MATHEMATICAL STATISTICS
c 2000 by Chapman & Hall/CRC
CHAPMAN & HALLKRC Texts in Statistical Science Series Series E...
87 downloads
1279 Views
3MB Size
Report
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!
Report copyright / DMCA form
MATHEMATICAL STATISTICS
c 2000 by Chapman & Hall/CRC
CHAPMAN & HALLKRC Texts in Statistical Science Series Series Editors C. Chatfield, University of Bath, UK J. Zidek, University of British Columbia, Canada The Analysis ofTime Series An Introduction, Fifth Edition C. Chatfield Applied Bayesian Forecasting and Time Series Analysis A. Pole, M. West and J. Harrison Applied Nonpkametric Statistical Methods, Second Edition, P. Sprent Applied Statistics Principles and Examples D.R. Cox and E.J. Snell Bayesian Data Analysis A. Gelman, J. Carlin, H. Stern and D. Rubin Beyond ANOVA Basics of Applied Statistics R.G. Miller, Jr. ComputerAided Multivariate Analysis Third Edition A.A. Afifi and V.A. Clark A Course in Large SampleTheory T.S. Ferguson Data Driven Statistical Methods, €? Sprent DecisionAnalysis A Bayesian Approach J . Q Smith Elementary Applications of Probability Theory, Second Edition H.C. Tuckwell Elements of Simulation B.J.T. Morgan Epidemiology Study Design and Data Analysis M. Woodward Essential Statistics, Third Edition D.G. Rees Interpreting Data A First Course in Statistics A.J.B. Anderson An Introduction to Generalized Linear Models A.J. Dobson c 2000 by Chapman & Hall/CRC
Introduction to Multivariate Analysis C. Chatfield and A.J. Collins Introduction to Optimization Methods and their Applications in Statistics B.S. Everitt Large Sample Methods in Statistics P.K. Sen and J. da Motta Singer Markov Chain Monte Carlo Stochastic Simulation for Bayesian Inference D. Gamerman Modeling and Analysis of Stochastic Systems V. Kulkarni Modelling Binary Data D. Collett Modelling Survival Data in Medical Research D. Collett Multivariate Analysis of Variance and Repeated Measures A Practical Approach for Behavioural Scientists D.J. Hand and C.C. Taylor Multivariate Statistics A Practical Approach B. Flury and H. Riedwyl Practical Data Analysis for Designed Experiments B.S. Yandell Practical Longitudinal Data Analysis D.J. Hand and M. Crowder Practical Statistics for Medical Research D.G. Altman ProbabilityMethods and Measurement A. OHagan Problem SolvingA Statistician’sGuide, Second Edition C. Chatfield Randomization, Bootstrap and Monte Carlo Methods in Biology, Second Edition B.F.J. Manly Readings in Decision Analysis S. French
StatisticalAnalysis of Reliability Data M.J. Crowder, A.C. Kimber, T.J. Sweeting and R.L. Smith Statistical Methods for SPC andTQM D. Bissell Statistical Methods in Agriculture and Experimental Biology, Second Edition R. Mead, R.N. Curnow and A.M. Hasted StatisticalProcess ControlTheory Practice,Third Edition G.B. Wetherill and D.W. Brown
and
StatisticalTheory, Fourth Edition B.W. Lindgren StatisticsforAccountants,Fourth Edition S. Letchford
c 2000 by Chapman & Hall/CRC
Statistics for Technology A Course in Applied Statistics,Third Edition C. Chatfield Statistics in Engineering A Approach A.V. Metcalfe
Practical
Statistics in Research and Development, Second Edition R. Caulcutt TheTheory of Linear Models B. Jorgensen Mathematical Statistics K. Knight A Course in Categorical Data Analysis T. Leonard
MATHEMATICAL STATISTICS Keith Knight Department of Statistics University of Toronto Ontario, Canada
CHAPMAN & HALL/CRC Boca Raton London New York Washington, D.C.
Library of Congress CataloginginPublication Data Knight, Keith. Mathematical statistics / Keith Knight. p. cm.  (Texts in statistical science series) Includes bibliographical references and index. ISBN 158488178X (alk. paper) I. Mathematical statistics. I. Title. II. Texts in statistical science. QA276 .K565 I999 519.5–dc21
99056997
This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use Apart from any fair dealing for the purpose of research or private study, or criticism or review, as permitted under the UK Copyright Designs and Patents Act, 1988, this publication may not be reproduced, stored or transmitted, in any form or by any means, electronic or mechanical, including photocopying, microfilming, and recording, or by any information storage or retrieva1 system, without the prior permission in writing of the publishers, or in the case of reprographic reproduction only in accordance with the terms of the licenses issued by the Copyright Licensing Agency in the UK, or in accordance with the terms of the license issued by the appropriate Reproduction Rights Organization outside the UK The consent of CRC Press LLC does not extend to copying for general distribution, for promotion, for creating new works, or for resale. Specific permission must be obtained in writing from CRC Press LLC for such copying. Direct all inquiries to CRC Press LLC, 2000 N.W. Corporate Blvd., Boca Raton, Florida 33431. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identitication and explanation, without intent to infringe.
Visit the CRC Press Web site at www.crcpress.com © 2000 by Chapman & Hall/CRC No claim to original U.S. Government works International Standard Book Number I58488178X Library of Congress Card Number 99056997 Printed in the United States of America 2 3 4 5 6 7 8 9 0 Printed on acidfree paper
To my parents
c 2000 by Chapman & Hall/CRC
Contents
Preface 1 Introduction to Probability 1.1 Random experiments 1.2 Probability measures 1.3 Conditional probability and independence 1.4 Random variables 1.5 Transformations of random variables 1.6 Expected values 1.7 Problems and complements 2 Random vectors and joint distributions 2.1 Introduction 2.2 Discrete and continuous random vectors 2.3 Conditional distributions and expected values 2.4 Distribution theory for Normal samples 2.5 Poisson processes 2.6 Generating random variables 2.7 Problems and complements 3 Convergence of Random Variables 3.1 Introduction 3.2 Convergence in probability and distribution 3.3 Weak Law of Large Numbers 3.4 Proving convergence in distribution 3.5 Central Limit Theorems 3.6 Some applications 3.7 Convergence with probability 1 3.8 Problems and complements 4 Principles of Point Estimation 4.1 Introduction c 2000 by Chapman & Hall/CRC
4.2 Statistical models 4.3 Suﬃciency 4.4 Point estimation 4.5 The substitution principle 4.6 Inﬂuence curves 4.7 Standard errors and their estimation 4.8 Asymptotic relative eﬃciency 4.9 The jackknife 4.10 Problems and complements 5 LikelihoodBased Estimation 5.1 Introduction 5.2 The likelihood function 5.3 The likelihood principle 5.4 Asymptotic theory for MLEs 5.5 Misspeciﬁed models 5.6 Nonparametric maximum likelihood estimation 5.7 Numerical computation of MLEs 5.8 Bayesian estimation 5.9 Problems and complements 6 Optimality in Estimation 6.1 Introduction 6.2 Decision theory 6.3 Minimum variance unbiased estimation 6.4 The Cram´erRao lower bound 6.5 Asymptotic eﬃciency 6.6 Problems and complements 7 Interval Estimation and Hypothesis Testing 7.1 Conﬁdence intervals and regions 7.2 Highest posterior density regions 7.3 Hypothesis testing 7.4 Likelihood ratio tests 7.5 Other issues 7.6 Problems and complements 8 Linear and Generalized Linear Model 8.1 Linear models 8.2 Estimation in linear models 8.3 Hypothesis testing in linear models c 2000 by Chapman & Hall/CRC
8.4 8.5 8.6 8.7
Nonnormal errors Generalized linear models QuasiLikelihood models Problems and complements
9 GoodnessofFit 9.1 Introduction 9.2 Tests based on the Multinomial distribution 9.3 Smooth goodnessofﬁt tests 9.4 Problems and complements References
c 2000 by Chapman & Hall/CRC
Preface This book is intended as a textbook (or reference) for a full year Master’s level (or senior level undergraduate) course in mathematical statistics aimed at students in statistics, biostatistics, and related ﬁelds. This book grew from lecture notes and handouts that I developed for a course in mathematical statistics that I ﬁrst taught in 199293 at the University of Toronto. In teaching this course, I realized that many students viewed the course as largely irrelevant to their education. To me this seemed strange since much of mathematical statistics is directly relevant to statistical practice; for example, what statistician has not used a χ2 approximation at some point in their life? At the same time, I could also sympathize with their point of view. To a student ﬁrst encountering the subject, the traditional syllabus of a mathematical statistics course does seem heavily weighed down with optimality theory of various ﬂavours that was developed in the 1940s and 1950s; while this is interesting (and certainly important), it does leave the impression that mathematical statistics has little to oﬀer beyond some nice mathematics. My main objective in writing this book was to provide a set of useful tools that would allow students to understand the theoretical underpinnings of statistical methodology. At the same time, I wanted to be as mathematically rigorous as possible within certain constraints. I have devoted a chapter to convergence for sequences of random variables (and random vectors) since, for better or for worse, these concepts play an important role in the analysis of estimation and other inferential procedures in statistics. I have concentrated on inferential procedures within the framework of parametric models; however, in recognition of the fact that models are typically misspeciﬁed, estimation is also viewed from a nonparametric perspective by considering estimation of functional parameters (or statistical functionals, as they are often called). This book also places greater emphasis on “classical” (that is, c 2000 by Chapman & Hall/CRC
Frequentist) methodology than it does on Bayesian methodology although this should not be interpreted as a claim of superiority for the Frequentist approach. The mathematical background necessary for this book is multivariate calculus and linear algebra; some exposure to real analysis (in particular, δ proofs) is also useful but not absolutely necessary. I have tried to make the book as selfcontained as possible although I have also implicitly assumed that the reader has some familiarity with basic probability theory and, more importantly, has had some exposure to statistical methodology so as to provide some context for this book. In teaching a course based on drafts of this book, I found it very useful to encourage the use of statistical software packages (such as SPlus and SAS) as well as other mathematical software packages (such as MATLAB, Maple, and Mathematica). When used appropriately, these packages can greatly enhance the eﬀectiveness of this course by increasing the scope of problems that can be considered by students. To facilitate this to some extent, I have included a few sections on computational issues, in particular, generating random variables and numerical computation of estimates. Moreover, some of the problems given in the book are most easily approached using some sort of mathematical or statistical software. Unlike many other textbooks in mathematical statistics, I decided not to include tables of the commonlyused distributions in statistics (Normal, χ2 , and so on). My reason for this is simple; most readers will have access to some statistical software that renders obsolete even the most detailed set of tables. My ﬁrst exposure to mathematical statistics was as a graduate student at the University of Washington where I was fortunate to have a number of outstanding and inspiring teachers, including Andreas Buja, Peter Guttorp, Doug Martin, Ron Pyke, Paul Sampson, and Jon Wellner. Since then, I have beneﬁted from the collective wisdom of many, including David Andrews, Richard Davis, Andrey Feuerverger, Nancy Heckman, Stephan Morgenthaler, John Petkau, Peter Phillips, Nancy Reid, Sid Resnick, Rob Tibshirani, and Jim Zidek. In preparing successive drafts of this book, I received many useful comments from anonymous reviewers and from students at the University of Toronto. I would like to acknowledge the assistance of the editors at Chapman and Hall/CRC Press in completing this c 2000 by Chapman & Hall/CRC
project. A particular vote of thanks is due Stephanie Harding for her tenacity in getting me to ﬁnish this book. I would also like to acknowledge the support of the Natural Sciences and Engineering Research Council of Canada. Last, but not least, I would like to thank my wife Luisa for her patience and her gentle but constructive criticisms of various drafts of this book. Keith Knight Toronto September, 1999
c 2000 by Chapman & Hall/CRC
CHAPTER 1
Introduction to Probability 1.1 Random experiments In simple terms, a random experiment (or experiment) is a process whose outcome is uncertain. It is often useful to think of this process as being repeatable but, in practice, this is seldom the case. For example, a football game may be regarded as a random experiment in the sense that the outcome of the game is uncertain a priori; however, this experiment (the game) is not repeatable as we surely could not guarantee that each game would be played under uniform conditions. Nonetheless, it is often plausible that a given random experiment is conceptually repeatable; for example, we might be willing to assume that team A would win 40% of its games against team B under a certain set of conditions. This “conceptual repeatability” is important as it allows us to interpret probabilities in terms of longrun frequencies. For a given random experiment, we can deﬁne the following terms: • The sample space is the set of all possible outcomes of a random experiment. We will denote the sample space by Ω. • A subset of the sample space Ω is called an event. We say that an event A occurs if the true outcome ω lies in A (that is, ω ∈ A). An event consisting of no outcomes is called the empty set and will be denoted by ∅. Operations on events Let A and B be arbitrary events deﬁned on a sample space Ω. • The union of A and B (denoted by A∪B) consists of all outcomes that belong to at least one of A and B. That is, ω ∈ A ∪ B if, and only if, ω ∈ A or ω ∈ B. • The intersection of A and B (denoted by A ∩ B) consists of all c 2000 by Chapman & Hall/CRC
outcomes that belong to both A and B. That is, ω ∈ A ∩ B if, and only if, ω ∈ A and ω ∈ B. • The complement of A (denoted by Ac ) consists of all outcomes in Ω that do not belong to A. That is, ω ∈ Ac if, and only if, ω ∈ A. • A and B are disjoint (or mutually exclusive) if A ∩ B = ∅. We can also derive the following properties involving union, intersection and complement. • A ∪ B ∪ C = (A ∪ B) ∪ C = A ∪ (B ∪ C). A ∩ B ∩ C = (A ∩ B) ∩ C = A ∩ (B ∩ C). • A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C). A ∪ (B ∩ C) = (A ∪ B) ∩ (A ∪ C). • (A ∪ B)c = Ac ∩ B c . (A ∩ B)c = Ac ∪ B c . 1.2 Probability measures Given a random experiment with a sample space Ω, we would like to deﬁne a function or measure P (·) on the subsets (events) of Ω that assigns a real number to each event; this number will represent the probability that a given event occurs. Clearly, these probabilities must satisfy certain “consistency” conditions; for example, if A ⊂ B then we should have P (A) ≤ P (B). However, from a mathematical point of view, some care must be taken in deﬁning probability, and there have been a number of axiomatic approaches to deﬁning probability. The approach that we will use in this book is due to Kolmogorov (1933) and eﬀectively deﬁnes probability from a measure theoretic point of view; see Billingsley (1995) for more technical details. In fact, Kolmogorov’s treatment was facilitated by the ideas of von Mises (1931) who introduced the notion of a sample space. DEFINITION. P (·) is called a probability measure if the following axioms are satisﬁed: 1. P (A) ≥ 0 for any event A. 2. P (Ω) = 1. 3. If A1 , A2 , · · · are disjoint events then P
∞
i=1 c 2000 by Chapman & Hall/CRC
Ai
=
∞ i=1
P (Ai ).
There are a number of ways of interpreting probabilities. Perhaps the easiest to conceptualize is the interpretation of probabilities as longrun frequencies from a sequence of repeatable experiments. If we assume that a given random experiment is inﬁnitely repeatable then we can interpret P (A) as the relative frequency of occurrences of the event A; that is, if the experiment is repeated N times (where N is large) and A occurs k times then P (A) ≈ k/N . However, other equally valid interpretations and axiomatic deﬁnitions of probability are possible. For example, P (A) could be deﬁned to be a person’s degree of belief in the occurrence of the event A; that is, if B is judged more likely to occur than A, we have P (B) ≥ P (A). This type of probability is sometimes called subjective probability or personal probability (Savage, 1972). Consequences of the axioms The three axioms given above allow us to derive a number of simple but useful properties of probability measures. PROPOSITION 1.1 The following are consequence of the axioms of probability: (a) P (Ac ) = 1 − P (A). (b) P (A ∩ B) ≤ min(P (A), P (B)). (c) P (A ∪ B) = P (A) + P (B) − P (A ∩ B). (d) Suppose that {An } is a nested, increasing sequence of events (in the sense that An ⊂ An+1 ) and let A = ∞ A k=1 k . Then P (An ) → P (A) as n → ∞. (e) Let A1 , A2 , · · · be any events. Then P
∞ k=1
Ak
≤
∞
P (Ak ).
k=1
Proof. (a) Since Ac ∪ A = Ω and A and Ac are disjoint, it follows that 1 = P (Ω) = P (A ∪ Ac ) = P (A) + P (Ac ) and so P (Ac ) = 1 − P (A). (b) B = (B ∩ A) ∪ (B ∩ Ac ). Since B ∩ A and B ∩ Ac are disjoint, we have P (B) = P (B ∩ A) + P (B ∩ Ac ) ≥ P (B ∩ A). A similar argument gives P (A) ≥ P (A ∩ B). c 2000 by Chapman & Hall/CRC
(c) A ∪ B = A ∪ (B ∩ Ac ). Since A and B ∩ Ac are disjoint, we have P (A ∪ B) = P (A) + P (B ∩ Ac ) and
P (B ∩ Ac ) = P (B) − P (B ∩ A).
Thus P (A ∪ B) = P (A) + P (B) − P (A ∩ B). (d) Deﬁne B1 = A1 and Bk = Ak ∩ Ack−1 for k ≥ 2. Then B1 , B2 , · · · are disjoint and An =
n
Bk
and A =
k=1
∞
Bk .
k=1
Hence P (A) =
∞
n
P (Bk ) = lim
n→∞
k=1
P (Bk ) = lim P (An ). n→∞
k=1
(e) First, it follows from (c) that P (A1 ∪ A2 ) ≤ P (A1 ) + P (A2 ) and so for any n < ∞, we have P
n
Ak
k=1
Now let Bn = nk=1 Ak , B = Thus applying (d),
≤
∞
n→∞
P (Ak ).
k=1
k=1 Ak
P (B) = lim P (Bn ) ≤ lim n→∞
n
n k=1
and note that Bn ⊂ Bn+1 .
P (Ak ) =
∞
P (Ak )
k=1
which completes the proof. EXAMPLE 1.1: For any events A1 , A2 , and A3 , we have P (A1 ∪ A2 ∪ A3 ) = P (A1 ) + P (A2 ) + P (A3 ) −P (A1 ∩ A2 ) − P (A1 ∩ A3 ) − P (A2 ∩ A3 ) +P (A1 ∩ A2 ∩ A3 ). To see this, note that P (A1 ∪ A2 ∪ A3 ) = P (A1 ∪ A2 ) + P (A3 ) − P (A3 ∩ (A1 ∪ A2 )) c 2000 by Chapman & Hall/CRC
Since A3 ∩ (A1 ∪ A2 ) = (A1 ∩ A3 ) ∪ (A2 ∩ A3 ), it follows that P (A3 ∩ (A1 ∪ A2 )) = P (A1 ∩ A3 ) + P (A2 ∩ A3 ) − P (A1 ∩ A2 ∩ A3 ) and the conclusion follows since P (A1 ∪ A2 ) = P (A1 ) + P (A2 ) − P (A1 ∩ A2 ). ✸ EXAMPLE 1.2: Suppose that A1 , A2 , · · · is a collection of events. Then ∞
P
Ai
i=1
≥1−
∞
P (Aci ).
i=1
To see this, note that P
∞
Ai
= 1−P
i=1
≥ 1−
∞
∞
Aci
i=1
P (Aci ).
i=1
This inequality is known as Bonferroni’s inequality.
