficiency of θˆn w.r.t. X, given by (5.76). Show that limτ →∞ e(F ) = ∞. Find the value of e(F ) when ǫ = 0, σ = 1, and C = 1.5. 87. Consider the ψ function in Example 5.7(ii). Show that under some conditions on F , ψ satisfies the conditions given in Theorem 5.13(i) or (ii). Obtain σF2 in (5.98) in this case. 88. In the proof of Theorem 5.14, show that (a) (5.101) holds; (b) (5.103) holds; (c) (5.104) implies (5.102). (Hint: use Theorem 1.9(iii).) 89. Prove the claim in Example 5.14, assuming some necessary moment conditions. 90. Derive the asymptotic distribution of the MQLE (the GEE estimator based on (5.90)), assuming that Xi = (Xi1 , ..., Xidi ), E(Xit ) = meηi /(1 + eηi ), Var(Xit ) = mφi eηi /(1 + eηi )2 , and (4.57) holds with t g(t) = log 1−t .

5.6. Exercises

391

ηi 91. Repeat the previous exercise under the assumption that E(Xit ) = √e , ηi Var(Xit ) = φi e , and (4.57) holds with g(t) = log t or g(t) = 2 t.

˜ i is replaced 92. In Theorem 5.14, show that result (5.99) still holds if R ˆ i satisfying maxi≤n kR ˆ i − Ui k = op (1), where Ui ’s by an estimator R are correlation matrices. 93. Show that (5.106) holds if and only if one of the following holds: (a) λ− →p 1 and λ+ →p 1, where λ− and λ+ are respectively the −1/2 ˆ −1/2 smallest and largest eigenvalues of Vn . Vn Vn (b) lnτ Vˆn ln /lnτ Vn ln →p 1, where {ln } is any sequence of nonzero vectors in Rk . −1/2 ˆ (θn − θ) →d Nk (0, Ik ). 94. Show that (5.105) and (5.106) imply Vˆn

95. Suppose that X1 , ..., Xn are independent (not necessarily identically distributed) random d-vectors with E(Xi ) = µ for all i. Suppose also that supi EkXi k2+δ < ∞ for some δ > 0. Let µ = E(X1 ), θ = g(µ), ¯ Show that and θˆn = g(X). Pn (a) (5.105) holds with Vn = n−2 [∇g(µ)]τ i=1 Var(Xi )∇g(µ); (b) Vˆn in (5.108) is consistent for Vn in part (a). 96. Consider the ratio estimator in Example 3.21. Derive the estimator Vˆn given by (5.108) and show that Vˆn is consistent for the asymptotic variance of the ratio estimator. ˆ 97. Derive a consistent variance estimator for R(t) in Example 3.23. 98. Prove the claims in Example 5.16. 99. Let σF2 n be given by (5.79) with F replaced by the empirical c.d.f. Fn . (a) Show that σF2 n /n is the same as Vˆn in (5.110) for an L-estimator with influence function φF . (b) Show directly (without using Theorem 5.15) σF2 n →a.s. σF2 in (5.79), under the conditions in Theorem 5.6(i) or (ii) (with EX12 < ∞). 100. Derive a consistent variance estimator for a U-statistic satisfying the conditions in Theorem 3.5(i). 101. Derive a consistent variance estimator for Huber’s M-estimator discussed in Example 5.13. 102. Assume the conditions in Theorem 5.8. Let r ∈ (0, 21 ). (a) Show that nr λF (T(Fn ) + n−r ) →p λF (T(F )). (b) Show that nr [λFn (T(Fn ) + n−r ) − λF (T(Fn ) + n−r )] →p 0. (c) Derive a consistent estimator of the asymptotic variance of T(Fn ), using the results in (a) and (b).

392

5. Estimation in Nonparametric Models

103. Prove Theorem 5.16. ¯ 2 . Show that the 104. Let X1 , ..., Xn be random variables and θˆ = X ¯ 2 cˆ2 ¯ cˆ3 cˆ4 −ˆ c22 4X 4X jackknife estimator in (5.112) equals n−1 − (n−1)2 + (n−1) 3 , where cˆj ’s are the sample central moments defined by (3.52). 105. Prove Theorem 5.17 for the case where g is from Rd to Rk and k ≥ 2. 106. Prove (5.114). 107. In the proof of Theorem 5.18, prove (5.116). 108. Show that θˆ−i ’s in (5.118) satisfy (5.117), under the conditions of Theorem 5.14. 109. Prove Theorem 5.19. 110. Prove (5.119). ¯ 2 . Show that the 111. Let X1 , ..., Xn be random variables and θˆ = X ∗ bootstrap variance estimator based on i.i.d. Xi ’s from Fn is equal to ¯2 ¯ cˆ3 cˆ4 −ˆ c22 ˆj ’s are the sample central moments VˆB = 4Xn cˆ2 + 4X n2 + n3 , where c defined by (3.52). 112. Let G, G1 , G2 ,..., be c.d.f.’s on R. Suppose that ̺∞ (Gj , G) → 0 as j → ∞ and G′ (x) exists and is positive for all x ∈ R. Show that −1 G−1 (p) for any p ∈ (0, 1). j (p) → G 113. Let X1 , ..., Xn be i.i.d. from a c.d.f. F on Rd with a finite Var(X1 ). Let X1∗ , ..., Xn∗ be i.i.d. from the empirical √ ¯ ∗c.d.f.¯ Fn . Show that for almost all given sequences X1 , X2 , ..., n(X − X) →d Nd (0, Var(X1 )). (Hint: verify Lindeberg’s condition.) 114. Let X1 , ..., Xn be i.i.d. from a c.d.f. F on Rd , X1∗ , ..., Xn∗ be i.i.d. from the empirical c.d.f. Fn , and let Fn∗ be the empirical c.d.f. based on Xi∗ ’s. Using DKW’s inequality (Lemma 5.1), show that (a) ̺∞ (Fn∗ , F ) →a.s. 0; (b) ̺∞ (Fn∗ , F ) = Op (n−1/2 ); (c) ̺Lp (Fn∗ , F ) = Op (n−1/2 ), under the condition in Theorem 5.20(ii). 115. Using the results from the previous two exercises, prove Theorem 5.20(ii). 116. Under the conditions in Theorem 5.11, establish a Bahadur’s representation for the bootstrap sample quantile θˆp∗ .

Chapter 6

Hypothesis Tests A general theory of testing hypotheses is presented in this chapter. Let X be a sample from a population P in P, a family of populations. Based on the observed X, we test a given hypothesis H0 : P ∈ P0 versus H1 : P ∈ P1 , where P0 and P1 are two disjoint subsets of P and P0 ∪ P1 = P. Notational conventions and basic concepts (such as two types of errors, significance levels, and sizes) given in Example 2.20 and §2.4.2 are used in this chapter.

6.1 UMP Tests A test for a hypothesis is a statistic T (X) taking values in [0, 1]. When X = x is observed, we reject H0 with probability T (x) and accept H0 with probability 1−T (x). If T (X) = 1 or 0 a.s. P, then T (X) is a nonrandomized test. Otherwise T (X) is a randomized test. For a given test T (X), the power function of T (X) is defined to be βT (P ) = E[T (X)],

P ∈ P,

(6.1)

which is the type I error probability of T (X) when P ∈ P0 and one minus the type II error probability of T (X) when P ∈ P1 . As we discussed in §2.4.2, with a sample of a fixed size, we are not able to minimize two error probabilities simultaneously. Our approach involves maximizing the power βT (P ) over all P ∈ P1 (i.e., minimizing the type II error probability) and over all tests T satisfying sup βT (P ) ≤ α,

(6.2)

P ∈P0

where α ∈ [0, 1] is a given level of significance. Recall that the left-hand side of (6.2) is defined to be the size of T . 393

394

6. Hypothesis Tests

Definition 6.1. A test T∗ of size α is a uniformly most powerful (UMP) test if and only if βT∗ (P ) ≥ βT (P ) for all P ∈ P1 and T of level α. If U (X) is a sufficient statistic for P ∈ P, then for any test T (X), E(T |U ) has the same power function as T and, therefore, to find a UMP test we may consider tests that are functions of U only. The existence and characteristics of UMP tests are studied in this section.

6.1.1 The Neyman-Pearson lemma A hypothesis H0 (or H1 ) is said to be simple if and only if P0 (or P1 ) contains exactly one population. The following useful result, which has already been used once in the proof of Theorem 4.16, provides the form of UMP tests when both H0 and H1 are simple. Theorem 6.1 (Neyman-Pearson lemma). Suppose that P0 = {P0 } and P1 = {P1 }. Let fj be the p.d.f. of Pj w.r.t. a σ-finite measure ν (e.g., ν = P0 + P1 ), j = 0, 1. (i) (Existence of a UMP test). For every α, there exists a UMP test of size α, which is equal to f1 (X) > cf0 (X) 1 T∗ (X) = (6.3) γ f1 (X) = cf0 (X) 0 f1 (X) < cf0 (X),

where γ ∈ [0, 1] and c ≥ 0 are some constants chosen so that E[T∗ (X)] = α when P = P0 (c = ∞ is allowed). (ii) (Uniqueness). If T∗∗ is a UMP test of size α, then 1 f1 (X) > cf0 (X) T∗∗ (X) = a.s. P. (6.4) 0 f1 (X) < cf0 (X) Proof. The proof for the case of α = 0 or 1 is left as an exercise. Assume now that 0 < α < 1. (i) We first show that there exist γ and c such that E0 [T∗ (X)] = α, where Ej is the expectation w.r.t. Pj . Let γ(t) = P0 (f1 (X) > tf0 (X)). Then γ(t) is nonincreasing, γ(0) = 1, and γ(∞) = 0 (why?). Thus, there exists a c ∈ (0, ∞) such that γ(c) ≤ α ≤ γ(c−). Set ( α−γ(c) γ(c−) 6= γ(c) γ(c−)−γ(c) γ= 0 γ(c−) = γ(c). Note that γ(c−) − γ(c) = P (f1 (X) = cf0 (X)). Then E0 [T∗ (X)] = P0 f1 (X) > cf0 (X) + γP0 f1 (X) = cf0 (X) = α.

395

6.1. UMP Tests

Next, we show that T∗ in (6.3) is a UMP test. Suppose that T (X) is a test satisfying E0 [T (X)] ≤ α. If T∗ (x) − T (x) > 0, then T∗ (x) > 0 and, therefore, f1 (x) ≥ cf0 (x). If T∗ (x) − T (x) < 0, then T∗ (x) < 1 and, therefore, f1 (x) ≤ cf0 (x). In any case, [T∗ (x) − T (x)][f1 (x) − cf0 (x)] ≥ 0 and, therefore, Z [T∗ (x) − T (x)][f1 (x) − cf0 (x)]dν ≥ 0, i.e.,

Z

[T∗ (x) − T (x)]f1 (x)dν ≥ c

Z

[T∗ (x) − T (x)]f0 (x)dν.

(6.5)

The left-hand side of (6.5) is E1 [T∗ (X)] − E1 [T (X)] and the right-hand side of (6.5) is c{E0 [T∗ (X)] − E0 [T (X)]} = c{α − E0 [T (X)]} ≥ 0. This proves the result in (i). (ii) Let T∗∗ (X) be a UMP test of size α. Define A = {x : T∗ (x) 6= T∗∗ (x), f1 (x) 6= cf0 (x)}. Then [T∗ (x)−T∗∗ (x)][f1 (x)−cf0 (x)] > 0 when x ∈ A and = 0 when x ∈ Ac , and Z [T∗ (x) − T∗∗ (x)][f1 (x) − cf0 (x)]dν = 0,

since both T∗ and T∗∗ are UMP tests of size α. By Proposition 1.6(ii), ν(A) = 0. This proves (6.4).

Theorem 6.1 shows that when both H0 and H1 are simple, there exists a UMP test that can be determined by (6.4) uniquely (a.s. P) except on the set B = {x : f1 (x) = cf0 (x)}. If ν(B) = 0, then we have a unique nonrandomized UMP test; otherwise UMP tests are randomized on the set B and the randomization is necessary for UMP tests to have the given size α; furthermore, we can always choose a UMP test that is constant on B. Example 6.1. Suppose that X is a sample of size 1, P0 = {P0 }, and P1 = {P1 }, where P0 is N (0, 1) and P1 is the double exponential distribution DE(0, 2) with the p.d.f. 4−1 e−|x|/2. Since P (f1 (X) = cf0 (X)) = 0, there is a unique nonrandomized UMP test. From (6.3), the UMP test T∗ (x) = 1 2 if and only if π8 ex −|x| > c2 for some c > 0, which is equivalent to |x| > t or |x| < 1 − t for some t > 21 . Suppose that α < 13 . To determine t, we use α = E0 [T∗ (X)] = P0 (|X| > t) + P0 (|X| < 1 − t).

(6.6)

If t ≤ 1, then P0 (|X| > t) ≥ P0 (|X| > 1) = 0.3374 > α. Hence t should be larger than 1 and (6.6) becomes α = P0 (|X| > t) = Φ(−t) + 1 − Φ(t).

396

6. Hypothesis Tests

Thus, t = Φ−1 (1 − α/2) and T∗ (X) = I(t,∞) (|X|). Note that it is not necessary to find out what c is. Intuitively, the reason why the UMP test in this example rejects H0 when |X| is large is that the probability of getting a large |X| is much higher under H1 (i.e., P is the double exponential distribution DE(0, 2)). The power of T∗ when P ∈ P1 is Z 1 t −|x|/2 e dx = e−t/2 . E1 [T∗ (X)] = P1 (|X| > t) = 1 − 4 −t Example 6.2. Let X1 , ..., Xn be i.i.d. binary random variables with p = P (X1 = 1). Suppose that H0 : p = p0 and H1 : p = p1 , where 0 < p0 < p1 < 1. By Theorem 6.1, a UMP test of size α is λ(Y ) > c 1 T∗ (Y ) = γ λ(Y ) = c 0 λ(Y ) < c, Pn where Y = i=1 Xi and n−Y Y 1 − p1 p1 . λ(Y ) = p0 1 − p0 Since λ(Y ) is increasing in Y , there is an integer m > 0 such that Y >m 1 T∗ (Y ) = γ Y =m 0 Y < m,

where m and γ satisfy α = E0 [T∗ (Y )] = P0 (Y > m) + γP0 (Y = m). Since Y has the binomial distribution Bi(p, n), we can determine m and γ from n X n j n m α= p0 (1 − p0 )n−j + γ p0 (1 − p0 )n−m . (6.7) j m j=m+1 Unless α=

n X n j p0 (1 − p0 )n−j j j=m+1

for some integer m, in which case we can choose γ = 0, the UMP test T∗ is a randomized test. An interesting phenomenon in Example 6.2 is that the UMP test T∗ does not depend on p1 . In such a case, T∗ is in fact a UMP test for testing H0 : p = p0 versus H1 : p > p0 .

6.1. UMP Tests

397

Lemma 6.1. Suppose that there is a test T∗ of size α such that for every P1 ∈ P1 , T∗ is UMP for testing H0 versus the hypothesis P = P1 . Then T∗ is UMP for testing H0 versus H1 . Proof. For any test T of level α, T is also of level α for testing H0 versus the hypothesis P = P1 with any P1 ∈ P1 . Hence βT ∗ (P1 ) ≥ βT (P1 ). We conclude this section with the following generalized Neyman-Pearson lemma. Its proof is left to the reader. Other extensions of the NeymanPearson lemma can be found in Exercises 8 and 9 in §6.6. Proposition 6.1. Let f1 , ..., fm+1 be Borel functions on Rp that are integrable w.r.t. a σ-finite measure ν. For given constants t1 , ..., tm , let T be the class of Borel functions φ (from Rp to [0, 1]) satisfying Z φfi dν ≤ ti , i = 1, ..., m, (6.8) and T0 be the set of φ’s in T satisfying (6.8) with all inequalities replaced by equalities. If there are constants c1 , ..., cm such that 1 fm+1 (x) > c1 f1 (x) + · · · + cm fm (x) φ∗ (x) = (6.9) 0 fm+1 (x) < c1 f1 (x) + · · · + cm fm (x) R is a member of T0 , then φR∗ maximizes φfm+1 dν over φ ∈ T0 . If ci ≥ 0 for all i, then φ∗ maximizes φfm+1 dν over φ ∈ T .

The existence of constants ci ’s in (6.9) is considered in the following lemma whose proof can be found in Lehmann (1986, pp. 97-99).

Lemma 6.2. R Let f1 , ...,Rfm and ν be given by Proposition 6.1. Then the set M = ( φf1 dν, ..., φfm dν) : φ is from Rp to [0, 1] is convex and closed. If (t1 , ..., tm ) is an interior point of M , then there exist constants c1 , ..., cm such that the function defined by (6.9) is in T0 .

6.1.2 Monotone likelihood ratio The case where both H0 and H1 are simple is mainly of theoretical interest. If a hypothesis is not simple, it is called composite. As we discussed in §6.1.1, UMP tests for composite H1 exist in the problem discussed in Example 6.2. We now extend this result to a class of parametric problems in which the likelihood functions have a special property. Definition 6.2. Suppose that the distribution of X is in P = {Pθ : θ ∈ Θ}, a parametric family indexed by a real-valued θ, and that P is dominated by a σ-finite measure ν. Let fθ = dPθ /dν. The family P is said to have

398

6. Hypothesis Tests

monotone likelihood ratio in Y (X) (a real-valued statistic) if and only if, for any θ1 < θ2 , fθ2 (x)/fθ1 (x) is a nondecreasing function of Y (x) for values x at which at least one of fθ1 (x) and fθ2 (x) is positive. The following lemma states a useful result for a family with monotone likelihood ratio. Lemma 6.3. Suppose that the distribution of X is in a parametric family P indexed by a real-valued θ and that P has monotone likelihood ratio in Y (X). If ψ is a nondecreasing function of Y , then g(θ) = E[ψ(Y )] is a nondecreasing function of θ. Proof. Let θ1 < θ2 , A = {x : fθ1 (x) > fθ2 (x)}, a = supx∈A ψ(Y (x)), B = {x : fθ1 (x) < fθ2 (x)}, and b = inf x∈B ψ(Y (x)). Since P has monotone likelihood ratio in Y (X) and ψ is nondecreasing in Y , b ≥ a. Then the result follows from Z g(θ2 ) − g(θ1 ) = ψ(Y (x))(fθ2 − fθ1 )(x)dν Z Z ≥ a (fθ2 − fθ1 )(x)dν + b (fθ2 − fθ1 )(x)dν A B Z = (b − a) (fθ2 − fθ1 )(x)dν B

≥ 0.

Before discussing UMP tests in families with monotone likelihood ratio, let us consider some examples of such families. Example 6.3. Let θ be real-valued and η(θ) be a nondecreasing function of θ. Then the one-parameter exponential family with fθ (x) = exp{η(θ)Y (x) − ξ(θ)}h(x)

(6.10)

has monotone likelihood ratio in Y (X). From Tables 1.1-1.2 (§1.3.1), this includes the binomial family {Bi(θ, r)}, the Poisson family {P (θ)}, the negative binomial family {N B(θ, r)}, the log-distribution family {L(θ)}, the normal family {N (θ, c2 )} or {N (c, θ)}, the exponential family {E(c, θ)}, the gamma family {Γ(θ, c)} or {Γ(c, θ)}, the beta family {B(θ, c)} or {B(c, θ)}, and the double exponential family {DE(c, θ)}, where r or c is known. Example 6.4. Let X1 , ..., Xn be i.i.d. from the uniform distribution on (0, θ), where θ > 0. The Lebesgue p.d.f. of X = (X1 , ..., Xn ) is fθ (x) = θ−n I(0,θ) (x(n) ), where x(n) is the value of the largest order statistic X(n) . For θ1 < θ2 , fθ2 (x) θn I(0,θ2 ) (x(n) ) = 1n , fθ1 (x) θ2 I(0,θ1 ) (x(n) )

6.1. UMP Tests

399

which is a nondecreasing function of x(n) for x’s at which at least one of fθ1 (x) and fθ2 (x) is positive, i.e., x(n) < θ2 . Hence the family of distributions of X has monotone likelihood ratio in X(n) . Example 6.5. The following families have monotone likelihood ratio: (a) the double exponential distribution family {DE(θ, c)} with a known c; (b) the exponential distribution family {E(θ, c)} with a known c; (c) the logistic distribution family {LG(θ, c)} with a known c; (d) the uniform distribution family {U (θ, θ + 1)}; (e) the hypergeometric distribution family {HG(r, θ, N − θ)} with known r and N (Table 1.1, page 18). An example of a family that does not have monotone likelihood ratio is the Cauchy distribution family {C(θ, c)} with a known c. Hypotheses of the form H0 : θ ≤ θ0 (or H0 : θ ≥ θ0 ) versus H1 : θ > θ0 (or H1 : θ < θ0 ) are called one-sided hypotheses for any given constant θ0 . The following result provides UMP tests for testing one-sided hypotheses when the distribution of X is in a parametric family with monotone likelihood ratio. Theorem 6.2. Suppose that X has a distribution in P = {Pθ : θ ∈ Θ} (Θ ⊂ R) that has monotone likelihood ratio in Y (X). Consider the problem of testing H0 : θ ≤ θ0 versus H1 : θ > θ0 , where θ0 is a given constant. (i) There exists a UMP test of size α, which is given by Y (X) > c 1 T∗ (X) = (6.11) γ Y (X) = c 0 Y (X) < c,

where c and γ are determined by βT∗ (θ0 ) = α, and βT (θ) = E[T (X)] is the power function of a test T . (ii) βT∗ (θ) is strictly increasing for all θ’s for which 0 < βT∗ (θ) < 1. (iii) For any θ < θ0 , T∗ minimizes βT (θ) (the type I error probability of T ) among all tests T satisfying βT (θ0 ) = α. (iv) Assume that Pθ (fθ (X) = cfθ0 (X)) = 0 for any θ > θ0 and c ≥ 0, where fθ is the p.d.f. of Pθ . If T is a test with βT (θ0 ) = βT∗ (θ0 ), then for any θ > θ0 , either βT (θ) < βT∗ (θ) or T = T∗ a.s. Pθ . (v) For any fixed θ1 , T∗ is UMP for testing H0 : θ ≤ θ1 versus H1 : θ > θ1 , with size βT∗ (θ1 ). Proof. (i) Consider the hypotheses θ = θ0 versus θ = θ1 with any θ1 > θ0 . From Theorem 6.1, a UMP test is given by (6.3) with fj = the p.d.f. of Pθj , j = 0, 1. Since P has monotone likelihood ratio in Y (X), this UMP test can be chosen to be the same as T∗ in (6.11) with possibly different c and γ satisfying βT∗ (θ0 ) = α. Since T∗ does not depend on θ1 , it follows from