✸
Finite sample spaces When an experiment has a ﬁnite sample space, it is sometimes possible to assume that each outcome is equally likely. That is, if Ω = {ω1 , ω2 , · · · , ωN } then we may be able to assume that P (ωk ) = 1/N for k = 1, · · · , N . This assumption is particularly appropriate in games of chance (for example, card games and lotteries). However, some care should be taken before assuming equally likely outcomes. In the event of equally likely outcomes, for any event A we have P (A) =
number of outcomes in A . N
In some cases, it may be possible to enumerate all possible outcomes, but in general such enumeration is physically impossible; for example, enumerating all possible 5 card poker hands dealt from a deck of 52 cards would take several months under the most favourable conditions. Thus we need to develop methods for counting the number of outcomes in a given event A or the sample space itself. c 2000 by Chapman & Hall/CRC
Many experiments can be broken down into a sequence of subexperiments for which the number of outcomes is easily counted. For example, consider dealing 5 cards from a deck of 52 cards. This particular “experiment” is a sequence of 5 subexperiments corresponding to the 5 cards being dealt. EXAMPLE 1.3: Consider a simple experiment that consists of rolling a pair of dice. The outcomes can be denoted by an ordered pair (i, j) with i representing the outcome of the ﬁrst die and j representing the outcome of the second; these can be thought of as the outcome of two subexperiments. We can then represent the sample space by Ω = {(1, 1), (1, 2), · · · , (1, 6), (2, 1), · · · , (2, 6), · · · , (6, 1), · · · , (6, 6)}. For each outcome of the ﬁrst die, there are 6 outcomes of the second die. Since there are 6 outcomes for the ﬁrst die, it follows that the sample space has 6 × 6 = 36 outcomes. ✸ Example 1.3 suggests that if an experiment consists of a sequence of subexperiments E1 , · · · , Ek having, respectively, n1 , · · · , nk possible outcomes then the total number of outcomes is N = n1 ×· · ·×nk . However, some care should be taken in deﬁning the numbers n1 , · · · , nk . In many cases, the outcome of subexperiment Ei depends on the outcome of the previous i − 1 subexperiments. For example, if we are dealing 5 cards from a deck of 52 cards, once the ﬁrst card is dealt, there are only 51 possible outcomes for the second card dealt; however, there are in fact 52 possible outcomes for the second card if we ignore the outcome of the ﬁrst card. In deﬁning ni for i ≥ 2, we must take into account the outcomes of the previous i − 1 subexperiments. A general rule for counting the number of outcomes in an experiment can be described as follows: Suppose a random experiment E consists of subexperiments E1 , · · · , Ek where the outcome of subexperiment Ei may depend on the outcomes of subexperiments E1 , · · · , Ei−1 but the number of outcomes of Ei is ni independent of the outcomes of E1 , · · · , Ei−1 . Then the number of possible outcomes of E is N = n 1 × · · · × nk . This is sometimes called the product rule for determining the number of outcomes in the sample space. EXAMPLE 1.4: Consider an urn consisting of the integers 0 to c 2000 by Chapman & Hall/CRC
9. We will select three integers from the urn to form a three digit number. The integers can be selected in two possible ways: “with replacement” or “without replacement”. If the integers are selected with replacement, an integer is replaced in the urn after it has been selected and so can be selected again in subsequent draws. On the other hand, if the integers are selected without replacement, an integer is removed from the urn after its selection and therefore cannot be selected subsequently. If we draw the integers with replacement, we have 10 × 10 × 10 = 1000 possible three digit numbers since we have 10 choices at each stage. On the other hand, if the selection is done without replacement, we have 10×9×8 = 720 possible sequences since we have removed one integer from the urn at the second stage and two integers at the third stage. ✸ In many cases, the outcomes of an experiment can be represented as sets of k elements drawn without replacement from a set of n elements; such sets are called either permutations or combinations depending on whether we distinguish between distinct orderings of the elements of the sets. Using the product rule stated above, we can determine the number of permutations and combinations. Consider a set consisting of n distinct elements; for convenience, we will represent this set by S = {1, 2, · · · , n}. Suppose we look at all sequences of k ≤ n elements drawn (without replacement) from S; each of these sequences is called a permutation of length k from S. Using the product rule, it follows that the number of permutations of length k from S is n × (n − 1) × · · · × (n − k + 1) =
n! . (n − k)!
We can also consider all subsets of size k in S; such a subset is called a combination of size k. A combination diﬀers from a permutation in the sense that the ordering of elements is unimportant; that is, the sets {1, 2, 3}, {2, 1, 3} and {3, 1, 2} represent the same combination but three distinct permutations. Note that for each subset (or combination) of size k, we have k! permutations of length k and so the number of combinations of size k in S is n! = k!(n − k)!
n . k
We can also think of a combination of size k from S as splitting S into two disjoint subsets, one of size k and one of size n − k. Thus c 2000 by Chapman & Hall/CRC
n
k is the number of ways of splitting S into two disjoint subsets of size n and n − k respectively. We can extend the argument above to the problem of ﬁnding the number of ways of splitting S into m disjoint subsets of size k1 , · · · , km where k1 + · · · + km = n; this number is simply
n! = k1 ! · · · km !
n k1 , · · · , km
.
The quantities deﬁned above are often called binomial and multinomial coeﬃcients, respectively, due to their presence in the following theorems. THEOREM 1.2 (Binomial Theorem) If n is a nonnegative integer and a, b real numbers then n
(a + b) =
n k=0
n k n−k a b . k
THEOREM 1.3 (Multinomial Theorem) If n is a nonnegative integer and a1 , · · · , am real numbers then
(a1 + a2 + · · · + am ) = n
k1 +···+km =n
n k1 , · · · , km
ak11 · · · akmm ;
the sum above extends over all nonnegative integers k1 , · · · , km whose sum is n. EXAMPLE 1.5: Lotto games typically involve “random” selecting k numbers from the integers 1 through n. For example, in the Canadian Lotto 6/49 game, 6 numbers are chosen from the numbers 1 through 49; prizes are awarded if a player selects 3 or more of the 6 numbers. (A 7th “bonus” is also drawn but does not aﬀect the probability calculations given below.) The order of selection of the numbers is not important so the sample space for this “experiment” consists of all combinations of 6 numbers from the set 49 S = {1, · · · , 49}; there are 6 = 13983816 such combinations. Assuming that each outcome is equally likely, we can determine the probability that a given player who has selected 6 numbers has won a prize. Suppose we want to determine the probability that a player selects k numbers correctly. By choosing 6 numbers, the player has split S into two disjoint subsets one of which consists of the player’s 6 numbers and another consisting of the 43 other c 2000 by Chapman & Hall/CRC
numbers. Selecting k numbers correctly means that the outcome of the draw consisted of k of the player’s 6 numbers while the other 6 − k numbers were among the 43 he did not select. There are k6 43 combinations of k numbers from the player’s selected 6 and 6−k combinations of 6−k numbers from his unselected 43. Applying the product rule, we ﬁnd that the number of outcomes with exactly k of the selected numbers is
N (k) =
6 k
43 6−k
and so p(k) = P (k correct numbers) =
N (k) . 13983816
Substituting for k, we obtain p(3) p(4) p(5) and p(6)
= = = =
1.77 × 10−2 ≈ 1/57, 9.69 × 10−4 ≈ 1/1032, 1.84 × 10−5 ≈ 1/54201 1/13983816 ≈ 7.15 × 10−8 .
The number of correct selections is an example of a random variable, which will be formally deﬁned in section 1.4. ✸ 1.3 Conditional probability and independence So far in deﬁning probabilities, we have assumed no information over and above that available from specifying the sample space and the probability measure on this space. However, knowledge that a particular event has occurred will change our assessment of the probabilities of other events. DEFINITION. Suppose that A and B are events deﬁned on some sample space Ω. If P (B) > 0 then P (AB) =
P (A ∩ B) P (B)
is called the conditional probability of A given B. The restriction to events B with P (B) > 0 in the deﬁnition of conditional probability may not seem too restrictive at this point in the discussion. However, as we will see later, it does often make sense to consider P (AB) where P (B) = 0 and B = ∅. One possible c 2000 by Chapman & Hall/CRC
approach to deﬁning conditional probability in this case would be to take a sequence of events {Bn } decreasing to B (in the sense that Bn+1 ⊂ Bn and B = ∞ n=1 Bn ) such that P (Bn ) > 0 for all n. We could deﬁne P (AB) by P (A ∩ Bn ) . n→∞ P (Bn )
P (AB) = lim
However, it is not clear that the limit (if indeed it exists) is independent of the sequence of events {Bn }. A more fundamental problem is the fact that this deﬁnition fails the most basic test. Suppose that A = B with P (B) = 0; any reasonable deﬁnition of conditional probability would seem to give P (AB) = 1. However, for any sequence of events {Bn } with P (Bn ) > 0 for all n, we have P (ABn ) = 0 and so P (ABn ) → 0. It is easy to see that, for a ﬁxed B, P (·B) satisﬁes the axioms of a probability measure and thus shares the same properties as any other probability measure. Given P (AB), we can write P (A ∩ B) = P (B)P (AB). This rearrangement is important because in many situations, it is convenient to specify conditional probabilities and from these derive probabilities of given events. For example, as we mentioned earlier, many experiments can be thought of occurring as a ﬁnite (or countable) number of stages with the outcome of the kth stage dependent (or conditional) on the outcome of the previous k − 1 stages. If Ak is some event that refers speciﬁcally to the kth stage of the experiment, we can write P (A1 ∩ · · · ∩ An ) = P (A1 ∩ · · · ∩ An−1 )P (An A1 ∩ · · · ∩ An−1 ) = P (A1 )P (A2 A1 ) × · · · ×P (An A1 ∩ · · · ∩ An−1 ). The following classic example (the “birthday problem”) illustrates the application of this formula. EXAMPLE 1.6: Suppose a room is ﬁlled with n people who represent a random sample from the population as a whole. What is the probability that at least two of the people in the room share the same birthday? To answer this question, we must ﬁrst make a simplifying assumption: we will assume that birthdays are uniformly distributed over the 365 days of a nonleap year. Deﬁne Bn to be the event where no two people (out of n) share the same birthday; the probability we want is simply P (Bnc ) = c 2000 by Chapman & Hall/CRC
1 − P (Bn ). In order to evaluate P (Bn ), it is convenient to think of the people entering the room one at a time; we can then deﬁne the events A2 = 2nd person’s birthday is diﬀerent from 1st person’s A3 = 3rd person’s birthday is diﬀerent from previous two .. .. . . An = nth person’s birthday is diﬀerent from previous n − 1. Clearly now, Bn = A2 ∩ · · · ∩ An and so P (A2 )P (A3 A2 ) · · · P (An A2 ∩ · · · ∩ An−1 ). From our assumption of uniformity of birthdays, it follows that P (A2 ) = 364/365, P (A3 A2 ) = 363/365, P (A4 A2 ∩ A3 ) = 362/365 and, in general, P (Ak A2 ∩ · · · ∩ Ak−1 ) =
366 − k 365
since the occurrence of the event A2 ∩ · · · ∩ Ak−1 implies that the ﬁrst k − 1 people in the room have k − 1 distinct birthdays and hence there are 365 − (k − 1) = 366 − k “unclaimed” birthdays. Thus 364 363 366 − n P (Bn ) = × × ··· × . 365 365 365 Substituting into the expression above, it is straightforward to evaluate the probability of at least one match, namely P (Bnc ). For n = 5, P (Bnc ) = 0.027 while for n = 30, P (Bnc ) = 0.706 and n = 70, P (Bnc ) = 0.999; in fact, it is easy to verify that P (Bnc ) ≥ 0.5 for n ≥ 23, a fact that is somewhat counterintuitive to many people. A useful approximation is
P (Bnc )
n(n − 1) ≈ 1 − exp − , 730
which is valid if n/365 is small.
✸
Bayes’ Theorem As mentioned above, conditional probabilities are often naturally speciﬁed by the problem at hand. In many problems, we are given P (AB1 ), · · · , P (ABk ) where B1 , · · · , Bk are disjoint events whose union is the sample space; however, we would like to compute c 2000 by Chapman & Hall/CRC
P (Bj A) for some Bj . Provided P (A) > 0, we have P (A ∩ Bj ) . P (A)
P (Bj A) =
The following result gives a simple formula for P (A). PROPOSITION 1.4 (Law of total probability) If B , B2 , · · · are disjoint events with P (Bk ) > 0 for all k and ∞ 1 k=1 Bk = Ω then P (A) = ∞
Proof. Since A = A ∩ (
∞
P (Bk )P (ABk ).
k=1
k=1 Bk ),
we have
∞
A∩
P (A) = P = P = =
∞
∞ k=1 ∞
Bk
k=1
(A ∩ Bk )
k=1
P (A ∩ Bk ) P (Bk )P (ABk )
k=1
since A ∩ B1 , A ∩ B2 , · · · are disjoint. A simple corollary of the law of total probability is Bayes’ Theorem. PROPOSITION 1.5 (Bayes’ Theorem) Supposethat B1 , B2 , B3 , · · · are disjoint sets with P (Bk ) > 0 for all k and ∞ k=1 Bk = Ω. Then for any event A, P (Bj )P (ABj ) P (Bj A) = ∞ . k=1 P (Bk )P (ABk ) Proof. By deﬁnition, P (Bj A) = P (A ∩ Bj )/P (A) and P (A ∩ Bj ) = P (Bj )P (ABj ). The conclusion follows by applying the law of total probability to P (A). EXAMPLE 1.7: Suppose that the incidence of a certain disease in a population is 0.001. A diagnostic test for this disease exists c 2000 by Chapman & Hall/CRC
but has a false positive rate of 0.05 and a false negative rate of 0.01; that is, 5% of tests on nondiseased people will indicate the presence of the disease while 1% of tests on people with the disease will not indicate the presence of the disease. If a person drawn at random from the population tests positive, what is the probability that that person has the disease? Deﬁne the event D to indicate the presence of disease and A to indicate a positive test for the disease. Then P (D) = 0.001, P (ADc ) = 0.05 and P (Ac D) = 0.01. By Bayes’ Theorem, we have P (D)P (AD) P (D)P (AD) + P (Dc )P (ADc ) (0.001)(0.99) = (0.001)(0.99) + (0.999)(0.05) = 0.0194.
P (DA) =
This example illustrates the potential danger of mandatory testing in a population where the false positive rate exceeds the incidence of the disease; this danger may be particularly acute in situations where a positive test carries a signiﬁcant social stigma. ✸ EXAMPLE 1.8: In epidemiology, one is often interested in measuring the relative risk of disease in one group relative to another. (For example, we might be interested in the relative risk of lung cancer in smokers compared to nonsmokers.) If we have two distinct groups, A and B, then the relative risk of disease in group A compared to group B is deﬁned to be rr =
incidence of disease in group A . incidence of disease in group B
The incidences of disease in groups A and B can be thought of as the conditional probabilities P (diseaseA) and P (diseaseB). If the overall incidence of disease in the population (P (disease)) is small then the relative risk can be diﬃcult to estimate if one draws samples from groups A and B. However, using Bayes’ Theorem and a little algebra, one obtains rr =
P (Adisease)[θP (Bdisease) + (1 − θ)P (Bdiseasefree)] P (Bdisease)[θP (Adisease) + (1 − θ)P (Adiseasefree)]
c 2000 by Chapman & Hall/CRC
where θ = P (disease), the overall incidence of disease in the population. If θ is close to 0 then rr ≈
P (Adisease)P (Bdiseasefree) . P (Bdisease)P (Adiseasefree)
This suggests that the relative risk can be estimated by drawing samples not from the groups A and B but rather from the “disease” and “diseasefree” portions of the population; even though the absolute number of subjects in the “disease” group is small, they can typically be easily sampled (since they are usually undergoing treatment for their illness). Such studies are know as casecontrol studies and are very common in epidemiology. ✸ Independence In discussing conditional probability, we noted that our assessment of the probability of an event A may change if we have knowledge about the occurrence of another event B; if our assessment of the probability of A changes then we can say that there is some dependence between A and B. In some cases, this knowledge will not change our assessment of the probability of A, that is, P (AB) = P (A). In this case, we say that the events A and B are independent. Since we have deﬁned P (AB) only for events B with P (B) > 0, the “deﬁnition” of independence given in the previous paragraph is not completely satisfactory as it does not allow us to deal with events that have probability 0. However, if both P (A) and P (B) are positive then this deﬁnition is consistent in the sense that P (AB) = P (A) is equivalent to P (BA) = P (B). Also note that if P (AB) = P (A) then P (A ∩ B) = P (A)P (B); since the probabilities in the latter equality are always welldeﬁned, we will use it as a formal deﬁnition of independence when at least one of P (A) or P (B) is positive. DEFINITION. Events A and B are said to be independent if P (A ∩ B) = P (A)P (B). Some care must be taken in interpreting independence of A and B when P (A) or P (B) is equal to 0 or 1. For example, if B = Ω then A ∩ B = A and hence P (A ∩ B) = P (A) = P (A)P (B) which implies that A is always independent of the sample space. This is somewhat counterintuitive since in this case, it follows that the c 2000 by Chapman & Hall/CRC
event A implies the event B = Ω. This would suggest that A and B are independent but intuitively A should not be independent of the sample space. The following example illustrates the problem. EXAMPLE 1.9: Suppose that Ω = [0, 1] and that probabilities of events are assigned so that if A = [a, b] (with 0 ≤ a ≤ b ≤ 1), P (A) = b−a. (It turns out that this simple speciﬁcation is suﬃcient to deﬁne probabilities for virtually all events of interest on Ω.) From this, it follows that for events C consisting of a single outcome (that is, C = {c} where c is a real number), we have P (C) = 0. Let A = {1/4, 3/4} and B = {3/4} so P (A) = P (B) = 0; clearly, knowing that the outcome of the experiment belongs to A gives us fairly signiﬁcant information about the occurrence of B. In fact, heuristic considerations suggest that if A is known to have occurred then the two outcomes of A are equally likely, in which case P (BA) = 1/2 = P (B). (In any event, it seems reasonable that P (BA) > 0.) ✸ We can extend our notion of independence to a ﬁnite or countably inﬁnite collection of events. It is tempting to deﬁne independence of A1 , · · · , An to mean that the probability of the intersection of these events equals the product of their probabilities. However, this deﬁnition is defective in the sense that if P (Aj ) = 0 (for some j) then P (A1 ∩ · · · ∩ An ) = 0 regardless of the other Aj ’s; this would mean, for example, that independence of A1 , · · · , An does not imply independence of any given pair of events, which seems somewhat illogical. Therefore, in deﬁning independence of a ﬁnite or countably inﬁnite collection of events, we look at all possible ﬁnite intersections. DEFINITION. A1 , · · · , An (or A1 , A2 , · · ·) are (mutually) independent events if for any ﬁnite subcollection Ai1 , · · · , Aik , P (Ai1 ∩ Ai2 ∩ · · · ∩ Aik ) =
k
P (Aij ).
j=1
Thus if A1 , · · · , An are independent then Ai is independent of Aj for i = j; that is, mutual independence implies pairwise independence. However, the converse is not true as the following example indicates. EXAMPLE 1.10: Consider an experiment with the following c 2000 by Chapman & Hall/CRC
sample space Ω = {abc, bac, cab, bca, acb, cba, aaa, bbb, ccc} where each outcome is equally likely. Deﬁne events Ak = a in kth position (for k = 1, 2, 3). It is easy to see that P (Ak ) = 1/3 for k = 1, 2, 3 and P (Aj ∩ Ak ) = 1/9 for j = k; thus Aj and Ak are independent for all j = k. However, P (A1 ∩ A2 ∩ A3 ) = 1/9 = 1/27 and so A1 , A2 and A3 are not independent. This example shows that pairwise independence does not imply mutual independence. ✸ 1.4 Random variables Consider an experiment where a coin is tossed 20 times. We can represent the sample space as the sets of all 220 ≈ 106 sequences of heads and tails: Ω = {HH · · · H, T HH · · · H, HT H · · · H, · · · , T T · · · T } For such experiments, we are usually not interested in every event deﬁned on Ω but rather just those events involving, for example, the number of heads in the 20 tosses. In this case, we could “redeﬁne” to be the set of all possible values of the number of heads: Ω = {0, 1, 2, · · · , 20} Note that each outcome ω ∈ Ω corresponds to an event deﬁned on Ω; for example, the outcome {0} ∈ Ω corresponds to the single outcome {HH · · · H} ∈ Ω while the outcome {1} ∈ Ω corresponds to the event {T H · · · H, HT H · · · H, · · · , H · · · HT }. Thus we can deﬁne a function X mapping Ω to Ω that counts the number of heads in 20 tosses of the coin; such a function is called a random variable. DEFINITION. A random variable X is a function that maps the sample space to the real line; that is, for each ω ∈ Ω, X(ω) is a real number. EXAMPLE 1.11: Consider an experiment where a coin is tossed until the ﬁrst heads comes up. The sample space for this experiment can be represented as Ω = {H, T H, T T H, T T T H, · · ·} c 2000 by Chapman & Hall/CRC
where H and T represent heads and tails. We can deﬁne a random variable X that counts the number of tails until the ﬁrst heads: X(H) = 0 X(T H) = 1 X(T T H) = 2 and so on. In this case, each value of the random variable X corresponds to a single outcome in Ω. ✸ Random variables allow us, in some sense, to ignore the sample space; more precisely, we can redeﬁne the sample space to be the range of a random variable or a collection of random variables. Probability distributions Suppose that X is a random variable deﬁned on a sample space Ω. If we deﬁne the event [a ≤ X ≤ b] = {ω ∈ Ω : a ≤ X(ω) ≤ b} = A then P (a ≤ X ≤ b) = P (A). DEFINITION. Let X be a random variable (deﬁned on some sample space Ω). The distribution function of X is deﬁned by F (x) = P (X ≤ x) = P ({ω ∈ Ω : X(ω) ≤ x}) . (The distribution function is often referred to as the cumulative distribution function.) Distribution functions satisfy the following basic properties: • If x ≤ y then F (x) ≤ F (y). (F is a nondecreasing function.) • If y ↓ x then F (y) ↓ F (x). (F is a rightcontinuous function although it is not necessarily a continuous function.) • limx→−∞ F (x) = 0; limx→∞ F (x) = 1. Figure 1.1 shows a generic distribution function on the interval [0, 1]. Note that the distribution function has a jump at x = 0.25; the height of this jump is the probability that the random variable is equal to 0.25. The distribution function is also ﬂat over [0.25, 0.5) indicating that X can take no values in this interval. If X has a distribution function F then knowledge of F (x) for all x allows us to compute P (X ∈ A) for any given set A. It is c 2000 by Chapman & Hall/CRC
1.0 0.8 0.6 0.0
0.2
0.4
F(x)
0.0
0.2
0.4
0.6
0.8
1.0
x
Figure 1.1 A distribution function on the interval [0, 1].
important to note that we could also deﬁne the distribution function to be P (X < x); this function would be leftcontinuous rather than rightcontinuous. Suppose that X is a random variable with distribution function F . Then • P (a < X ≤ b) = F (b) − F (a); • P (X < a) = F (a−) = limx↑a F (x); (F (a−) is called the lefthand limit of F at the point a; if F is continuous at a then F (a−) = F (a).) • P (X > a) = 1 − F (a); • P (X = a) = F (a) − F (a−). If F is continuous at a then P (X = a) = 0. These simple properties are useful for computing P (X ∈ A). c 2000 by Chapman & Hall/CRC
Discrete random variables DEFINITION. A random variable X is discrete if its range is a ﬁnite or countably inﬁnite set. That is, there exists a set S = {s1 , s2 , · · ·} such that P (X ∈ S) = 1. From the deﬁnition above, we can deduce that the probability distribution of a discrete random variable is completely determined by specifying P (X = x) for all x. DEFINITION. The frequency function of a discrete random variable X is deﬁned by f (x) = P (X = x). The frequency function of a discrete random variable is known by many other names: some examples are probability mass function, probability function and density function. We will reserve the term “density function” for continuous random variables. Given the frequency function f (x), we can determine the distribution function: F (x) = P (X ≤ x) =
f (t).
t≤x
Thus F (x) is a step function with jumps of height f (x1 ), f (x2 ), · · · occurring at the points x1 , x2 , · · ·. Likewise, we have P (X ∈ A) =
f (x);
x∈A
in the special case where A = (−∞, ∞), we obtain 1 = P (−∞ < X < ∞) =
f (x).
−∞<x 0, we have ∞
1=
−∞
1
f (x) dx =
kx3 dx =
0
k 4
and so k = 4. We can also determine the distribution function by integrating the density from −∞ to any point x; we obtain 0
F (x) =
for x < 0 x4 for 0 ≤ x ≤ 1 . 1 for x > 1
As we noted above, we can modify the density function at a countable number of points and still obtain the same distribution function; for example, if we deﬁne ∗
f (x) =
4x3 for 0 < x < 1 0 otherwise
then we obtain the same distribution function although f ∗ (x) diﬀers from f (x) at x = 0 and x = 1. ✸ Typically, it turns out to be more convenient to specify the probability distribution of a continuous random variable via its density function rather than via its distribution function. The density function essentially describes the probability that X takes a value in a “neighbourhood” of a point x (and thus it is analogous to the frequency function of a discrete random variable); from the deﬁnition above, we have P (x ≤ X ≤ x + ∆) =
x+∆ x
f (t) dx ≈ ∆f (x)
if ∆ is small and f is continuous at x. Thus f (x) > f (y) suggests that we are more likely to see values of X close to x than values of X close to y. EXAMPLE 1.15: (Normal distribution) A random variable X is said to have a Normal distribution with parameters µ and σ 2 (X ∼ N (µ, σ 2 )) if its density is
1 (x − µ)2 f (x) = √ exp − 2σ 2 σ 2π
.
It is easy to see that f (x) has its maximum value at x = µ (which is called the mode of the distribution). When µ = 0 and σ 2 = 1, c 2000 by Chapman & Hall/CRC
X is said to have a standard Normal distribution; we will denote the distribution function of X ∼ N (0, 1) by 1 Φ(x) = √ 2π If X ∼
N (µ, σ 2 )
x
−∞
exp −t2 /2 dt.
then its distribution function is
F (x) = Φ
x−µ . σ
Tables of Φ(x) can be found in many books, and Φ(x) can be evaluated numerically using practically any statistical software package. ✸ EXAMPLE 1.16: (Uniform distribution) Suppose that X is a continuous random variable with density function
f (x) =
(b − a)−1 for a ≤ x ≤ b . 0 otherwise
We say that X has a Uniform distribution on the interval [a, b]; we will sometimes write X ∼ Unif(a, b). Simple integration shows that the distribution function is given by F (x) = (x − a)/(b − a) for a ≤ x ≤ b (F (x) = 0 for x < a and F (x) = 1 for x > b). ✸ An important special case of the Uniform family of distributions is the Uniform distribution on [0, 1]. The following two results are quite useful. THEOREM 1.6 Suppose that X is a continuous random variable with distribution function F (x) and let U = F (X). Then U ∼ Unif(0, 1). Proof. Since 0 ≤ F (x) ≤ 1, we have P (0 ≤ U ≤ 1) = 1. Thus we need to show that P (U < y) = y for 0 < y < 1. (This, in turn, implies that P (U ≤ y) = y.) For given y, choose x such that F (x) = y; such an x exists since F is continuous. Then P (U < y) = P (U < y, X < x) + P (U < y, X ≥ x). Since F is a nondecreasing function, U = F (X) and y = F (x), it follows that the event [X < x] is a subset of the event [U < y] and so P (U < y, X < x) = P (X < x) = F (x) = y. Furthermore the event [U < y, X ≥ x] is the empty set and so P (U < y, X ≥ x) = 0. Thus P (U < y) = y for 0 < y < 1. c 2000 by Chapman & Hall/CRC
1.0 0.8 0.6 0.4 0.0
0.2
inverse of F
0.0
0.2
0.4
0.6
0.8
1.0
t
Figure 1.2 The inverse distribution function F −1 (t) corresponding to F (x) in Figure 1.1.
An easier proof of Theorem 1.6 can be obtained if we assume that F is strictly increasing. In this case, we can deﬁne a continuous inverse F −1 of F satisfying F (F −1 (y)) = y for 0 < y < 1 and so P (U ≤ y) = P (X ≤ F −1 (y)) = F (F −1 (y)) = y. We can also invert this argument. If U ∼ Unif(0, 1) and F is a strictly increasing continuous distribution function then the random variable X = F −1 (U ) has distribution function F : P (X ≤ x) = P (U ≤ F (x)) = F (x). In fact, this latter result holds for an arbitrary distribution function F provided that the inverse function is deﬁned appropriately. THEOREM 1.7 Let F be any distribution function and deﬁne F −1 (t) = inf{x : F (x) ≥ t} to be its inverse function for 0 < t < 1. If U ∼ Unif(0, 1) and X = F −1 (U ) then the distribution function of X is F . Before proceeding to the proof of Theorem 1.7, we will brieﬂy discuss the function F −1 , which is also called the quantile function c 2000 by Chapman & Hall/CRC
of X with F −1 (α) being the α quantile (or 100α percentile) of the distribution of X. F −1 (t) is simplest to deﬁne when F (x) is continuous and strictly increasing; in this case, we can deﬁne F −1 (t) via the equation F (F −1 (t)) = t, that is, F −1 agrees with the usual deﬁnition of inverse function. A somewhat more complicated case occurs when F is the distribution function of a discrete random variable in which case F is piecewise constant with jumps at certain points. If t is such that F (x) = t for a ≤ x < b then F −1 (t) = a; if t is such that F (a) > t and F (a−) < t (so that t falls in the middle of a jump in F ) then F −1 (t) = a since a is the smallest value of x with F (x) ≥ t. Figure 1.2 illustrates the form of F −1 (t) for the generic distribution function F in Figure 1.1. It is also important to note that F −1 (t) cannot always be deﬁned for t = 0 or t = 1; in the context of deﬁning X = F −1 (U ) (where U ∼ Unif(0, 1)) this is not important since P (U = 0) = P (U = 1) = 0. Proof. (Theorem 1.7) From the deﬁnition of F −1 (t) and the fact that F (x) is a rightcontinuous function, we have that F −1 (t) ≤ x if, and only if, t ≤ F (x). Thus
P (X ≤ x) = P F −1 (U ) ≤ x = P (U ≤ F (x)) = F (x) and so the distribution function of X is F . A common application of Theorem 1.7 is the simulation of random variables with a particular distribution. If we can simulate a random variable U ∼ Unif(0, 1) then given any distribution function F with inverse F −1 , X = F −1 (U ) will have distribution function F . A drawback with this method is the fact that F −1 (t) is not necessarily easily computable. The quantiles are sometimes used to describe various features of a distribution. For example, F −1 (1/2) is called the median of the distribution and is sometimes useful as a measure of the centre of the distribution. Another example is the interquartile range, F −1 (3/4) − F −1 (1/4), which is used as a measure of dispersion.
EXAMPLE 1.17: (Exponential distribution) Suppose that X is a continuous random variable with density
f (x) =
λ exp(−λx) for x ≥ 0 . 0 for x < 0
c 2000 by Chapman & Hall/CRC
X is said to have an Exponential distribution with parameter λ (X ∼ Exp(λ)). The distribution function of X is
F (x) =
0 for x < 0 . 1 − exp(−λx) for x ≥ 0
Since F (x) is strictly increasing over the set where F (x) > 0, we can determine the inverse F −1 (t) by solving the equation
1 − exp −λF −1 (t) = t, which yields
1 F −1 (t) = − ln(1 − t). λ Thus if U ∼ Unif(0, 1) then Y = −λ−1 ln(1 − U ) ∼ Exp(λ). Note that since 1 − U has the same distribution as U , we also have −λ−1 ln(U ) ∼ Exp(λ). ✸ EXAMPLE 1.18: Suppose that X has a Binomial distribution with parameters n = 3 and θ = 1/2. We then have F (x) = 1/8 for 0 ≤ x < 1, F (x) = 1/2 for 1 ≤ x < 2 and F (x) = 7/8 for 2 ≤ x < 3 (F (x) = 0 and F (x) = 1 for x < 0 and x ≥ 3 respectively). The inverse of F is given by F −1 (t) =
0 for 0 < t ≤ 1/8
1 for 1/8 < t ≤ 1/2
2 for 1/2 < t ≤ 7/8
.
3 for 7/8 < t < 1
If U ∼ Unif(0, 1) then it is simple to see that F −1 (U ) ∼ Bin(3, 0.5). ✸ Hazard functions Suppose that X is a nonnegative, continuous random variable with density function f (x); it is useful to think of X as a lifetime of some object, for example, a human or an electronic component. Suppose we know that the object is still alive at time x, that is, we know that X > x; given this information, we would like to determine the probability of death in the next instant (of length ∆) following time x: P (x < X ≤ x + ∆X > x) = c 2000 by Chapman & Hall/CRC
P (x < X ≤ x + ∆, X > x) P (X > x)
P (x < X ≤ x + ∆) P (X > x) F (x + ∆) − F (x) 1 − F (x)
= =
Since F is continuous, the conditional probability above tends to 0 as ∆ tends to 0. However, by dividing the conditional probability by ∆ and letting ∆ ↓ 0, we obtain (if F (x) = f (x)), f (x) 1 P (x < X ≤ x + ∆X > x) = . ∆↓0 ∆ 1 − F (x) lim
DEFINITION. Suppose that a nonnegative random variable X has distribution function F and density function f . The function λ(x) =
f (x) 1 − F (x)
is called the hazard function of X. (If F (x) = 1 then λ(x) is deﬁned to be 0.) The hazard function is useful in applications (such as insurance and reliability) where we are interested in the probability of failure (for example, death) given survival up to a certain point in time. For example, if X represents a lifetime and has hazard function λ(x) then P (x ≤ X ≤ X + ∆X ≥ x) ≈ ∆λ(x) if ∆ is suﬃciently small. Given the hazard function of nonnegative continuous random variable X, we can determine both the distribution function and the density of X. Integrating the hazard function from 0 to x, we get x
x
f (t) dt 0 1 − F (t) F (x) du (where u = F (x)) = 1 −u 0 = − ln(1 − F (x))
λ(t) dt = 0
and so
F (x) = 1 − exp − c 2000 by Chapman & Hall/CRC
x
λ(t) dt . 0
Diﬀerentiating F (x), we get
f (x) = λ(x) exp −
x
λ(t) dt . 0
Thus the hazard function gives another way of specifying the distribution of a nonnegative continuous random variable. EXAMPLE 1.19: Suppose that a nonnegative random variable X has a hazard function λ(x) = λβxβ−1 for some λ > 0 and β > 0; when β < 1, the hazard function is decreasing over time while if β > 1, the hazard function is increasing over time. The distribution function of X is
F (x) = 1 − exp −λxβ while its density function is
f (x) = λβxβ−1 exp −λxβ
(x > 0)
(x > 0).
The distribution of X is called a Weibull distribution and is commonly used as a model in reliability and survival analysis. The Exponential distribution is a special case of the Weibull with β = 1; it is easy to see that when β = 1, the hazard function is constant and X has the “memoryless” property P (X > x + tX > x) = P (X > t) ✸
for any x > 0 and t > 0. 1.5 Transformations of random variables
Suppose X is a random variable and Y = g(X) for some function g. Given the distribution of X, what is the distribution of Y ? First consider the case where X is discrete. If Y = g(X), it follows that Y is also discrete since Y can assume no more values than X. To determine the frequency function of Y , deﬁne A(y) = {x : g(x) = y}; then if fX (x) is the frequency function of X, we have fY (y) =
fX (x).
x∈A(y)
The continuous case is somewhat more complicated but the underlying principle is essentially the same as for discrete random c 2000 by Chapman & Hall/CRC
variables. First we will assume that g is strictly monotone (that is, either strictly increasing or strictly decreasing) and diﬀerentiable (at least over a set A such that P (X ∈ A) = 1). To ﬁnd the density of Y , we will ﬁrst ﬁnd FY (y) = P (Y ≤ y) and try to express this y as FY (y) = fY (t) dt −∞
or, alternatively, diﬀerentiate FY to determine fY . For example, if g is strictly increasing then P (Y ≤ y) = P (g(X) ≤ y) = P (X ≤ g −1 (y)) = FX (g −1 (y)) g−1 (y)
=
−∞
fX (x) dx
where g −1 (y) is the inverse of g (g(g −1 (y)) = y). Making a change of variables u = g(x), we get y
FY (y) = The derivative of
g −1 (y)
−∞
fX (g −1 (t))
d −1 g (t). dt
is 1/g (g −1 (y)) and so
fY (y) =
fX (g −1 (y)) . g (g −1 (y))
The argument in the case where g is strictly decreasing is similar except in this case, P (Y ≤ y) = P (X ≥ g −1 (y)) = 1 − FX (g −1 (y)) and so
fX (g −1 (y)) . g (g −1 (y)) Note that g (x) < 0 when g is strictly decreasing and so the density fY will always be nonnegative. If we put the two cases (strictly increasing and strictly decreasing) together, we get fY (y) = −
fY (y) =
fX (g −1 (y)) . g (g −1 (y))
EXAMPLE 1.20: Suppose that X ∼ N (0, 1) (see Example 1.15) c 2000 by Chapman & Hall/CRC
and deﬁne Y = µ + σX for some σ > 0. Then
y−µ 1 fY (y) = fX σ σ and so 1 (x − µ)2 . fY (y) = √ exp − 2σ 2 σ 2π Thus Y ∼ N (µ, σ 2 ).
✸
If g is not a monotone function then a somewhat more careful analysis is needed to ﬁnd the density of Y = g(X). However, the underlying principle is the same as when g is monotone; ﬁrst, we express P (Y ≤ y) as P (X ∈ A(y)) where A(y) = {x : g(x) ≤ y}. Then P (X ∈ A(y)) is diﬀerentiated with respect to y to obtain fY (y). The following examples illustrate the approach. EXAMPLE 1.21: Suppose that X is a continuous random variable with density
fX (x) =
x for −1 ≤ x ≤ 1 0 otherwise
and deﬁne Y = X 2 ; note that the function g(x) = x2 is not monotone for −1 ≤ x ≤ 1. Since P (−1 ≤ X ≤ 1) = 1, it follows that P (0 ≤ Y ≤ 1) = 1. Thus for 0 ≤ y ≤ 1, we have √ √ P (Y ≤ y) = P (− y ≤ X ≤ y) √ √ = FX ( y) − FX (− y). Now diﬀerentiating with respect to y, we get
1 −1/2 1 −1/2 √ √ + fX (− y) y y fY (y) = fX ( y) 2 2 1 1 = + = 1 for 0 ≤ y ≤ 1. 2 2 Thus Y ∼ Unif(0, 1).
✸
1.6 Expected values Suppose that X is a random variable with distribution function F ; given F (x) for all values of x, we know everything we need to c 2000 by Chapman & Hall/CRC
know about X. However, it is sometimes useful to summarize the distribution of X by certain characteristics of the distribution (for example, the median of the distribution). Expected values are one way to characterize a distribution; essentially, the expected value of X is the “centre of mass” of its distribution. EXAMPLE 1.22: Deﬁne X to be a discrete random variable taking values x1 , · · · , xk with probabilities p1 , · · · , pk . Imagine an inﬁnite (weightless) beam with masses p1 , · · · , pk suspended at the points x1 , · · · , xk . Suppose that we try to balance the beam about a fulcrum placed at a point µ; the force exerted downwards by the mass placed at xi is proportional to pi xi − µ and so, in order for the beam to be in balance at the fulcrum µ, we must have
or
pi xi − µ =
xi µ
pi (µ − xi ) =
xi µ
Solving for µ, we get µ=
k
xi pi =
i=1
k
xi P (X = xi )
i=1
to be the centre of mass; µ is called the expected value of X.
✸
Note if X takes only two values, 0 and 1, then the expected value of X (according to Example 1.22) is simply P (X = 1). Thus given some event A, if we deﬁne a random variable X = I(A), where the indicator function I(A) = 1 if A occurs and I(A) = 0 otherwise, then E(X) = E[I(A)] = P (A). For example, E[I(Y ≤ y)] = F (y) for any random variable Y where F is the distribution function of Y. Extending the “centre of mass” deﬁnition of expected value to more general distributions is possible although it does involve some mathematical subtleties: it is not clear that the centre of mass will exist if X can take an inﬁnite (whether countable or uncountable) number of values. EXAMPLE 1.23: Suppose that X is a discrete random variable with frequency function f (x) =
6 π 2 x2
c 2000 by Chapman & Hall/CRC
for x = 1, 2, 3, · · ·.
Then following Example 1.22, we have µ=
∞
x f (x) =
x=1
∞ 6 x−1 = ∞ π 2 x=1
and so this distribution does not have a ﬁnite expected value. ✷ The fact that the expected value in Example 1.23 is inﬁnite does not pose any problems from a mathematical point of view; since X is positive, the inﬁnite summation deﬁning the expected value is unambiguously deﬁned. However, when the random variable takes both positive and negative values, the deﬁnition of the expected value becomes somewhat more delicate as an inﬁnite summation with both positive and negative summands need not have a welldeﬁned value. We will split the random variable X into its positive and negative parts X + and X − where X + = max(X, 0)
and X − = max(−X, 0);
then X = X + − X − and X = X + + X − . Provided that at least one of E(X + ) and E(X − ) is ﬁnite, we can deﬁne E(X) = E(X + ) − E(X − ); otherwise (that is, if both E(X + ) and E(X − ) are inﬁnite), E(X) is undeﬁned. We will now give a (somewhat nonintuitive) deﬁnition of the expected value in the general case. DEFINITION. Suppose that X is a nonnegative random variable with distribution function F . The expected value or mean of X (denoted by E(X)) is deﬁned to be ∞
E(X) = 0
(1 − F (x)) dx,
which may be inﬁnite. In general, if X = X + − X − , we deﬁne E(X) = E(X + ) − E(X − ) provided that at least one of E(X + ) and E(X − ) is ﬁnite; if both are inﬁnite then E(X) is undeﬁned. If E(X) is welldeﬁned then ∞
E(X) = 0
(1 − F (x)) dx −
0 −∞
F (x) dx.