400

6. Hypothesis Tests

Lemma 6.1 that T∗ is UMP for testing the hypothesis θ = θ0 versus H1 . Note that if T∗ is UMP for testing θ = θ0 versus H1 , then it is UMP for testing H0 versus H1 , provided that βT∗ (θ) ≤ α for all θ ≤ θ0 , i.e., the size of T∗ is α. But this follows from Lemma 6.3, i.e., βT∗ (θ) is nondecreasing in θ. This proves (i). (ii) See Exercise 2 in §6.6. (iii) The result can be proved using Theorem 6.1 with all inequalities reversed. (iv) The proof for (iv) is left as an exercise. (v) The proof for (v) is similar to that of (i). By reversing inequalities throughout, we can obtain UMP tests for testing H0 : θ ≥ θ0 versus H1 : θ < θ0 . A major application of Theorem 6.2 is to problems with one-parameter exponential families. Corollary 6.1. Suppose that X has the p.d.f. given by (6.10) w.r.t. a σ-finite measure, where η is a strictly monotone function of θ. If η is increasing, then T∗ given by (6.11) is UMP for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 , where γ and c are determined by βT∗ (θ0 ) = α. If η is decreasing or H0 : θ ≥ θ0 (H1 : θ < θ0 ), the result is still valid by reversing inequalities in (6.11). Example 6.6. Let X1 , ..., Xn be i.i.d. from the N (µ, σ 2 ) distribution with an unknown µ ∈ R and a known σ 2 . Consider H0 : µ ≤ µ0 versus H1 : µ > µ0 , where µ0 is a fixed constant. The p.d.f. of X = (X1 , ..., Xn ) is of ¯ and η(µ) = nµ/σ 2 . By Corollary 6.1 and the form (6.10) with Y (X) = X 2 ¯ is N (µ, σ /n), the UMP test is T∗ (X) = I(c ,∞) (X), ¯ where the fact that √ X α cα = σz1−α / n + µ0 and za = Φ−1 (a) (see also Example 2.28). To derive a UMP test for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 when X has the p.d.f. (6.10), it is essential to know the distribution of Y (X). Typically, a nonrandomized test can be obtained if the distribution of Y is continuous; otherwise UMP tests are randomized. Example 6.7. Let X1 , ..., Xn be i.i.d. binary random variables with p = P 1 = 1). The p.d.f. of X = (X1 , ..., Xn ) is of the form (6.10) with Y = P(X n p i=1 Xi and η(p) = log 1−p . Note that η(p) is a strictly increasing function of p. By Corollary 6.1, a UMP test for H0 : p ≤ p0 versus H1 : p > p0 is given by (6.11), where c and γ are determined by (6.7) with c = m. Example 6.8. Let X1 , ..., Xn be i.i.d. random variables from the Poisson distribution P (θ) with an unknown θ > 0. The p.d.f. of X = (X1 , ..., Xn )

401

6.1. UMP Tests

P is of the form (6.10) with Y (X) = ni=1 Xi and η(θ) = log θ. Note that Y has the Poisson distribution P (nθ). By Corollary 6.1, a UMP test for H0 : θ ≤ θ0 versus H1 : θ > θ0 is given by (6.11) with c and γ satisfying α=

∞ X enθ0 (nθ0 )j enθ0 (nθ0 )c +γ . j! c! j=c+1

Example 6.9. Let X1 , ..., Xn be i.i.d. random variables from the uniform distribution U (0, θ), θ > 0. Consider the hypotheses H0 : θ ≤ θ0 and H1 : θ > θ0 . Since the p.d.f. of X = (X1 , ..., Xn ) is in a family with monotone likelihood ratio in Y (X) = X(n) (Example 6.4), by Theorem 6.2, a UMP test is of the form (6.11). Since X(n) has the Lebesgue p.d.f. nθ−n xn−1 I(0,θ) (x), the UMP test in (6.11) is nonrandomized and Z θ0 n cn xn−1 dx = 1 − n . α = βT∗ (θ0 ) = n θ0 c θ0 Hence c = θ0 (1 − α)1/n . The power function of T∗ when θ > θ0 is Z θ n θn (1 − α) βT∗ (θ) = n xn−1 dx = 1 − 0 n . θ c θ In this problem, however, UMP tests are not unique. (Note that the condition Pθ (fθ (X) = cfθ0 (X)) = 0 in Theorem 6.2(iv) is not satisfied.) It can be shown (exercise) that the following test is also UMP with size α: 1 X(n) > θ0 T (X) = α X(n) ≤ θ0 .

6.1.3 UMP tests for two-sided hypotheses The following hypotheses are called two-sided hypotheses: H0 : θ ≤ θ1 or θ ≥ θ2 H0 : θ 1 ≤ θ ≤ θ 2

H0 : θ = θ 0

versus H1 : θ1 < θ < θ2 ,

(6.12)

versus H1 : θ < θ1 or θ > θ2 ,

(6.13)

versus H1 : θ 6= θ0 ,

(6.14)

where θ0 , θ1 , and θ2 are given constants and θ1 < θ2 .

Theorem 6.3. Suppose that X has the p.d.f. given by (6.10) w.r.t. a σfinite measure, where η is a strictly increasing function of θ. (i) For testing hypotheses (6.12), a UMP test of size α is c1 < Y (X) < c2 1 T∗ (X) = (6.15) γ Y (X) = ci , i = 1, 2 i 0 Y (X) < c1 or Y (X) > c2 ,

402

6. Hypothesis Tests

where ci ’s and γi ’s are determined by βT∗ (θ1 ) = βT∗ (θ2 ) = α.

(6.16)

(ii) The test defined by (6.15) minimizes βT (θ) over all θ < θ1 , θ > θ2 , and T satisfying βT (θ1 ) = βT (θ2 ) = α. (iii) If T∗ and T∗∗ are two tests satisfying (6.15) and βT∗ (θ1 ) = βT∗∗ (θ1 ) and if the region {T∗∗ = 1} is to the right of {T∗ = 1}, then βT∗ (θ) < βT∗∗ (θ) for θ > θ1 and βT∗ (θ) > βT∗∗ (θ) for θ < θ1 . If both T∗ and T∗∗ satisfy (6.15) and (6.16), then T∗ = T∗∗ a.s. P. Proof. (i) The distribution of Y has a p.d.f. gθ (y) = exp{η(θ)y − ξ(θ)}

(6.17)

(Theorem 2.1). Since Y is sufficient for θ, we only need to consider tests of the form T (Y ). Let θ1 < θ3 < θ2 . Consider the problem of testing θ = θ1 or θ = θ2 versus θ = θ3 . Clearly, (α, α) is an interior point of the set of all points (βT (θ1 ), βT (θ2 )) as T ranges over all tests of the form T (Y ). By (6.17) and Lemma 6.2, there are constants c˜1 and c˜2 such that 1 a1 eb1 Y + a2 eb2 Y < 1 T∗ (Y ) = 0 a1 eb1 Y + a2 eb2 Y > 1 satisfies (6.16), where ai = c˜i eξ(θ3 )−ξ(θi ) and bi = η(θi ) − η(θ3 ), i = 1, 2. Clearly ai ’s cannot both be ≤ 0. If one of the ai ’s is ≤ 0 and the other is > 0, then a1 eb1 Y + a2 eb2 Y is strictly monotone (since b1 < 0 < b2 ) and T∗ or 1 − T∗ is of the form (6.11), which has a strictly monotone power function (Theorem 6.2) and, therefore, cannot satisfy (6.16). Thus, both ai ’s are positive. Then, T∗ is of the form (6.15) (since b1 < 0 < b2 ) and it follows from Proposition 6.1 that T∗ is UMP for testing θ = θ1 or θ = θ2 versus θ = θ3 . Since T∗ does not depend on θ3 , it follows from Lemma 6.1 that T∗ is UMP for testing θ = θ1 or θ = θ2 versus H1 . To show that T∗ is a UMP test of size α for testing H0 versus H1 , it remains to show that βT∗ (θ) ≤ α for θ ≤ θ1 or θ ≥ θ2 . But this follows from part (ii) of the theorem by comparing T∗ with the test T (Y ) ≡ α. (ii) The proof is similar to that in (i) and is left as an exercise. (iii) The first claim in (iii) follows from Lemma 6.4, since the function T∗∗ − T∗ has a single change of sign. The second claim in (iii) follows from the first claim. Lemma 6.4. Suppose that X has a p.d.f. in {fθ (x) : θ ∈ Θ}, a parametric family of p.d.f.’s w.r.t. a single σ-finite measure ν on R, where Θ ⊂ R. Suppose that this family has monotone likelihood ratio in X. Let ψ be a function with a single change of sign. (i) There exists θ0 ∈ Θ such that Eθ [ψ(X)] ≤ 0 for θ < θ0 and Eθ [ψ(X)] ≥ 0

403

6.1. UMP Tests

for θ > θ0 , where Eθ is the expectation w.r.t. fθ . (ii) Suppose that fθ (x) > 0 for all x and θ, that fθ1 (x)/fθ (x) is strictly increasing in x for θ < θ1 , and that ν({x : ψ(x) 6= 0}) > 0. If Eθ0 [ψ(X)] = 0, then Eθ [ψ(X)] < 0 for θ < θ0 and Eθ [ψ(X)] > 0 for θ > θ0 . Proof. (i) Suppose that there is an x0 ∈ R such that ψ(x) ≤ 0 for x < x0 and ψ(x) ≥ 0 for x > x0 . Let θ1 < θ2 . We first show that Eθ1 [ψ(X)] > 0 implies Eθ2 [ψ(X)] ≥ 0. If fθ2 (x0 )/fθ1 (x0 ) = ∞, then fθ1 (x) = 0 for x ≥ x0 and, therefore, Eθ1 [ψ(X)] ≤ 0. Hence fθ2 (x0 )/fθ1 (x0 ) = c < ∞. Then ψ(x) ≥ 0 on the set A = {x : fθ1 (x) = 0 and fθ2 (x) > 0}. Thus, Z fθ Eθ2 [ψ(X)] ≥ ψ 2 fθ1 dν fθ1 c Z ZA cψfθ1 dν + cψfθ1 dν (6.18) ≥ x<x0

x≥x0

= cEθ1 [ψ(X)].

The result follows by letting θ0 = inf{θ : Eθ [ψ(X)] > 0}. (ii) Under the assumed conditions, fθ2 (x0 )/fθ1 (x0 ) = c < ∞. The result follows from the proof in (i) with θ1 replaced by θ0 and the fact that ≥ should be replaced by > in (6.18) under the assumed conditions. Part (iii) of Theorem 6.3 shows that the ci ’s and γi ’s are uniquely determined by (6.15) and (6.16). It also indicates how to select the ci ’s and (0) (0) (0) (0) γi ’s. One can start with some trial values c1 and γ1 , find c2 and γ2 such that βT∗ (θ1 ) = α, and compute βT∗ (θ2 ). If βT∗ (θ2 ) < α, by Theorem 6.3(iii), the correct rejection region {T∗ = 1} is to the right of the one (1) (0) (1) (0) (1) (0) chosen so that one should try c1 > c1 or c1 = c1 and γ1 < γ1 ; the converse holds if βT∗ (θ2 ) > α. Example 6.10. Let X1 , ..., Xn be i.i.d. from N (θ, 1). By Theorem 6.3, a ¯ where ci ’s are deterUMP test for testing (6.12) is T∗ (X) = I(c1 ,c2 ) (X), mined by √ √ Φ n(c2 − θ1 ) − Φ n(c1 − θ1 ) = α and

Φ

√ √ n(c2 − θ2 ) − Φ n(c1 − θ2 ) = α.

When the distribution of X is not given by (6.10), UMP tests for hypotheses (6.12) exist in some cases (see Exercises 17 and 26). Unfortunately, a UMP test does not exist in general for testing hypotheses (6.13) or (6.14) (Exercises 28 and 29). A key reason for this phenomenon is that UMP tests for testing one-sided hypotheses do not have level α for testing (6.12); but they are of level α for testing (6.13) or (6.14) and there does not exist a single test more powerful than all tests that are UMP for testing one-sided hypotheses.

404

6. Hypothesis Tests

6.2 UMP Unbiased Tests When a UMP test does not exist, we may use the same approach used in estimation problems, i.e., imposing a reasonable restriction on the tests to be considered and finding optimal tests within the class of tests under the restriction. Two such types of restrictions in estimation problems are unbiasedness and invariance. We consider unbiased tests in this section. The class of invariant tests is studied in §6.3.

6.2.1 Unbiasedness, similarity, and Neyman structure A UMP test T of size α has the property that βT (P ) ≤ α,

P ∈ P0

and

βT (P ) ≥ α,

P ∈ P1 .

(6.19)

This means that T is at least as good as the silly test T ≡ α. Thus, we have the following definition. Definition 6.3. Let α be a given level of significance. A test T for H0 : P ∈ P0 versus H1 : P ∈ P1 is said to be unbiased of level α if and only if (6.19) holds. A test of size α is called a uniformly most powerful unbiased (UMPU) test if and only if it is UMP within the class of unbiased tests of level α. Since a UMP test is UMPU, the discussion of unbiasedness of tests is useful only when a UMP test does not exist. In a large class of problems for which a UMP test does not exist, there do exist UMPU tests. Suppose that U is a sufficient statistic for P ∈ P. Then, similar to the search for a UMP test, we need to consider functions of U only in order to find a UMPU test, since, for any unbiased test T (X), E(T |U ) is unbiased and has the same power function as T . Throughout this section, we consider the following hypotheses: H0 : θ ∈ Θ 0

versus

H1 : θ ∈ Θ 1 ,

(6.20)

where θ = θ(P ) is a functional from P onto Θ and Θ0 and Θ1 are two disjoint Borel sets with Θ0 ∪ Θ1 = Θ. Note that Pj = {P : θ ∈ Θj }, j = 0, 1. For instance, X1 , ..., Xn are i.i.d. from F but we are interested in testing H0 : θ ≤ 0 versus H1 : θ > 0, where θ = EX1 or the median of F . Definition 6.4. Consider the hypotheses specified by (6.20). Let α be a ¯ 01 be the common boundary of Θ0 and given level of significance and let Θ Θ1 , i.e., the set of points θ that are points or limit points of both Θ0 and ¯ 01 if and only if Θ1 . A test T is similar on Θ ¯ 01 . βT (P ) = α, θ∈Θ (6.21)

405

6.2. UMP Unbiased Tests

It is more convenient to work with (6.21) than to work with (6.19) when the hypotheses are given by (6.20). Thus, the following lemma is useful. For a given test T , the power function βT (P ) is said to be continuous in θ if and only if for any {θj : j = 0, 1, 2, ...} ⊂ Θ, θj → θ0 implies βT (Pj ) → βT (P0 ), where Pj ∈ P satisfying θ(Pj ) = θj , j = 0, 1,.... Note that if βT is a function of θ, then this continuity property is simply the continuity of βT (θ). Lemma 6.5. Consider hypotheses (6.20). Suppose that, for every T , βT (P ) is continuous in θ. If T∗ is uniformly most powerful among all tests satisfying (6.21) and has size α, then T∗ is a UMPU test. Proof. Under the continuity assumption on βT , the class of tests satisfying (6.21) contains the class of tests satisfying (6.19). Since T∗ is uniformly at least as powerful as the test T ≡ α, T∗ is unbiased. Hence, T∗ is a UMPU test. Using Lemma 6.5, we can derive a UMPU test for testing hypotheses given by (6.13) or (6.14), when X has the p.d.f. (6.10) in a one-parameter exponential family. (Note that a UMP test does not exist in these cases.) We do not provide the details here, since the results for one-parameter exponential families are special cases of those in §6.2.2 for multiparameter exponential families. To prepare for the discussion in §6.2.2, we introduce the following result that simplifies (6.21) when there is a statistic sufficient ¯ 01 }. and complete for P ∈ P¯ = {P : θ(P ) ∈ Θ Let U (X) be a sufficient statistic for P ∈ P¯ and let P¯U be the family of ¯ If T is a test satisfying distributions of U as P ranges over P. E[T (X)|U ] = α

a.s. P¯U ,

(6.22)

then

¯ E[T (X)] = E{E[T (X)|U ]} = α P ∈ P, ¯ 01 . A test satisfying (6.22) is said to have Neyman i.e., T is similar on Θ ¯ 01 have Neyman structure w.r.t. structure w.r.t. U . If all tests similar on Θ U , then working with (6.21) is the same as working with (6.22). ¯ Then a necLemma 6.6. Let U (X) be a sufficient statistic for P ∈ P. ¯ essary and sufficient condition for all tests similar on Θ01 to have Neyman ¯ structure w.r.t. U is that U is boundedly complete for P ∈ P. ¯ Let Proof. (i) Suppose first that U is boundedly complete for P ∈ P. ¯ ¯ T (X) be a test similar on Θ01 . Then E[T (X) − α] = 0 for all P ∈ P. From the boundedness of T (X), E[T (X)|U ] is bounded (Proposition 1.10). Since ¯ (6.22) holds. E{E[T (X)|U ] − α} = E[T (X) − α] = 0 for all P ∈ P, ¯ Then (ii) Suppose now that U is not boundedly complete for P ∈ P. ¯ and there is a function h such that |h(u)| ≤ C, E[h(U )] = 0 for all P ∈ P, ¯ h(U ) 6= 0 with positive probability for some P ∈ P. Let T (X) = α + ch(U ),

406

6. Hypothesis Tests

where c = min{α, 1 − α}/C. The result follows from the fact that T is a ¯ 01 but does not have Neyman structure w.r.t. U . test similar on Θ

6.2.2 UMPU tests in exponential families Suppose that the distribution of X is in a multiparameter natural exponential family (§2.1.3) with the following p.d.f. w.r.t. a σ-finite measure: fθ,ϕ(x) = exp {θY (x) + ϕτ U (x) − ζ(θ, ϕ)} ,

(6.23)

where θ is a real-valued parameter, ϕ is a vector-valued parameter, and Y (real-valued) and U (vector-valued) are statistics. It follows from Theorem 2.1(i) that the p.d.f. of (Y, U ) (w.r.t. a σ-finite measure) is in a natural exponential family of the form exp {θy + ϕτ u − ζ(θ, ϕ)} and, given U = u, the p.d.f. of the conditional distribution of Y (w.r.t. a σ-finite measure νu ) is in a natural exponential family of the form exp {θy − ζu (θ)}. Theorem 6.4. Suppose that the distribution of X is in a multiparameter natural exponential family given by (6.23). (i) For testing H0 : θ ≤ θ0 versus H1 : θ > θ0 , a UMPU test of size α is Y > c(U ) 1 (6.24) T∗ (Y, U ) = γ(U ) Y = c(U ) 0 Y < c(U ),

where c(u) and γ(u) are Borel functions determined by Eθ0 [T∗ (Y, U )|U = u] = α

for every u, and Eθ0 is the expectation w.r.t. fθ0 ,ϕ . (ii) For testing hypotheses (6.12), a UMPU test of size α is c1 (U ) < Y < c2 (U ) 1 T∗ (Y, U ) = Y = ci (U ), i = 1, 2, γi (U ) 0 Y < c1 (U ) or Y > c2 (U ),

(6.25)

(6.26)

where ci (u)’s and γi (u)’s are Borel functions determined by Eθ1 [T∗ (Y, U )|U = u] = Eθ2 [T∗ (Y, U )|U = u] = α

for every u. (iii) For testing hypotheses (6.13), a UMPU test of size α is Y < c1 (U ) or Y > c2 (U ) 1 T∗ (Y, U ) = Y = ci (U ), i = 1, 2, γ (U ) i 0 c1 (U ) < Y < c2 (U ),

(6.27)

(6.28)

407

6.2. UMP Unbiased Tests

where ci (u)’s and γi (u)’s are Borel functions determined by (6.27) for every u. (iv) For testing hypotheses (6.14), a UMPU test of size α is given by (6.28), where ci (u)’s and γi (u)’s are Borel functions determined by (6.25) and Eθ0 [T∗ (Y, U )Y |U = u] = αEθ0 (Y |U = u)

(6.29)

for every u. Proof. Since (Y, U ) is sufficient for (θ, ϕ), we only need to consider tests that are functions of (Y, U ). Hypotheses in (i)-(iv) are of the form (6.20) ¯ 01 = {(θ, ϕ) : θ = θ0 } or = {(θ, ϕ) : θ = θi , i = 1, 2}. In case (i) or with Θ (iv), U is sufficient and complete for P ∈ P¯ and, hence, Lemma 6.6 applies. In case (ii) or (iii), applying Lemma 6.6 to each {(θ, ϕ) : θ = θi } also shows that working with (6.21) is the same as working with (6.22). By Theorem 2.1, the power functions of all tests are continuous and, hence, Lemma 6.5 applies. Thus, for (i)-(iii), we only need to show that T∗ is UMP among all tests T satisfying (6.25) (for part (i)) or (6.27) (for part (ii) or (iii)) with T∗ replaced by T . For (iv), any unbiased T should satisfy (6.25) with T∗ replaced by T and ∂ Eθ,ϕ [T (Y, U )] = 0, ∂θ

¯ 01 . θ∈Θ

(6.30)

By Theorem 2.1, the differentiation can be carried out under the expectation sign. Hence, one can show (exercise) that (6.30) is equivalent to Eθ,ϕ [T (Y, U )Y − αY ] = 0,

¯ 01 . θ∈Θ

(6.31)

Using the argument in the proof of Lemma 6.6, one can show (exercise) that (6.31) is equivalent to (6.29) with T∗ replaced by T . Hence, to prove (iv) we only need to show that T∗ is UMP among all tests T satisfying (6.25) and (6.29) with T∗ replaced by T . Note that the power function of any test T (Y, U ) is Z Z βT (θ, ϕ) = T (y, u)dPY |U=u (y) dPU (u). Thus, it suffices to show that for every fixed u and θ ∈ Θ1 , T∗ maximizes Z T (y, u)dPY |U =u (y) over all T subject to the given side conditions. Since PY |U=u is in a one-parameter exponential family, the results in (i) and (ii) follow from Corollary 6.1 and Theorem 6.3, respectively. The result in (iii) follows from Theorem 6.3(ii) by considering 1 − T∗ with T∗ given by (6.15). To

408

6. Hypothesis Tests

prove the result in (iv), it suffices to show that if Y has the p.d.f. given by (6.10) and if U is treated as a constant in (6.25), (6.28), and (6.29), T∗ in (6.28) is UMP subject to conditions (6.25) and (6.29). We now omit U in the following proof for (iv), which is very similar to the proof of Theorem 6.3. First, (α, αEθ0 (Y )) is an interior point of the set of points (Eθ0 [T (Y )], Eθ0 [T (Y )Y ]) as T ranges over all tests of the form T (Y ) (exercise). By Lemma 6.2 and Proposition 6.1, for testing θ = θ0 versus θ = θ1 , the UMP test is equal to 1 when (k1 + k2 y)eθ0 y < C(θ0 , θ1 )eθ1 y ,

(6.32)

where ki ’s and C(θ0 , θ1 ) are constants. Note that (6.32) is equivalent to a1 + a2 y < eby for some constants a1 , a2 , and b. This region is either one-sided or the outside of an interval. By Theorem 6.2(ii), a one-sided test has a strictly monotone power function and therefore cannot satisfy (6.29). Thus, this test must have the form (6.28). Since T∗ in (6.28) does not depend on θ1 , by Lemma 6.1, it is UMP over all tests satisfying (6.25) and (6.29); in particular, the test ≡ α. Thus, T∗ is UMPU. Finally, it can be shown that all the c- and γ-functions in (i)-(iv) are Borel functions (see Lehmann (1986, p. 149)). Example 6.11. A problem arising in many different contexts is the comparison of two treatments. If the observations are integer-valued, the problem often reduces to testing the equality of two Poisson distributions (e.g., a comparison of the radioactivity of two substances or the car accident rate in two cities) or two binomial distributions (when the observation is the number of successes in a sequence of trials for each treatment). Consider first the Poisson problem in which X1 and X2 are independently distributed as the Poisson distributions P (λ1 ) and P (λ2 ), respectively. The p.d.f. of X = (X1 , X2 ) is e−(λ1 +λ2 ) exp {x2 log(λ2 /λ1 ) + (x1 + x2 ) log λ1 } x1 !x2 !