The deﬁnition given above may appear to be inconsistent with our previous discussion of expected values for discrete random variables; in fact, the two deﬁnitions are exactly the same in this case. c 2000 by Chapman & Hall/CRC
If X is a discrete random variable with frequency function f (x) then E(X) = x f (x) x
provided that at least one of E(X + ) =
x f (x)
and E(X − ) = −
x>0
x f (x)
x 0
where λ > 0, α > 0 and Γ(α) is the Gamma function deﬁned by ∞
Γ(α) =
tα−1 exp(−t) dt.
0
X is said to a Gamma distribution with shape parameter α and scale parameter λ (X ∼ Gamma(α, λ)). Using properties of the function Γ(α), we have ∞
E(X) = =
x f (x) dx
−∞ ∞ α α λ x
exp (−λx) dx Γ(α) Γ(α + 1) ∞ λα+1 xα+1−1 exp (−λx) dx λΓ(α) 0 Γ(α + 1) Γ(α + 1) λΓ(α) α λ 0
= = =
since Γ(α + 1) = αΓ(α). Note that the Exponential distribution (see Example 1.17) is a special case of the Gamma distribution with α = 1 and so if X ∼ Exp(λ) then E(X) = 1/λ. Figure 1.3 shows three Gamma density functions with diﬀerent shape parameters α with the scale parameters λ = α chosen so that each distribution has expected value equal to 1. ✸ EXAMPLE 1.25: Suppose that X has a Binomial distribution with parameters n and θ. Since X is discrete, we have E(X) =
n x=0
x
n x θ (1 − θ)n−x x
c 2000 by Chapman & Hall/CRC
1.5 1.0 0.0
0.5
density
0
1
2
3
4
x
Figure 1.3 Gamma density functions for α = 1, λ = 1 (solid line), α = 0.5, λ = 0.5 (dotted line) and α = 2, λ = 2 (dashed line).
=
n
n! θx (1 − θ)n−x (n − x)!(x − 1)! x=1
= nθ = nθ
n
(n − 1)! θx−1 (1 − θ)n−x (n − x)!(x − 1)! x=1 n−1
y=0
n−1 y θ (1 − θ)n−1−y y
= nθ where the last equality follows by applying the Binomial Theorem. ✸ EXAMPLE 1.26: Suppose that X is a continuous random variable with density f (x) =
1 π(1 + x2 )
for −∞ < x < ∞;
X is said to have a Cauchy distribution. Note that X is symmetric c 2000 by Chapman & Hall/CRC
around 0 so that E(X) should be 0 if E(X) exists. However, note that E(X + ) =
∞
xf (x) dx 0
1 ∞ x dx π 0 1 + x2 1 = lim ln(1 + x2 ) 2π x→∞ = ∞. =
Likewise E(X − ) = ∞ and so it follows that E(X) is not welldeﬁned. ✸ We can also deﬁne the expected value of a function of X. If Y = h(X) for some function h then we can deﬁne E[h(X)] by applying the deﬁnition of expected value for the random variable Y ; to do this, of course, requires that we determine the distribution function of Y . However, if X has a distribution function F with F = pFc + (1 − p)Fd as before then ∞
E[h(X)] = p
−∞
h(x)fc (x) dx + p
h(x)fd (x)
x
provided the integral and sum are welldeﬁned. Suppose that h(x) = h1 (x) + h2 (x). Then if both E[h1 (X)] and E[h2 (X)] are ﬁnite, it follows that E[h(X)] = E[h1 (X)] + E[h2 (X)]. If h(x) = a g(x) where a is constant then E[h(X)] = aE[g(X)]; the expected value of h(X) is welldeﬁned if, and only if, the expected value of g(X) is. From these two properties of expected values, we can deduce E[a g(X) + b) = aE[g(X)] + b if a and b are constants. In general, E[g(X)] = g(E(X)) only if g(x) is a linear function, that is, if g(x) = ax + b. If g(x) is a convex function, that is, g(tx + (1 − t)y) ≤ t g(x) + (1 − t) g(y) for 0 ≤ t ≤ 1, then we have Jensen’s inequality, which says that g(E(X)) ≤ E[g(X)]; a simple example of Jensen’s inequality is c 2000 by Chapman & Hall/CRC
E(X)r ≤ E[Xr ], which holds for r ≥ 1 since the function g(x) = xr is convex for r ≥ 1. At this point, we will introduce a convenient shorthand for writing expected values. Given a random variable X with distribution function F , we will write ∞
E[h(X)] =
−∞
h(x) dF (x) ,
that is, E[h(X)] is the integral of h(x) with respect to the distribution function F . For example, if we write h(x) = I(x ≤ t) then E[h(X)] = P (X ≤ t) = F (t) and so ∞
F (t) =
−∞
I(x ≤ t) dF (x) =
t −∞
dF (x).
From a mathematical point of view, this integral can be interpreted as a RiemannStieltjes integral or (more usefully) as a LebesgueStieltjes integral; precise deﬁnitions of these integrals can be found in Rudin (1976). If F is a discrete distribution function then the integral above is interpreted as a summation while if F is a continuous distribution function with a density function then the integral is interpreted as the Riemann integral of calculus. This integral representation can be manipulated usefully as the following example suggests. EXAMPLE 1.27: Suppose that X is a positive random variable with distribution function F and set γ = E[w(X)] < ∞ for some nonnegative function w(x). Given F and γ, we can deﬁne a new distribution function G by G(x) =
1 γ
x 0
w(t) dF (t) for x ≥ 0.
G is sometimes called the selection biased distribution of F . A random variable with distribution G might arise as follows. Suppose that X is a discrete random variable uniformly distributed on the set S = {a1 , · · · , an } where the ai > 0 for i = 1, · · · , n. Rather than sampling uniformly from S, we sample so that the probability that ai is selected is proportional to ai (this is called length biased sampling); that is, we sample a discrete random variable Y so that P (Y = ai ) = kai for some constant k. It is easy to see that for Y c 2000 by Chapman & Hall/CRC
to have a valid probability distribution, we must have ai ai P (Y = ai ) = n = . nE(X) j=1 aj The distribution function of Y has a jump of ai /[nE(X)] at ai (for i = 1, · · · , n) and so can be written as P (Y ≤ x) =
1 E(X)
x
t dF (t) 0
where F is the distribution function of X. Returning to the general case, it is fairly easy to see that, given G and the function w, it is possible to determine F and hence evaluate E[h(X)] for any function h. For example, in the case of length biased sampling (w(x) = x), we can determine µ = E(X) by noting that ∞
E[g(Y )] = 0
1 g(y) dG(y) = µ
∞
y g(y) dF (y). 0
Substituting g(y) = 1/y above, we get E(1/Y ) =
1 µ
∞
dF (x) = 0
1 µ
and so µ = [E(1/Y )]−1 .
✸
Variance and moment generating function DEFINITION. Let X be a random variable with µ = E(X). Then the variance of X, Var(X), is deﬁned to be Var(X) = E[(X − µ)2 ]. If E(X 2 ) < ∞ then Var(X) < ∞; if E(X 2 ) = ∞ then we will deﬁne Var(X) = ∞ (even if E(X) is not ﬁnite or is not deﬁned). Given σ 2 = Var(X), we can deﬁne the standard deviation of X to be SD(X) = σ = Var(X). We can also obtain the following properties of the variance and standard deviation of a random variable. • Var(X) ≥ 0 and Var(X) = 0 if, and only if, P (X = µ) = 1 where µ = E(X). • Var(X) = E(X 2 ) − µ2 . c 2000 by Chapman & Hall/CRC
• For any constants a and b, Var(aX + b) = a2 Var(X) and so it follows that SD(aX + b) = aSD(X). The last property above is particularly important as it indicates that both the variance and standard deviation somehow measure the amount of “dispersion” in the distribution of a random variable and are unaﬀected by changes in the “location” of a random variable (since Var(X + b) = Var(X)). The standard deviation has the same units of measurement as the random variable itself and so has a more natural interpretation than the variance (whose units are the square of the original units). However, as we will see, the variance has very nice algebraic and computational properties. It is important to note that there is nothing particularly special about the variance and standard deviation as measures of dispersion; one problem is the fact that both can be distorted by small amounts of probability occurring in the “tails” of the distribution. EXAMPLE 1.28: Suppose that X has a Uniform distribution on [a, b]. Then 1 E(X) = b−a
b
b2 − a2 1 = (a + b) 2(b − a) 2
x dx = a
and 1 E(X ) = b−a 2
b
x2 dx =
a
b3 − a3 1 = (a2 + ab + b2 ). 3(b − a) 3
Thus Var(X) = E(X 2 ) − [E(X)]2 = (b − a)2 /12.
✸
EXAMPLE 1.29: Suppose that X has a Binomial distribution with parameters n and θ. To compute the variance of X, it is convenient to introduce the formula Var(X) = E[X(X − 1)] + E(X) − [E(X)]2 , which follows from the identity Var(X) = E(X 2 ) − [E(X)]2 . We now have E[X(X − 1)] = =
n x=0 n
x(x − 1)
n x θ (1 − θ)n−x x
n! θx (1 − θ)n−x (x − 2)!(n − x)! x=2
c 2000 by Chapman & Hall/CRC
= n(n − 1)θ
2
n−2 y=0
n−2 y θ (1 − θ)n−2−y y
= n(n − 1)θ . 2
Using the fact that E(X) = nθ (as shown in Example 1.25), it follows that Var(X) = nθ(1 − θ). ✸ EXAMPLE 1.30: Suppose that X has a standard Normal distribution. Since the density of X is symmetric around 0, it follows that E(X) = 0, provided of course that the integral is welldeﬁned. Since ∞ ∞ 1 1 1 2 √ √ x exp(−x /2) dx = exp(−y) dy = √ 2π 0 2π 0 2π is ﬁnite, E(X) is welldeﬁned and equals 0. To compute the variance of X, note that Var(X) = E(X 2 ) and so
Var(X) = = = = = since
∞ 1 √ x2 exp −x2 /2 dx 2π −∞ ∞ 2 √ x2 exp −x2 /2 dx 2π 0 ∞ 2 √ 2 2y exp(−y) dy 2π 0 ∞ 1 √ √ y exp(−y) dy π 0 1
∞ √
y exp(−y) dy =
√
π.
0
If Y = µ + σX then Y ∼ N (µ, σ 2 ) and it follows that E(Y ) = µ + σE(X) = µ and Var(Y ) = σ 2 Var(X) = σ 2 . ✸ An useful tool for computing means and variances is the moment generating function, which, when it exists, uniquely characterizes a probability distribution. DEFINITION. Let X be a random variable and deﬁne m(t) = E[exp(tX)]. If m(t) < ∞ for t ≤ b > 0 then m(t) is called the moment generating function of X. c 2000 by Chapman & Hall/CRC
It is important to note that the moment generating function is not deﬁned for all random variables and, when it is, m(t) is not necessarily ﬁnite for all t. However, if the moment generating function m(t) is welldeﬁned then it essentially determines the probability distribution of the random variable. THEOREM 1.9 Suppose that X and Y are random variables such that mX (t) = E[exp(tX)] = E[exp(tY )] = mY (t) for all t ≤ b > 0. Then X and Y have the same probability distribution; that is, P (X ≤ x) = P (Y ≤ x) for all x. The proof of Theorem 1.9 is quite diﬃcult and will not be pursued here. However, the proof relies on a parallel result to that given here: If we deﬁne the characteristic function of X to be the complexvalued function φ(t) = E[exp(itX)] where exp(is) = cos(s)+i sin(s) then φ(t) speciﬁes the probability distribution in the sense that equality of characteristic functions implies equality of distribution functions. It can be shown that under the hypothesis of Theorem 1.9, we obtain equality of the characteristic functions of X and Y and hence equality in distribution. It should be noted that the characteristic function of a random variable is always welldeﬁned. EXAMPLE 1.31: Suppose that X is an Exponential random variable with parameter λ. The moment generating function (if it exists) is deﬁned by ∞
m(t) = E[exp(tX)] = 0
λ exp(−(λ − t)x) dx.
Note that the integral deﬁning m(t) is ﬁnite if, and only if, λ−t > 0 or, equivalently, if t < λ. Since λ > 0, we then have m(t) < ∞ for t in a neighbourhood around 0 and so the moment generating function exists. Simple integration gives m(t) = for t < λ. c 2000 by Chapman & Hall/CRC
λ λ−t ✸
EXAMPLE 1.32: Suppose that X has a Binomial distribution with parameters n and θ. The moment generating function is given by m(t) = E[exp(tX)] = =
n
exp(tx)
x=0 n
n x θ (1 − θ)n−x x
n (θ exp(t))x (1 − θ)n−x x
x=0
= [1 + θ(exp(t) − 1)]n where the ﬁnal line follows by the Binomial Theorem. Note that m(t) < ∞ for all t. ✸ EXAMPLE 1.33: Suppose that X has a Poisson distribution with parameter λ > 0 (X ∼ Pois(λ)); this is a discrete distribution with frequency function f (x) =
exp(−λ)λx x!
for x = 0, 1, 2, · · ·.
The moment generating function of X is m(t) =
∞
exp(tx)
x=0
= exp(−λ)
exp(−λ)λx x!
∞ [λ exp(t)]x x=0
x!
= exp(−λ) exp[λ exp(t)] = exp[λ(exp(t) − 1)] since
∞
x x=0 a /x!
= exp(a) for any a.
✸
EXAMPLE 1.34: Suppose that X has a standard Normal distribution. The moment generating function of X is
m(t) = = =
∞ 1 √ exp(tx) exp −x2 /2 dx 2π −∞ ∞ 1 √ exp tx − x2 /2 dx 2π −∞ ∞ 1 √ exp −(x − t)2 /2 + t2 /2 dx 2π −∞
c 2000 by Chapman & Hall/CRC
exp(t2 /2) √ = 2π = exp(t2 /2).
∞ −∞
exp −(x − t)2 /2 dx
If Y = µ + σX so that Y is normally distributed with mean µ and variance σ 2 then the moment generating function of Y is mY (t) = exp(µt + σ 2 t2 /2). ✸ The moment generating function has no real useful interpretation; it is used almost exclusively as a technical device for computing certain expected values and as a device for proving limit theorems for sequences of distribution functions. Also given the moment generating function of a random variable, it is possible to invert the moment generating function (or the characteristic function) to obtain the distribution or density function; however, using these inversion formulas can involve some very delicate analytical and numerical techniques. As the name suggests, moment generating functions are useful for computing the moments of a random variable. The moments of X are deﬁned to be the expected values E(X), E(X 2 ), E(X 3 ), · · ·. E(X k ) is deﬁned to be the kth moment of X. THEOREM 1.10 Suppose that X is a random variable with moment generating function m(t). Then for any r > 0, E(Xr ) < ∞ and E(X) = m (0), E(X 2 ) = m (0) and in general E(X k ) = m(k) (0) where m(k) (t) denotes the kth derivative of m at t. Proof. If X has moment generating function then X has a moment generating function m∗ (t). Then for any t > 0, we have xr ≤ exp(tx) for x suﬃciently large (say, for x ≥ c). We then have E[Xr ] = E[Xr I(X ≤ c)] + E[Xr I(X > c)] ≤ E[Xr I(X ≤ c)] + E[exp(tX)I(X > c)] ≤ cr + m∗ (t) < ∞ and so E(Xr ) < ∞. Now using the expansion exp(tx) =
∞ k k t x k=0
we obtain m(t) = E[exp(tX)] =
k! ∞ k t E(X k ) k=0
c 2000 by Chapman & Hall/CRC
k!
.
(It can be shown in this instance that taking expectations inside the inﬁnite summation is valid; however, this is not true in general.) Diﬀerentiating with respect to t, we get m (t) =
∞ ktk−1 E(X k ) k=1
k!
= E(X) +
∞ k−1 t E(X k ) k=2
(k − 1)!
and setting t = 0, it follows that E(X) = 0. Repeatedly diﬀerentiating, we get m(k) (t) = E(X k ) +
∞ tj−k E(X j ) j=k+1
(j − k)!
and so m(k) (0) = E(X k ). (Again, in this case, we are justiﬁed in diﬀerentiating inside the inﬁnite summation although this is not true in general.) EXAMPLE 1.35: Suppose that X is an Exponential random variable with parameter λ. In Example 1.31, we found the moment generating function to be m(t) = λ/(λ − t) (for t < λ). Diﬀerentiating we obtain m (t) = λ/(λ − t)2 , m (t) = 2λ/(λ − t)3 and, in general, m(k) (t) =
k!λ . (λ − t)k+1
Thus, setting t = 0, we get E(X k ) = k!λ−k ; in particular, E(X) = λ−1 and Var(X) = E(X 2 ) − [E(X)]2 = λ−2 . ✸ EXAMPLE 1.36: Suppose that X ∼ Geom(θ) (Example 1.12); recall that the frequency function of this random variable is f (x) = θ(1 − θ)x
for x = 0, 1, 2, · · ·.
Thus the moment generating function is given by m(t) = E[exp(tX)] =
∞
exp(tx)θ(1 − θ)x
x=0 ∞
= θ
[(1 − θ) exp(t)]x .
x=0
The inﬁnite series above converges if, and only if, (1−θ) exp(t) < 1 or (since (1−θ) exp(t) > 0) for t < − ln(1−θ). Since − ln(1−θ) > 0, the moment generating function of X is θ m(t) = 1 − (1 − θ) exp(t) c 2000 by Chapman & Hall/CRC
for t < − ln(1 − θ) (which includes a neighbourhood of 0) and diﬀerentiating, we get m (t) = and m (t) =
θ(1 − θ) exp(t) (1 − (1 − θ) exp(t))2 θ(1 − θ) exp(t)(1 − exp(t)θ(1 − θ)) . (1 − (1 − θ) exp(t))3
We obtain E(X) = (1 − θ)/θ and Var(X) = (1 − θ)/θ2 by setting t = 0. ✸ The converse to Theorem 1.10 is not true; that is, the ﬁniteness of E(X k ) for all k > 0 does not imply the existence of the moment generating function of X. The following example illustrates this. EXAMPLE 1.37: Deﬁne X to be a continuous random variable such that ln(X) has a standard Normal distribution; the density of X is 1 f (x) = √ exp −(ln(x))2 /2 for x > 0 x 2π (X is said to have a logNormal distribution). Since Y = ln(X) is a standard Normal random variable, it follows that E(X k ) = E [exp(kY )] = exp(k 2 /2) (using the moment generating function of the standard Normal distribution) and so E(X k ) < ∞ for any real k. However, X does not have a moment generating function. To see this, note that
E [exp(tX)] = =
∞ 1 √ exp(tx)x−1 exp −(ln(x))2 /2 dx 2π 0 ∞ 1 √ exp t exp(y) − y 2 /2 dy; 2π −∞
for t > 0, the integral is inﬁnite since t exp(y) − y 2 /2 → ∞ as y → ∞ (for t > 0). Thus E [exp(tX)] = ∞ for t > 0 and so the moment generating function of X does not exist. ✸ Taken together, Theorems 1.9 and 1.10 imply that if X and Y are two random variables such that E(X k ) = E(Y k ) for k = 1, 2, 3, · · · and X has a moment generating function then X and Y have the same distribution. However, the fact that X and Y have similar moment generating functions does not mean that X and Y will c 2000 by Chapman & Hall/CRC
have similar distributions; for example, McCullagh (1994) gives an example of two moment generating functions that are virtually indistinguishable (they diﬀer by at most 3 × 10−9 ) but correspond to very diﬀerent distributions (see also Waller, 1995). Moreover, equality of E(X k ) and E(Y k ) (for k = 1, 2, · · ·) does not imply equality of distributions if moment generating functions do not exist as the following example illustrates. EXAMPLE 1.38: Suppose that X and Y are continuous random variables with density functions 1 fX (x) = √ exp −(ln(x))2 /2 x 2π
for x > 0
and
1 fY (x) = √ exp −(ln(x))2 /2 (1 + sin(2π ln(x))) x 2π
for x > 0.
(X has the logNormal distribution of Example 1.37.) For any k = 1, 2, 3, · · ·, we have E(Y k ) = E(X k ) 1 +√ 2π
∞
(ln(x))2 exp − 2
k−1
x 0
sin(2π ln(x))) dx
and making the substitution y = ln(x) − k, we have ∞ 0
xk−1 exp −(ln(x))2 /2 sin(2π ln(x))) dx ∞
=
−∞
2
∞
= exp k /2 = 0
exp k 2 + ky − (y + k)2 /2 sin(2π(y + k)) dy −∞
exp −y 2 /2 sin(2πy) dy
since the integrand exp −y 2 /2 sin(2πy) is an odd function. Thus E(X k ) = E(Y k ) for k = 1, 2, 3, · · · even though X and Y have diﬀerent distributions. ✸ 1.7 Problems and complements 1.1: Show that P (A1 ∪ · · · ∪ An ) =
n i=1
c 2000 by Chapman & Hall/CRC
P (Ai ) −
i<j
P (Ai ∩ Aj )
+
i<j xn+1 for c 2000 by Chapman & Hall/CRC
all n) whose limit is x0 . Deﬁne An = [X ≤ xn ]. Show that ∞
An = [X ≤ x0 ]
n=1
and hence (using part (a)) that P (X ≤ xn ) → P (X ≤ x0 ). (c) Now let {xn } be a strictly increasing sequence of numbers (that is, xn < xn+1 for all n) whose limit is x0 . Again deﬁning An = [X ≤ xn ] show that ∞
An = [X < x0 ]
n=1
and hence that P (X ≤ xn ) → P (X < x0 ). 1.6: Let A, B, C be events. We say that A and B are conditionally independent given C if P (A ∩ BC) = P (AC)P (BC). Suppose that • A and B are conditionally independent given C, and • A and B are conditionally independent given C c . Show that A and B are not necessarily independent but that A and B are independent if C is independent of either A or B. 1.7: Suppose that F1 (x), · · · , Fk (x) are distribution functions. (a) Show that G(x) = p1 F1 (x) + · · · + pk Fk (x) is a distribution function provided that pi ≥ 0 (i = 1, · · · , k) and p1 +· · ·+pk = 1. (b) If F1 (x), · · · , Fk (x) have density (frequency) functions f1 (x), · · · , fk (x), show that G(x) deﬁned in (a) has density (frequency) function g(x) = p1 f1 (x) + · · · + pk fk (x). 1.8: (a) Let X be a nonnegative discrete random variable taking values x1 , x2 , · · · with probabilities f (x1 ), f (x2 ), · · ·. Show that ∞
E(X) =
P (X > x) dx = 0
∞
xk f (xk ).
k=1
(Hint: Note that P (X > x) has downward jumps of height f (xk ) at x = xk and is constant between the jumps.) (b) Suppose that X only takes nonnegative integer values 0, 1, c 2000 by Chapman & Hall/CRC
2, · · ·. Show that E(X) =
∞
P (X > k) =
k=0
∞
P (X ≥ k).
k=1
1.9: Suppose that X is a random variable with distribution function F and inverse (or quantile function) F −1 . Show that 1
E(X) =
F −1 (t) dt
0
if E(X) is welldeﬁned. 1.10: Suppose that X is stochastically greater than Y in the sense that P (X > x) ≥ P (Y > x) for all x. (a) Suppose that at least one of E(X) or E(Y ) is ﬁnite. Show that E(X) ≥ E(Y ). (b) Let FX−1 (t) and FY−1 (t) be the quantile functions of X and Y , respectively. Show that FX−1 (t) ≥ FY−1 (t) for all t. 1.11: Let X be a random variable with ﬁnite expected value E(X) and suppose that g(x) is a convex function: g(tx + (1 − t)y) ≤ t g(x) + (1 − t) g(y) for 0 ≤ t ≤ 1. (a) Show that for any x0 , there exists a linear function h(x) = ax + b such that h(x0 ) = g(x0 ) and h(x) ≤ g(x) for all x. (b) Prove Jensen’s inequality: g(E(X)) ≤ E[g(X)]. (Hint: Set x0 = E(X) in part (a); then h(X) ≤ g(X) with probability 1 and so E[h(X)] ≤ E[g(X)].) 1.12: The Gamma function Γ(x) is deﬁned for x > 0 via the integral ∞
Γ(x) =
tx−1 exp(−t) dt
0
Prove the following facts about Γ(x). (a) Γ(x + 1) = xΓ(x) for any x > 0. (Hint: Integrate Γ(x + 1) by parts.) (b) Γ(k) = (k − 1)! for k = 1, 2, 3, · · ·. √ (c) Γ(1/2) = π. (Hint: Note that ∞
Γ(1/2) =
t−1/2 exp(−t) dt
0
∞
= 2 0 c 2000 by Chapman & Hall/CRC
exp(−s2 ) ds
and so Γ2 (1/2) = 4
∞ ∞ 0
0
exp −(s2 + t2 ) ds dt.
Then make a change of variables to polar coordinates.) 1.13: Suppose that X ∼ Gamma(α, λ). Show that (a) E(X r ) = Γ(r + α)/(λ2 Γ(α)) for r > −α; (b) Var(X) = α/λ2 . 1.14: Suppose that X ∼ Gamma(k, λ) where k is a positive integer. Show that k−1 exp(−λx)(λx)j P (X > x) = j! j=0 for x > 0; thus we can evaluate the distribution function of X in terms of a Poisson distribution. (Hint: Use integration by parts.) 1.15: Suppose that X ∼ N (0, 1). (a) Show that E(X k ) = 0 if k is odd. (b) Show that E(X k ) = 2k/2 Γ((k + 1)/2)/Γ(1/2) if k is even. 1.16: Suppose that X is a continuous random variable with density function f (x) = k(p) exp(−xp )
for −∞ < x < ∞
where p > 0. (a) Show that k(p) = p/(2Γ(1/p)). (b) Show that E[Xr ] = Γ((r + 1)/p)/Γ(1/p) for r > −1. 1.17: Let m(t) = E[exp(tX)] be the moment generating function of X. c(t) = ln mX (t) is often called the cumulant generating function of X. (a) Show that c (0) = E(X) and c (0) = Var(X). (b) Suppose that X has a Poisson distribution with parameter λ as in Example 1.33. Use the cumulant generating function of X to show that E(X) = Var(X) = λ. (c) The mean and variance are the ﬁrst two cumulants of a c 2000 by Chapman & Hall/CRC
distribution; in general, the kth cumulant is deﬁned to be c(k) (0). Show that the third and fourth cumulants are c(3) (0) = E(X 3 ) − 3E(X)E(X 2 ) + 2[E(X)]3 c(4) (0) = E(X 4 ) − 4E(X 3 )E(X 3 ) + 12E(X 2 )[E(X)]2 −3[E(X 2 )]2 − 6[E(X)]4 . (d) Suppose that X ∼ N (µ, σ 2 ). Show that all but the ﬁrst two cumulants of X are exactly 0. 1.18: Suppose that X is an integervalued random variable with moment generating function m(t). The probability generating function of X is deﬁned to be p(t) = E(tX ) for t > 0. Note that p(t) = m(ln(t)). (a) Show that p (1) = E(X). (b) Let p(k) (t) be the kth derivative of p. Show that p(k) (1) = E[X(X − 1) × · · · × (X − k + 1)]. (c) Suppose that X is nonnegative. Show that P (X = k) = p(k) (0)/k!. 1.19: The Gompertz distribution is sometimes used as a model for the length of human life; this model is particular good for modeling survival beyond 40 years. Its distribution function is F (x) = 1 − exp[−β(exp(αx) − 1)]
for x ≥ 0
where α, β > 0. (a) Find the hazard function for this distribution. (b) Suppose that X has distribution function F . Show that
exp(β) ∞ exp(−βt) E(X) = dt α t 1 while the median of F is 1 F −1 (1/2) = ln (1 + ln(2)/β) . α (c) Show that F −1 (1/2) ≥ E(X) for all α > 0, β > 0. 1.20: Suppose that X is a nonnegative random variable. c 2000 by Chapman & Hall/CRC
(a) Show that E(X r ) = r
∞ 0
xr−1 (1 − F (x)) dx.
for any r > 0. (b) Since X is nonnegative, it follows that φ(t) = E[exp(−tX)] is ﬁnite for t ≥ 0. Show that for any r > 0, E(1/X r ) =
1 Γ(r)
∞
tr−1 φ(t) dt.
0
(Hint: Write φ(t) as an integral involving the distribution function of X.) 1.21: Suppose that X is a nonnegative random variable where E(X r ) is ﬁnite for some r > 0. Show that E(X s ) is ﬁnite for 0 ≤ s ≤ r. 1.22: Let X be a continuous random variable with distribution function FX (x). Suppose that Y = g(X) where g is a strictly increasing continuous function. If FX−1 (t) is the inverse of FX show that the inverse of the distribution function of Y is FY−1 (t) = g(FX−1 (t)). 1.23: Suppose that X is a nonnegative random variable with distribution function F (x) = P (X ≤ x). Show that E(X r ) = r
∞ 0
xr−1 (1 − F (x)) dx.
for any r > 0. 1.24: Let U be a Uniform random variable on [0, 1] and deﬁne X = tan(π(U − 1/2)). Show that the density of X is f (x) =
1 . π(1 + x2 )
(This is the Cauchy distribution in Example 1.26.) 1.25: Suppose that X has a distribution function F (x) with inverse F −1 (t). (a) Suppose also that E(X) < ∞ and deﬁne g(t) = E[X − t]. Show that g is minimized at t = F −1 (1/2). (b) The assumption that E(X) < ∞ in (a) is unnecessary if we redeﬁne g(t) = E[X − t − X]. Show that g(t) is ﬁnite for all t and that t = F −1 (1/2) minimizes g(t). c 2000 by Chapman & Hall/CRC
(c) Deﬁne ρα (x) = αxI(x ≥ 0) + (α − 1)xI(x < 0) for some 0 < α < 1. Show that g(t) = E[ρα (X −t)−ρα (X)] is minimized at t = F −1 (α). 1.26: Suppose that X is a random variable with E[X 2 ] < ∞. Show that g(t) = E[(X − t)2 ] is minimized at t = E(X). 1.27: Let X be a positive random variable with distribution function F . Show that E(X) < ∞ if, and only if, ∞ k=1
for any > 0.
c 2000 by Chapman & Hall/CRC
P (X > k) < ∞
CHAPTER 2
Random vectors and joint distributions 2.1 Introduction To this point, we have considered only a single random variable deﬁned on a sample space. It is, of course, possible to deﬁne more than one random variable for a given experiment. Consider, for example, sampling from a ﬁnite population consisting of N people. Each person in the population has a number of attributes that can be measured: height, weight, age, and so on. Clearly, we can deﬁne random variables that measure each of these attributes for a randomly chosen person. If we take a sample of n < N people from this population, we can deﬁne random variables X1 , · · · , Xn that measure a certain attribute (for example, weight) for each of the n people in the sample. Suppose we have random variables X1 , · · · , Xk deﬁned on some sample space. We then call the vector X = (X1 , · · · , Xk ) a random vector. DEFINITION. The joint distribution function of a random vector (X1 , · · · , Xk ) is F (x1 , · · · , xk ) = P (X1 ≤ x1 , · · · , Xk ≤ xk ) where the event [X1 ≤ x1 , · · · , Xk ≤ xk ] is the intersection of the events [X1 ≤ x1 ], · · · , [Xk ≤ xk ]. Given the joint distribution function of random vector X, we can determine P (X ∈ A) for any set A ⊂ Rk . Not surprisingly, it is often convenient to think of random vectors as (random) elements of a vector space; this allows us, for example, to manipulate random vectors via the operations of linear algebra. When this is the case, we will assume (by default) that the random vector is, in fact, a column vector unless explicitly stated otherwise. Hopefully, this will be clear from the context. c 2000 by Chapman & Hall/CRC
2.2 Discrete and continuous random vectors For single random variables, we noted earlier that it is possible to describe the probability distribution by means of a density function or frequency function, depending on whether the random variable is continuous or discrete. It is possible to deﬁne analogous functions in the case of random vectors. DEFINITION. Suppose that X1 , · · · , Xk are discrete random variables deﬁned on the same sample space. Then the joint frequency function of X = (X1 , · · · , Xk ) is deﬁned to be f (x1 , · · · , xk ) = P (X1 = x1 , · · · , Xk = xk ). If X1 , · · · , Xk are discrete then the joint frequency function must exist. DEFINITION. Suppose that X1 , · · · , Xn are continuous random variables deﬁned on the same sample space and that P [X1 ≤ x1 , · · · , Xk ≤ xk ] =
xk
−∞
···
x1
−∞
f (t1 , · · · , tk ) dt1 · · · dtk
for all x1 , · · · , xk . Then f (x1 , · · · , xk ) is the joint density function of (X1 , · · · , Xk ) (provided that f (x1 , · · · , xk ) ≥ 0). To avoid confusion, we will sometimes refer to the density or frequency function of a random variable Xi as its marginal density or frequency function. The joint density and frequency functions must satisfy the following conditions:
f (x1 , · · · , xk ) = 1
(x1 ,···,xk )
if f (x1 , · · · , xk ) is a frequency function, and ∞
−∞
···
∞
−∞
f (x1 , · · · , xk ) dx1 · · · dxk = 1
if f (x1 , · · · , xk ) is a density function. Moreover, we can determine the probability that (X1 , · · · , Xn ) lies in a given set A by summing or integrating f (x1 , · · · , xn ) over A. EXAMPLE 2.1: Suppose that X and Y are continuous random variables with joint density function f (x, y) = c 2000 by Chapman & Hall/CRC
1 π
for x2 + y 2 ≤ 1.
X and Y thus have a Uniform distribution on a disk of radius 1 centered at the origin. Suppose we wish to determine P (X ≤ u) for −1 ≤ u ≤ 1. We can do this by integrating the joint density f (x, y) over the region {(x, y) : −1 ≤ x ≤ u, x2 + y 2 ≤ 1}. Thus we have P (X ≤ u) = = =
u √1−x2 1 √ −1 − 1−x2 1 u
π
dy dx
2 1 − x2 dx π −1 1 1 1 + u 1 − u2 + sin−1 (u). 2 π π
Note that to ﬁnd the marginal density of X, we can diﬀerentiate P (X ≤ u) with respect to u: fX (x) =
2 1 − x2 π
for x ≤ 1.
It is easy to see (by symmetry) that Y has the same marginal density. ✸ The following result indicates how to obtain the joint density (frequency) function of a subset of X1 , · · · , Xk or the marginal density function of one of the Xi ’s. THEOREM 2.1 (a) Suppose that X = (X1 , · · · , Xk ) has joint frequency function f (x). For : < k, the joint frequency function of (X1 , · · · , X ) is g(x1 , · · · , x ) =
f (x1 , · · · , xk ).
x+1 ,···,xk
(b) Suppose that X = (X1 , · · · , Xk ) has joint density function f (x). For : < k, the joint density function of (X1 , · · · , X ) is g(x1 , · · · , x ) =
∞
−∞
···
∞
−∞
f (x1 , · · · , xk ) dx+1 · · · dxk .
Proof. (a) This result follows trivially since P (X1 = x1 , · · · , X = x ) =
(x+1 ,···,xk )
for any x1 , · · · , x . c 2000 by Chapman & Hall/CRC
P (X1 = x1 , · · · , Xk = xk )
(b) For any x1 , · · · , x , P (X1 ≤ x1 , · · · , X ≤ x ) x1
= =
−∞ x1 −∞
··· ···
x ∞
−∞ −∞ x −∞
···
∞ −∞
f (t1 , · · · , tk ) dtk · · · dt+1 dt · · · dt1
g(t1 , · · · , t ) dt · · · dt1
and so g(x1 , · · · , x ) is the joint density of (X1 , · · · , X ). Theorem 2.1 can be applied to ﬁnd the marginal density √ function of X in Example 2.1. Since f (x, y) = 0 for y > 1 − x2 and x ≤ 1, it follows that the marginal density of X is √
fX (x) = =
2
1−x 1 dy π −√1−x2 2 1 − x2 π
for x ≤ 1 as given in Example 2.1. As is the case for continuous random variables, the joint density function of a continuous random vector is not uniquely deﬁned; for example, we could change f (x) at a countably inﬁnite number of points and the resulting function would still satisfy the deﬁnition. However, just as it is useful to think of the density function as the derivative of the distribution function of a random variable, we can think of the joint density as a partial derivative of the joint distribution function. More precisely, if F (x1 , · · · , xk ) is diﬀerentiable then we can write f (x1 , · · · , xk ) =
∂k F (x1 , · · · , xk ). ∂x1 · · · ∂xk
It is also important to note that the joint density function of (X1 , · · · , Xk ) need not exist even if each of X1 , · · · , Xk have their own marginal density function. For example, suppose that X2 = g(X1 ) for some continuous function g. Existence of a joint density f (x1 , x2 ) of (X1 , X2 ) implies that the range of (X1 , X2 ) contains an open rectangle of the form {(x1 , x2 ) : a1 < x1 < b1 , a2 < x2 < b2 }. However, since the range of (X1 , X2 ) is at most {(x1 , x2 ) : −∞ < x1 < ∞, x2 = g(x1 )} c 2000 by Chapman & Hall/CRC
which does not contain an open rectangle, we can conclude that (X1 , X2 ) does not have a joint density function. Independent random variables DEFINITION. Let X1 , · · · , Xk be random variables deﬁned on the same sample space. X1 , · · · , Xk are said to be independent if the events [a1 < X1 ≤ b1 ], [a2 < X2 ≤ b2 ], · · · , [ak < Xk ≤ bk ] are independent for all ai < bi , i = 1, · · · , k. An inﬁnite collection X1 , X2 , · · · of random variables are independent if every ﬁnite collection of random variables is independent. If (X1 , · · · , Xk ) have a joint density or joint frequency function then there is a simple equivalent condition for independence. THEOREM 2.2 If X1 , · · · , Xk are independent and have joint density (or frequency) function f (x1 , · · · , xk ) then f (x1 , · · · , xk ) =
k
fi (xi )
i=1
where fi (xi ) is the marginal density (frequency) function of Xi . Conversely, if the joint density (frequency) function is the product of marginal density (frequency) functions then X1 , · · · , Xk are independent. Proof. We will give the proof only for the case where f (x1 , · · · , xn ) is a joint density function; the proof for the frequency function case is similar. If X1 , · · · , Xk are independent, P (a1 < X1 ≤ b1 , · · · , ak < Xk ≤ bk ) = =
k
P (ai < Xi ≤ bi )
i=1 k bi i=1 ai
bk
= ak
···
fi (xi ) dxi b1 k a1 i=1
fi (xi ) dx1 · · · dxk
and so f1 (x1 ) × · · · × fk (xk ) is the joint density of (X1 , · · · , Xk ). Conversely, if f (x1 , · · · , xk ) = f1 (x1 ) · · · fk (xk ) then it is easy to verify that [a1 < X1 ≤ b1 ], · · · , [ak < Xk ≤ bk ] are independent for all choices of the ai ’s and bi ’s. c 2000 by Chapman & Hall/CRC
Independence is an important assumption for many statistical models. Assuming independence essentially allows us to concentrate on the marginal distributions of random variables; given these marginal distributions together with independence, we can determine the joint distribution of the random variables. If X1 , · · · , Xn are independent random variables with the same marginal distribution, we say that X1 , · · · , Xn are independent, identically distributed (i.i.d.) random variables. EXAMPLE 2.2: Suppose that X1 , · · · , Xn are i.i.d. continuous random variables with common (marginal) density f (x) and distribution function F (x). Given X1 , · · · , Xn , we can deﬁne two new random variables U = min(X1 , · · · , Xn )
and V = max(X1 , · · · , Xn ),
which are the minimum and maximum of the i.i.d. sample. It is fairly simple to determine the marginal densities of U and V . We note that U > x if, and only if, Xi > x for all i and also that V ≤ x if, and only if, Xi ≤ x for all i. Thus P (U ≤ x) = 1 − P (U > x) = 1 − P (X1 > x, · · · , Xn > x) = 1 − [1 − F (x)]n and P (V ≤ x) = P (X1 ≤ x, · · · , Xn ≤ n) = [F (x)]n . From this it follows that the marginal densities of U and V are fU (x) = n[1 − F (x)]n−1 f (x) and fV (x) = n[F (x)]n−1 f (x). The joint density of (U, V ) is somewhat more complicated. Deﬁne fU,V (u, v) to be this joint density. Since U ≤ V , it follows that fU,V (u, v) = 0 for u > v. For u ≤ v, we have P (U ≤ u, V ≤ v) = P (V ≤ v) − P (U > u, V ≤ v) = P (V ≤ v) −P (u < X1 ≤ v, · · · , u < Xn ≤ v) = [F (v)]n − [F (v) − F (u)]n . c 2000 by Chapman & Hall/CRC
Now if P (U ≤ u, V ≤ v) is twicediﬀerentiable with respect to u and v, we can evaluate the joint density of (U, V ) by ∂2 P (U ≤ u, V ≤ v) ∂u∂v = n(n − 1)[F (v) − F (u)]n−2 f (u)f (v).
fU,V (u, v) =
In general, we have
fU,V (u, v) =
n(n − 1)[F (v) − F (u)]n−2 f (u)f (v) for u ≤ v 0 otherwise ✸
as the joint density of (U, V ). Transformations
Suppose that X = (X1 , · · · , Xk ) is a random vector with some joint distribution. Deﬁne new random variables Yi = hi (X) (i = 1, · · · , k) where h1 , · · · , hk are realvalued functions. We would like to determine • the (marginal) distribution of Yi , and • the joint distribution of Y = (Y1 , · · · , Yk ). The two questions above are, of course, related; if the joint distribution of Y is known then we can (at least in theory) ﬁnd the marginal distribution of a single Yi . On the other hand, it would seem to be easier to determine the marginal distribution of a single Yi than the joint distribution of Y ; while this is true in some cases, in many cases (particularly when the Yi ’s are continuous), the marginal distribution is most easily determined by ﬁrst determining the joint distribution. We will ﬁrst consider determining the marginal distribution of a single Yi . Given a random vector X = (X1 , · · · , Xk ), deﬁne Y = h(X) for some (realvalued) function h. To simplify the discussion somewhat, we will assume that X has a joint density or joint frequency function. The following general algorithm can be used for determining the distribution of Y : • Find P (Y ≤ y) (or P (Y = y) if this is nonzero) by integrating (summing) the joint density (frequency) function over the appropriate region. • If Y is a continuous random variable, write P (Y ≤ y) = c 2000 by Chapman & Hall/CRC
y
−∞
fY (t) dt
to obtain the density function fY (y) of Y ; typically, this can be done by diﬀerentiating P (Y ≤ y). This algorithm is conceptually simple but not always feasible in practice unless h is a fairly simple function. In the case where Y is a sum of two random variables (independent or not), the algorithm is very simple to apply. EXAMPLE 2.3: Suppose that X1 , X2 are discrete random variables with joint frequency function fX (x1 , x2 ) and let Y = X1 +X2 . Then fY (y) = P (Y = y) =
P (X1 = x, X2 = y − x)
x
=
fX (x, y − x)
x
✸
is the frequency function of Y .