(6.33)

w.r.t. the counting measure on {(i, j) : i = 0, 1, 2, ..., j = 0, 1, 2, ...}. Let θ = log(λ2 /λ1 ). Then hypotheses such as λ1 = λ2 and λ1 ≥ λ2 are equivalent to θ = 0 and θ ≤ 0, respectively. The p.d.f. in (6.33) is of the form (6.23) with ϕ = log λ1 , Y = X2 , and U = X1 + X2 . Thus, Theorem 6.4 applies. To obtain various tests in Theorem 6.4, it is enough to derive the conditional distribution of Y = X2 given U = X1 + X2 = u. Using the fact that X1 + X2 has the Poisson distribution P (λ1 + λ2 ), one can show that u y P (Y = y|U = u) = p (1 − p)u−y I{0,1,...,u} (y), u = 0, 1, 2, ..., y

409

6.2. UMP Unbiased Tests

where p = λ2 /(λ1 + λ2 ) = eθ /(1 + eθ ). This is the binomial distribu¯ 01 , θ = θj (a known value) and the tion Bi(p, u). On the boundary set Θ distribution PY |U =u is known. The previous result can obviously be extended to the case where two independent samples, Xi1 , ..., Xini , i = 1, 2, are i.i.d. from the Poisson distributions P (λi ), i = 1, 2, respectively. Consider next the binomial problem in which Xj , j = 1, 2, are independently distributed as the binomial distributions Bi(pj , nj ), j = 1, 2, respectively, where nj ’s are known but pj ’s are unknown. The p.d.f. of X = (X1 , X2 ) is o n n2 n1 p1 1) (1 − p1 )n1 (1 − p2 )n2 exp x2 log pp21 (1−p + (x1 + x2 ) log (1−p (1−p ) ) 2 1 x1 x2 w.r.t. the counting measure on {(i, j) : i = 0, 1, ..., n1 , j = 0, 1, ..., n2 }. This 1) p.d.f. is of the form (6.23) with θ = log pp21 (1−p (1−p2 ) , Y = X2 , and U = X1 +X2 . Thus, Theorem 6.4 applies. Note that hypotheses such as p1 = p2 and p1 ≥ p2 are equivalent to θ = 0 and θ ≤ 0, respectively. Using the joint distribution of (X1 , X2 ), one can show (exercise) that n2 θy n1 e IA (y), P (Y = y|U = u) = Ku (θ) u−y y

u = 0, 1, ..., n1 + n2 ,

where A = {y : y = 0, 1, ..., min{u, n2 }, u − y ≤ n1 } and −1 X n1 n2 eθy . Ku (θ) = u−y y

(6.34)

y∈A

If θ = 0, this distribution reduces to a known distribution: the hypergeometric distribution HG(u, n2 , n1 ) (Table 1.1, page 18). Example 6.12 (2 × 2 contingency tables). Let A and B be two different events in a probability space related to a random experiment. Suppose that n independent trials of the experiment are carried out and that we observe the frequencies of the occurrence of the events A ∩ B, A ∩ B c , Ac ∩ B, and Ac ∩ B c . The results can be summarized in the following 2 × 2 contingency table:

B Bc Total

A X11 X21 m1

Ac X12 X22 m2

Total n1 n2 n

410

6. Hypothesis Tests

The distribution of X = (X11 , X12 , X21 , X22 ) is multinomial (Example 2.7) with probabilities p11 , p12 , p21 , and p22 , where pij = E(Xij )/n. Thus, the p.d.f. of X is o n n! pn22 exp x11 log pp11 + x12 log pp12 + x21 log pp21 22 22 22 x11 !x12 !x21 !x22 ! w.r.t. the counting measure on the range of X. This p.d.f. is clearly of the form (6.23). By Theorem 6.4, we can derive UMPU tests for any parameter of the form θ = a0 log pp11 + a1 log pp12 + a2 log pp21 , 22 22 22 where ai ’s are given constants. In particular, testing independence of A and B is equivalent to the hypotheses H0 : θ = 0 versus H1 : θ 6= 0 when a0 = 1 and a1 = a2 = −1 (exercise). For hypotheses concerning θ with a0 = 1 and a1 = a2 = −1, the p.d.f. of X can be written as (6.23) with Y = X11 and U = (X11 + X12 , X11 + X21 ). A direct calculation shows that P (Y = y|X11 + X12 = n1 , X11 + X21 = m1 ) is equal to n2 n1 eθ(m1 −y) IA (y), Km1 (θ) y m1 − y where A = {y : y = 0, 1, ..., min{m1 , n1 }, m1 − y ≤ n2 } and Ku (θ) is given by (6.34). This distribution is known when θ = θj is known. In particular, for testing independence of A and B, θ = 0 implies that PY |U =u is the hypergeometric distribution HG(m1 , n1 , n2 ), and the UMPU test in Theorem 6.4(iv) is also known as Fisher’s exact test. Suppose that Xij ’s in the 2 × 2 contingency table are from two binomial distributions, i.e., Xi1 is from the binomial distribution Bi(pi , ni ), Xi2 = ni − Xi1 , i = 1, 2, and that Xi1 ’s are independent. Then the UMPU test for independence of A and B previously derived is exactly the same as the UMPU test for p1 = p2 given in Example 6.11. The only difference is that ni ’s are fixed for testing the equality of two binomial distributions, whereas ni ’s are random for testing independence of A and B. This is also true for the general r × c contingency tables considered in §6.4.3.

6.2.3 UMPU tests in normal families An important application of Theorem 6.4 to problems with continuous distributions in exponential families is the derivation of UMPU tests in normal families. The results presented here are the basic justifications for tests in elementary textbooks concerning parameters in normal families. We start with the following lemma, which is useful especially when X is from a population in a normal family.

6.2. UMP Unbiased Tests

411

Lemma 6.7. Suppose that X has the p.d.f. (6.23) and that V (Y, U ) is a statistic independent of U when θ = θj , where θj ’s are known values given in the hypotheses in (i)-(iv) of Theorem 6.4. (i) If V (y, u) is increasing in y for each u, then the UMPU tests in (i)-(iii) of Theorem 6.4 are equivalent to those given by (6.24)-(6.28) with Y and (Y, U ) replaced by V and with ci (U ) and γi (U ) replaced by constants ci and γi , respectively. (ii) If there are Borel functions a(u) > 0 and b(u) such that V (y, u) = a(u)y + b(u), then the UMPU test in Theorem 6.4(iv) is equivalent to that given by (6.25), (6.28), and (6.29) with Y and (Y, U ) replaced by V and with ci (U ) and γi (U ) replaced by constants ci and γi , respectively. Proof. (i) Since V is increasing in y, Y > ci (u) is equivalent to V > di (u) for some di . The result follows from the fact that V is independent of U so that di ’s and γi ’s do not depend on u when Y is replaced by V . (ii) Since V = a(U )Y + b(U ), the UMPU test in Theorem 6.4(iv) is the same as V < c1 (U ) or V > c2 (U ) 1 T∗ (V, U ) = (6.35) V = ci (U ), i = 1, 2, γi (U ) 0 c1 (U ) < V < c2 (U ), subject to Eθ0 [T∗ (V, U )|U = u] = α and V − b(U ) V − b(U ) U = αEθ0 U . Eθ0 T∗ (V, U ) a(U ) a(U )

(6.36)

Under Eθ0 [T∗ (V, U )|U = u] = α, (6.36) is the same as Eθ0 [T∗ (V, U )V |U ] = αEθ0 (V |U ). Since V and U are independent when θ = θ0 , ci (u)’s and γi (u)’s do not depend on u and, therefore, T∗ in (6.35) does not depend on U. If the conditions of Lemma 6.7 are satisfied, then UMPU tests can be derived by working with the distribution of V instead of PY |U=u . In exponential families, a V (Y, U ) independent of U can often be found by applying Basu’s theorem (Theorem 2.4). When we consider normal families, γi ’s can be chosen to be 0 since the c.d.f. of Y given U = u or the c.d.f. of V is continuous. One-sample problems Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ ∈ R and σ 2 > 0, where n ≥ 2. The joint p.d.f. of X = (X1 , ..., Xn ) is ) ( n n 1 X 2 µ X nµ2 1 exp − 2 x + 2 xi − 2 . 2σ i=1 i σ i=1 2σ (2πσ 2 )n/2

412

6. Hypothesis Tests

Consider first hypotheses concerning σ 2 . The Pnp.d.f. of X has the¯ form (6.23) with θ = −(2σ 2 )−1 , ϕ = nµ/σ 2 , Y = i=1 Xi2 , and U = X. By ¯ (Example 2.18), Basu’s theorem, V = (n − 1)S 2 is independent of U = X where S 2 is the sample variance. Also, n X i=1

2

¯ 2, Xi2 = (n − 1)S 2 + nX

i.e., V = Y − nU . Hence the conditions of Lemma 6.7 are satisfied. Since V /σ 2 has the chi-square distribution χ2n−1 (Example 2.18), values of ci ’s for hypotheses in (i)-(iii) of Theorem 6.4 are related to quantiles of χ2n−1 . For testing H0 : θ = θ0 versus H1 : θ 6= θ0 (which is equivalent to testing H0 : σ 2 = σ02 versus H1 : σ 2 6= σ02 ), di = ci /σ02 , i = 1, 2, are determined by Z d2 Z d2 fn−1 (v)dv = 1 − α and vfn−1 (v)dv = (n − 1)(1 − α), d1

d1

where fm is the Lebesgue p.d.f. of the chi-square distribution χ2m . Since vfn−1 (v) = (n − 1)fn+1 (v), d1 and d2 are determined by Z d2 Z d2 fn−1 (v)dv = fn+1 (v)dv = 1 − α. d1

d1

If n − 1 ≈ n + 1, then d1 and d2 are nearly the (α/2)th and (1 − α/2)th quantiles of χ2n−1 , respectively, in which case the UMPU test in Theorem 6.4(iv) is the same as the “equal-tailed” chi-square test for H0 in elementary textbooks. Consider next hypotheses concerning µ. The p.d.f. of X has the form ¯ U = Pn (Xi − µ0 )2 , θ = n(µ − µ0 )/σ 2 , and ϕ = (6.23) with Y = X, i=1 −(2σ 2 )−1 . For testing √ ¯ hypotheses H0 : µ ≤ µ0 versus H1 : µ > µ0 , we take V to be t(X) = n(X − µ0 )/S. By Basu’s theorem, t(X) is independent of U when µ = µ0 . Hence it satisfies the conditions in Lemma 6.7(i). From Examples 1.16 and 2.18, t(X) has the t-distribution tn−1 when µ = µ0 . Thus, c(U ) in Theorem 6.4(i) is the (1 − α)th quantile of tn−1 . For the two-sided√hypotheses H0 : µ = µ0 versus H1 : µ 6= µ0 , the statistic V = ¯ (X−µ 0 )/ U satisfies the conditions in Lemma 6.7(ii) and has a distribution symmetric about 0 when µ = µ0 . Then the UMPU test in Theorem 6.4(iv) rejects H0 when |V | > d, where d satisfies P (|V | > d) = α when µ = µ0 . Since p p t(X) = (n − 1)nV (X) 1 − n[V (X)]2 , the UMPU test rejects H0 if and only if |t(X)| > tn−1,α/2 , where tn−1,α is the (1 − α)th quantile of the t-distribution tn−1 . The UMPU tests derived here are the so-called one-sample t-tests in elementary textbooks. The power function of a one-sample t-test is related to the noncentral t-distribution introduced in §1.3.1 (see Exercise 36).

413

6.2. UMP Unbiased Tests

Two-sample problems The problem of comparing the parameters of two normal distributions arises in the comparison of two treatments, products, and so on (see also Example 6.11). Suppose that we have two independent samples, Xi1 , ..., Xini , i = 1, 2, i.i.d. from N (µi , σi2 ), i = 1, 2, respectively, where ni ≥ 2. The joint p.d.f. of Xij ’s is ni 2 2 X X X 1 ni µi 2 C(µ1 , µ2 , σ12 , σ22 ) exp − x + x ¯i , ij 2σi2 σi2 i=1

j=1

i=1

where x¯i is the sample mean based on xi1 , ..., xini and C(·) is a known function. Consider first the hypothesis H0 : σ22 /σ12 ≤ ∆0 or H0 : σ22 /σ12 = ∆0 . The p.d.f. of Xij ’s is of the form (6.23) with 1 1 1 n1 µ1 n2 µ2 , θ= − , ϕ = − , , 2∆0 σ12 2σ22 2σ12 σ12 σ22 n2 n2 n1 X X X 1 2 2 2 ¯1, X ¯2 . Y = X2j , U = X1j + X2j , X ∆ 0 j=1 j=1 j=1

To apply Lemma 6.7, consider V =

(Y − n2 U3 )/∆0 (n2 − 1)S22 /∆0 = , 2 2 (n1 − 1)S1 + (n2 − 1)S2 /∆0 U1 − n1 U2 − n2 U3 /∆0

where Si2 is the sample variance based on Xi1 , ..., Xini and Uj is the jth component of U . By Basu’s theorem, V and U are independent when θ = 0 (σ22 = ∆0 σ12 ). Since V is increasing and linear in Y , the conditions of Lemma 6.7 are satisfied. Thus, a UMPU test rejects H0 : θ ≤ 0 (which is equivalent to H0 : σ22 /σ12 ≤ ∆0 ) when V > c0 , where c0 satisfies P (V > c0 ) = α when θ = 0; and a UMPU test rejects H0 : θ = 0 (which is equivalent to H0 : σ22 /σ12 = ∆0 ) when V < c1 or V > c2 , where ci ’s satisfy P (c1 < V < c2 ) = 1 − α and E[V T∗ (V )] = αE(V ) when θ = 0. Note that V =

(n2 − 1)F n1 − 1 + (n2 − 1)F

with

F=

S22 /∆0 . S12

It follows from Example 1.16 that F has the F-distribution Fn2 −1,n1 −1 (Table 1.2, page 20) when θ = 0. Since V is a strictly increasing function of F, a UMPU test rejects H0 : θ ≤ 0 when F > Fn2 −1,n1 −1,α , where Fa,b,α is the (1 − α)th quantile of the F-distribution Fa,b . This is the F-test in elementary textbooks.

414

6. Hypothesis Tests

When θ = 0, V has the beta distribution B((n2 − 1)/2, (n1 − 1)/2) and E(V ) = (n2 − 1)/(n1 + n2 − 2) (Table 1.2). Then, E[V T∗ (V )] = αE(V ) when θ = 0 is the same as Z c2 (1 − α)(n2 − 1) = vf(n2 −1)/2,(n1 −1)/2 (v)dv, n1 + n2 − 2 c1 where fa,b is the p.d.f. of the beta distribution B(a, b). Using the fact that vf(n2 −1)/2,(n1 −1)/2 (v) = (n1 + n2 − 2)−1 (n2 − 1)f(n2 +1)/2,(n1 −1)/2 (v), we conclude that a UMPU test rejects H0 : θ = 0 when V < c1 or V > c2 , where c1 and c2 are determined by Z c2 Z c2 1−α= f(n2 −1)/2,(n1 −1)/2 (v)dv = f(n2 +1)/2,(n1 −1)/2 (v)dv. c1

c1

If n2 − 1 ≈ n2 + 1 (i.e., n2 is large), then this UMPU test can be approximated by the F-test that rejects H0 : θ = 0 if and only if F < Fn2 −1,n1 −1,1−α/2 or F > Fn2 −1,n1 −1,α/2 . Consider next the hypothesis H0 : µ1 ≥ µ2 or H0 : µ1 = µ2 . If σ12 6= σ22 , the problem is the so-called Behrens-Fisher problem and is not accessible by the method introduced in this section. We now assume that σ12 = σ22 = σ 2 but σ 2 is unknown. The p.d.f. of Xij ’s is then ni 2 X 1 X n1 µ1 n2 µ2 x2ij + 2 x ¯1 + 2 x ¯2 , C(µ1 , µ2 , σ 2 ) exp − 2 2σ σ σ i=1 j=1

which is of the form (6.23) with µ2 − µ1 , θ = −1 2 (n1 + n−1 2 )σ ¯1, ¯2 − X Y =X

ϕ=

1 n1 µ1 + n2 µ2 , − 2 (n1 + n2 )σ 2 2σ

¯ 1 + n2 X ¯2, U = n1 X

ni 2 X X i=1 j=1

,

2 Xij .

For testing H0 : θ ≤ 0 (i.e., µ1 ≥ µ2 ) versus H1 : θ > 0, we consider V in Lemma 6.7 to be q −1 ¯2 − X ¯1) (X n1 + n−1 2 t(X) = p . (6.37) 2 2 [(n1 − 1)S1 + (n2 − 1)S2 ]/(n1 + n2 − 2) When θ = 0, t(X) is independent of U (Basu’s theorem) and satisfies the conditions in Lemma 6.7(i); the numerator and the denominator of t(X) (after division by σ) are independently distributed as N (0, 1) and

415

6.2. UMP Unbiased Tests

the chi-square distribution χ2n1 +n2 −2 , respectively. Hence t(X) has the tdistribution tn1 +n2 −2 and a UMPU test rejects H0 when t(X) > tn1 +n2 −2,α , where tn1 +n2 −2,α is the (1 − α)th quantile of the t-distribution tn1 +n2 −2 . This is the so-called (one-sided) two-sample t-test. For testing H0 : θ = 0 (i.e., µ1 = µ2 ) versus H1 : θ 6= 0, it follows from a similar argument used in the derivation of the (two-sided) one-sample t-test that a UMPU test rejects H0 when |t(X)| > tn1 +n2 −2,α/2 (exercise). This is the (two-sided) two-sample t-test. The power function of a two-sample t-test is related to a noncentral t-distribution. Normal linear models Consider linear model (3.25) with assumption A1, i.e., X = (X1 , ..., Xn ) is Nn (Zβ, σ 2 In ),

(6.38)

where β is a p-vector of unknown parameters, Z is the n × p matrix whose ith row is the vector Zi , Zi ’s are the values of a p-vector of deterministic covariates, and σ 2 > 0 is an unknown parameter. Assume that n > p and the rank of Z is r ≤ p. Let l ∈ R(Z) (the linear space generated by the rows of Z) and θ0 be a fixed constant. We consider the hypotheses H0 : l τ β ≤ θ 0

versus

H1 : lτ β > θ0

(6.39)

H0 : l τ β = θ 0

versus

H1 : lτ β 6= θ0 .

(6.40)

HΓ = ( Γ1 0 ),

(6.41)

or τ

−

τ

Since H = Z(Z Z) Z is a projection matrix of rank r, there exists an n × n orthogonal matrix Γ such that Γ = ( Γ 1 Γ2 )

and

where Γ1 is n×r and Γ2 is n×(n−r). Let Yj = Γτj X, j = 1, 2. Consider the transformation (Y1 , Y2 ) = Γτ X. Since Γτ Γ = In and X is Nn (Zβ, σ 2 In ), (Y1 , Y2 ) is Nn (Γτ Zβ, σ2 In ). It follows from (6.41) that E(Y2 ) = E(Γτ2 X) = Γτ2 Zβ = Γτ2 HZβ = 0. Let η = Γτ1 Zβ = E(Y1 ). Then the p.d.f. of (Y1 , Y2 ) is τ 1 η Y1 kY1 k2 + kY2 k2 kηk2 . exp − − σ2 2σ 2 2σ 2 (2πσ 2 )n/2

(6.42)

Since l in (6.39) or (6.40) is in R(Z), there exists λ ∈ Rn such that l = Z τ λ. Then lτ βˆ = λτ HX = λτ ΓΓτ HX = λτ Γ1 Γτ1 X = λτ Γ1 Y1 , (6.43)

416

6. Hypothesis Tests

where βˆ is the LSE defined by (3.27). By (6.43) and Theorem 3.6(ii), ˆ = lτ β = λτ Γ1 E(Y1 ) = aτ η, E(lτ β) where a = Γτ1 λ. Let η = (η1 , ..., ηr ) and a = (a1 , ..., ar ). Without loss of generality, we assume that a1 6= 0. Then the p.d.f. in (6.42) is of the form (6.23) with 1 η2 ηr a τ η − θ0 , Y = Y11 , , ϕ = − , , ..., θ= a1 σ 2 2σ 2 σ 2 σ2 2θ0 Y11 a2 Y11 ar Y11 , U = kY1 k2 + kY2 k2 − , Y12 − , ..., Y1r − a1 a1 a1

where Y1j is the jth component of Y1 . By Basu’s theorem, √ n − r(aτ Y1 − θ0 ) t(X) = kY2 k kak

is independent of U when aτ η = lτ β = θ0 . Note that kY2 k2 = SSR in (3.35) and kak2 = λτ Γ1 Γτ1 λ = λτ Hλ = lτ (Z τ Z)− l. Hence, by (6.43), lτ βˆ − θ0 p t(X) = p , lτ (Z τ Z)− l SSR/(n − r)

which has the t-distribution tn−r (Theorem 3.8). Using the same arguments in deriving the one-sample or two-sample t-test, we obtain that a UMPU test for the hypotheses in (6.39) rejects H0 when t(X) > tn−r,α , and that a UMPU test for the hypotheses in (6.40) rejects H0 when |t(X)| > tn−r,α/2 . Testing for independence in the bivariate normal family Suppose that X1 , ..., Xn are i.i.d. from a bivariate normal distribution, i.e., the p.d.f. of X = (X1 , ..., Xn ) is o n kY1 −µ1 k2 ρ(Y1 −µ1 )τ (Y2 −µ2 ) kY2 −µ2 k2 1 √ , (6.44) exp − + − 2 2 2 2 2 σ1 σ2 (1−ρ ) 2σ (1−ρ ) 2σ (1−ρ ) 2 n (2πσ1 σ2

1−ρ )

1

2

where Yj = (X1j , ..., Xnj ) and Xij is the jth component of Xi , j = 1, 2. Testing for independence of the two components of X1 (or Y1 and Y2 ) is equivalent to testing H0 : ρ = 0 versus H1 : ρ 6= 0. In some cases, one may also be interested in the one-sided hypotheses H0 : ρ ≤ 0 versus H1 : ρ > 0. It can be shown (exercise) that the p.d.f. in (6.44) is of the form (6.23) with ρ θ = σ1 σ2 (1−ρ 2 ) and ! n n n n n X X X X X 2 2 Xi1 Xi2 , U= Xi1 , Xi2 , Xi1 , Xi2 . Y = i=1

i=1

i=1

i=1

i=1

417

6.3. UMP Invariant Tests

The hypothesis ρ ≤ 0 is equivalent to θ ≤ 0. The sample correlation coefficient is #1/2 "X n n n X X 2 2 ¯ 1 )(Xi2 − X ¯2) ¯1) ¯2 ) R= (Xi1 − X (Xi1 − X (Xi2 − X , i=1

i=1

i=1

¯ j is the sample mean of X1j , ..., Xnj , and is independent of U when where X ρ = 0 (Basu’s theorem), j = 1, 2. To apply Lemma 6.7, we consider p √ V = n − 2R/ 1 − R2 . (6.45) It can be shown (exercise) that R is linear in Y and that V has the tdistribution tn−2 when ρ = 0. Hence, a UMPU test for H0 : ρ ≤ 0 versus H1 : ρ > 0 rejects H0 when V > tn−2,α and a UMPU test for H0 : ρ = 0 versus H1 : ρ 6= 0 rejects H0 when |V | > tn−2,α/2 , where tn−2,α is the (1 − α)th quantile of the t-distribution tn−2 .