EXAMPLE 2.4: Suppose that X1 , X2 be continuous random variables with joint density function fX (x1 , x2 ) and let Y = X1 + X2 . We can obtain P (Y ≤ y) by integrating the joint density over the region {(x1 , x2 ) : x1 + x2 ≤ y}. Thus P (Y ≤ y) =
∞ y−x2 −∞ −∞
∞ y
=
−∞ −∞
fX (x1 , x2 ) dx1 dx2 fX (t − x2 , x2 ) dt dx2
(setting t = x1 + x2 )
y ∞
=
−∞ −∞
fX (t − x2 , x2 ) dx2 dt.
Thus it follows that Y has a density function ∞
fY (y) =
−∞
fX (y − u, u) du
or equivalently ∞
fY (y) =
−∞
fX (u, y − u) du.
Note that if both X and Y are nonnegative random variables then fX (y − u, u) = 0 if u < 0 or if u > y. In this case, the density c 2000 by Chapman & Hall/CRC
simpliﬁes to
y
fY (y) =
0
fX (y − u, u) du.
Notice that the expression for the density function of Y is similar to the expression for the frequency function of Y in Example 2.3 with the integral replacing the summation. ✸ We now turn our attention to ﬁnding the joint density of Y = (Y1 , · · · , Yk ) where Yi = hi (X1 , · · · , Xk ) (i = 1, · · · , k) and X = (X1 , · · · , Xk ) has a joint density fX . We start by deﬁning a vectorvalued function h whose elements are the functions h1 , · · · , hk :
h(x) =
h1 (x1 , · · · , xk ) h2 (x1 , · · · , xk ) .. .
.
hk (x1 , · · · , xk ) We will assume (for now) that h is a onetoone function with inverse h−1 (that is, h−1 (h(x)) = x). Next we will deﬁne the Jacobian matrix of h to be a k × k whose ith row and jth column element is ∂ hi (x1 , · · · , xk ) ∂xj with the Jacobian of h, Jh (x1 , · · · , xk ), deﬁned to be the determinant of this matrix. THEOREM 2.3 Suppose that P (X ∈ S) = 1 for some open set S ⊂ Rk . If (a) h has continuous partial derivatives on S, (b) h is onetoone on S, (c) Jh (x) = 0 for x ∈ S then (Y1 , · · · , Yk ) has joint density function fX (h−1 (y)) Jh (h−1 (y)) = fX (h−1 (y))Jh−1 (y)
fY (y) =
for y ∈ h(S). (Jh−1 is the Jacobian of h−1 .) The proof of Theorem 2.3 follows from the standard changeofvariables formula for integrals in multivariate calculus. c 2000 by Chapman & Hall/CRC
EXAMPLE 2.5: Suppose that X1 , X2 are independent Gamma random variables with common scale parameters: X1 ∼ Gamma(α, λ)
and X2 ∼ Gamma(β, λ).
Deﬁne Y1 = X1 + X2 X1 Y2 = X1 + X2 and note that Y1 takes values in the interval (0, ∞) while Y2 takes its values in (0, 1). The functions h and h−1 in this case are
h(x1 , x2 ) = −1
and h
(y1 , y2 ) =
x1 + x2 x1 /(x1 + x2 ) y1 y2 y1 (1 − y2 )
and the Jacobian of h−1 is Jh−1 (y1 , y2 ) = −y1 . The joint density of X1 and X2 can be determined by multiplying the marginal densities together: fX (x1 , x2 ) =
xβ−1 exp[−λ(x1 + x2 )] λα+β xα−1 1 2 Γ(α)Γ(β)
for x1 , x2 > 0
Now applying Theorem 2.3 and doing some rearranging, we get fY (y1 , y2 ) = g1 (y1 )g2 (y2 ) where g1 (y1 ) = and g2 (y2 ) =
λα+β y1α+β−1 exp(−λy1 ) Γ(α + β) Γ(α + β) α−1 (1 − y2 )β−1 y Γ(α)Γ(β) 2
for 0 < y1 < ∞ and 0 < y2 < 1. There are several things to note here. First, Y1 is independent of Y2 ; second, Y1 has a Gamma distribution with shape parameter α+β and scale parameter λ; ﬁnally, Y2 has a Beta distribution with parameters α and β (Y2 ∼ Beta(α, β)). ✸ Theorem 2.3 can be extended to the case where the transformation h is not onetoone. Suppose that P [X ∈ S] = 1 for some open c 2000 by Chapman & Hall/CRC
set and that S is a disjoint union of open sets S1 , · · · , Sm where h is onetoone on each of the Sj ’s (with inverse h−1 on Sj ). Then j deﬁning (Y1 , · · · , Yk ) as before, the joint density of (Y1 , · · · , Yk ) is fY (y) =
m
−1 fX (h−1 j (y))Jh−1 (y)I(hj (y ∈ Sj ) j
j=1
where Jh−1 is the Jacobian of h−1 j . j
EXAMPLE 2.6: Suppose that X1 , · · · , Xn are i.i.d. random variables with density function f (x). Reorder the Xi ’s so that X(1) < X(2) < · · · < X(n) ; these latter random variables are called the order statistics of X1 , · · · , Xn . Deﬁne Yi = X(i) and notice that the transformation taking X1 , · · · , Xn to the order statistics is not onetoone; given a particular ordering y1 < y2 < · · · < yn , (x1 , · · · , xn ) could be any one of the n! permutations of (y1 , · · · , yn ). To determine the joint density of the order statistics, we divide the range of (X1 , · · · , Xn ) into the n! disjoint subregions S1 = {(x1 , · · · , xn ) : x1 < x2 < · · · < xn } S2 = {(x1 , · · · , xn ) : x2 < x1 < x3 < · · · < xn } .. .. .. . . . Sn! = {(x1 , · · · , xn ) : xn < xn−1 < · · · < x1 } corresponding to the n! orderings of (x1 , · · · , xn ). Note that the transformation h from the Xi ’s to the order statistics is onetoone on each Sj ; moreover, the Jacobian Jh−1 = ±1 for each j
j = 1, · · · , n!. Since the joint density of (X1 , · · · , Xn ) is fX (x1 , · · · , xn ) = f (x1 ) × · · · × f (xn ), it follows that the joint density of the order statistics is fY (y1 , · · · , yn ) =
n!
fX (h−1 j (y1 , · · · , yn ))
j=1
= n!
n
f (yi )
i=1
for y1 < y2 < · · · < yn . c 2000 by Chapman & Hall/CRC
✸
Expected values Suppose that X1 , · · · , Xn are random variables deﬁned on some sample space and let Y = h(X1 , · · · , Xk ) for some realvalued function h. In section 1.6, we deﬁned the expected value of Y to be ∞
P (Y > y) dy −
E(Y ) = 0
0 −∞
P (Y ≤ y) dy.
This formula implies that we need to ﬁrst determine the distribution function of Y (or equivalently its density or frequency function) in order to evaluate E(Y ). Fortunately, evaluating E(Y ) is typically not so complicated if X = (X1 , · · · , Xk ) has a joint density or frequency function; more precisely, we can deﬁne E[h(X)] =
h(x)f (x)
x
if X has joint frequency function f (x) and ∞
E[h(X)] =
−∞
···
∞
−∞
h(x)f (x) dx1 · · · dxk
if X has joint density function f (x). (Of course, in order for the expected value to be ﬁnite we require that E[h(X)] < ∞.) PROPOSITION 2.4 Suppose that X1 , · · · , Xk are random variables with ﬁnite expected values. (a) If X1 , · · · , Xk are deﬁned on the same sample space then E(X1 + · · · + Xk ) =
k
E(Xi ).
i=1
(b) If X1 , · · · , Xk are independent random variables then E
k
i=1
Xi
=
k
E(Xi ).
i=1
The proofs of parts (a) and (b) of Proposition 2.4 are quite simple if we assume that (X1 , · · · , Xk ) has a joint density or frequency function; this is left as an exercise. More generally, the proofs are a bit more diﬃcult but can be found in a more advanced probability text such as Billingsley (1995). EXAMPLE 2.7: Suppose that X1 , · · · , Xn are independent random variables with moment generating functions m1 (t), · · · , mn (t) c 2000 by Chapman & Hall/CRC
respectively. Deﬁne S = X1 + · · · + Xn ; the moment generating function of S is mS (t) = E[exp(tS)] = E[exp(t(X1 + · · · + Xn ))] = E[exp(tX1 ) exp(tX2 ) × · · · × exp(tXn )] = =
n i=1 n
E[exp(tXi )] mi (t)
i=1
where we use the independence of X1 , · · · , Xn . Thus the moment generating function of a sum of independent random variables is the product of the individual moment generating functions. ✸ EXAMPLE 2.8: Suppose that X1 , · · · , Xn are independent Normal random variables with E(Xi ) = µi and Var(Xi ) = σi2 . Again deﬁne S = X1 + · · · + Xn . The moment generating function of Xi
is 1 2 2 mi (t) = exp µi t + σi t 2 and so the moment generating function of S is
n
n
n t2 mi (t) = exp t µi + σi2 . mS (t) = 2 i=1 i=1 i=1
Since moment generating functions characterize distributions, it is
n easy to see that S has a Normal distribution with mean i=1 µi
and variance ni=1 σi2 . ✸ EXAMPLE 2.9: Suppose that X1 , · · · , Xn are i.i.d. random variables with moment generating function m(t). Deﬁne n ¯n = 1 X Xi n i=1
and let
¯ n )] = [m(t/n)]n mn (t) = E[exp(tX be its moment generating function. What happens to mn (t) as n → ∞? Note that we can write m(t) = 1 + tE(X1 ) + c 2000 by Chapman & Hall/CRC
t2 t3 E(X12 ) + E(X13 ) + · · · 2 6
and so t2 t3 t E(X1 ) + 2 E(X12 ) + 3 E(X13 ) + · · · n 2n 6n t t t2 2 3 = 1+ E(X1 ) + E(X1 ) + 2 E(X1 ) + · · · n 2n 6n t = 1 + an (t) n
m(t/n) = 1 +
where an (t) → E(X1 ) for all t as n → ∞. Now using the fact that
lim
n→∞
if cn → c, it follows that
lim mn (t) = lim
n→∞
cn 1+ n
n→∞
= exp(c)
tan (t) 1+ n
n
= exp(tE(X1 )).
Note that exp(tE(X1 )) is the moment generating function of a random variable that takes the single value E(X1 ). This suggests ¯ n } converges in some sense to that sequence of random variables {X the constant E(X1 ). We will elaborate on this in Chapter 3. ✸ Covariance and Correlation DEFINITION. Suppose X and Y are random variables with E(X 2 ) and E(Y 2 ) both ﬁnite and let µX = E(X) and µY = E(Y ). The covariance between X and Y is Cov(X, Y ) = E[(X − µX )(Y − µY )] = E(XY ) − µX µY . The covariance is a measure of linear dependence between two random variables. Using properties of expected values it is quite easy to derive the following properties. 1. For any constants a, b, c, and d, Cov(aX + b, cY + d) = a c Cov(X, Y ). 2. If X and Y are independent random variables (with E(X) and E(Y ) ﬁnite) then Cov(X, Y ) = 0. The converse to 2 is not true. In fact, it is simple to ﬁnd an example where Y = g(X) but Cov(X, Y ) = 0. EXAMPLE 2.10: Suppose that X has a Uniform distribution on the interval [−1, 1] and let Y = −1 if X < 1/2 and Y = 1 if c 2000 by Chapman & Hall/CRC
X ≥ 1/2. Since E(X) = E(Y ) = 0, Cov(X, Y ) = E(XY ) and
XY = Thus 1 E(XY ) = 2
−1/2 −1
−X if X < 1/2 X if X ≥ 1/2.
1 x dx + 2
1/2
1 (−x) dx + 2 −1/2
1
x dx = 0 1/2
✸
and Cov(X, Y ) = 0 even though Y = g(X).
PROPOSITION 2.5 Suppose that X1 , · · · , Xn are random variables with E(Xi2 ) < ∞ for all i. Then Var
n
ai Xi
=
i=1
n
a2i Var(Xi ) + 2
i=1
n j−1
ai aj Cov(Xi , Xj ).
j=2 i=1
Proof. Deﬁne µi = E(Xi ). Then Var
n
ai Xi
n n = E ai aj (Xi − µi )(Xj − µj )
i=1
=
i=1 j=1 n n
ai aj E[(Xi − µi )(Xj − µj )]
i=1 j=1
=
n
a2i Var(Xi ) + 2
i=1
n j−1
ai aj Cov(Xi , Xj )
j=2 i=1
since Cov(Xi , Xj ) = Cov(Xj , Xi ). ✷ Note that Proposition 2.5 implies that if X1 , · · · , Xn are independent random variables then the variance of X1 + · · · + Xn is simply the sum of the variances of the Xi ’s. EXAMPLE 2.11: Suppose we are sampling without replacement from a ﬁnite population consisting of N items a1 , · · · , aN . Let Xi denote the result of the ith draw; we then have P (Xi = ak ) =
1 N
and P (Xi = ak , Xj = a ) =
1 N (N − 1)
where 1 ≤ i, j, k, : ≤ N , i = j and k = :. Suppose we deﬁne Sn =
n i=1
c 2000 by Chapman & Hall/CRC
Xi
where n ≤ N ; what are the mean and variance of Sn ? First of all, we deﬁne µa =
N 1 ak N k=1
and σa2 =
N 1 (ak − µa )2 , N k=1
which are the mean and variance of the population; it is easy to see that E(Xi ) = µa and Var(Xi ) = σa2 . Thus it follows that E(Sn ) =
n
E(Xi ) = nµa .
i=1
Var(Sn ) is somewhat trickier; note that Var(Sn ) = nσa2 + n(n − 1)Cov(X1 , X2 ) since Cov(Xi , Xj ) will be the same for all i = j. (Cov(Xi , Xj ) depends on the joint distribution of (Xi , Xj ) and this is the same for all i = j.) We can determine Cov(X1 , X2 ) using the joint distribution of X1 and X2 ; however, an easier approach is to
N consider the case n = N . SN = k=1 ak , a constant, and so Var(SN ) = 0. Thus 0 = N σa2 + N (N − 1)Cov(X1 , X2 ) and so
σa2 . N −1 Substituting into the expression for Var(Sn ) above, we get Cov(X1 , X2 ) = −
Var(Sn ) = nσa2
N −n . N −1
Note that if we sampled with replacement, the random variables X1 , X2 , · · · would be independent and so Var(Sn ) = nσa2 , which is greater than the variance when the sampling is without replacement (provided n ≥ 2). The extra factor (N − n)/(N − 1) that appears in Var(Sn ) when the sampling is without replacement is called the ﬁnite population correction. ✸ Given random variables X1 , · · · , Xn , it is often convenient to represent the variances and covariances of the Xi ’s via a n × n matrix. Set X = (X1 , · · · , Xn )T (a column vector); then we deﬁne the variancecovariance matrix (or covariance matrix) of X to be an n × n matrix C = Cov(X) whose diagonal elements are c 2000 by Chapman & Hall/CRC
Cii = Var(Xi ) (i = 1, · · · , n) and whose oﬀdiagonal elements are Cij = Cov(Xi , Xj ) (i = j). Variancecovariance matrices can be manipulated usefully for linear transformations of X: If Y = BX + a for some m × n matrix B and vector a of length m then Cov(Y ) = BCov(X)B T . Likewise, if we deﬁne the mean vector of X to be
E(X1 ) .. E(X) = . E(Xn ) then E(Y ) = BE(X) + a. While the covariance gives some indication of the linear association between two random variables, its value is dependent on the scale of the two random variables. DEFINITION. Suppose that X and Y are random variables where both E(X 2 ) and E(Y 2 ) are ﬁnite. Then the correlation between X and Y is Cov(X, Y ) Corr(X, Y ) = . [Var(X)Var(Y )]1/2 The advantage of the correlation is the fact that it is essentially invariant to linear transformations (unlike covariance). That is, if U = aX + b and V = cY + d then Corr(U, V ) = Corr(X, Y ) if a and c have the same sign; if a and c have diﬀerent signs then Corr(U, V ) = −Corr(X, Y ). PROPOSITION 2.6 Suppose that X and Y are random variables where both E(X 2 ) and E(Y 2 ) are ﬁnite. Then (a) −1 ≤ Corr(X, Y ) ≤ 1; (b) Corr(X, Y ) = 1 if, and only if, Y = aX + b for some a > 0; Corr(X, Y ) = −1 if, and only if, Y = aX + b for some a < 0. We will leave the proof of this result as an exercise. If X and Y are independent random variables (with E(X 2 ) and E(Y 2 ) ﬁnite) then Corr(X, Y ) = 0 since Cov(X, Y ) = 0. However, as with the covariance, a correlation of 0 does not imply independence. Correlation is merely a measure of linear dependence between random variables; it essentially measures the degree to which we may approximate one random variable by a linear function of another. c 2000 by Chapman & Hall/CRC
PROPOSITION 2.7 Suppose that X and Y are random variables where both E(X 2 ) and E(Y 2 ) are ﬁnite and deﬁne g(a, b) = E[(Y − a − bX)2 ]. Then g(a, b) is minimized at
b0 and
a0
Var(Y ) Cov(X, Y ) = = Corr(X, Y ) Var(X) Var(X) = E(Y ) − b0 E(X)
1/2
with g(a0 , b0 ) = Var(Y )(1 − Corr2 (X, Y )). The proof is left as an exercise. Proposition 2.7 can be interpreted by considering predicting Y as a linear function h of X and considering the mean square prediction error E[(Y − h(X))2 ]. If we take h(x) to be a constant, then as a function of a, E[(Y − a)2 ] is minimized at a = E(Y ) with E[(Y − E(Y ))2 ] = Var(Y ). Taking h(x) to be a linear function, the minimum mean square prediction error (according to Proposition 2.7) is Var(Y )(1 − Corr2 (X, Y )). Thus the reduction in the mean square prediction error when predicting Y by a linear function of X depends explicitly on the correlation. With some imagination, it is possible to derive a more useful measure of dependence between two random variables. Let X and Y be random variables and consider Corr(φ(X), ψ(Y )). If X and Y are independent then this correlation (when welldeﬁned) will always be 0 since φ(X) and ψ(Y ) will always be independent. Other the other hand, if Y = φ(X) then Corr(φ(X), Y ) = 1 even if Corr(X, Y ) = 0. This suggests that we can deﬁne the maximal correlation between X and Y to be maxCorr(X, Y ) = sup Corr(φ(X), ψ(Y )) φ,ψ
where the supremum is taken over all functions φ and ψ with Var(φ(X)) = Var(ψ(Y )) = 1. The condition that Var(φ(X)) = Var(ψ(Y )) = 1 is needed to rule out constant transformations φ and ψ. Clearly, maxCorr(X, Y ) ≥ 0 with maxCorr(X, Y ) = 0 if, and only if, X and Y are independent. (Of course, the functions φ and c 2000 by Chapman & Hall/CRC
1.0 0.5 0.0 1.0
0.5
y
•• • •• • • • •• •• • • • • • • •• • • • ••••• • • • •• •• ••• •• • • •••• •• • • • • ••••• ••••• • • ••• • • • • • ••• • • •••• • • •• • • •• •• • •• ••• • ••• • •• • •• •• •• ••• ••• • •• • • • •• • • • • • • • • •• • •• • ••• • • • •• •••• •• • • • • • • • • • • • • • • •• • •• • ••••••• • • •••• • • • •• • • ••• • • ••• • •• • • • • • • • • ••• • ••••• •• • • • • •• •• • •• ••• • •• •••• • • • • •• •• • • •• •• • • ••• • • ••• • • • • • • • • • ••• • • ••• • • •• • • • • ••• • •• • • • • •• • • • • • • •••• • • • •• • • • • • •• • •• • • • • • • •• • •• • • •••••••••••• •• •• • •• • •••• • • • •• • • •• •• ••• ••• ••• • •• • • • • • • • • • • • • •• • • • • • • • • • • • • • •• •• • •• • • • • • • •• • ••• ••••• •• •• •• •• •• •• •• ••• • •• ••• • • ••••• •• • •• • • ••• • ••••••• • • •• •• ••• •• • •• • • • • • • • • • • • • • • • • • • • • • • •• • • • • •• • ••• • • ••• • ••• • • •••• • • • • •••••• ••• • • • • • •• ••• • •• ••••••• •• • •• •• • •• • •••••••• • • • • • • •• •• • ••• • • •• • •••• • •• •• • •• • • • ••• •• • •• • • •• • • • • • • • • • • • • • ••• • • • •••• • • •••• • • • • • • •• •• • • • • •• • •• •• • • • •• • ••• ••• • • •• • • • • ••• ••••• • • • •• • •• • •• • • • • • • • • • • • • • • •• • • • •• • • •••••••• •••• •• • •• •• • • •• ••• • • •• • • • • • • • • ••••• ••• •••• •••• •• •• • • ••• ••••• •••• •• •• • • •• • • • ••• • • •• •• • • • ••• • • • • •• • ••• • •• ••• • • •• • •• • • • • • • •••• •••••••• ••• • • • • • • • • • • • • • •• •• • • •• •• • • • •• 1.0
0.5
0.0
0.5
1.0
x
Figure 2.1 1000 independent pairs of random variables uniformly distributed on the unit circle.