6.3 UMP Invariant Tests In the previous section the unbiasedness principle is considered to derive an optimal test within the class of unbiased tests when a UMP test does not exist. In this section, we study the same problem with unbiasedness replaced by invariance under a given group of transformations. The principles of unbiasedness and invariance often complement each other in that each is successful in cases where the other is not.

6.3.1 Invariance and UMPI tests The invariance principle considered here is similar to that introduced in §2.3.2 (Definition 2.9) and in §4.2. Although a hypothesis testing problem can be treated as a particular statistical decision problem (see, e.g., Example 2.20), in the following definition we define invariant tests without using any loss function which is a basic element in statistical decision theory. However, the reader is encouraged to compare Definition 2.9 with the following definition. Definition 6.5. Let X be a sample from P ∈ P and G be a group (Definition 2.9(i)) of one-to-one transformations of X. (i) We say that the problem of testing H0 : P ∈ P0 versus H1 : P ∈ P1 is invariant under G if and only if both P0 and P1 are invariant under G in the sense of Definition 2.9(ii). (ii) In an invariant testing problem, a test T (X) is said to be invariant

418

6. Hypothesis Tests

under G if and only if T (g(x)) = T (x)

for all x and g.

(6.46)

(iii) A test of size α is said to be a uniformly most powerful invariant (UMPI) test if and only if it is UMP within the class of level α tests that are invariant under G. (iv) A statistic M (X) is said to be maximal invariant under G if and only if (6.46) holds with T replaced by M and M (x1 ) = M (x2 )

implies x1 = g(x2 ) for some g ∈ G.

(6.47)

The following result indicates that invariance reduces the data X to a maximal invariant statistic M (X) whose distribution may depend only on a functional of P that shrinks P. Proposition 6.2. Let M (X) be maximal invariant under G. (i) A test T (X) is invariant under G if and only if there is a function h such that T (x) = h(M (x)) for all x. (ii) Suppose that there is a functional θ(P ) on P satisfying θ(¯ g (P )) = θ(P ) for all g ∈ G and P ∈ P and θ(P1 ) = θ(P2 )

implies P1 = g¯(P2 ) for some g ∈ G

(i.e., θ(P ) is “maximal invariant”), where g¯(PX ) = Pg(X) is given in Definition 2.9(ii). Then the distribution of M (X) depends only on θ(P ). Proof. (i) If T (x) = h(M (x)) for all x, then T (g(x)) = h(M (g(x))) = h(M (x)) = T (x) so that T is invariant. If T is invariant and if M (x1 ) = M (x2 ), then x1 = g(x2 ) for some g and T (x1 ) = T (g(x2 )) = T (x2 ). Hence T is a function of M . (ii) Suppose that θ(P1 ) = θ(P2 ). Then P2 = g¯(P1 ) for some g ∈ G and for any event B in the range of M (X), P2 M (X) ∈ B = g¯(P1 ) M (X) ∈ B = P1 M (g(X)) ∈ B = P1 M (X) ∈ B . Hence the distribution of M (X) depends only on θ(P ).

In applications, maximal invariants M (X) and θ = θ(P ) are frequently real-valued. If the hypotheses of interest can be expressed in terms of θ, then there may exist a test UMP among those depending only on M (X) (e.g., when the distribution of M (X) is in a parametric family having monotone likelihood ratio). Such a test is then a UMPI test.

419

6.3. UMP Invariant Tests

Example 6.13 (Location-scale families). Suppose that X has the Lebesgue p.d.f. fi,µ (x) = fi (x1 − µ, ..., xn − µ), where n ≥ 2, µ ∈ R is unknown, and fi , i = 0, 1, are known Lebesgue p.d.f.’s. We consider the problem of testing H0 : X is from f0,µ

versus

H1 : X is from f1,µ .

(6.48)

Consider G = {gc : c ∈ R} with gc (x) = (x1 + c, ..., xn + c). For any gc ∈ G, it induces a transformation g¯c (fi,µ ) = fi,µ+c and the problem of testing H0 versus H1 in (6.48) is invariant under G. We now show that a maximal invariant under G is D(X) = (D1 , ..., Dn−1 ) = (X1 − Xn , ..., Xn−1 − Xn ). First, it is easy to see that D(X) is invariant under G. Let x = (x1 , ..., xn ) and y = (y1 , ..., yn ) be two points in the range of X. Suppose that xi − xn = yi − yn for i = 1, ..., n − 1. Putting c = yn − xn , we have yi = xi + c for all i. Hence, D(X) is maximal invariant under G. R By Proposition 1.8, D has the p.d.f. fi (d1 + t, ..., dn−1 + t, t)dt under Hi , i = 0, 1, which does not depend on µ. In fact, in this case Proposition 6.2 applies with M (X) = D(X) and θ(fi,µ ) = i. If we consider tests that are functions of D(X), then the problem of testing the hypotheses in (6.48) becomes one of testing a simple hypothesis versus a simple hypothesis. By Theorem 6.1, the test UMP among functions of D(X), which is then the UMPI test, rejects H0 in (6.48) when R R f1 (x1 + t, ..., xn + t)dt f (d + t, ..., dn−1 + t, t)dt R 1 1 R = > c, f0 (d1 + t, ..., dn−1 + t, t)dt f0 (x1 + t, ..., xn + t)dt where c is determined by the size of the UMPI test. The previous result can be extended to the case of a location-scale family where the p.d.f. of X is one of fi,µ,σ = σ1n fi x1σ−µ , ..., xnσ−µ , i = 0, 1, fi,µ,σ is symmetric about µ, the hypotheses of interest are given by (6.48) with fi,µ replaced by fi,µ,σ , and G = {gc,r : c ∈ R, r 6= 0} with gc,r (x) = (rx1 +c, ..., rxn +c). When n ≥ 3, it can be shown that a maximal invariant under G is W (X) = (W1 , ..., Wn−2 ), where Wi = (Xi − Xn )/(Xn−1 − Xn ), and that the p.d.f. of W does not depend on (µ, σ). A UMPI test can then be derived (exercise). The next example considers finding a maximal invariant in a problem that is not a location-scale family problem. Example 6.14. Let G be the set of n! permutations of the components of x ∈ Rn . Then a maximal invariant is the vector of order statistics. This is because a permutation of the components of x does not change the values of these components and two x’s with the same set of ordered components can be obtained from each other through a permutation of coordinates.

420

6. Hypothesis Tests

Suppose that P contains continuous c.d.f.’s on Rn . Let G be the class of all transformations of the form g(x) = (ψ(x1 ), ..., ψ(xn )), where ψ is continuous and strictly increasing. For x = (x1 , ..., xn ), let R(x) = (R1 , ..., Rn ) be the vector of ranks (§5.2.2), i.e., xi = x(Ri ) , where x(j) is the jth smallest value of xi ’s. Clearly, R(g(x)) = R(x) for any g ∈ G. For any x and y in Rn with R(x) = R(y), define ψ(t) to be linear between x(j) and x(j+1) , j = 1, ..., n−1, ψ(t) = t+(y(1) −x(1) ) for t ≤ x(1) , and ψ(t) = t+(y(n) −x(n) ) for t ≥ x(n) . Then ψ(xi ) = ψ(yi ), i = 1, ..., n. This shows that the vector of rank statistics is maximal invariant. When there is a sufficient statistic U (X), it is convenient first to reduce the data to U (X) before applying invariance. If there is a test T (U ) UMP among all invariant tests depending only on U , one would like to conclude that T (U ) is a UMPI test. Unfortunately, this may not be true in general, since it is not clear that for any invariant test based on X there is an equivalent invariant test based only on U (X). The following result provides a sufficient condition under which it is enough to consider invariant tests depending only on U (X). Its proof is omitted and can be found in Lehmann (1986, pp. 297-302). Proposition 6.3. Let G be a group of transformations on X (the range of X) and (G, BG , λ) be a measure space with a σ-finite λ. Suppose that the testing problem under consideration is invariant under G, that for any set A ∈ BX , the set of points (x, g) for which g(x) ∈ A is in σ(BX × BG ), and that λ(B) = 0 implies λ({h ◦ g : h ∈ B}) = 0 for all g ∈ G. Suppose further that there is a statistic U (X) sufficient for P ∈ P and that U (x1 ) = U (x2 ) implies U (g(x1 )) = U (g(x2 )) for all g ∈ G so that G induces a group GU of transformations on the range of U through gU (U (x)) = U (g(x)). Then, for any test T (X) invariant under G, there exists a test based on U (X) that is invariant under G (and GU ) and has the same power function as T (X). In many problems g(x) = ψ(x, g), where g ranges over a set G in Rm and ψ is a Borel function on Rn+m . Then the measurability condition in Proposition 6.3 is satisfied by choosing BG to be the Borel σ-field on G. In such cases it is usually not difficult to find a measure λ satisfying the condition in Proposition 6.3. Example 6.15. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ ∈ R and σ 2 > 0. The problem of testing H0 : σ 2 ≥ σ02 versus H1 : σ 2 < σ02 is invariant under G = {gc : c ∈ R} with gc (x) = (x1 + c, ..., xn + c). It ¯ S2) can be shown (exercise) that G and the sufficient statistic U = (X, satisfy the conditions in Proposition 6.3 with GU = {hc : c ∈ R} and hc (u1 , u2 ) = (u1 + c, u2 ), and that S 2 is maximal invariant under GU . It follows from Proposition 6.3, Corollary 6.1, and the fact that (n − 1)S 2 /σ02

6.3. UMP Invariant Tests

421

has the chi-square distribution χ2n−1 when σ 2 = σ02 that a UMPI test of size α rejects H0 when (n − 1)S 2 /σ02 ≤ χ2n−1,1−α , where χ2n−1,α is the (1 − α)th quantile of the chi-square distribution χ2n−1 . This test coincides with the UMPU test given in §6.2.3. Example 6.16. Let Xi1 , ..., Xini , i = 1, 2, be two independent samples i.i.d. from N (µi , σi2 ), i = 1, 2, respectively. The problem of testing H0 : σ22 /σ12 ≤ ∆0 versus H1 : σ22 /σ12 > ∆0 is invariant under G = {gc1 ,c2 ,r : ci ∈ R, i = 1, 2, r > 0} with gc1 ,c2 ,r (x1 , x2 ) = (rx11 + c1 , ..., rx1n1 + c1 , rx21 + c2 , ..., rx2n2 + c2 ). ¯1, X ¯ 2 , S12 , S22 ) It can be shown (exercise) that the sufficient statistic U = (X and G satisfy the conditions in Proposition 6.3 with GU = {hc1 ,c2 ,r : ci ∈ R, i = 1, 2, r > 0} and hc1 ,c2 ,r (u1 , u2 , u3 , u4 ) = (ru1 + c1 , ru2 + c2 , ru3 , ru4 ). A maximal invariant under GU is S2 /S1 . Let ∆ = σ22 /σ12 . Then (S22 /S12 )/∆ has an F-distribution and, therefore, V = S22 /S12 has a Lebesgue p.d.f. of the form f∆ (v) = C(∆)v (n2 −3)/2 [∆ + (n2 − 1)v/(n1 − 1)]−(n1 +n2 −2)/2 I(0,∞) (v), where C(∆) is a known function of ∆. It can be shown (exercise) that the family {f∆ : ∆ > 0} has monotone likelihood ratio in V so that a UMPI test of size α rejects H0 when V > Fn2 −1,n1 −1,α , where Fa,b,α is the (1 − α)th quantile of the F-distribution Fa,b . Again, this UMPI test coincides with the UMPU test given in §6.2.3. The following result shows that, in Examples 6.15 and 6.16, the fact that UMPI tests are the same as the UMPU tests is not a simple coincidence. Proposition 6.4. Consider a testing problem invariant under G. If there exists a UMPI test of size α, then it is unbiased. If there also exists a UMPU test of size α that is invariant under G, then the two tests have the same power function on P ∈ P1 . If either the UMPI test or the UMPU test is unique a.s. P, then the two tests are equal a.s. P. Proof. We only need to prove that a UMPI test of size α is unbiased. This follows from the fact that the test T ≡ α is invariant under G.

422

6. Hypothesis Tests

The next example shows an application of invariance in a situation where a UMPU test may not exist. Example 6.17. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ and σ 2 . Let θ = (µ − u)/σ, where u is a known constant. Consider the problem of testing H0 : θ ≤ θ0 versus H1 : θ > θ0 . Note that H0 is the same as P (X1 ≤ u) ≥ p0 for a known constant p0 = Φ(−θ0 ). Without loss of generality, we consider the case of u = 0. The problem is invariant under G = {gr : r > 0} with gr (x) = rx. By Proposition 6.3, we can consider tests that are functions of the sufficient ¯ S 2 ) only. A maximal invariant under G is t(X) = √nX/S. ¯ statistic (X, To find a UMPI test, it remains to find a test UMP among all tests that are functions of t(X). From √ the discussion in §1.3.1, t(X) has the noncentral t-distribution tn−1 ( nθ). Let fθ (t) be the Lebesgue p.d.f. √ of t(X), i.e., fθ is given by (1.32) with n replaced by n − 1 and δ = nθ. It can be shown (exercise) that the family of p.d.f.’s, {fθ (t) : θ ∈ R}, has monotone likelihood ratio in t. Hence, by Theorem 6.2, a UMPI test of√size α rejects H0 when t(X) > c, where c is the (1 − α)th quantile of tn−1 ( nθ0 ). In some problems, we may have to apply both unbiasedness and invariance principles. For instance, suppose that in the current problem we would like to test H0 : θ = θ0 versus H1 : θ 6= θ0 . The problem is still invariant under G. Following the previous discussion, we only need to consider tests that are functions of t(X). But a test UMP among functions of t(X) does not exist in this case. A test UMP among all unbiased tests of level α that are functions of t(X) rejects H0 when t(X) < c1 or t(X) > c2 , where c1 and c2 are determined by Z c 2 Z c2 d fθ0 (t)dt = 1 − α and fθ (t)dt =0 dθ c1 c1 θ=θ0

(see Exercise 26). This test is then UMP among all tests that are invariant and unbiased of level α. Whether it is also UMPU without the restriction to invariant tests is an open problem.

6.3.2 UMPI tests in normal linear models Consider normal linear model (6.38): X = Nn (Zβ, σ 2 In ), where β is a p-vector of unknown parameters, σ 2 > 0 is unknown, and Z is a fixed n × p matrix of rank r ≤ p < n. In §6.2.3, UMPU tests for testing

423

6.3. UMP Invariant Tests

(6.39) or (6.40) are derived. A frequently encountered problem in practice is to test H0 : Lβ = 0 versus H1 : Lβ 6= 0, (6.49) where L is an s × p matrix of rank s ≤ r and all rows of L are in R(Z). However, a UMPU test for (6.49) does not exist if s > 1. We now derive a UMPI test for testing (6.49). We use without proof the following result from linear algebra: there exists an orthogonal matrix Γ such that (6.49) is equivalent to H0 : η1 = 0

versus

H1 : η1 6= 0,

(6.50)

where η1 is the s-vector containing the first s components of η, η is the r-vector containing the first r components of ΓZβ, and the last n − r components of ΓZβ are 0’s. Let Y = ΓX. Then Y = Nn ((η, 0), σ 2 In ) with the p.d.f. given by (6.42). Let Y = (Y1 , Y2 ), where Y1 is an r-vector, and let Y1 = (Y11 , Y12 ), where Y11 is an s-vector. Define G = {gΛ,c,γ : c ∈ Rr−s , γ > 0, Λ is an s × s orthogonal matrix} with gΛ,c,γ (Y ) = γ(ΛY11 , Y12 + c, Y2 ). Testing (6.50) is invariant under G. By Proposition 6.3, we can restrict our attention to the sufficient statistic U = (Y1 , kY2 k2 ). The statistic M (U ) = kY11 k2 /kY2 k2

(6.51)

is invariant under GU , the group of transformations on the range of U defined by g˜Λ,c,γ (U (Y )) = U (gΛ,c,γ (Y )). We now show that M (U ) is maximal invariant under GU . Let li ∈ Rs , li 6= 0, and ti ∈ (0, ∞), i = 1, 2. If kl1 k2 /t21 = kl2 k2 /t22 , then t1 = γt2 with γ = kl1 k/kl2 k. Since l1 /kl1 k and l2 /kl2 k are two points having the same distance from the origin, there exists an orthogonal matrix Λ such that l1 /kl1 k = Λl2 /kl2 k, i.e., l1 = γΛl2 . (j) (j) This proves that if M (u(1) ) = M (u(2) ) with u(j) = (y11 , y12 , t2j ), then (1)

(2)

y11 = γΛy11 and t1 = γt2 for some γ > 0 and orthogonal matrix Λ and, (1) (2) therefore, u(1) = g˜Λ,c,γ (u(2) ) with c = γ −1 y12 − y12 . Thus, M (U ) is maximal invariant under GU . It can be shown (exercise) that W = M (U )(n − r)/s has the noncentral F-distribution Fs,n−r (θ) with θ = kη1 k2 /σ 2 (see §1.3.1). Let fθ (w) be the Lebesgue p.d.f. of W , i.e., fθ is given by (1.33) with n1 = s, n2 = n − r, and δ = θ. Note that under H0 , θ = 0 and fθ reduces to the p.d.f. of the central F-distribution Fs,n−r (Table 1.2, page 20). Also, it can be shown (exercise) that the ratio fθ1 (w)/f0 (w) is an increasing function of w for any given θ1 > 0. By Theorem 6.1, a UMPI test of size α for testing H0 : θ = 0

424

6. Hypothesis Tests

versus H1 : θ = θ1 rejects H0 when W > Fs,n−r,α , where Fs,n−r,α is the (1 − α)th quantile of the F-distribution Fs,n−r . Since this test does not depend on θ1 , by Lemma 6.1, it is also a UMPI test of size α for testing H0 : θ = 0 versus H1 : θ > 0, which is equivalent to testing (6.50). In applications it is not convenient to carry out the test by finding explicitly the orthogonal matrix Γ. Hence, we now express the statistic W in terms of X. Since Y = ΓX and E(Y ) = ΓE(X) = ΓZβ, kY1 − ηk2 + kY2 k2 = kX − Zβk2 and, therefore, min kY1 − ηk2 + kY2 k2 = min kX − Zβk2 , η

β

which is the same as ˆ 2 = SSR, kY2 k2 = kX − Z βk where βˆ is the LSE defined by (3.27). Similarly, kY11 k2 + kY2 k2 = min kX − Zβk2 . β:Lβ=0

If we define βˆH0 to be a solution of kX − Z βˆH0 k2 = min kX − Zβk2 , β:Lβ=0

which is called the LSE of β under H0 or the LSE of β subject to Lβ = 0, then ˆ 2 )/s (kX − Z βˆH0 k2 − kX − Z βk W = . (6.52) ˆ 2 /(n − r) kX − Z βk Thus, the UMPI test for (6.49) can be used without finding Γ. When s = 1, the UMPI test derived here is the same as the UMPU test for (6.40) given in §6.2.3. Example 6.18. Consider the one-way ANOVA model in Example 3.13: Xij = N (µi , σ 2 ),

j = 1, ..., ni , i = 1, ..., m,

and Xij ’s are independent. A common testing problem in applications is the test for homogeneity of means, i.e., H0 : µ1 = · · · = µm

versus

H1 : µi 6= µk for some i 6= k.

(6.53)

One can easily find a matrix L for which (6.53) is equivalent to (6.49). But it is not necessary to find such a matrix in order to compute the

425

6.3. UMP Invariant Tests

statistic W that defines the UMPI test. Note that the LSE of (µ1 , ..., µm ) ¯ 1· , ..., X ¯ m· ), where X ¯ i· is the sample mean based on Xi1 , ..., Xini , and is (X ¯ the sample mean based on all Xij ’s. Thus, the LSE under H0 is simply X, ˆ 2= SSR = kX − Z βk

ni m X X i=1 j=1

SST = kX − Z βˆH0 k2 = and SSA = SST − SSR = Then W =

¯ i· )2 , (Xij − X

ni m X X i=1 j=1

m X i=1

¯ 2, (Xij − X)

¯ i· − X) ¯ 2. ni (X

SSA/(m − 1) , SSR/(n − m)

Pm where n = i=1 ni . The name ANOVA comes from the fact that the UMPI test is carried out by comparing two sources of variation: the variation within each group of observations (measured by SSR) and the variation among m groups (measured by SSA), and that SSA + SSR = SST is the total variation in the data set. In this case, the distribution of W can also be derived using Cochran’s theorem (Theorem 1.5). See Exercise 75. Example 6.19. Consider the two-way balanced ANOVA model in Example 3.14: Xijk = N (µij , σ 2 ),

i = 1, ..., a, j = 1, ..., b, k = 1, ..., c,

Pa Pb Pa Pb where µij = µ+αi +βj +γij , i=1 αi = j=1 βj = i=1 γij = j=1 γij = 0, and Xijk ’s are independent. Typically the following hypotheses are of interest: H0 : αi = 0 for all i

versus

H1 : αi 6= 0 for some i,

(6.54)

H0 : βj = 0 for all j

versus

H1 : βj 6= 0 for some j,

(6.55)

H0 : γij = 0 for all i, j

versus

H1 : γij 6= 0 for some i, j.