ψ maximizing the correlation are not unique since Corr(aφ(X) + b, cψ(Y ) + d) is the same as Corr(φ(X), ψ(Y )) if a and c have the same sign.) Unfortunately, the maximal correlation is typically not easy to compute. The following example is an exception. EXAMPLE 2.12: Suppose that X and Y are random variables with joint density function f (x, y) =
1 π
if x2 + y 2 ≤ 1.
Thus (X, Y ) have a Uniform distribution over the region {(x, y) : x2 + y 2 ≤ 1}; it can be shown that Cov(X, Y ) = Corr(X, Y ) = 0. Buja (1990) shows that Cov(φ(X), ψ(Y )) is maximized when φ(x) = x2
and ψ(y) = −y 2 .
The maximal correlation between X and Y is 1/3. Figure 2.1 is a plot of 1000 independent pairs of observations from the density f (x, y); Figure 2.2 shows the observations transformed to give the maximal correlation. ✸ c 2000 by Chapman & Hall/CRC
0.0 1.0
0.8
0.6
ψ
0.4
0.2
• •• • ••• • ••••• • • •••••• ••• • ••• ••••• •••••••••••••••••• • •• •••• ••• •• •••••• ••••••••• ••• •• ••••••••• • • •••• ••• ••• •• • • • • •• • • ••••••••••••••••••••••••••••••••• •••••••••• ••••• • •••••• •• •••••••••••••••• •• •••• • •••••••••••••• •• ••• ••• ••• •••• • • • •• • • • • •• • •• ••• •• •• •••••••••• ••••••••• ••••• •••• • •• ••• •• •• •• •• •• ••• •• ••• •• • •• •••• •• • • • • • •• • • • • •••••••••••• ••••• •• • •••• • • • •• • • • •• • • ••••• • • •• • •••• • • • •• • • • • •••••••• •• • ••••••• • •• •••••• •••• • • • •• • • • • • •••• ••• ••• • • •• • • • • •• • • •• • •• • • • • • •••••••••• •••••••• •• ••• •• ••••• • • ••• •• •• • • ••• •••• ••• ••••• •• • •• •• • •• •• • •• •• ••• •• • • •• •• • •• • • •••••••••••• •• • •••• ••• •• • • •• ••• • • •• • • • • • •• ••• • • • • • • • • • • • • • • • • • • •••••••• ••••••• •••••••• • ••• • • •• • • •••• •• • • ••• • • • • • • • •• • • ••• •• • • • ••• ••• • • • • • • • • • •• • • • • • • • • •• ••••••••• •• • •• ••••••• • •• • • • •• •• • •• • •••••• •••••• ••••• • ••••••• • ••• • •• • • •• ••• ••• • • ••• • • • •• ••••• • •• •••• •••• • • • • • • •• • •• •• ••• ••• • ••••• • ••• • ••• • ••• • •• • •• • • ••• 0.0
0.2
0.4
0.6
0.8
1.0
φ
Figure 2.2 Data in Figure 2.1 transformed to have maximal correlation.
2.3 Conditional distributions and expected values We are often interested in the probability distribution of a random variable (or random variables) given knowledge of some event A. If the conditioning event A has positive probability then we can deﬁne conditional distributions, conditional density functions (marginal and joint) and conditional frequency functions using the deﬁnition of conditional probability, for example, P (X1 ≤ x1 , · · · , Xk ≤ xk A) =
P (X1 ≤ x1 , · · · , Xk ≤ xk , A) . P (A)
In the case of discrete random variables, it is straightforward to deﬁne the conditional frequency function of (say) X1 , · · · , Xj given the event Xj+1 = xj+1 , · · · , Xk = xk as f (x1 , · · · , xj xj+1 , · · · , xk ) = P (X1 = x1 , · · · , Xj = xj Xj+1 = xj+1 , · · · , Xk = xk ) =
P (X1 = x1 , · · · , Xj = xj , Xj+1 = xj+1 , · · · , Xk = xk ) , P (Xj+1 = xj+1 , · · · , Xk = xk )
c 2000 by Chapman & Hall/CRC
which is simply the joint frequency function of X1 , · · · , Xk divided by the joint frequency function of Xj+1 , · · · , Xk . In many problems, it is in fact the conditional distributions that are most naturally speciﬁed and from these, the joint distribution can be obtained. The following example is a nontrivial illustration of this. EXAMPLE 2.13: Mark/recapture experiments are used to estimate the size of animal populations. Suppose that the size of the population is N (unknown). Initially, m0 members of the populations are captured and tagged for future identiﬁcation before being returned to the population. Subsequently, a similar process is repeated k times: mi members are captured at stage i and we deﬁne a random variable Xi to be the number of captured members who were tagged previously; the mi −Xi nontagged members are tagged and all mi members are returned to the population. If we assume that the population size does not change over the course of the experiment then it is possible to derive the joint distribution of (X1 , · · · , Xk ). Given our assumptions, it is reasonable to assume that X1 has a Hypergeometric distribution; that is, & m0 N − m0 N . P (X1 = x1 ) = x1 m1 − x1 m1 Similarly, given X1 = x1 , X2 has a Hypergeometric distribution
P (X2 = x2 X1 = x1 ) =
n1 x2
N − n1 n1 − x2
&
N n1
where n1 = m0 + (m1 − x1 ) is the number of tagged members of the population prior to the second sampling stage. Similarly, we can ﬁnd the conditional distribution of Xj+1 given X1 = x1 , · · · , Xj = xj . Setting nj = m0 + (m1 − x1 ) + · · · + (mj − xj ), we have P (Xj+1 = xj+1 X1 = x1 , · · · , Xj = xj )
=
nj xj+1
N − nj mj+1 − xj+1
&
N . mj+1
The joint frequency function of (X1 , · · · , Xk ) is now obtained by multiplying the respective conditional frequency functions: P (X1 = x1 , X2 = x2 ) = P (X1 = x1 )P (X2 = x2 X1 = x1 ) c 2000 by Chapman & Hall/CRC
and for k ≥ 3, P (X1 = x1 , · · · , Xk = xk ) = P (X1 = x1 , · · · , Xk−1 = xk−1 ) ×P (Xk = xk X1 = x1 , · · · , Xk−1 = xk−1 ). From a statistical point of view, the goal in this problem is to use the information in (X1 , · · · , Xk ) to estimate the population size N . ✸ Deﬁning conditional distributions when the conditioning event has probability 0 is much more diﬃcult but nonetheless an important problem. For example, if we have two continuous random variables X and Y , we might be interested in the conditional distribution of Y given X = x where (since X is continuous) P (X = x) = 0. DEFINITION. Suppose that (X1 , · · · , Xk ) has the joint density function g(x1 , · · · , xk ). Then the conditional density function of X1 , · · · , Xj given Xj+1 = xj+1 , · · · , Xk = xk is deﬁned to be f (x1 , · · · , xj xj+1 , · · · , xk ) =
g(x1 , · · · , xj , xj+1 , · · · , xk ) h(xj+1 , · · · , xk )
provided that h(xj+1 , · · · , xk ), the joint density of Xj+1 , · · · , Xk , is strictly positive. This conditional density function (viewed as a function of x1 , · · · , xj for ﬁxed xj+1 , · · · , xk ) has the same properties as any other density function; we can use this conditional density function to evaluate conditional probabilities (given Xj+1 = xj+1 , · · · , Xk = xk ) by integration as well as evaluate conditional expected values. DEFINITION. Given an event A with P (A) > 0 and a random variable X with E[X] < ∞, we deﬁne ∞
E(XA) = 0
P (X > xA) dx −
0 −∞
P (X < xA) dx
to be the conditional expected value of X given A. The assumption that E[X] < ∞ is not always necessary to ensure that E(XA) is welldeﬁned; for example, the condition A may imply that X is bounded on A in which case the integral deﬁning E(XA) would be welldeﬁned. The following result extends Proposition 1.4 (law of total probability) to conditional expectations. c 2000 by Chapman & Hall/CRC
THEOREM 2.8 Suppose that A1 , A2 , · · · are disjoint events with P (Ak ) > 0 for all k and ∞ k=1 Ak = Ω. Then if E[X] < ∞, E(X) =
∞
E(XAk )P (Ak ).
k=1
Proof. Assume X ≥ 0 with probability 1. (Otherwise, split X into its positive and negative parts.) Then ∞
E(XAk ) = and so ∞
E(XAk )P (Ak ) =
k=1
=
P (X > xAk ) dx
0
∞ ∞ k=1 0 ∞ ∞ 0
∞
P (Ak )P (X > xAk ) dx P (Ak )P (X > xAk ) dx
k=1
P (X > x) dx
= 0
= E(X). We can interchange the order of summation and integration because P (Ak )P (X > xAk ) ≥ 0. Theorem 2.8 also holds if some of the Ak ’s have P (Ak ) = 0 provided we take care of the fact that E(XAk ) is not necessarily welldeﬁned. This can be done by assigning E(XAk ) an arbitrary (but ﬁnite) value, which is then annihilated by multiplying by P (Ak ) = 0. There is an interesting interpretation of Theorem 2.8 that will be useful in a more general setting. Given E(XAk ) (k = 1, 2, · · ·), deﬁne a random variable Y such that Y (ω) = E(XAk ) if ω ∈ Ak (k = 1, 2, · · ·). Y is now a discrete random variable whose expected value is E(Y ) = =
y ∞
y P (Y = y) E(XAk )P (Ak )
k=1
= E(X) c 2000 by Chapman & Hall/CRC
by Theorem 2.8 (provided that E[X] < ∞). More generally, given a continuous random vector X, we would like to deﬁne E(Y X = x) for a random variable Y with E[Y ] < ∞. Since the event [X = x] has probability 0, this is somewhat delicate from a technical point of view, although if Y has a conditional density function given X = x, f (yx) then we can deﬁne ∞ E(Y X = x) = y f (yx) dy. −∞
We can obtain similar expressions for E[g(Y )X = x] provided that we can deﬁne the conditional distribution of Y given X = x in a satisfactory way. In general, E[g(Y )X = x] is a function of x. Moreover, if h(x) = E[g(Y )X = x] then E[h(X)] = E[E[E(g(Y )X]] = E[g(Y )] as we had in Theorem 2.8. The following result records some of the key properties of conditional expected values. PROPOSITION 2.9 Suppose that X and Y are random vectors. Then (a) if E[g1 (Y )] and E[g2 (Y )] are ﬁnite, E[a g1 (Y ) + b g2 (Y )X = x] = a E[g1 (Y )X = x] + b E[g2 (Y )X = x] (b) E[g1 (X)g2 (Y )X = x] = g1 (x)E[g2 (Y )X = x] if E[g2 (Y )] is ﬁnite; (c) If h(x) = E[g(Y )X = x] then E[h(X)] = E[g(Y )] if E[g(Y )] is ﬁnite. A rigorous proof of Proposition 2.9 follows from a more technically rigorous deﬁnition of conditional expectation; see, for example, Billingsley (1995). In special cases, for example when we assume a conditional density, the proof is straightforward. The following result provides a useful decomposition for the variance of a random variable. THEOREM 2.10 Suppose that Y is a random variable with ﬁnite variance. Then Var(Y ) = E[Var(Y X)] + Var[E(Y X)] where Var(Y X) = E[(Y − E(Y X))2 X]. c 2000 by Chapman & Hall/CRC
Proof. Deﬁne h(X) = E(Y X) and µ = E(Y ). Then Var(Y ) = E[(Y − h(X) + h(X) − µ)2 ]
= E E[(Y − h(X) + h(X) − µ)2 X]
= E E[(Y − h(X))2 X] + E E[(h(X) − µ)2 X] +2E [E[(Y − h(X))(h(X) − µ)X]] . Now
E E[(Y − h(X))2 X] and E E[(h(X) − µ)2 X]
= E[Var(Y X)] = E[(h(X) − µ)2 ] = Var[E(Y X)].
Finally, for the “crossproduct” term, we have E[(Y = = =
− h(X))(h(X) − µ)X = x] (h(x) − µ)E[Y − h(X)X = x] (h(x) − µ)[h(x) − h(x)] 0.
Thus
E [E[(Y − h(X))(h(X) − µ)X]] = 0, which completes the proof.
EXAMPLE 2.14: Suppose that X1 , X2 , · · · are i.i.d. random variables with mean µ and variance σ 2 ; let N be a Poisson random variable (with mean λ) that is independent of the Xi ’s. Deﬁne the N random variable S= Xi i=1
where S = 0 if N = 0. We would like to determine the mean and variance of S. First of all, note that E(S) = E[E(SN )] and E(SN = n) = nµ since N is independent of the Xi ’s. Thus E(S) = µE(N ) = λµ. Likewise, we have Var(S) = Var[E(SN )] + E[Var(SN )]. We already know that E(SN ) = µN and since the Xi ’s are i.i.d. and independent of N , we have Var(SN = n) = nσ 2 . Thus Var(S) = µ2 Var(N ) + σ 2 E(N ) = λ(µ2 + σ 2 ). Note that µ2 + σ 2 = E(Xi2 ). c 2000 by Chapman & Hall/CRC
The random variable S is said to have a compound Poisson distribution. Such distributions are often used in actuarial science to model the monetary value of claims against various types of insurance policies in a given period of time; the motivation here is that N represents the number of claims with the Xi ’s representing the monetary value of each claim. ✸ 2.4 Distribution theory for Normal samples In Example 1.15 we introduced the Normal distribution. Recall that a random variable X has a Normal distribution with mean µ and variance σ 2 if its density is
1 (x − µ)2 f (x) = √ exp − 2σ 2 σ 2π
.
The Normal distribution is very important in probability and statistics as it is frequently used in statistical models and is often used to approximate the distribution of certain random variables. A number of distributions arise naturally in connection with samples from a Normal distribution. These distributions include the χ2 (chisquare) distribution, Student’s t distribution and the F distribution. The Multivariate Normal Distribution Suppose that X1 , · · · , Xp are i.i.d. Normal random variables with mean 0 and variance 1. Then the joint density function of X = (X1 , · · · , Xp )T is
fX (x) = =
1 1 exp − (x21 + · · · + x2p ) 2 (2π)p/2
1 1 T exp − x x . 2 (2π)p/2
The random vector X has a standard multivariate (or pvariate) Normal distribution. DEFINITION. Let A be a p × p matrix and µ = (µ1 , · · · , µp )T a vector of length p. Given a standard multivariate Normal random vector X, deﬁne Y = µ + AX. We say that Y has a multivariate Normal distribution with mean c 2000 by Chapman & Hall/CRC
vector µ and variancecovariance matrix C = AAT . (We will abbreviate this Y ∼ Np (µ, C).) Note that C = AAT is a symmetric, nonnegative deﬁnite matrix (that is, C T = C and v T Cv ≥ 0 for all vectors v). If A is an invertible matrix then the joint density of Y exists and is given by
1 1 fY (y) = exp − (y − µ)T C −1 (y − µ) . p/2 1/2 2 (2π)  det(C) Note that C −1 exists since A−1 exists. On the other hand, if A is not invertible then the joint density of Y does not exist; Y is deﬁned on a hyperplane in Rp of dimension r = rank(C) < p. We will now state some basic properties of the multivariate Normal distribution. Assume that Y ∼ Np (µ, C) and let Cij be the element of C in the ith row and jth column of C. Then 1. if B is an r × p matrix then BY ∼ Nr (Bµ, BCB T ); 2. any subcollection (Yi1 , · · · , Yik )T of Y has a multivariate Normal distribution; 3. Yi ∼ N (µi , σi2 ) where σi2 = Cii ; 4. Cov(Yi , Yj ) = Cij . Properties 2 and 3 follow from property 1 by choosing the matrix B appropriately. Property 4 follows by writing Y = µ + AX where X ∼ Np (0, I) and C = AAT ; if the ith and jth rows of A are aTi and aTj then Yi = µi + aTi X and Yj = µj + aTj X, and so it is easy to verify that Cov(Yi , Yj ) = aTi aj = Cij . The χ2 distribution and orthogonal transformations Suppose that X1 , · · · , Xp are i.i.d. N (0, 1) random variables so that X = (X1 , · · · , Xp )T ∼ Np (0, I) where I is the identity matrix. In some statistical applications, we are interested in the squared “length” of X, X2 = X T X. DEFINITION. Let X ∼ Np (0, I) and deﬁne V = X2 . The random variable V is said to have a χ2 (chisquare) distribution with p degrees of freedom. (V ∼ χ2 (p).) c 2000 by Chapman & Hall/CRC
The density of a χ2 distribution is quite easy to determine. We start with the case where p = 1; in this case, V = X12 and so √ √ P (V ≤ x) = P (− x ≤ X1 ≤ x) √ √ = Φ( x) − Φ(− x) where Φ is the N (0, 1) distribution function. Diﬀerentiating, we get the density x−1/2 fV (x) = √ exp(−x/2) 2π
(x > 0),
which is simply the density of a Gamma(1/2, 1/2) distribution; thus a χ2 (1) distribution is simply a Gamma(1/2, 1/2) distribution. The general case now is simple since V = X12 + · · · + Xp2 , a sum of p independent χ2 (1) random variables. It follows (from Example 2.5) that the density of V ∼ χ2 (p) is fV (x) =
xp/2−1 exp(−x/2) 2p/2 Γ(p/2)
(x > 0).