(6.56)

and

In applications, αi ’s are effects of a factor A (a variable taking finitely many values), βj ’s are effects of a factor B, and γij ’s are effects of the interaction of factors A and B. Hence, testing hypotheses in (6.54), (6.55), and (6.56)

426

6. Hypothesis Tests

are the same as testing effects of factor A, of factor B, and of the interaction between A and B, respectively. ¯ ··· , The LSE’s of µ, αi , βj , and γij are given by (Example 3.14) µ ˆ=X ˆ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ α ˆ i = Xi·· − X··· , βj = X·j· − X··· , γˆij = Xij· − Xi·· − X·j· + X··· , and a dot is used to denote averaging over the indicated subscript. Let SSR =

b X c a X X

¯ ij· )2 , (Xijk − X

i=1 j=1 k=1

SSA = bc

a X i=1

SSB = ac

b X j=1

and SSC = c

b a X X i=1 j=1

¯ i·· − X ¯ ··· )2 , (X ¯ ·j· − X ¯ ··· )2 , (X

¯ ij· − X ¯ i·· − X ¯ ·j· + X ¯ ··· )2 . (X

Then, one can show (exercise) that for testing (6.54), (6.55), and (6.56), the statistics W in (6.52) (for the UMPI tests) are, respectively, SSA/(a − 1) , SSR/[(c − 1)ab]

SSB/(b − 1) , SSR/[(c − 1)ab]

and

SSC/[(a − 1)(b − 1)] . SSR/[(c − 1)ab]

We end this section with a discussion of testing for random effects in the following balanced one-way random effects model (Example 3.17): Xij = µ + Ai + eij ,

i = 1, ..., a, j = 1, ..., b,

(6.57)

where µ is an unknown parameter, Ai ’s are i.i.d. random effects from N (0, σa2 ), eij ’s are i.i.d. measurement errors from N (0, σ 2 ), and Ai ’s and eij ’s are independent. Consider the problem of testing H0 : σa2 /σ 2 ≤ ∆0

versus

H1 : σa2 /σ 2 > ∆0

(6.58)

for a given ∆0 . When ∆0 is small, hypothesis H0 in (6.58) means that the random effects are negligible relative to the measurement variation. Let (Yi1 , ..., Yib ) = Γ(Xi1 , ..., Xib ), where Γ is a√b × b orthogonal matrix whose elements in the first row are all equal to 1/ b. Then √ √ ¯ i· = b(µ + Ai + e¯i· ), i = 1, ..., a, Yi1 = bX

427

6.3. UMP Invariant Tests

√ are i.i.d. from N ( bµ, σ 2 + bσa2 ), Yij , i = 1, ..., a, j = 2, ..., b, are i.i.d. from N (0, σ 2 ), and Yij ’s are independent. The reason why E(Yij ) = 0 when j > 1 is because row j of Γ is orthogonal to the first row of Γ. Let Λ be an√a × a orthogonal matrix whose elements in the first√row are all √ equal to 1/ a and (U11 , ..., Ua1 ) = Λ(Y11 , ..., Ya1 ). Then U11 = aY¯·1 is N ( abµ, σ 2 + bσa2 ), Ui1 , i = 2, ..., a, are from N (0, σ 2 + bσa2 ), and Ui1 ’s are independent. Let Uij = Yij for j = 2, ..., b, i = 1, ..., a. The problem of testing (6.58) is invariant under the group of transformations that transform U11 to rU11 +c and Uij to rUij , (i, j) 6= (1, 1), where r > 0 and c ∈ R. It can be shown (exercise) that the maximal invariant under this group of transformations is SSA/SSR, where SSA =

a X

2 Ui1

and

i=2

SSR =

b a X X

Uij2 .

i=1 j=2

Note that H0 in (6.58) is equivalent to (σ 2 + bσa2 )/σ 2 ≤ 1 + b∆0 . Also, SSA/(σ 2 + bσa2 ) has the chi-square distribution χ2a−1 and SSR/σ 2 has the chi-square distribution χ2a(b−1) . Hence, the p.d.f. of the statistic W =

SSA/(a − 1) 1 1 + b∆0 SSR/[a(b − 1)]

is in a parametric family (indexed by the parameter (σ 2 + bσa2 )/σ 2 ) with monotone likelihood ratio in W . Thus, a UMPI test of size α for testing (6.58) rejects H0 when W > Fa−1,a(b−1),α , where Fa−1,a(b−1),α is the (1 − α)th quantile of the F-distribution Fa−1,a(b−1) . It remains to express W in terms of Xij ’s. Note that b b a X a b a X X X X X ¯ i· )2 SSR = Yij2 = e2ij − b¯ e2i· = (Xij − X i=1 j=2

i=1

j=1

i=1 j=1

and

SSA =

a X i=1

2 2 Ui1 − U11 =

a X i=1

Yi12 − aY¯·12 = b

a X i=1

¯ i· − X ¯ ·· )2 . (X

The SSR and SSA derived here are the same as those in Example 6.18 when ni = b for all i and m = a. It can also be seen that if ∆0 = 0, then testing (6.58) is equivalent to testing H0 : σa2 = 0 versus H1 : σa2 > 0 and the derived UMPI test is exactly the same as that in Example 6.18, although the testing problems are different in these two cases. Extensions to balanced two-way random effects models can be found in Lehmann (1986, §7.12).

428

6. Hypothesis Tests

6.4 Tests in Parametric Models A UMP, UMPU, or UMPI test often does not exist in a particular problem. In the rest of this chapter, we study some methods for constructing tests that have intuitive appeal and frequently coincide with optimal tests (UMP or UMPU tests) when optimal tests do exist. We consider tests in parametric models in this section, whereas tests in nonparametric models are studied in §6.5. When the hypothesis H0 is not simple, it is often difficult or even impossible to obtain a test that has exactly a given size α, since it is hard to find a population P that maximizes the power function of the test over all P ∈ P0 . In such cases a common approach is to find tests having asymptotic significance level α (Definition 2.13). This involves finding the limit of the power of a test at P ∈ P0 , which is studied in this section and §6.5. Throughout this section, we assume that a sample X is from P ∈ P = θ {Pθ : θ ∈ Θ}, Θ ⊂ Rk , fθ = dP dν exists w.r.t. a σ-finite measure ν for all θ, and the testing problem is H0 : θ ∈ Θ0

versus

H1 : θ ∈ Θ 1 ,

(6.59)

where Θ0 ∪ Θ1 = Θ and Θ0 ∩ Θ1 = ∅.

6.4.1 Likelihood ratio tests When both H0 and H1 are simple (i.e., both Θ0 = {θ0 } and Θ1 = {θ1 } are single-point sets), Theorem 6.1 applies and a UMP test rejects H0 when fθ1 (X) > c0 fθ0 (X)

(6.60)

for some c0 > 0. When c0 ≥ 1, (6.60) is equivalent to (exercise) fθ0 (X) θ0 , there is an LR test whose rejection region is the same as that of the UMP test T∗ given by (6.11). (ii) For testing the hypotheses in (6.12), there is an LR test whose rejection region is the same as that of the UMP test T∗ given by (6.15). (iii) For testing the hypotheses in (6.13) or (6.14), there is an LR test whose rejection region is equivalent to Y (X) < c1 or Y (X) > c2 for some constants c1 and c2 . Proof. (i) Let θˆ be the MLE of θ. Note that ℓ(θ) is increasing when θ ≤ θˆ ˆ Thus, and decreasing when θ > θ. ( 1 θˆ ≤ θ0 λ(X) = ℓ(θ0 ) θˆ > θ0 . ˆ ℓ(θ) ˆ < c. From the Then λ(X) < c is the same as θˆ > θ0 and ℓ(θ0 )/ℓ(θ) discussion in §4.4.2, θˆ is a strictly increasing function of Y . It can be ˆ − log ℓ(θ0 ) is strictly increasing in Y when θˆ > θ0 and shown that log ℓ(θ) strictly decreasing in Y when θˆ < θ0 (exercise). Hence, for any d ∈ R, ˆ < c is equivalent to Y > d for some c ∈ (0, 1). θˆ > θ0 and ℓ(θ0 )/ℓ(θ) (ii) The proof is similar to that in (i). Note that ( 1 θˆ < θ1 or θˆ > θ2 λ(X) = max{ℓ(θ1 ),ℓ(θ2 )} θ1 ≤ θˆ ≤ θ2 . ˆ ℓ(θ)

430

6. Hypothesis Tests

Hence λ(X) < c is equivalent to c1 < Y < c2 . (iii) The proof for (iii) is left as an exercise. Proposition 6.5 can be applied to problems concerning one-parameter exponential families such as the binomial, Poisson, negative binomial, and normal (with one parameter known) families. The following example shows that the same result holds in a situation where Proposition 6.5 is not applicable. Example 6.20. Consider the testing problem H0 : θ = θ0 versus H1 : θ 6= θ0 based on i.i.d. X1 , ..., Xn from the uniform distribution U (0, θ). We now show that the UMP test with rejection region X(n) > θ0 or X(n) ≤ θ0 α1/n given in Exercise 19(c) is an LR test. Note that ℓ(θ) = θ−n I(X(n) ,∞) (θ). Hence (X(n) /θ0 )n X(n) ≤ θ0 λ(X) = 0 X(n) > θ0 and λ(X) < c is equivalent to X(n) > θ0 or X(n) /θ0 < c1/n . Taking c = α ensures that the LR test has size α.

More examples of this kind can be found in §6.6. The next example considers multivariate θ. Example 6.21. Consider normal linear model (6.38) and the hypotheses in (6.49). The likelihood function in this problem is n/2 1 ℓ(θ) = 2πσ exp − 2σ1 2 kX − Zβk2 , 2

where θ = (β, σ2 ). Let βˆ be the LSE defined by (3.27). Since kX − Zβk2 ≥ ˆ 2 for any β, kX − Z βk o n n/2 1 ˆ 2 . exp − 2σ1 2 kX − Z βk ℓ(θ) ≤ 2πσ 2

Treating the right-hand side of the previous expression as a function of σ 2 , ˆ 2 /n and, it is easy to show that it has a maximum at σ 2 = σ ˆ 2 = kX − Z βk therefore, sup ℓ(θ) = (2πˆ σ 2 )−n/2 e−n/2 . θ∈Θ 2 ˆH = kX − Z βˆH0 k2 /n. Then Similarly, let βˆH0 be the LSE under H0 and σ 0 2 −n/2 −n/2 sup ℓ(θ) = (2πˆ σH ) e . 0

θ∈Θ0

Thus, λ(X) = (ˆ σ

2

2 n/2 /ˆ σH ) 0

=

ˆ 2 kX − Z βk kX − Z βˆH k2 0

!n/2

=

sW +1 n−r

−n/2

,

431

6.4. Tests in Parametric Models

where W is given in (6.52). This shows that LR tests are the same as the UMPI tests derived in §6.3.2. The one-sample or two-sample two-sided t-tests derived in §6.2.3 are special cases of LR tests. For a one-sample problem, we define β = µ and 2 ¯ σ Z = Jn , the n-vector of ones. Note that βˆ = X, ˆ 2 = (n − 1)S 2 /n, βˆH =0 0 2 2 2 2 ¯ . Hence (H0 : β = 0), and σ ˆH0 = kXk /n = (n − 1)S /n + X λ(X) = 1 +

¯2 nX (n − 1)S 2

−n/2 [t(X)]2 1+ , n−1

−n/2

=

Jn1 0

0 Jn2

√ ¯ has the t-distribution tn−1 under H0 . Thus, λ(X) < where t(X) = nX/S c is equivalent to |t(X)| > c0 , which is the rejection region of a one-sample two-sided t-test. For a two-sample problem, we let n = n1 + n2 , β = (µ1 , µ2 ), and Z=

.

Testing H0 : µ1 = µ2 versus H1 : µ1 6= µ2 is the same as testing (6.49) with ¯ and βˆ = (X ¯1, X ¯ 2 ), where X ¯ 1 and X ¯ 2 are L = ( 1 −1 ). Since βˆH0 = X the sample means based on X1 , ..., Xn1 and Xn1 +1 , ..., Xn , respectively, we have nˆ σ2 =

n1 X i=1

¯ 1 )2 + (Xi − X

n X

¯ 2 )2 = (n1 − 1)S12 + (n2 − 1)S22 (Xi − X

i=n1 +1

and 2 ¯1 − X ¯ 2 )2 + (n1 − 1)S 2 + (n2 − 1)S 2 . nˆ σH = (n − 1)S 2 = n−1 n1 n2 (X 1 2 0

Therefore, λ(X) < c is equivalent to |t(X)| > c0 , where t(X) is given by (6.37), and LR tests are the same as the two-sample two-sided t-tests in §6.2.3.

6.4.2 Asymptotic tests based on likelihoods As we can see from Proposition 6.5 and the previous examples, an LR test is often equivalent to a test based on a statistic Y (X) whose distribution under H0 can be used to determine the rejection region of the LR test with size α. When this technique fails, it is difficult or even impossible to find an LR test with size α, even if the c.d.f. of λ(X) is continuous. The following result shows that in the i.i.d. case we can obtain the asymptotic distribution (under H0 ) of the likelihood ratio λ(X) so that an LR test

432

6. Hypothesis Tests

having asymptotic significance level α can be obtained. Assume that Θ0 is determined by H0 : θ = g(ϑ), (6.63) where ϑ is a (k − r)-vector of unknown parameters and g is a continuously differentiable function from Rk−r to Rk with a full rank ∂g(ϑ)/∂ϑ. For example, if Θ = R2 and Θ0 = {(θ1 , θ2 ) ∈ Θ : θ1 = 0}, then ϑ = θ2 , g1 (ϑ) = 0, and g2 (ϑ) = ϑ. Theorem 6.5. Assume the conditions in Theorem 4.16. Suppose that H0 is determined by (6.63). Under H0 , −2 log λn →d χ2r , where λn = λ(X) and χ2r is a random variable having the chi-square distribution χ2r . Con2 sequently, the LR test with rejection region λn < e−χr,α /2 has asymptotic significance level α, where χ2r,α is the (1 − α)th quantile of the chi-square distribution χ2r . Proof. Without loss of generality, we assume that there exist an MLE θˆ and an MLE ϑˆ under H0 such that λn =

ˆ supθ∈Θ0 ℓ(θ) ℓ(g(ϑ)) = . ˆ supθ∈Θ ℓ(θ) ℓ(θ)

Following the proof of Theorem 4.17 in §4.5.2, we can obtain that √ nI1 (θ)(θˆ − θ) = n−1/2 sn (θ) + op (1), where sn (θ) = ∂ log ℓ(θ)/∂θ and I1 (θ) is the Fisher information about θ contained in X1 , and that ˆ − log ℓ(θ)] = n(θˆ − θ)τ I1 (θ)(θˆ − θ) + op (1). 2[log ℓ(θ) Then ˆ − log ℓ(θ)] = n−1 [sn (θ)]τ [I1 (θ)]−1 sn (θ) + op (1). 2[log ℓ(θ) Similarly, under H0 , ˆ − log ℓ(g(ϑ))] = n−1 [˜ 2[log ℓ(g(ϑ)) sn (ϑ)]τ [I˜1 (ϑ)]−1 s˜n (ϑ) + op (1), where s˜n (ϑ) = ∂ log ℓ(g(ϑ))/∂ϑ = D(ϑ)sn (g(ϑ)), D(ϑ) = ∂g(ϑ)/∂ϑ, and I˜1 (ϑ) is the Fisher information about ϑ (under H0 ) contained in X1 . Combining these results, we obtain that ˆ − log ℓ(g(ϑ))] ˆ −2 log λn = 2[log ℓ(θ)

= n−1 [sn (g(ϑ))]τ B(ϑ)sn (g(ϑ)) + op (1)

under H0 , where B(ϑ) = [I1 (g(ϑ))]−1 − [D(ϑ)]τ [I˜1 (ϑ)]−1 D(ϑ).

433

6.4. Tests in Parametric Models

By the CLT, n−1/2 [I1 (θ)]−1/2 sn (θ) →d Z, where Z = Nk (0, Ik ). Then, it follows from Theorem 1.10(iii) that, under H0 , −2 log λn →d Z τ [I1 (g(ϑ))]1/2 B(ϑ)[I1 (g(ϑ))]1/2 Z. Let D = D(ϑ), B = B(ϑ), A = I1 (g(ϑ)), and C = I˜1 (ϑ). Then (A1/2 BA1/2 )2 = A1/2 BABA1/2 = A1/2 (A−1 − Dτ C −1 D)A(A−1 − Dτ C −1 D)A1/2

= (Ik − A1/2 Dτ C −1 DA1/2 )(Ik − A1/2 Dτ C −1 DA1/2 ) = Ik − 2A1/2 Dτ C −1 DA1/2 + A1/2 Dτ C −1 DADτ C −1 DA1/2 = Ik − A1/2 Dτ C −1 DA1/2 = A1/2 BA1/2 ,

where the fourth equality follows from the fact that C = DADτ . This shows that A1/2 BA1/2 is a projection matrix. The rank of A1/2 BA1/2 is tr(A1/2 BA1/2 ) = tr(Ik − Dτ C −1 DA) = k − tr(C −1 DADτ ) = k − tr(C −1 C) = k − (k − r) = r.

Thus, by Exercise 51 in §1.6, Z τ [I1 (g(ϑ))]1/2 B(ϑ)[I1 (g(ϑ))]1/2 Z = χ2r . As an example, Theorem 6.5 can be applied to testing problems in Example 4.33 where the exact rejection region of the LR test of size α is difficult to obtain but the likelihood ratio λn can be calculated numerically. Tests whose rejection regions are constructed using asymptotic theory (so that these tests have asymptotic significance level α) are called asymptotic tests, which are useful when a test of exact size α is difficult to find. There are two popular asymptotic tests based on likelihoods that are asymptotically equivalent to LR tests. Note that the hypothesis in (6.63) is equivalent to a set of r ≤ k equations: H0 : R(θ) = 0,

(6.64)

where R(θ) is a continuously differentiable function from Rk to Rr . Wald (1943) introduced a test that rejects H0 when the value of ˆ τ {[C(θ)] ˆ τ [In (θ)] ˆ −1 C(θ)} ˆ −1 R(θ) ˆ Wn = [R(θ)] is large, where C(θ) = ∂R(θ)/∂θ, In (θ) is the Fisher information matrix based on X1 , ..., Xn , and θˆ is an MLE or RLE of θ. For testing H0 : θ = θ0

434

6. Hypothesis Tests

with a known θ0 , R(θ) = θ − θ0 and Wn simplifies to ˆ θˆ − θ0 ). Wn = (θˆ − θ0 )τ In (θ)( Rao (1947) introduced a score test that rejects H0 when the value of ˜ τ [In (θ)] ˜ −1 sn (θ) ˜ Rn = [sn (θ)] is large, where sn (θ) = ∂ log ℓ(θ)/∂θ is the score function and θ˜ is an MLE or RLE of θ under H0 in (6.64). Theorem 6.6. Assume the conditions in Theorem 4.16. (i) Under H0 given by (6.64), Wn →d χ2r and, therefore, the test rejects H0 if and only if Wn > χ2r,α has asymptotic significance level α, where χ2r,α is the (1 − α)th quantile of the chi-square distribution χ2r . (ii) The result in (i) still holds if Wn is replaced by Rn . Proof. (i) Using Theorems 1.12 and 4.17, √ ˆ − R(θ)] →d Nr 0, [C(θ)]τ [I1 (θ)]−1 C(θ) , n[R(θ)

where I1 (θ) is the Fisher information about θ contained in X1 . Under H0 , R(θ) = 0 and, therefore, ˆ τ {[C(θ)]τ [I1 (θ)]−1 C(θ)}−1 R(θ) ˆ →d χ2 n[R(θ)] r (Theorem 1.10). Then the result follows from Slutsky’s theorem (Theorem 1.11) and the fact that θˆ →p θ and I1 (θ) and C(θ) are continuous at θ. (ii) From the Lagrange multiplier, θ˜ satisfies ˜ + C(θ)λ ˜ n=0 sn (θ)

and

˜ = 0. R(θ)

Using Taylor’s expansion, one can show (exercise) that under H0 ,

and

[C(θ)]τ (θ˜ − θ) = op (n−1/2 )

(6.65)

sn (θ) − In (θ)(θ˜ − θ) + C(θ)λn = op (n1/2 ),

(6.66)

where In (θ) = nI1 (θ). Multiplying [C(θ)]τ [In (θ)]−1 to the left-hand side of (6.66) and using (6.65), we obtain that [C(θ)]τ [In (θ)]−1 C(θ)λn = −[C(θ)]τ [In (θ)]−1 sn (θ) + op (n−1/2 ),

(6.67)

which implies λτn [C(θ)]τ [In (θ)]−1 C(θ)λn →d χ2r

(6.68) ˜ n= (exercise). Then the result follows from (6.68) and the fact that C(θ)λ ˜ −sn (θ), In (θ) = nI1 (θ), and I1 (θ) is continuous at θ.

435

6.4. Tests in Parametric Models

Thus, Wald’s tests, Rao’s score tests, and LR tests are asymptotically ˆ not θ˜ = g(ϑ), ˆ equivalent. Note that Wald’s test requires computing θ, ˜ ˆ whereas Rao’s score test requires computing θ, not θ. On the other hand, an LR test requires computing both θˆ and θ˜ (or solving two maximization problems). Hence, one may choose one of these tests that is easy to compute in a particular application. The results in Theorems 6.5 and 6.6 can be extended to non-i.i.d. situations (e.g., the GLM in §4.4.2). We state without proof the following result. Theorem 6.7. Assume the conditions in Theorem 4.18. Consider the problem of testing H0 in (6.64) (or equivalently, (6.63)) with θ = (β, φ). Then the results in Theorems 6.5 and 6.6 still hold. Example 6.22. Consider the GLM (4.55)-(4.58) with ti ’s in a fixed interval (t0 , t∞ ), 0 < t0 ≤ t∞ < ∞. Then the Fisher information matrix In (θ) =

φ−1 Mn (β) 0

0 I˜n (β, φ)

,

where Mn (β) is given by (4.60) and I˜n (β, φ) is the Fisher information about φ. Consider the problem of testing H0 : β = β0 versus H1 : β 6= β0 , where ˆ φ) ˆ be the MLE (or β0 is a fixed vector. Then R(β, φ) = β − β0 . Let (β, RLE) of (β, φ). Then, Wald’s test is based on ˆ βˆ − β0 ) Wn = φˆ−1 (βˆ − β0 )τ Mn (β)( and Rao’s score test is based on ˜ sn (β0 )]τ [Mn (β0 )]−1 s˜n (β0 ), Rn = φ[˜ where s˜n (β) is given by (4.65) and φ˜ is a solution of ∂ log ℓ(β0 , φ)/∂φ = 0. It follows from Theorem 4.18 that both Wn and Rn are asymptotically distributed as χ2p under H0 . By Slutsky’s theorem, we may replace φˆ or φ˜ by any consistent estimator of φ. Wald’s tests, Rao’s score tests, and LR tests are typically consistent according to Definition 2.13(iii). They are also Chernoff-consistent (Definition 2.13(iv)) if α is chosen to be αn → 0 and χ2r,αn = o(n) as n → ∞ (exercise). Other asymptotic optimality properties of these tests are discussed in Wald (1943); see also Serfling (1980, Chapter 10).

436

6. Hypothesis Tests

6.4.3 χ2 -tests A test that is related to the asymptotic tests described in §6.4.2 is the so-called χ2 -test for testing cell probabilities in a multinomial distribution. Consider a sequence of n independent trials with k possible outcomes for each trial. Let pj > 0 be the cell probability of occurrence of the jth outcome in any given trial and Xj be the number of occurrences of the jth outcome in n trials. Then X = (X1 , ..., Xk ) has the multinomial distribution (Example 2.7) with the parameter p = (p1 , ..., pk ). Let ξi = (0, ..., 0, 1, 0, ..., 0), where the single nonzero component 1 is located in the jth position if thePith trial yields the jth outcome. Then ξ1 , ..., ξn are n i.i.d. and X/n = ξ¯ = i=1 ξi /n. By the CLT, √ √ n(ξ¯ − p) →d Nk (0, Σ), (6.69) Zn (p) = n X n −p =

√ where Σ = Var(X/ n) is a symmetric k × k matrix whose ith diagonal element is pi (1 − pi ) and (i, j)th off-diagonal element is −pi pj . Consider the problem of testing H0 : p = p0

versus

H1 : p 6= p0 ,

(6.70)

where p0 = (p01 , ..., p0k ) is a known vector of cell probabilities. A popular test for (6.70) is based on the following χ2 -statistic: χ2 =

k X (Xj − np0j )2 = kD(p0 )Zn (p0 )k2 , np 0j j=1

(6.71)

where Zn (p) is given by (6.69) and D(c) with c = (c1 , ..., ck ) is the k × k −1/2 diagonal matrix whose jth diagonal element is cj . Another popular test is based on the following modified χ2 -statistic: χ ˜2 =

k X (Xj − np0j )2 = kD(X/n)Zn (p0 )k2 . X j j=1

(6.72)

Note that X/n is an unbiased estimator of p. √ √ Theorem 6.8. Let φ = ( p1 , ..., pk ) and Λ be a k × k projection matrix. (i) If Λφ = aφ, then [Zn (p)]τ D(p)ΛD(p)Zn (p) →d χ2r , where χ2r has the chi-square distribution χ2r with r = tr(Λ) − a. (ii) The same result holds if D(p) in (i) is replaced by D(X/n).

437

6.4. Tests in Parametric Models

Proof. (i) Let D = D(p), Zn = Zn (p), and Z = Nk (0, Ik ). From (6.69) and Theorem 1.10, Znτ DΛDZn →d Z τ AZ

with A = Σ1/2 DΛDΣ1/2 .

From Exercise 51 in §1.6, the result in (i) follows if we can show that A2 = A (i.e., A is a projection matrix) and r = tr(A). Since Λ is a projection matrix and Λφ = aφ, a must be either 0 or 1. Note that DΣD = Ik − φφτ . Then A3 = Σ1/2 DΛDΣDΛDΣDΛDΣ1/2 = Σ1/2 D(Λ − aφφτ )(Λ − aφφτ )ΛDΣ1/2 = Σ1/2 D(Λ − 2aφφτ + a2 φφτ )ΛDΣ1/2 = Σ1/2 D(Λ − aφφτ )ΛDΣ1/2

= Σ1/2 DΛDΣDΛDΣ1/2 = A2 ,

which implies that the eigenvalues of A must be 0 or 1. Therefore, A2 = A. Also, tr(A) = tr[Λ(DΣD)] = tr(Λ − aφφτ ) = tr(Λ) − a. (ii) The result in (ii) follows from the result in (i) and X/n →p p.

Note that the χ2 -statistic in (6.71) and the modified χ2 -statistic in (6.72) are special cases of the statistics in Theorem 6.8(i) and (ii), respectively, with Λ = Ik satisfying Λφ = φ. Hence, a test of asymptotic significance level α for testing (6.70) rejects H0 when χ2 > χ2k−1,α (or χ ˜2 > χ2k−1,α ), where χ2k−1,α is the (1 − α)th quantile of χ2k−1 . These tests are called (asymptotic) χ2 -tests. Example 6.23 (Goodness of fit tests). Let Y1 , ..., Yn be i.i.d. from F . Consider the problem of testing H0 : F = F0

versus

H1 : F 6= F0 ,

(6.73)

where F0 is a known c.d.f. For instance, F0 = N (0, 1). One way to test (6.73) is to partition the range of Y1 into k disjoint events A1 , ..., Ak and test (6.70) with pj = PF (Aj ) and p0j = PF0 (Aj ), j = 1, ..., k. Let Xj be the number of Yi ’s in Aj , j = 1, ..., k. Based on Xj ’s, the χ2 -tests discussed previously can be applied to this problem and they are called goodness of fit tests. In the goodness of fit tests discussed in Example 6.23, F0 in H0 is known so that p0j ’s can be computed. In some cases, we need to test the following hypotheses that are slightly different from those in (6.73): H0 : F = Fθ

versus

H1 : F 6= Fθ ,

(6.74)

438

6. Hypothesis Tests

where θ is an unknown parameter in Θ ⊂ Rs . For example, Fθ = N (µ, σ 2 ), θ = (µ, σ 2 ). If we still try to test (6.70) with pj = PFθ (Aj ), j = 1, ..., k, the result in Example 6.23 is not applicable since p is unknown under H0 . A generalized χ2 -test for (6.74) can be obtained using the following result. Let p(θ) = (p1 (θ), ..., pk (θ)) be a k-vector of known functions of θ ∈ Θ ⊂ Rs , where s < k. Consider the testing problem H0 : p = p(θ)

H1 : p 6= p(θ).

versus

(6.75)

Note that (6.70) is the special case of (6.75) with s = 0, i.e., θ is known. Let θˆ be an MLE of θ under H0 . Then, by Theorem 6.5, the LR test that rejects H0 when −2 log λn > χ2k−s−1,α has asymptotic significance level α, where χ2k−s−1,α is the (1 − α)th quantile of χ2k−s−1 and λn =

k Y

j=1

ˆ Xj [pj (θ)]

(Xj /n)Xj .

ˆ Using the fact that pj (θ)/(X j /n) →p 1 under H0 and log(1 + x) = x − x2 /2 + o(|x|2 )

as |x| → 0,

we obtain that ! ˆ pj (θ) −2 log λn = −2 −1 Xj log 1 + Xj /n j=1 ! k k ˆ X X pj (θ) −1 + = −2 Xj Xj Xj /n j=1 j=1 k X

ˆ pj (θ) −1 Xj /n

!2

+ op (1)

k X ˆ 2 [Xj − npj (θ)] + op (1) = Xj j=1

k X ˆ 2 [Xj − npj (θ)] + op (1), = ˆ npj (θ) j=1

Pk ˆ = Pk Xj /n = 1. Dewhere the third equality follows from j=1 pj (θ) j=1 ˜2 to be the χ2 and χ ˜2 in (6.71) fine the generalized χ2 -statistics χ2 and χ ˆ We then have the and (6.72), respectively, with p0j ’s replaced by pj (θ)’s. following result. Theorem 6.9. Under H0 given by (6.75), the generalized χ2 -statistics converge in distribution to χ2k−s−1 . The χ2 -test with rejection region χ2 > χ2k−s−1,α (or χ ˜2 > χ2k−s−1,α ) has asymptotic significance level α, where χ2k−s−1,α is the (1 − α)th quantile of χ2k−s−1 .

439

6.4. Tests in Parametric Models

Theorem 6.9 can be applied to derive a goodness of fit test for hypotheses (6.74). However, one has to formulate (6.75) and compute an MLE of θ under H0 : p = p(θ), which is different from an MLE under H0 : F = Fθ unless (6.74) and (6.75) are the same; see Moore and Spruill (1975). The next example is the main application of Theorem 6.9. Example 6.24 (r × c contingency tables). The following r × c contingency table is a natural extension of the 2 × 2 contingency table considered in Example 6.12: B1 B2 ··· Br Total

A1 X11 X21 ··· Xr1 m1

A2 X12 X22 ··· Xr2 m2

··· ··· ··· ··· ··· ···

Ac X1c X2c ··· Xrc mc

Total n1 n2 ··· nr n

where Ai ’s are disjoint events with A1 ∪ · · · ∪ Ac = Ω (the sample space of a random experiment), Bi ’s are disjoint events with B1 ∪ · · · ∪ Br = Ω, and Xij is the observed frequency of the outcomes in Aj ∩ Bi . Similar to the case of the 2 × 2 contingency table discussed in Example 6.12, there are two important applications in this problem. We first consider testing independence of {Aj : j = 1, ..., c} and {Bi : i = 1, ..., r} with hypotheses H0 : pij = pi· p·j for all i, j

versus

H1 : pij 6= pi· p·j for some i, j,

where pij = P (Aj ∩ Bi ) = E(Xij )/n, pi· = P (Bi ), and p·j = P (Aj ), i = 1, ..., r, j = 1, ..., c. In this case, X = (Xij , i = 1, ..., r, j = 1, ..., c) has the multinomial distribution with parameters pij , i = 1, ..., r, j = ¯ i· = ni /n and X ¯ ·j = mj /n, 1, ..., c. Under H0 , MLE’s of pi· and p·j are X respectively, i = 1, ..., r, j = 1, ..., c (exercise). By Theorem 6.9, the χ2 -test rejects H0 when χ2 > χ2(r−1)(c−1),α , where χ2 =

c r X X ¯ i· X ¯ ·j )2 (Xij − nX ¯ i· X ¯ ·j nX i=1 j=1

(6.76)

and χ2(r−1)(c−1),α is the (1 − α)th quantile of the chi-square distribution χ2(r−1)(c−1) (exercise). One can also obtain the modified χ2 -test by replacing ¯ i· X ¯ ·j by Xij in the denominator of each term of the sum in (6.76). nX Next, suppose that (X1j , ..., Xrj ), j = 1, ..., c, are c independent random vectors having the multinomial distributions with parameters (p1j , ..., prj ), j = 1, ..., c, respectively. Consider the problem of testing whether c multinomial distributions are the same, i.e., H0 : pij = pi1 for all i, j

versus

H1 : pij 6= pi1 for some i, j.

440

6. Hypothesis Tests

It turns out that the rejection region of the χ2 -test given in Theorem 6.9 is still χ2 > χ2(r−1)(c−1),α with χ2 given by (6.76) (exercise). One can also obtain the LR test in this problem. When r = c = 2, the LR test is equivalent to Fisher’s exact test given in Example 6.12, which is a UMPU test. When r > 2 or c > 2, however, a UMPU test does not exist in this problem.

6.4.4 Bayes tests An LR test actually compares supθ∈Θ0 ℓ(θ) with supθ∈Θ1 ℓ(θ) for testing (6.59). Instead of comparing two maximum values, one may compare two R R averages such as π ˆj = Θj ℓ(θ)dΠ(θ)/ Θ ℓ(θ)dΠ(θ), j = 0, 1, where Π(θ) is ˆ1 > π ˆ0 . If Π is treated as a prior c.d.f., a c.d.f. on Θ, and reject H0 when π then π ˆj is the posterior probability of Θj , and this test is a particular Bayes action (see Exercise 18 in §4.6) and is called a Bayes test. In Bayesian analysis, one often considers the Bayes factor defined to be β=

posterior odds ratio π ˆ0 /ˆ π1 = , prior odds ratio π0 /π1

where πj = Π(Θj ) is the prior probability of Θj . Clearly, if there is a statistic sufficient for θ, then the Bayes test and Bayes factor depend only on the sufficient statistic. Consider the special case where Θ0 = {θ0 } and Θ1 = {θ1 } are simple hypotheses. For given X = x, π ˆj =

πj fθj (x) . π0 fθ0 (x) + π1 fθ1 (x)

Rejecting H0 when π ˆ1 > π ˆ0 is the same as rejecting H0 when π0 fθ1 (x) > . fθ0 (x) π1

(6.77)

This is equivalent to the UMP test T∗ in (6.3) (Theorem 6.1) with c = π0 /π1 and γ = 0. The Bayes factor in this case is β=

π ˆ0 π1 fθ (x) . = 0 π ˆ1 π0 fθ1 (x)

Thus, the UMP test T∗ in (6.3) is equivalent to the test that rejects H0 when the Bayes factor is small. Note that the rejection region given by (6.77) depends on prior probabilities, whereas the Bayes factor does not. When either Θ0 or Θ1 is not simple, however, Bayes factors also depend on the prior Π.

441

6.4. Tests in Parametric Models

If Π is an improper prior, the Bayes test is still defined as long as the posterior probabilities π ˆj are finite. However, the Bayes factor may not be well defined when Π is improper. Example 6.25. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with an unknown µ ∈ R and a known σ 2 > 0. Let the prior of µ be N (ξ, τ 2 ). Then the posterior of µ is N (µ∗ (x), c2 ), where µ∗ (x) =

σ2 nτ 2 ξ+ 2 x ¯ 2 +σ nτ + σ 2

nτ 2

c2 =

and

τ 2 σ2 nτ 2 + σ 2

(see Example 2.25). Consider first the problem of testing H0 : µ ≤ µ0 versus H1 : µ > µ0 . Let Φ be the c.d.f. of the standard normal. Then the posterior probability of Θ0 and the Bayes factor are, respectively, ξ−µ0 µ −µ (x) Φ 0 c∗ Φ µ0 −µ∗ (x) µ0τ−ξ . and β= π ˆ0 = Φ µ∗ (x)−µ0 c Φ

c

Φ

τ

It is interesting to see that if we let τ → ∞, which is the same as considering the improper prior Π = the Lebesgue measure on R, then µ0√ −¯ x , π ˆ0 → Φ σ/ n

which is exactly the p-value α ˆ (x) derived in Example 2.29. Consider next the problem of testing H0 : µ = µ0 versus H1 : µ 6= µ0 . In this case the prior c.d.f. cannot be continuous at µ0 . We consider Π(µ) = π0 I[µ0 ,∞) (µ) + (1 − π0 )Φ µ−ξ . Let ℓ(µ) be the likelihood function based τ on x¯. Then Z x ¯−ξ 1 ′ √ √ m1 (x) = , = ℓ(µ)dΦ µ−ξ Φ τ 2 2 2 2 µ6=µ0

τ +σ /n

τ +σ /n

where Φ′ (t) is the p.d.f. of the standard normal distribution, and −1 1 − π0 π0 ℓ(µ0 ) = 1+ , π ˆ0 = π0 ℓ(µ0 ) + (1 − π0 )m1 (x) π0 β where

√ ¯−µ 2 + σ 2 Φ′ x √0 nτ ℓ(µ0 ) σ/ n = β= m1 (x) σΦ′ √ 2x¯−ξ2 τ +σ /n

is the Bayes factor.

More discussions about Bayesian hypothesis tests can be found in Berger (1985, §4.3.3).

442

6. Hypothesis Tests

6.5 Tests in Nonparametric Models In a nonparametric problem, a UMP, UMPU, or UMPI test usually does not exist. In this section we study some nonparametric tests that have size α, limiting size α, or asymptotic significance level α. Consistency (Definition 2.13) of these nonparametric tests is also discussed. Nonparametric tests are derived using some intuitively appealing ideas. They are commonly referred to as distribution-free tests, since almost no assumption is imposed on the population under consideration. But a nonparametric test may not be as good as a parametric test (in terms of its power) when the parametric model is correct. This is very similar to the case where we consider parametric estimation methods versus nonparametric estimation methods.

6.5.1 Sign, permutation, and rank tests Three popular classes of nonparametric tests are introduced here. The first one is the class of sign tests. Let X1 , ..., Xn be i.i.d. random variables from F , u be a fixed constant, and p = F (u). Consider the problem of testing H0 : p ≤ p0 versus H1 : p > p0 , or testing H0 : p = p0 versus H1 : p 6= p0 , where p0 is a fixed constant in (0, 1). Let ∆i =

1 0

Xi − u ≤ 0 Xi − u > 0,

i = 1, ..., n.

Then ∆1 , ..., ∆n are i.i.d. binary random variables with p = P (∆i = 1). For testing H0 : p ≤ p0 versus H1 : p > p0 , it follows from Corollary 6.1 that the test Y >m 1 T∗ (Y ) = (6.78) γ Y =m 0 Y <m Pn is of size α and UMP among tests based on ∆i ’s, where Y = i=1 ∆i and m and γ satisfy (6.7). Although T∗ is of size α, we cannot conclude immediately that T∗ is a UMP test, since ∆1 , ..., ∆n may not be sufficient for F . However, it can be shown that T∗ is in fact a UMP test (Lehmann, 1986, pp. 106-107) in this particular case. Note that no assumption is imposed on F . For testing H0 : p = p0 versus H1 : p 6= p0 , it follows from Theorem 6.4 that the test Y < c1 or Y > c2 1 T∗ (Y ) = (6.79) γ Y = ci , i = 1, 2, i 0 c1 < Y < c2

443

6.5. Tests in Nonparametric Models

is of size α and UMP among unbiased tests based on ∆i ’s, where γ and ci ’s are chosen so that E(T∗ ) = α and E(T∗ Y ) = αnp0 when p = p0 . This test is in fact a UMPU test (Lehmann, 1986, p. 166). Since Y is equal to the number of nonnegative signs of (u − Xi )’s, tests based on T∗ in (6.78) or (6.79) are called sign tests. One can easily extend the sign tests to the case where p = P (X1 ∈ B) with any fixed event B. Another extension is to the case where we observe i.i.d. (X1 , Y1 ), ..., (Xn , Yn ) (matched pairs). By using ∆i = Xi − Yi − u, one can obtain sign tests for hypotheses concerning P (X1 − Y1 ≤ u). Next, we introduce the class of permutation tests. Let Xi1 , ..., Xini , i = 1, 2, be two independent samples i.i.d. from Fi , i = 1, 2, respectively, where Fi ’s are c.d.f.’s on R. In §6.2.3, we showed that the two-sample t-tests are UMPU tests for testing hypotheses concerning the means of Fi ’s, under the assumption that Fi ’s are normal with the same variance. Such types of testing problems arise from the comparison of two treatments. Suppose now we remove the normality assumption and replace it by a much weaker assumption that Fi ’s are in the nonparametric family F containing all continuous c.d.f.’s on R. Consider the problem of testing H0 : F1 = F2

versus

H1 : F1 6= F2 ,

(6.80)

which is the same as testing the equality of the means of Fi ’s when Fi ’s are normal with the same variance. Let X = (Xij , j = 1, ..., ni , i = 1, 2), n = n1 + n2 , and α be a given significance level. A test T (X) satisfying 1 X T (z) = α n!

(6.81)

z∈π(x)

is called a permutation test, where π(x) is the set of n! points obtained from x ∈ Rn by permuting the components of x. Permutation tests are of size α (exercise). Under the assumption that F1 (x) = F2 (x − θ) and F1 ∈ F containing all c.d.f.’s having Lebesgue p.d.f.’s that are continuous a.e., which is still much weaker than the assumption that Fi ’s are normal with the same variance, the class of permutation tests of size α is exactly the same as the class of unbiased tests of size α; see, for example, Lehmann (1986, p. 231). Unfortunately, a test UMP among all permutation tests of size α does not exist. In applications, we usually choose a Lebesgue p.d.f. h and define a permutation test h(X) > hm 1 T (X) = (6.82) γ h(X) = hm 0 h(X) < hm ,

444

6. Hypothesis Tests

where hm is the (m + 1)th largest value of the set {h(z) : z ∈ π(x)}, m is the integer part of αn!, and γ = αn! − m. This permutation test is optimal in some sense (Lehmann, 1986, §5.11). While the class of permutation tests is motivated by the unbiasedness principle, the third class of tests introduced here is motivated by the invariance principle. Consider first the one-sample problem in which X1 , ..., Xn are i.i.d. random variables from a continuous c.d.f. F and we would like to test H0 : F is symmetric about 0 versus H1 : F is not symmetric about 0. Let G be the class of transformations g(x) = (ψ(x1 ), ..., ψ(xn )), where ψ is ˜ continuous, odd, and strictly increasing. Let R(X) be the vector of ranks of ˜ |Xi |’s and R+ (X) (or R− (X)) be the subvector of R(X) containing ranks corresponding to positive (or negative) Xi ’s. It can be shown (exercise) that (R+ , R− ) is maximal invariant under G. Furthermore, sufficiency permits o a reduction from R+ and R− to R+ , the vector of ordered components of o R+ . A test based on R+ is called a (one-sample) signed rank test. Similar to the case of permutation tests, there is no UMP test within the class of signed rank tests. A common choice is the signed rank test that o rejects H0 when W (R+ ) is too large or too small, where o o o W (R+ ) = J(R+1 /n) + · · · + J(R+n /n), ∗

(6.83)

o J is a continuous and strictly increasing function on [0, 1], R+i is the ith o component of R+ , and n∗ is the number of positive Xi ’s. This is motivated by the fact that H0 is unlikely to be true if W in (6.83) is too large or too small. Note that W/n is equal to T(Fn ) with T given by (5.53) and J(t) = t, and the test based on W in (6.83) is the well-known one-sample Wilcoxon signed rank test. o Under H0 , P (R+ = y) = 2−n for each y ∈ Y containing 2n n∗ -tuples y = (y1 , ..., yn∗ ) satisfying 1 ≤ y1 < · · · < yn∗ ≤ n. Then, the following signed rank test is of size α:

1 T (X) = γ 0

o o W (R+ ) < c1 or W (R+ ) > c2 o W (R+ ) = ci , i = 1, 2 o ) < c2 , c1 < W (R+

(6.84)

where c1 and c2 are the (m + 1)th smallest and largest values of the set {W (y) : y ∈ Y}, m is the integer part of α2n /2, and γ = α2n /2 − m. Consider next the two-sample problem of testing (6.80) based on two independent samples, Xi1 , ..., Xini , i = 1, 2, i.i.d. from Fi , i = 1, 2, respectively. Let G be the class of transformations g(x) = (ψ(xij ), j = 1, ..., ni , i =

6.5. Tests in Nonparametric Models

445

1, 2), where ψ is continuous and strictly increasing. Let R(X) be the vector of ranks of all Xij ’s. In Example 6.14, we showed that R is maximal invariant under G. Again, sufficiency permits a reduction from R to R1o , the vector of ordered values of the ranks of X11 , ..., X1n1 . A test for (6.80) based on R1o is called a two-sample rank test. Under H0 , P (R1o = y) = n −1 for each y ∈ Y containing nn1 n1 -tuples y = (y1 , ..., yn1 ) satisfying n1 o o 1 ≤ y1 < · · · < yn1 ≤ n. Let R1o = (R11 , ..., R1n ). Then a commonly 1 o used two-sample rank test is given by (6.83)-(6.84) with R+i , n∗ , and 2n n o replaced by R1i , n1 , and n1 , respectively. When n1 = n2 , the statistic W/n is equal to T(Fn ) with T given by (5.55). When J(t) = t − 21 , this reduces to the well-known two-sample Wilcoxon rank test. A common feature of the permutation and rank tests previously introduced is that tests of size α can be obtained for each fixed sample size n, but the computation involved in determining the rejection regions {T (X) = 1} may be cumbersome if n is large. Thus, one may consider approximations to permutation and rank tests when n is large. Permutation tests can often be approximated by the two-sample t-tests derived in §6.2.3 (Lehmann, 1986, §5.13). Using the results in §5.2.2, we now derive one-sample signed rank tests having limiting size α (Definition 2.13(ii)), which can be viewed as signed rank tests of size approximately α when n is large. From the discussion in §5.2.2, W/n = T(Fn ) with a ̺∞ -Hadamard differentiable functional T given by (5.53) and, by Theorem 5.5, √ n[W/n − T(F )] →d N (0, σF2 ), where σF2 = E[φF (X1 )]2 , Z ∞ J ′ (F˜ (y))(˜διx − F˜ )(y)dF (y) + J(F˜ (x)) − T(F ) φF (x) = 0

(see (5.54)), and διx denotes the c.d.f. degenerated at x. Since F is continuous, F˜ (x) = F (x) − F (−x). Under H0 , F (x) = 1 − F (−x). Hence, σF2 under H0 is equal to v1 + v2 + 2v12 , where Z 1 ∞ v1 = Var J(F˜ (X1 )) = [J(F˜ (x))]2 dF˜ (x), 2 0 Z

∞

˜ ˜ ˜ J (F (y))(διX1 − F )(y)dF (y) ′

v2 = Var 0 Z ∞Z ∞ =E J ′ (F˜ (y))J ′ (F˜ (z))(˜διX1 − F˜ )(y)(˜διX1 − F˜ )(z)dF (y)dF (z) 0 0 Z Z 1 ∞ ∞ ′ ˜ = J (F (y))J ′ (F˜ (z))[F˜ (min{y, z}) − F˜ (y)F˜ (z)]dF˜ (y)dF˜ (z) 4 0 0 Z 1 J ′ (F˜ (y))J ′ (F˜ (z))F˜ (z)[1 − F˜ (y)]dF˜ (y)dF˜ (z), = 2 0 σ0 z1−α/2 → α,

i.e., T has limiting size α. Two-sample rank tests having limiting size α can be similarly derived (exercise).

6.5.2 Kolmogorov-Smirnov and Cram´ er-von Mises tests In this section we introduce two types of tests for hypotheses concerning continuous c.d.f.’s on R. Let X1 , ..., Xn be i.i.d. random variables from a continuous c.d.f. F . Suppose that we would like to test hypotheses (6.73), i.e., H0 : F = F0 versus H1 : F 6= F0 with a fixed F0 . Let Fn be the empirical c.d.f. and Dn (F ) = sup |Fn (x) − F (x)|,

(6.86)

x∈R

which is in fact the distance ̺∞ (Fn , F ). Intuitively, Dn (F0 ) should be small if H0 is true. From the results in §5.1.1, we know that Dn (F0 ) →a.s. 0 if and

6.5. Tests in Nonparametric Models

447

only if H0 is true. The statistic Dn (F0 ) is called the Kolmogorov-Smirnov statistic. Tests with rejection region Dn (F0 ) > c are called KolmogorovSmirnov tests. In some cases we would like to test “one-sided” hypotheses H0 : F = F0 versus H1 : F ≥ F0 , F 6= F0 , or H0 : F = F0 versus H1 : F ≤ F0 , F 6= F0 . The corresponding Kolmogorov-Smirnov statistic is Dn+ (F0 ) or Dn− (F0 ), where Dn+ (F ) = sup [Fn (x) − F (x)] (6.87) x∈R

and Dn− (F ) = sup [F (x) − Fn (x)]. x∈R

The rejection regions of one-sided Kolmogorov-Smirnov tests are, respectively, Dn+ (F0 ) > c and Dn− (F0 ) > c. Let X(1) < · · · < X(n) be the order statistics and define X(0) = −∞ and X(n+1) = ∞. Since Fn (x) = i/n when X(i) ≤ x < X(i+1) , i = 0, 1, ..., n, i + − F (x) sup Dn (F ) = max 0≤i≤n X(i) ≤x<X(i+1) n i = max − inf F (x) 0≤i≤n n X(i) ≤x<X(i+1) i − F (X(i) ) . = max 0≤i≤n n When F is continuous, F (X(i) ) is the ith order statistic of a sample of size n from the uniform distribution U (0, 1) irrespective of what F is. Therefore, the distribution of Dn+ (F ) does not depend on F , if we restrict our attention to continuous c.d.f.’s on R. The distribution of Dn− (F ) is the same as that of Dn+ (F ) because of symmetry (exercise). Since Dn (F ) = max{Dn+ (F ), Dn− (F )}, the distribution of Dn (F ) does not depend on F . This means that the distributions of Kolmogorov-Smirnov statistics are known under H0 . Theorem 6.10. Let Dn (F ) and Dn+ (F ) be defined by (6.86) and (6.87), respectively, for a continuous c.d.f. F on R. (i) For any fixed n, 0 t≤0 n Z un−i+2 Y n! du1 · · · dun 0 0.

j=−∞

6.5.3 Empirical likelihood ratio tests The method of likelihood ratio is useful in deriving tests under parametric models. In nonparametric problems, we now introduce a similar method based on the empirical likelihoods introduced in §5.1.2 and §5.1.4. Suppose that a sample X is from a population determined by a c.d.f. F ∈ F, where F is a class of c.d.f.’s on Rd . Consider the problem of testing H0 : T(F ) = t0

versus

H1 : T(F ) 6= t0 ,

(6.89)

where T is a functional from F to Rk and t0 is a fixed vector in Rk . Let ℓ(G), G ∈ F, be a given empirical likelihood, Fˆ be an MELE of F , and FˆH0 be an MELE of F under H0 , i.e., FˆH0 is an MELE of F subject to T(F ) = t0 . Then the empirical likelihood ratio is defined as λn (X) = ℓ(FˆH0 )/ℓ(Fˆ ). A test with rejection region λn (X) < c is called an empirical likelihood ratio test.

450

6. Hypothesis Tests

As a specific example, consider the following empirical likelihood (or nonparametric likelihood) when X = (X1 , ..., Xn ) with i.i.d. Xi ’s: ℓ(G) =

n Y

subject to pi ≥ 0,

pi

i=1

n X

pi = 1,

i=1

R where pi = PG ({xi }), i = 1, ..., n. Suppose that T(G) = u(x)dG(x) with a known function u(x) from Rd to Rr . Then Fˆ = Fn ; H0 in (6.89) with t0 = 0 is the same as the case where assumption (5.9) holds; FˆH0 is the MELE given by (5.11); and the empirical likelihood ratio is λn (X) = nn

n Y

pˆi ,

(6.90)

i=1

where pˆi is given by (5.12). An empirical likelihood ratio test with asymptotic significance level α can be obtained using the following result. Theorem 6.11. Assume the conditions in Theorem 5.4. Under the hypothesis H0 in (6.89) with t0 = 0 (i.e., (5.9) holds), −2 log λn →d χ2r , where λn = λn (X) is given by (6.90) and χ2r has the chi-square distribution χ2r . The proof of this result can be found in Owen (1988, 1990). In fact, the result in Theorem 6.11 holds for some other functionals T such as the median functional. We can also derive tests based on the profile empirical likelihoods discussed in §5.4.1. Consider an empirical likelihood ℓ(G) =

n Y

pi

subject to

i=1

pi ≥ 0,

n X i=1

pi = 1,

n X

pi ψ(xi , θ) = 0,

i=1

where θ is a k-vector of unknown parameters and ψ is a known function. Let θ = (ϑ, ϕ), where ϑ is an r-vector and ϕ is a (k − r)-vector. Suppose that we would like to test H0 : ϑ = ϑ0

versus

H1 : ϑ 6= ϑ0 ,

where ϑ0 is a fixed r-vector. Let θˆ be a maximum of the profile empirical likelihood ℓP (θ) given by (5.36) and let ϕˆ be a maximum of ℓP (ϕ) = ℓP (ϑ0 , ϕ). Then a profile empirical likelihood ratio test rejects H0 when λn (X) < c, where λn (X) =

n Y

ˆ τ ψ(xi , θ) ˆ 1 + [ξn (θ)] , 1 + [ζn (ϑ0 , ϕ)] ˆ τ ψ(xi , ϑ0 , ϕ) ˆ i=1

(6.91)

6.5. Tests in Nonparametric Models

451

ˆ satisfies θˆ and ϕˆ are maximum profile empirical likelihood estimators, ξn (θ) n X

ˆ ψ(xi , θ) = 0, τ ˆ ˆ 1 + [ξn (θ)] ψ(xi , θ)

i=1

ˆ satisfies and ζn (ϑ0 , ϕ) n X i=1

ψ(xi , ϑ0 , ϕ) ˆ = 0. τ 1 + [ζn (ϑ0 , ϕ)] ˆ ψ(xi , ϑ0 , ϕ) ˆ

Pn From the discussion in §5.4.1, θˆ is a solution of the GEE i=1 ψ(Xi , θ) = 0 when the dimension of ψ is k. Under some regularity conditions (e.g., the conditions in Proposition 5.3), Qin and Lawless (1994) showed that the result in Theorem 6.11 holds with λn (X) given by (6.91). Thus, a profile empirical likelihood ratio test with asymptotic significance level α can be obtained. Example 6.26. Let Y1 , ..., Yn be i.i.d. random 2-vectors from F . Consider the problem of testing H0 : µ1 = µ2 versus H1 : µ1 6= µ2 , where (µ1 , µ2 ) = E(Y1 ). Let Yi = (Yi1 , Yi2 ), Xi1 = Yi1 − Yi2 , Xi2 = Yi1 + Yi2 , and Xi = (Xi1 , Xi2 ), i = 1, ..., n. Then X1 , ..., Xn are i.i.d. with E(X1 ) = θ = (ϑ, ϕ), where ϑ = µ1 − µ2 and ϕ = µ1 + µ2 . The hypotheses of interest becomes H0 : ϑ = 0 versus H1 : ϑ 6= 0. To apply the profile empirical likelihood method, Pn we define ψ(x, θ) = x − θ, x ∈ R2 . Note that a solution of the GEE i=1 (Xi − θ) = 0 is the ¯ The profile empirical likelihood ratio is then given by sample mean θˆ = X. λn (X) =

n Y

¯ τ (Xi − X) ¯ 1 + [ξn (X)] , τ 1 + [ζn (0, ϕ)] ˆ [Xi − (0, ϕ)] ˆ i=1

¯ ζn (0, ϕ), where ξn (X), ˆ and ϕˆ satisfy n X i=1

n X i=1

¯ Xi − X τ ¯ ¯ = 0, 1 + [ξn (X)] (Xi − X)

Xi − (0, ϕ) ˆ = 0, τ 1 + [ζn (0, ϕ)] ˆ [Xi − (0, ϕ)] ˆ

and ℓP (0, ϕ) ˆ = maxϕ ℓP (0, ϕ) with ℓP (0, ϕ) =

n Y

1 . τ [X − (0, ϕ)]} n{1 + [ξ (0, ϕ)] n i i=1

Empirical likelihood ratio tests or profile empirical likelihood ratio tests in various other problems can be found, for example, in Owen (1988, 1990, 2001), Chen and Qin (1993), Qin (1993), and Qin and Lawless (1994).

452

6. Hypothesis Tests

6.5.4 Asymptotic tests We now introduce a simple method of constructing asymptotic tests (i.e., tests with asymptotic significance level α). This method works for almost all problems (parametric or nonparametric) in which the hypotheses being tested are H0 : θ = θ0 versus H1 : θ 6= θ0 , where θ is a vector of parameters, and an asymptotically normally distributed estimator of θ can be found. However, this simple method may not provide the best or even nearly best solution to the problem, especially when there are different asymptotically normally distributed estimators of θ. Let X be a sample of size n from a population P and θˆn be an estimator of θ, a k-vector of parameters related to P . Suppose that under H0 , Vn−1/2 (θˆn − θ) →d Nk (0, Ik ),

(6.92)

where Vn is the asymptotic covariance matrix of θˆn . If Vn is known when θ = θ0 , then a test with rejection region (θˆn − θ0 )τ Vn−1 (θˆn − θ0 ) > χ2k,α

(6.93)

has asymptotic significance level α, where χ2k,α is the (1 − α)th quantile of the chi-squared distribution χ2k . If the distribution of θˆn does not depend on the unknown population P under H0 and (6.92) holds, then a test with rejection region (6.93) has limiting size α. If Vn in (6.93) depends on the unknown population P even if H0 is true (θ = θ0 ), then we have to replace Vn in (6.93) by an estimator Vˆn . If, under H0 , Vˆn is consistent according to Definition 5.4, then the test having rejection region (6.93) with Vn replaced by Vˆn has asymptotic significance level α. Variance estimation methods introduced in §5.5 can be used to construct a consistent estimator Vˆn . In some cases result (6.92) holds for any P . Then, the following result shows that the test having rejection region (6.93) is asymptotically correct (§2.5.3), i.e., it is a consistent asymptotic test (Definition 2.13). Theorem 6.12. Assume that (6.92) holds for any P and that λ+ [Vn ] → 0, where λ+ [Vn ] is the largest eigenvalue of Vn . (i) The test having rejection region (6.93) (with a known Vn or Vn replaced by an estimator Vˆn that is consistent for any P ) is consistent. (ii) If we choose α = αn → 0 as n → ∞ and χ2k,1−αn λ+ [Vn ] = o(1), then the test in (i) is Chernoff-consistent. Proof. The proof of (ii) is left as an exercise. We only prove (i) for the −1/2 ˆ −1/2 case where Vn is known. Let Zn = Vn (θn − θ) and ln = Vn (θ − θ0 ). −1/2 Then kZn k = Op (1) and kln k = kVn (θ − θ0 )k → ∞ when θ 6= θ0 . The

6.5. Tests in Nonparametric Models

453

result follows from the fact that when θ 6= θ0 ,

(θˆn − θ0 )τ Vn−1 (θˆn − θ0 ) = kZn k2 + kln k2 + 2lnτ Zn ≥ kZn k2 + kln k2 − 2kln kkZn k = Op (1) + kln k2 [1 − op (1)]

and, therefore,

P (θˆn − θ0 )τ Vn−1 (θˆn − θ0 ) > χ2k,α → 1.

Example 6.27. Let X1 , ..., Xn be i.i.d. random variables from a symmetric c.d.f. F having finite variance and positive F ′ . Consider the problem of testing H0 : F is symmetric about 0 versus H1 : F is not symmetric about 0. Under H0 , there are many estimators satisfying (6.92). We consider the following five estimators: ¯ and θ = E(X1 ); (1) θˆn = X ˆ ˆ (2) θn = θ0.5 (the sample median) and θ = F −1 ( 12 ) (the median of F ); ¯ a (the a-trimmed sample mean defined by (5.77)) and θ = T(F ), (3) θˆn = X where T is given by (5.46) with J(t) = (1 − 2a)−1 I(a,1−a) (t), a ∈ (0, 12 ); (4) θˆn = the Hodges-Lehmann estimator (Example 5.8) and θ = F −1 ( 12 ); (5) θˆn = W/n − 12 , where W is given by (6.83) with J(t) = t, and θ = T(F ) − 12 with T given by (5.53). Although the θ’s in (1)-(5) are different in general, in all cases θ = 0 is equivalent to that H0 holds. ¯ it follows from the CLT that (6.92) holds with Vn = σ 2 /n for any For X, F , where σ 2 = Var(X1 ). From the SLLN, S 2 /n is a consistent estimator of ¯ and Vn for any F . Thus, the test having rejection region (6.93) with θˆn = X 2 Vn replaced by S /n is asymptotically correct. This test is asymptotically equivalent to the one-sample t-test derived in §6.2.3. From Theorem 5.10, θˆ0.5 satisfies (6.92) with Vn = 4−1 [F ′ (θ)]−2 n−1 for any F . A consistent estimator of Vn can be obtained using the bootstrap method considered in §5.5.3. Another consistent estimator of Vn can be obtained using Woodruff’s interval introduced in §7.4 (see Exercise 86 in §7.6). The test having rejection region (6.93) with θˆn = θˆ0.5 and Vn replaced by a consistent estimator is asymptotically correct. ¯ a satisfies (6.92) for any It follows from the discussion in §5.3.2 that X F . A consistent estimator of Vn can be obtained using formula (5.110) or the jackknife method in §5.5.2. The test having rejection region (6.93) ¯ a and Vn replaced by a consistent estimator is asymptotically with θˆn = X correct. From Example 5.8, the Hodges-Lehmann estimator satisfies (6.92) for R any F and Vn = 12−1 γ −2 n−1 under H0 , where γ = F ′ (x)dF (x). A

454

6. Hypothesis Tests

consistent estimator of Vn under H0 can be obtained using the result in Exercise 102 in §5.6. The test having rejection region (6.93) with θˆn = the Hodges-Lehmann estimator and Vn replaced by a consistent estimator is asymptotically correct. Note that all tests discussed so far are not of limiting size α, since the distributions of θˆn are still unknown under H0 . The test having rejection region (6.93) with θˆn = W/n − 21 and Vn = (12n)−1 is equivalent to the one-sample Wilcoxon signed rank test and is shown to have limiting size α (§6.5.1). Also, (6.92) is satisfied for any F (§5.2.2). Although Theorem 6.12 is not applicable, a modified proof of Theorem 6.12 can be used to show the consistency of this test (exercise). It is not clear which one of the five tests discussed here is to be preferred in general. The results for θˆn in (1)-(3) and (5) still hold for testing H0 : θ = 0 versus H1 : θ 6= 0 without the assumption that F is symmetric. An example of asymptotic tests for one-sided hypotheses is given in Exercise 123. Most tests in §6.1-§6.4 derived under parametric models are asymptotically correct even when the parametric model assumptions are removed. Some examples are given in Exercises 121-123. Finally, a study of asymptotic efficiencies of various tests can be found, for example, in Serfling (1980, Chapter 10).

6.6 Exercises 1. Prove Theorem 6.1 for the case of α = 0 or 1. 2. Assume the conditions in Theorem 6.1. Let β(P ) be the power function of a UMP test of size α ∈ (0, 1). Show that α < β(P1 ) unless P0 = P1 . 3. Let T∗ be given by (6.3) with c = c(α) for an α > 0. (a) Show that if α1 < α2 , then c(α1 ) ≥ c(α2 ). (b) Show that if α1 < α2 , then the type II error probability of T∗ of size α1 is larger than that of T∗ of size α2 . 4. Let H0 and H1 be simple and let α ∈ (0, 1). Suppose that T∗ is a UMP test of size α for testing H0 versus H1 and that β < 1, where β is the power of T∗ when H1 is true. Show that 1 − T∗ is a UMP test of size 1 − β for testing H1 versus H0 . 5. Let X be a sample of size 1 from a Lebesgue p.d.f. fθ . Find a UMP test of size α ∈ (0, 21 ) for H0 : θ = θ0 versus H1 : θ = θ1 when

6.6. Exercises

455

(a) fθ (x) = 2θ−2 (θ − x)I(0,θ) (x), θ0 < θ1 ; (b) fθ (x) = 2[θx + (1 − θ)(1 − x)]I(0,1) (x), 0 ≤ θ1 < θ0 ≤ 1; (c) fθ0 is the p.d.f. of N (0, 1) and fθ1 is the p.d.f. of the Cauchy distribution C(0, 1); (d) fθ0 (x) = 4xI(0, 21 ) (x) + 4(1 − x)I( 12 ,1) (x) and fθ1 (x) = I(0,1) (x); (e) fθ is the p.d.f. of the Cauchy distribution C(θ, 1) and θi = i; (f) fθ0 (x) = e−x I(0,∞) (x) and fθ1 (x) = 2−1 x2 e−x I(0,∞) (x). 6. Let X1 , ..., Xn be i.i.d. from a Lebesgue p.d.f. fθ . Find a UMP test of size α for H0 : θ = θ0 versus H1 : θ = θ1 in the following cases: (a) fθ (x) = e−(x−θ) I(θ,∞) (x), θ0 < θ1 ; (b) fθ (x) = θx−2 I(θ,∞) (x), θ0 6= θ1 . 7. Prove Proposition 6.1. 8. Let X ∈ Rn be a sample with a p.d.f. f w.r.t. a σ-finite measure ν. Consider the problem of testing H0 : f = fθ versus H1 : f = g, where θ ∈ Θ, fθ (x) is Borel on (Rn × Θ, σ(B n × F )), and (Θ, F , Λ) is a probability space. Let c > 0 be a constant and R 1 g(x) ≥ c Θ fθ (x)dΛ R φ∗ (x) = 0 g(x) < c Θ fθ (x)dΛ. R R Suppose that φ∗ (x)fθ (x)dν = supθ∈Θ φ∗ (x)fθ (x)dν = α for any θ ∈ Θ′ with Λ(Θ′ ) = 1. Show that φ∗ is a UMP test of size α. 9. Let f0 and f1 be Lebesgue integrable functions on R and φ∗ be the indicator function of the set R {x : f0 (x) < 0} ∪ {x : f0 (x) = 0, f1 (x) ≥ 0}. Show that φ∗ maximizes φ(x)f R 1 (x)dx over all RBorel functions φ on R satisfying 0 ≤ φ(x) ≤ 1 and φ(x)f0 (x)dx = φ∗ (x)f0 (x)dx.

10. Let F1 and F2 beR two c.d.f.’s on RR. Show that F1 (x) ≤ F2 (x) for all x if and only if g(x)dF2 (x) ≤ g(x)dF1 (x) for any nondecreasing function g. 11. Prove the claims in Example 6.5. 12. Show that the family {fθ : θ ∈ R} has monotone likelihood ratio, where fθ (x) = c(θ)h(x)I(a(θ),b(θ)) (x), h is a positive Lebesgue integrable function, and a and b are nondecreasing functions of θ. 13. Prove part (iv) and part (v) of Theorem 6.2. 14. Let X1 , ..., Xn be i.i.d. from a Lebesgue p.d.f. fθ , θ ∈ Θ ⊂ R. Find a UMP test of size α for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 when (a) fθ (x) = θ−1 e−x/θ I(0,∞) (x), θ > 0; (b) fθ (x) = θ−1 xθ−1 I(0,1) (x), θ > 0; (c) fθ (x) is the p.d.f. of N (1, θ); c (d) fθ (x) = θ−c cxc−1 e−(x/θ) I(0,∞) (x), θ > 0, where c > 0 is known.

456

6. Hypothesis Tests

15. Suppose that the distribution of X is in a family with monotone likelihood ratio in Y (X), where Y (X) has a continuous distribution. Consider the hypotheses H0 : θ ≤ θ0 versus H1 : θ > θ0 . Show that the p-value (§2.4.2) of the UMP test is given by Pθ0 (Y ≥ y), where y is the observed value of Y . 16. Let X1 , ..., Xm be i.i.d. from N (µx , σx2 ) and Y1 , ..., Yn be i.i.d. from N (µy , σy2 ). Suppose that Xi ’s and Yj ’s are independent. (a) When σx = σy = 1, find a UMP test of size α for testing H0 : µx ≤ µy versus H1 : µx > µy . (Hint: see Lehmann (1986, §3.9).) (b) When µx and µy are known, find a UMP test of size α for testing H0 : σx ≤ σy versus H1 : σx > σy . (Hint: see Lehmann (1986, §3.9).) 17. Let F and G be two known c.d.f.’s on R and X be a single observation from the c.d.f. θF (x) + (1 − θ)G(x), where θ ∈ [0, 1] is unknown. (a) Find a UMP test of size α for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 . (b) Show that the test T∗ (X) ≡ α is a UMP test of size α for testing H0 : θ ≤ θ1 or θ ≥ θ2 versus H1 : θ1 < θ < θ2 . 18. Let X1 , ..., Xn be i.i.d. from the uniform distribution U (θ, θ + 1), θ ∈ R. Suppose that n ≥ 2. (a) Find the joint distribution of X(1) and X(n) . (b) Show that a UMP test of size α for testing H0 : θ ≤ 0 versus H1 : θ > 0 is of the form 0 X(1) < 1 − α1/n , X(n) < 1 T∗ (X(1) , X(n) ) = 1 otherwise. (c) Does the family of all possible distributions of (X(1) , X(n) ) have monotone likelihood ratio? (Hint: see Lehmann (1986, p. 115).) 19. Suppose that X1 , ..., Xn are i.i.d. from the discrete uniform distribution DU (1, ..., θ) (Table 1.1, page 18) with an unknown θ = 1, 2, .... (a) Consider H0 : θ ≤ θ0 versus H1 : θ > θ0 . Show that 1 X(n) > θ0 T∗ (X) = α X(n) ≤ θ0 is a UMP test of size α. (b) Consider H0 : θ = θ0 versus H1 : θ 6= θ0 . Show that 1 X(n) > θ0 or X(n) ≤ θ0 α1/n T∗ (X) = 0 otherwise is a UMP test of size α. (c) Show that the results in (a) and (b) still hold if the discrete uniform distribution is replaced by the uniform distribution U (0, θ), θ > 0.

6.6. Exercises

457

20. Let X1 , ..., Xn be i.i.d. from the exponential distribution E(a, θ), a ∈ R, θ > 0. (a) Derive a UMP test of size α for testing H0 : a = a0 versus H1 : a 6= a0 , when θ is known. (b) For testing H0 : a = a0 versus H1 : a = a1 < a0 , show that any UMP test T∗ of size α satisfies βT∗ (a1 ) = 1 − (1 − α)e−n(a0 −a1 )/θ . (c) For testing H0 : a = a0 versus H1 : a = a1 < a0 , show that the power of any size α test that rejects H0 when Y ≤P c1 or Y ≥ c2 is the same as that in part (b), where Y = (X(1) − a0 )/ ni=1 (Xi − X(1) ). (d) Derive a UMP test of size α for testing H0 : a = a0 versus H1 : a 6= a0 . (e) Derive a UMP test of size α for testing H0 : θ = θ0 , a = a0 versus H1 : θ < θ0 , a < a0 . 21. Let X1 , ..., Xn be i.i.d. from the Pareto distribution P a(a, θ), θ > 0, a > 0. (a) Derive a UMP test of size α for testing H0 : a = a0 versus H1 : a 6= a0 when θ is known. (b) Derive a UMP test of size α for testing H0 : a = a0 , θ = θ0 versus H1 : θ > θ0 , a < a0 . 22. In Exercise 19(a) of §3.6, derive a UMP test of size α ∈ (0, 1) for testing H0 : θ ≤ θ0 versus H1 : θ > θ0 , where θ0 is known and θ0 > (1 − α)−1/n . 23. In Exercise 55 of §2.6, derive a UMP test of size α for testing H0 : θ ≥ θ0 versus H1 : θ < θ0 based on data X1 , ..., Xn , where θ0 > 0 is a fixed value. 24. Prove part (ii) of Theorem 6.3. 25. Consider Example 6.10. Suppose that θ2 = −θ1 . Show that c2 = −c1 and discuss how to find the value of c2 . 26. Suppose that the distribution of X is in a family of p.d.f.’s indexed by a real-valued parameter θ; there is a real-valued sufficient statistic U (X) such that fθ2 (u)/fθ1 (u) is strictly increasing in u for θ1 < θ2 , where fθ (u) is the Lebesgue p.d.f. of U (X) and is continuous in u for each θ; and that for all θ1 < θ2 < θ3 and u1 < u2 < u3 , fθ1 (u1 ) fθ1 (u2 ) fθ1 (u3 ) fθ (u1 ) fθ (u2 ) fθ (u3 ) > 0. 2 2 2 f (u ) f (u ) f (u ) θ3 1 θ3 2 θ3 3 Show that the conclusions of Theorem 6.3 remain valid.

458

6. Hypothesis Tests

27. (p-values). Suppose that X has a distribution Pθ , where θ ∈ R is unknown. Consider a family of nonrandomized level α tests for H0 : θ = θ0 (or θ ≤ θ0 ) with rejection region Cα such that Pθ0 (X ∈ Cα ) = α for all 0 < α < 1 and Cα1 = ∩α>α1 Cα for all 0 < α1 < 1. (a) Show that the p-value is α ˆ (x) = inf{α : x ∈ Cα }. (b) Show that when θ = θ0 , α(X) ˆ has the uniform distribution U (0, 1). (c) If the tests with rejection regions Cα are unbiased of level α, show that under H1 , Pθ (ˆ α(X) ≤ α) ≥ α. 28. Suppose that X has the p.d.f. (6.10). Consider hypotheses (6.13) or (6.14). Show that a UMP test does not exist. (Hint: this follows from a consideration of the UMP tests for the one-sided hypotheses H0 : θ ≥ θ1 and H0 : θ ≤ θ2 .) 29. Consider Exercise 17 with H0 : θ ∈ [θ1 , θ2 ] versus H1 : θ 6∈ [θ1 , θ2 ], where 0 < θ1 ≤ θ2 < 1. (a) Show that a UMP test does not exist. (b) Obtain a UMPU test of size α. 30. In the proof of Theorem 6.4, show that (a) (6.30) is equivalent to (6.31); (b) (6.31) is equivalent to (6.29) with T∗ replaced by T ; (c) when 0 < α < 1, (α, αEθ0 (Y )) is an interior point of the set of points (Eθ0 (T ), Eθ0 (T Y )) as T ranges over all tests of the form T = T (Y ); (d) the UMPU tests are unique a.s. P if attention is restricted to tests depending on (Y, U ) and (Y, U ) has a continuous c.d.f. 31. Consider the decision problem in Example 2.20 with the 0-1 loss. Show that if a UMPU test of size α exists and is unique (in the sense that decision rules that are equivalent in terms of the risk are treated the same), then it is admissible. 32. Let X1 , ..., Xn be i.i.d. binary random variables with p = P (X1 = 1). (a) Determine the ci ’s and γi ’s in (6.15) and (6.16) for testing H0 : p ≤ 0.2 or p ≥ 0.7 when α = 0.1 and n = 15. Find the power of the UMP test (6.15) when p = 0.4. (b) Derive a UMPU test of size α for H0 : p = p0 versus H1 : p 6= p0 when n = 10, α = 0.05, and p0 = 0.4. 33. Suppose that X has the Poisson distribution P (θ) with an unknown θ > 0. Show that (6.29) reduces to cX 2 −1

x=c1

2

θ0x−1 e−θ0 X θci −1 e−θ0 + = 1 − α, (1 − γi ) 0 (x − 1)! (ci − 1)! +1 i=1

provided that c1 > 1.

6.6. Exercises

459

34. Let X be a random variable from the geometric distribution G(p). Find a UMPU test of size α for H0 : p = p0 versus H1 : p 6= p0 . 35. In Exercise 33 of §2.6, derive a UMPU test of size α ∈ (0, 1) for testing H0 : p ≤ 12 versus H1 : p > 12 . 36. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ and σ 2 . (a) Show how the power of the one-sample t-test depends on a noncentral t-distribution. (b) Show that the power of the one-sample t-test is an increasing function of (µ − µ0 )/σ for testing H0 : µ ≤ µ0 versus H1 : µ > µ0 , and of |µ − µ0 |/σ for testing H0 : µ = µ0 versus H1 : µ 6= µ0 . 37. Let X1 , ..., Xn be i.i.d. from the gamma distribution Γ(θ, γ) with unknown θ and γ. (a) For testing H0 : θ ≤ θ0 versus H1 : θ > θ0 and H0 : θ = θ0 versus H1 : θ 6= θ0 , show that Qnthere exist UMPU tests whose rejection ¯ regions are based on V = i=1 (Xi /X). (b) For testing H0 : γP≤ γ0 versus H Q1n : γ > γ0 , show that a UMPU n test rejects H0 when i=1 Xi > C( i=1 Xi ) for some function C.

38. Let X1 and X2 be independently distributed as the Poisson distributions P (λ1 ) and P (λ2 ), respectively. (a) Find a UMPU test of size α for testing H0 : λ1 ≥ λ2 versus H1 : λ1 < λ2 . (b) Calculate the power of the UMPU test in (a) when α = 0.1, (λ1 , λ2 ) = (0.1, 0.2), (1,2), (10,20), and (0.1,0.4). 39. Consider the binomial problem in Example 6.11. (a) Prove the claim about P (Y = y|U = u). (b) Find a UMPU test of size α for testing H0 : p1 ≥ p2 versus H1 : p1 < p2 . (c) Repeat (b) for H0 : p1 = p2 versus H1 : p1 6= p2 . 40. Let X1 and X2 be independently distributed as the negative binomial distributions N B(p1 , n1 ) and N B(p2 , n2 ), respectively, where ni ’s are known and pi ’s are unknown. (a) Show that there exists a UMPU test of size α for testing H0 : p1 ≤ p2 versus H1 : p1 > p2 . (b) Determine the conditional distribution PY |U =u in Theorem 6.4 when n1 = n2 = 1. 41. Let (X0 , X1 , X2 ) be a random vector having a multinomial distribution (Example 2.7) with k = 2, p0 = 1 − p1 − p2 , and unknown p1 ∈ (0, 1) and p2 ∈ (0, 1). Derive a UMPU test of size α for testing H0 : p0 = p2 , p1 = 2p(1 − p), p2 = (1 − p)2 versus H1 : H0 is not true, where p ∈ (0, 1) is unknown.

460

6. Hypothesis Tests

42. Consider Example 6.12. (a) Show that A and B are independent if and only if log pp11 = 22 p21 + log . log pp12 p22 22 (b) Derive a UMPU test of size α for testing H0 : P (A) = P (B) versus H1 : P (A) 6= P (B). 43. Let X1 and X2 be independently distributed according to p.d.f.’s given by (6.10) with ξ, η, θ, Y , and h replaced by ξi , ηi , θi , Yi , and hi , i = 1, 2, respectively. Show that there exists a UMPU test of size α for testing (a) H0 : η2 (θ2 ) − η1 (θ1 ) ≤ η0 versus H1 : η2 (θ2 ) − η1 (θ1 ) > η0 ; (b) H0 : η2 (θ2 ) + η1 (θ1 ) ≤ η0 versus H1 : η2 (θ2 ) + η1 (θ1 ) > η0 . 44. Let Xj , j = 1, 2, 3, be independent from the Poisson distributions P (λj ), j = 1, 2, 3, respectively. Show that there exists a UMPU test of size α for testing H0 : λ1 λ2 ≤ λ23 versus H1 : λ1 λ2 > λ23 . 45. Let Xij , i = 1, 2, j = 1, 2, be independent from the Poisson distributions P (λi pij ), where λi > 0, 0 < pij < 1, and pi1 + pi2 = 0. Derive a UMPU test of size α for testing H0 : p11 ≤ p21 versus H1 : p11 > p21 . 46. Let Xij be independent random variables satisfying P (Xij = 0) = θi , P (Xij = k) = (1 − θi )(1 − pi )j−1 pi , k = 1, 2, ..., where 0 < θi < 1 and 0 < pi < 1, j = 1, ..., ni and i = 1, 2. Derive a UMPU test of size α for testing H0 : p1 ≤ p2 versus H1 : p1 > p2 . 47. Let X11 , ..., X1n1 and X21 , ..., X2n2 be two independent samples i.i.d. from the gamma distributions Γ(θ1 , γ1 ) and Γ(θ2 , γ2 ), respectively. (a) Assume that θ1 and θ2 are known. For testing H0 : γ1 ≤ γ2 versus H1 : γ1 > γ2 and H0 : γ1 = γ2 versus H1 : γ1 6= γ2 , show that there exist UMPU tests and that the rejection regions can be determined by using beta distributions. (b) If θi ’s are unknown in (a), show that there exist UMPU tests and describe their general forms. (c) Assume that γ1 = γ2 (unknown). For testing H0 : θ1 ≤ θ2 versus H1 : θ1 > θ2 and H0 : θ1 = θ2 versus H1 : θ1 6= θ2 , show that there exist UMPU tests and describe their general forms. 48. Let N be a random variable with the following discrete p.d.f.: P (N = n) = C(λ)a(n)λn I{0,1,2,...} (n), where λ > 0 is unknown and a and C are known functions. Suppose that given N = n, X1 , ..., Xn are i.i.d. from the p.d.f. given in (6.10). Show that, based on (N, X1 , ..., XN ), there exists a UMPU test of size α for H0 : η(θ) ≤ η0 versus H1 : η(θ) > η0 .

6.6. Exercises

461

49. Let Xi1 , ..., Xini , i = 1, 2, be two independent samples i.i.d. from N (µi , σ 2 ), respectively, ni ≥ 2. Show that a UMPU test of size α for H0 : µ1 = µ2 versus H1 : µ1 6= µ2 rejects H0 when |t(X)| > tn1 +n2 −1,α/2 , where t(X) is given by (6.37) and tn1 +n2 −1,α is the (1 − α)th quantile of the t-distribution tn1 +n2 −1 . Derive the power function of this test. 50. In the two-sample problem discussed in §6.2.3, show that when n1 = n2 , a UMPU test of size α for testing H0 : σ22 = ∆0 σ12 versus H1 : σ22 6= ∆0 σ12 rejects H0 when 2 1−c S2 ∆0 S12 , > , max ∆0 S12 S22 c Rc where 0 f(n1 −1)/2,(n1 −1)/2 (v)dv = α/2 and fa,b is the p.d.f. of the beta distribution B(a, b). 51. Suppose that Xi = β0 + β1 ti + εi , where ti ’s are fixed constants that are not all the same, εi ’s are i.i.d. from N (0, σ 2 ), and β0 , β1 , and σ 2 are unknown parameters. Derive a UMPU test of size α for testing (a) H0 : β0 ≤ θ0 versus H1 : β0 > θ0 ; (b) H0 : β0 = θ0 versus H1 : β0 6= θ0 ; (c) H0 : β1 ≤ θ0 versus H1 : β1 > θ0 ; (d) H0 : β1 = θ0 versus H1 : β1 6= θ0 . 52. In the previous exercise, derive the power function in each of (a)-(d) in terms of a noncentral t-distribution. 53. Consider the normal linear model in §6.2.3 (i.e., model (3.25) with ε = Nn (0, σ 2 In )). For testing H0 : σ 2 ≤ σ02 versus H1 : σ 2 > σ02 and H0 : σ 2 = σ02 versus H1 : σ 2 6= σ02 , show that UMPU tests of size α are functions of SSR and their rejection regions can be determined using chi-square distributions. 54. In the problem of testing for independence in the bivariate normal family, show that (a) the p.d.f. in (6.44) is of the form (6.23) and identify ϕ; (b) the sample correlation coefficient R is independent of U when ρ = 0; (c) R is linear in Y , and V in (6.45) has the t-distribution tn−2 when ρ = 0. 55. Let X1 , ..., with the p.d.f. in (6.44) and n be i.i.d. bivariate normal PX n ¯ j )2 and S12 = Pn (Xi1 − X ¯ 1 )(Xi2 − X ¯ 2 ). let Sj2 = i=1 (Xij − X i=1 (a) Show that a UMPU test for testing H0 : σ2 /σ1 = ∆0 versus H1 : σ2 /σ1 6= ∆0 rejects H0 when

462

6. Hypothesis Tests q 2 2 2 > c. R = |∆20 S12 − S22 | (∆0 S1 + S22 )2 − 4∆20 S12

(b) Find the p.d.f. of R in (a) when σ2 /σ1 = ∆0 . (c) Assume that σ1 = σ2 . Show that a UMPU test for H0 : µ1 = µ2 versus H1 : µ1 6= µ2 rejects H0 when q 2 ¯2 − X ¯1| V = |X S1 + S22 − 2S12 > c. (d) Find the p.d.f. of V in (c) when µ1 = µ2 .

56. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. random 2-vectors having the bivariate normal distribution with EX1 = EY1 = 0, Var(X1 ) = σx2 , Var(Y1 ) = σy2 , and Cov(X1 , Y1 ) = ρσx σy , where σx > 0, σy > 0, and ρ ∈ [0, 1) are unknown. Derive the form and exact distribution of a UMPU test of size α for testing H0 : ρ = 0 versus H1 : ρ > 0. 57. Let X1 , ..., Xn be i.i.d. from the P exponential distribution E(a, θ) with n unknown a and θ. Let V = 2 i=1 (Xi − X(1) ), where X(1) is the smallest order statistic. (a) For testing H0 : θ = 1 versus H1 : θ 6= 1, show that a UMPU test of size α rejects H0 when V < c1 or V > c2 , where ci ’s are determined by Z c2 Z c2 f2n−2 (v)dv = f2n (v)dv = 1 − α, c1

c1

and fm (v) is the p.d.f. of the chi-square distribution χ2m . (b) For testing H0 : a = 0 versus H1 : a 6= 0, show that a UMPU test of size α rejects H0 when X(1) < 0 or 2nX(1) /V > c, where c is determined by Z c (n − 1) (1 + v)−n dv = 1 − α. 0

58. Let X1 , ..., Xn be i.i.d. random variables from the uniform distribution U (θ, ϑ), −∞ < θ < ϑ < ∞. (a) Show that the conditional distribution of X(1) given X(n) = x is the distribution of the minimum of a sample of size n − 1 from the uniform distribution U (θ, x). (b) Find a UMPU test of size α for testing H0 : θ ≤ 0 versus H1 : θ > 0. 59. Let X1 , ..., Xn be independent random variables having the binomial distributions Bi(pi , ki ), i = 1, ..., n, respectively, where pi = ea+bti /(1 + ea+bti ), (a, b) ∈ R2 is unknown, and ti ’s are known covariate values that are not all the same. Derive the UMPU test of size α for testing (a) H0 : a ≥ 0 versus H1 : a < 0; (b) H0 : b ≥ 0 versus H1 : b < 0.

6.6. Exercises

463

60. In the previous exercise, derive approximations to the UMPU tests by considering the limiting distributions of the test statistics. 61. Let X = {x ∈ Rn : all components of x are nonzero} and G be the group of transformations g(x) = (cx1 , ..., cxn ), c > 0. Show that a maximal invariant under G is (sgn(xn ), x1 /xn , ..., xn−1 /xn ), where sgn(x) is 1 or −1 as x is positive or negative. 62. Let X1 , ..., Xn be i.i.d. with a Lebesgue p.d.f. σ −1 f (x/σ) and fi , i = 0, 1, be two known Lebesgue p.d.f.’s on R that are either 0 for x < 0 or symmetric about 0. Consider H0 : f = f0 versus H1 : f = f1 and G = {gr : r > 0} with gr (x) = rx. (a) Show that a UMPI test rejects H0 when R ∞ n−1 v f1 (vX1 ) · · · f1 (vXn )dv R0∞ > c. n−1 f0 (vX1 ) · · · f0 (vXn )dv 0 v

(b) Show that if f0 = N (0, 1) =Pe−|x|/2, then the UMPI Pnand f21 (x) n 1/2 test in (a) rejects H0 when ( i=1 Xi ) / i=1 |Xi | > c. (c) Show that if f0 (x) = I(0,1) (x) and f1 (x) = 2xI(0,1) (x), then the Q UMPI test in (a) rejects H0 when X(n) /( ni=1 Xi )1/n < c. (d) Find the value of c in part (c) when the UMPI test is of size α.

63. Consider the location-scale family problem (with unknown parameters µ and σ) in Example 6.13. (a) Show that W is maximal invariant under the given G. (b) Show that Proposition 6.2 applies and find the form of the functional θ(fi,µ,σ ). (c) Derive the p.d.f. of W (X) under Hi , i = 0, 1. (d) Obtain a UMPI test. 64. In Example 6.13, find the rejection region of the UMPI test when X1 , ..., Xn are i.i.d. and (a) f0,µ,σ is N (µ, σ 2 ) and f1,µ,σ is the p.d.f. of the uniform distribution U (µ − 21 σ, µ + 12 σ); (b) f0,µ,σ is N (µ, σ 2 ) and f1,µ,σ is the p.d.f. of the exponential distribution E(µ, σ); (c) f0,µ,σ is the p.d.f. of U (µ − 21 σ, µ + 12 σ) and f1,µ,σ is the p.d.f. of E(µ, σ); (d) f0,µ is N (µ, 1) and f1,µ (x) = exp{−ex−µ + x − µ}. 65. Prove the claims in Example 6.15. 66. Let X1 , ..., Xn be i.i.d. from N (µ, σ 2 ) with unknown µ and σ 2 . Consider the problem of testing H0 : µ = 0 versus H1 : µ 6= 0 and the group of transformations gc (Xi ) = cXi , c 6= 0.

464

6. Hypothesis Tests

(a) Show that the testing problem is invariant under G. (b) Show that the one-sample two-sided t-test in §6.2.3 is a UMPI test. 67. Prove the claims in Example 6.16. 68. Consider Example 6.16 with H0 and H1 replaced by H0 : µ1 = µ2 and H1 : µ1 6= µ2 , and with G changed to {gc1 ,c2 ,r : c1 = c2 ∈ R, r 6= 0}. (a) Show that the testing problem is invariant under G. (b) Show that the two-sample two-sided t-test in §6.2.3 is a UMPI test. 69. Show that the UMPU tests in Exercise 37(a) and Exercise 47(a) are also UMPI tests under G = {gr : r > 0} with gr (x) = rx. 70. In Example 6.17, show that t(X) has the noncentral t-distribution √ tn−1 ( nθ); the family {fθ (t) : θ ∈ R} has monotone likelihood ratio in t; and that for testing H0 : θ = θ0 versus H1 : θ 6= θ0 , a test that is UMP among all level α unbiased tests based on t(X) rejects H0 when t(X) < c1 or t(X) > c2 . (Hint: consider Exercise 26.) 71. Let X1 and X2 be independently distributed as the exponential distributions E(0, θi ), i = 1, 2, respectively. Define θ = θ1 /θ2 . (a) For testing H0 : θ ≤ 1 versus θ > 1, show that the problem is invariant under the group of transformations gc (x1 , x2 ) = (cx1 , cx2 ), c > 0, and that a UMPI test of size α rejects H0 when X2 /X1 > (1 − α)/α. (b) For testing H0 : θ = 1 versus θ 6= 1, show that the problem is invariant under the group of transformations in (a) and g(x1 , x2 ) = (x2 , x1 ), and that a UMPI test of size α rejects H0 when X1 /X2 > (2 − α)/α and X2 /X1 > (2 − α)/α. 72. Let X1 , ..., Xm and Y1 , ..., Yn be two independent samples i.i.d. from the exponential distributions E(a1 , θ1 ) and E(a2 , θ2 ), respectively. Let gr,c,d(x, y) = (rx1 + c, ..., rxm + c, ry1 + d, ..., ryn + d) and let G = {gr,c,d : r > 0, c ∈ R, d ∈ R}. (a) Show that a UMPI test of size P α for testing H0 : θ1P /θ2 ≥ ∆0 versus H1 : θ1 /θ2 < ∆0 rejects H0 when ni=1 (Yi −Y(1) ) > c m i=1 (Xi −X(1) ) for some constant c. (b) Find the value of c in (a). (c) Show that the UMPI test in (a) is also a UMPU test. 73. Let M (U ) be given by (6.51) and W = M (U )(n − r)/s. (a) Show that W has the noncentral F-distribution Fs,n−r (θ). (b) Show that fθ1 (w)/f0 (w) is an increasing function of w for any given θ1 > 0.

465

6.6. Exercises

74. Consider normal linear model (6.38). Show that (a) the UMPI test derived in §6.3.2 for testing (6.49) is the same as the UMPU test for (6.40) given in §6.2.3 when s = 1 and θ0 = 0; (b) the test with the rejection region W > Fs,n−r,α is a UMPI test of size α for testing H0 : Lβ = θ0 versus H1 : Lβ 6= θ0 , where W is given by (6.52), θ0 is a fixed constant, L is the same as that in (6.49), and Fs,n−r,α is the (1 − α)th quantile of the F-distribution Fs,n−r . 75. In Examples 6.18-6.19, (a) prove the claim in Example 6.19; (b) derive the distribution of W by applying Cochran’s theorem. 76. (Two-way additive model). Assume that Xij ’s are independent and Xij = N (µij , σ 2 ),

i = 1, ..., a, j = 1, ..., b, Pa Pb where µij = µ + αi + βj and i=1 αi = j=1 βj = 0. Derive the forms of the UMPI tests in §6.3.2 for testing (6.54) and (6.55). 77. (Three-way additive model). Assume that Xijk ’s are independent and Xijk = N (µijk , σ 2 ),

i = 1, ..., a, j = 1, ..., b, k = 1, ..., c, Pa Pb Pc where µijk = µ+αi +βj +γk and i=1 αi = j=1 βj = k=1 γk = 0. Derive the UMPI test based on the W in (6.52) for testing H0 : αi = 0 for all i versus H1 : αi 6= 0 for some i. 78. Let X1 , ..., Xm and Y1 , ..., Yn be independently normally distributed with a common unknown variance σ 2 and means E(Xi ) = µx + βx (ui − u ¯),

E(Yj ) = µy + βy (vj − v¯), Pm where ui ’s and vj ’s are known constants, u ¯ = m−1 i=1 ui , v¯ = P n n−1 i=1 vi , and µx , µy , βx , and βy are unknown. Derive the UMPI test based on the W in (6.52) for testing (a) H0 : βx = βy versus H1 : βx 6= βy ; (b) H0 : βx = βy and µx = µy versus H1 : βx 6= βy or µx 6= µy . 79. Let (X1 , Y1 ), ..., (Xn , Yn ) be i.i.d. from a bivariate normal distribution with unknown means, variances, and correlation coefficient ρ. (a) Show that the problem of testing H0 : ρ ≤ ρ0 versus H1 : ρ > ρ0 is invariant under G containing transformations rXi + c, sYi + d, i = 1, ..., n, where r > 0, s > 0, c ∈ R, and d ∈ R. Show that a UMPI test rejects H0 when R > c, where R is the sample