Orthogonal transformations are eﬀectively transformations that preserve the length of vectors; for this reason, they turn out to be a useful tool in connection with the multivariate Normal distribution. To deﬁne an orthogonal matrix, we start by deﬁning vectors (of length p) a1 , a2 , · · · , ap such that aTk ak = 1
for k = 1, · · · , p and aTj ak = 0
for j = k
The vectors a1 , · · · , ap are said to be orthogonal vectors and, in fact, form an orthonormal basis for Rp ; that is, any vector v ∈ Rp , we have p v=
ci ai .
i=1
Now deﬁne a p × p matrix O such that the kth row of O is aTk :
aT1 O = ... aTp Note that
aT1 OOT = ... (a1 · · · ap ) = I aTp
c 2000 by Chapman & Hall/CRC
and so O−1 = OT ; thus we also have OT O = I. The matrix O is called an orthogonal (or orthonormal) matrix. EXAMPLE 2.15: If p = 2, orthogonal matrices rotate vectors by some angle θ. To see this, let
O=
cos(θ) − sin(θ) sin(θ) cos(θ)
for some θ between 0 and 2π; it is easy to verify that O is an orthogonal matrix and, in fact, an 2 × 2 orthogonal matrix can be written in this form. Now let v be any vector and write
v=
v1 v2
=
r cos(φ) r sin(φ)
where r2 = v T v = v12 + v22 and φ is the angle between v and the vector (1, 0)T . Then
Ov =
r cos(θ) cos(φ) − r sin(θ) sin(φ) r sin(θ) cos(φ) + cos(θ) sin(φ)
=
r cos(θ + φ) r sin(θ + φ)
.
Thus the orthogonal matrix O rotates the vector v by an angle θ. ✸ Now take X ∼ Np (0, I) and deﬁne deﬁne Y = OX where O is an orthogonal matrix. It follows from the properties of the multivariate Normal distribution that Y ∼ Np (0, I) since OOT = I. Moreover, p
Yi2 = Y T Y
i=1
= X T OT OX = XT X =
p
2 i=1 Xi
p
p
Xi2
i=1
2 i=1 Yi
and both and will have χ2 distributions with p degrees of freedom. Consider the following application of this theory. Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance σ 2 and deﬁne n ¯= 1 Xi X n i=1 c 2000 by Chapman & Hall/CRC
and
n 1 ¯ 2; (Xi − X) S = n − 1 i=1 2
¯ and S 2 are called the sample mean and sample variance X ¯ ∼ N (µ, σ 2 /n). The following respectively. We know already that X ¯ is independent of S 2 and that the results indicates that X 2 distribution of S is related to a χ2 with n − 1 degrees of freedom. PROPOSITION 2.11 (n − 1)S 2 /σ 2 ∼ χ2 (n − 1) and is indepen¯ ∼ N (µ, σ 2 ). dent of X Proof. First note that (n − 1) 2 S = σ2 =
n 1 ¯ 2 (Xi − X) σ 2 i=1 n '
¯ − µ)/σ (Xi − µ)/σ − (X
(2
i=1
and so we can assume (without loss of generality) that µ = 0 and σ 2 = 1. Deﬁne an orthogonal matrix O whose ﬁrst row consists of n−1/2 repeated n times; the remaining rows can be determined by some orthogonalization procedure (such as the GramSchmidt procedure) but do not need to be speciﬁed here. Now let Y = OX where X = (X1 , · · · , Xn )T and note that n 1 Xi Y1 = √ n i=1
is independent of the remaining elements of Y , Y2 , · · · , Yn . We now have n
¯ (Xi − X)
2
=
i=1
=
n i=1 n
Xi2
−
n 1 √ Xi n i=1
2
Yi2 − Y12
i=1
=
n
Yi2
i=2
∼ χ2 (n − 1). Moreover,
n
i=1 (Xi
¯ 2 depends only on Y2 , · · · , Yn and hence is − X)
c 2000 by Chapman & Hall/CRC
√ ¯ independent of Y1 = nX. Thus in general we have (n − 1)S 2 /σ 2 2 ¯ (and S is independent of X). ¯ independent of X The t and F distributions 2 DEFINITION. Let Z ∼ N (0, 1) and V ∼ χ (n) be independent random variables. Deﬁne T = Z/ V /n; the random variable T is said to have Student’s t distribution with n degrees of freedom. (T ∼ T (n).) DEFINITION. Let V ∼ χ2 (n) and W ∼ χ2 (m) be independent random variables. Deﬁne F = (V /n)/(W/m); the random variable F is said to have an F distribution with n and m degrees of freedom. (F ∼ F(n, m).)
Alternatively, we could deﬁne the t and F distributions via their density functions. However, the representation of t and F random variables as functions of Normal and χ2 random variables turns out to be convenient in many situations (see Example 2.17 below). Nonetheless, it is fairly easy to ﬁnd the densities of the t and F distributions using the representations as well as some the techniques developed earlier for ﬁnding densities of functions of random variables. EXAMPLE 2.16: Suppose that Z ∼ N (0, 1) and V∼ χ2 (n) are independent random variables, and deﬁne T = Z/ V /n. To determine the density of T , we will introduce another random variable S, determine the joint density of (S, T ) and then integrate out this joint density to determine the marginal density of T . The choice of S is somewhat arbitrary; we will take S = V so that (S, T ) = h(V, Z) for some function h that is onetoone over the range of the random variables V and Z. Theinverse of h is h−1 (s, t) = (s, t s/n) and the Jacobian of h−1 is s/n. Thus the joint density of (S, T ) is (after substituting the appropriate terms into the joint density of (V, Z)) )
g(s, t) =
*
s(n−1)/2 s √ exp − (1 + t2 /n) (n+1)/2 2 2 nπΓ(n/2)
for s > 0 and −∞ < t < ∞. Integrating out over s, we get ∞
fT (t) =
g(s, t) ds 0
c 2000 by Chapman & Hall/CRC
=
Γ((n + 1)/2) t2 √ 1+ nπΓ(n/2) n
−(n+1)/2
Note that when n = 1, the t distribution is simply the Cauchy distribution given in Example 1.26. ✸ We can go through a similar procedure to determine the density of the F distribution but that will be left as an exercise. The density of an F distribution with n and m degrees of freedom is f (x) =
nn/2 Γ((n + m)/2) (n−2)/2 x (1 + ns/m)−(n+m)/2 n/2 m Γ(m/2)Γ(n/2)
for x > 0. EXAMPLE 2.17: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance σ 2 . Deﬁne the sample mean and variance of the Xi ’s: ¯ = X
n 1 Xi n i=1
S2 =
n 1 ¯ 2. (Xi − X) n − 1 i=1
√ ¯ − µ)/S; it is wellknown that T ∼ T (n − 1). Now deﬁne T = n(X To see this, note that we can rewrite T as √ ¯ n(X − µ)/σ T = S 2 /σ 2 √ ¯ where n(X − µ)/σ ∼ N (0, 1) and independent of (n − 1)S 2 /σ 2 ∼ χ2 (n−1) by Proposition 2.11. Thus using the deﬁnition of Student’s t distribution, T ∼ T (n − 1). ✸ Projection matrices Orthogonal matrices eﬀectively “rotate” vectors without changing their length. Another class of matrices that has important applications to probability and statistics is projection matrices. DEFINITION. Let H be a symmetric p×p matrix with H 2 = H. Then H is called a projection matrix. As its name suggests, a projection matrix “projects” onto a subspace S of Rp . More precisely, if v ∈ Rp is a vector then Hv lies c 2000 by Chapman & Hall/CRC
in the subspace S. Moreover, Hv is the “closest” vector in S to v in the sense that Hv minimizes v − u2 over all u ∈ S. EXAMPLE 2.18: Suppose that x1 , · · · , xr be r ≤ p linearly independent vectors in Rp . (Linear independence means that c1 x1 + · · · + cr xr = 0 implies c1 = · · · = cr = 0.) Deﬁne the matrices B = (x1 · · · xr ) and
H = B(B T B)−1 B T . Then H is a projection matrix onto the space spanned by the vectors x1 , · · · , xr . To see this, note that
H T = B(B T B)−1 B T
T
= B(B T B)−1 B T = H
and H 2 = B(B T B)−1 B T B(B T B)−1 B T = B(B T B)−1 B T = H. Moreover, for any vector v, Hv clearly lies in the space spanned by x1 , · · · , xr since we can write Hv = B(B T B)−1 B T v = Bv & and Bv & lies in the space spanned by x1 , · · · , xr .
✸
The projection matrix in Example 2.18 turns out to have special signiﬁcance in statistics, for example in linear regression analysis (see Chapter 8). If H is a projection matrix onto the subspace S and v ∈ S then Hv = v which implies that 1 is an eigenvalue of H. The following result shows that 0 and 1 are the only possible eigenvalues of H. PROPOSITION 2.12 Suppose that H is a projection matrix. Then 0 and 1 are the only possible eigenvalues of H. Proof. Let λ be an eigenvalue of H and v be a corresponding eigenvector; thus Hv = λv. Multiplying both sides by H, we get H 2 v = Hλv = λHv = λ2 v. However, H 2 = H and so λ2 is an eigenvalue of H with eigenvector v; thus λ2 = λ and so λ = 0 or 1. Since H is symmetric , we can ﬁnd eigenvectors a1 , · · · , ap of H, which form an orthonormal basis for Rp . Since 0 and 1 are the only c 2000 by Chapman & Hall/CRC
possible eigenvalues of H, we have
ak k = 1, · · · , r 0 k = r + 1, · · · , p
Hak =
where a1 , · · · , ar are the eigenvectors of H with eigenvalues equal to 1. (r is the rank of H.) Now take any vector v ∈ Rp . Since the eigenvectors of H, a1 , · · · , ap form an orthonormal basis for Rp , we have v=
p
ck ak
where ck = v T ak .
p
r
k=1
From this it follows that Hv =
ck Hak =
k=1
ck ak .
k=1
The space S onto which the matrix H projects is spanned by the eigenvectors a1 , · · · , ar . The space spanned by the vectors ar+1 , · · · , ap is called the orthogonal complement of S. If S ⊥ is the orthogonal complement of S, v ∈ S and v ⊥ ∈ S ⊥ then v T v ⊥ = 0; that is, v and v ⊥ are orthogonal vectors. PROPOSITION 2.13 Suppose that H is a projection matrix that projects onto the subspace S. Then I − H is a projection matrix projecting onto the orthogonal complement of S. Proof. I − H is a projection matrix since (I − H)T = I − H T = H T and (I − H)2 = I − 2H + H 2 = I − H. To see that I − H projects onto the orthogonal complement of S, write v=
p
ck ak
k=1
where a1 , · · · , ap are the eigenvectors of H. Then (I − H)v =
p
ck ak −
k=1
r
p
ck ak =
k=1
ck ak ,
k=r+1
which lies in the orthogonal complement of S. From this result, we obtain the almost trivial decomposition of a vector v into a sum of orthogonal vectors: v = Hv + (I − H)v = u + u⊥ c 2000 by Chapman & Hall/CRC
where u ∈ S and u⊥ ∈ S ⊥ . We will now consider an application of projection matrices using the multivariate Normal distribution. Suppose that X ∼ Np (0, I) and deﬁne the random variable V = X T HX for some projection matrix H. The following result shows that V has a χ2 distribution. PROPOSITION 2.14 V has a χ2 distribution with r = rank(H) degrees of freedom. Proof. First of all, note that X T HX = X T H 2 X = (HX)T (HX) = HX2 and HX =
r
(aTk X)ak
k=1
where a1 , · · · , ap are the (orthonormal) eigenvectors of H. Now we can deﬁne the orthogonal matrix
aT1 O = ... aTp and deﬁne Y = OX; Y ∼ Np (0, I). If Y = (Y1 , · · · , Yp )T then Yk = aTk X where Y1 , · · · , Yp are i.i.d. N (0, 1) random variables. Thus we have r HX =
Yk ak
k=1
and so (HX)T (HX) = X T HX = =
r r k=1 j=1 r Yk2 k=1 2
∼ χ (r)
since aTk ak = 1 and aTk aj = 0 for j = k. c 2000 by Chapman & Hall/CRC
Yj Yk aTj ak
EXAMPLE 2.19: Suppose that X ∼ Nn (0, I). If H is a projection matrix with rank(H) = r < n then (by Proposition 2.14) HX2 ∼ χ2 (r) and likewise (I − H)X2 ∼ χ2 (n − r); moreover, these two random variables are independent. Thus if we deﬁne HX2 /r W = (I − H)X2 /(n − r) then W ∼ F(r, n − r). ✸ 2.5 Poisson processes Suppose we are interested in the arrival patterns of customers entering a grocery store at diﬀerent times of the day. Let T1 , T2 , · · · be random variables representing the arrival times of the customers and, for a given time interval A, deﬁne N (A) =
∞
I(Ti ∈ A)
i=1
to be the number of arrivals in the interval A. Clearly, N (A) is a nonnegative integervalued random variable that gives the number of arrivals in the interval A. If B is another interval that is disjoint of A (that is, A ∩ B = ∅) then clearly N (A ∪ B) = N (A) + N (B). We can now specify a model for the “point process” N (·) by specifying the joint distribution of the random vector (N (A1 ), N (A2 ), · · · , N (Ak )) for any sets A1 , · · · , Ak . DEFINITION. Let S ⊂ Rk and suppose that for any A ⊂ S, N (A) is a nonnegative integervalued random variable. Then N (·) is called a point process (on S) if (a) N (∅) = 0; (b) N (A ∪ B) = N (A) + N (B) for any disjoint sets A and B. DEFINITION. A point process N (·) deﬁned on S ⊂ Rk is called a (homogeneous) Poisson process if (a) For any A = [a1 , b1 ] × · · · × [ak , bk ] ⊂ S, N (A) has a Poisson distribution with mean λ
k i=1
for some λ > 0; c 2000 by Chapman & Hall/CRC
(bi − ai )
(b) For any two disjoint sets A and B, N (A) and N (B) are independent random variables. The parameter λ is called the intensity of the Poisson process. Although the deﬁnition gives the distribution of N (A) only for “rectangular” sets A, it is quite easy to extend this to a general set A. Any set A can be expressed as a countable union of disjoint rectangles B1 , B2 , · · ·. Since N (B1 ), N (B2 ), · · · are independent Poisson random variables, we have that N (A) =
∞
N (Bi )
i=1
is Poisson. The mean of N (A) is λ vol(A) where vol(A) is the “volume” of the set A. EXAMPLE 2.20: A Poisson process on the positive real line can be constructed as follows. Let X1 , X2 , · · · be independent Exponential random variables with parameter λ and deﬁne Tk =
k
Xi ;
i=1
note that Tk has a Gamma distribution with shape parameter k and scale parameter λ. Then we can deﬁne a point process N (·) such that ∞
N (A) =
I(Tk ∈ A).
k=1
To see that N (·) is a Poisson process, note that N (A) and N (B) will be independent for disjoint A and B because of the independence of the Xi ’s and the “memorylessness” property of the Exponential distribution. Moreover, it follows from the memorylessness property that if A = [s, s + t], the distribution of N (A) is the same for all s ≥ 0. Thus it suﬃces to show that N (A) has a Poisson distribution for A = [0, t]. First of all, note that N (A) ≤ k if, and only if, Yk+1 > t. Since Yk+1 has a Gamma distribution with parameters k + 1 and λ, we have P (N (A) ≤ k) =
λk+1 k!
c 2000 by Chapman & Hall/CRC
∞ t
xk exp(−λx) dx
=
(λt)k exp(−λt) k! ∞ λk xk−1 exp(−λx) dx + (k − 1)! t
after integrating by parts. Repeating this procedure (reducing the power of x by one at each stage), we get P (N (A) ≤ k) =
k (λt)j exp(−λt)
j!
j=0
and so N (A) has a Poisson distribution with mean λt for A = [0, t]. Thus the point process N (·) is a Poisson process with intensity λ. ✸ The homogeneous Poisson process assumes a constant intensity λ. One can also deﬁne a nonhomogeneous Poisson process N (·) on S ⊂ Rk whose intensity varies over S. Given a function λ(t) deﬁned for t ∈ S, N (·) is a nonhomogeneous Poisson process with intensity λ(t) if N (A) and N (B) are independent for disjoint A and B with
N (A) ∼ Pois
λ(t) dt . A
An interesting type of nonhomogeneous Poisson process is the marked Poisson process. Let T1 , T2 , · · · be the points of a homogeneous Poisson process on S with intensity λ and deﬁne i.i.d. random variables X1 , X2 , · · · that are also independent of the Ti ’s. Then for A ⊂ S and B ⊂ R, deﬁne N ∗ (A × B) =
∞
I(Ti ∈ A, Xi ∈ B).
i=1
N ∗ (A
Then × B) ∼ Pois(λ vol(A)P (X1 ∈ B)). Moreover, if A1 and A2 are disjoint sets then N ∗ (A1 × B1 ) and N ∗ (A2 × B2 ) are independent for any B1 and B2 ; the same is true if B1 and B2 are disjoint. N ∗ is typically on nonhomogeneous Poisson process on S × R whose intensity depends on the distribution of the Xi ’s. EXAMPLE 2.21: In forestry, one measure of the density of a “stand” of trees is the basal area proportion, which is roughly deﬁned to be the proportion of the forest actually covered by the bases of trees. (In fact, the basal area of a given tree is typically c 2000 by Chapman & Hall/CRC
not measured at a ground level but rather at a certain height above ground level.) A commonlyused method of estimating the basal area proportion is the anglecount method, which is due to the Austrian forester Walter Bitterlich (see Bitterlich, 1956; Holgate, 1967). Suppose that the forest under consideration is a “Poisson” forest; that is, we assume that the tree centers form a Poisson process with intensity λ trees/meter2 . Furthermore, we will assume that each tree is circular in cross section. The anglecount method counts the number of trees whose diameter subtends an angle greater than θ from some randomly chosen point O. (θ is speciﬁed and typically quite small.) Thus a tree whose diameter is x will be sampled if its distance d from the point O satisﬁes x d< ; 2 sin(θ/2) that is, a tree with diameter x must lie within a circle of radius r(x) = x/(2 sin(θ/2)) around O. The number of trees in a circle of radius r(x) is a Poisson random variable with mean λπr2 (x). Suppose that the tree diameters are represented by i.i.d. random variables X1 , X2 , · · · with distribution function F . Thus we have a marked Poisson process. Assume ﬁrst that F is a discrete distribution putting probability pk at xk for k = 1, · · · , m. Deﬁne Nk to be the number of sampled trees with diameter xk and N = N1 + · · · + Nm to be the total number of sampled trees. Now N1 , · · · , Nm are independent Poisson random variables with E(Nk ) = pk λπr2 (xk ) = λπpk
x2k 4 sin2 (θ/2)
and so N is a Poisson random variable with mean m λπ λπ E(N ) = E(Xi2 ). pk x2k = 2 4 sin (θ/2) k=1 4 sin2 (θ/2)
More generally, it can be shown that N (the total number of sampled trees) has a Poisson distribution regardless of the distribution of the Xi ’s. What is interesting here is the proportionality between E(Xi2 ) and E(N ), which can be exploited to estimate E(Xi2 ). For our Poisson forest, we can deﬁne the “average” basal area proportion to be B = λπE(Xi2 )/4. (Note that πE(Xi2 )/4 is the mean basal area for a tree while λ is the density of trees.) Using the relationship between E(Xi2 ) and E(N ), we have that B = c 2000 by Chapman & Hall/CRC
E(N ) sin2 (θ/2). Using this fact, it is possible to estimate B based on N . The angle count method is, in fact, a special case of biased sampling (Example 1.27) as we are more likely to sample larger trees than smaller trees. ✸ EXAMPLE 2.22: Poisson process models are often used to model the amount of traﬃc on a network such as a telephone system. Suppose that telephone calls arrive as a homogeneous Poisson process (on the entire real line) with intensity λ calls/minute. In addition, assume that the lengths of each call are i.i.d. continuous random variables with density function f (x). If S1 , S2 , · · · represent the starting times of the calls and X1 , X2 , · · · the lengths of the calls then for a < b, we have ∞
I(a < Si ≤ b, Xi ∈ B) ∼ Pois λ(b − a)
i=1
f (x) dx . B
What is of most interest in this example is the number of calls being made at a given point in time t; we will call this random variable N (t). A given call represented by (Si , Xi ) (its starting time and length) will be active at time t if both Si ≤ t and Si + Xi ≥ t. Thus N (t) =
∞
I(Si ≤ t, Si + Xi ≥ t) =
i=1
∞
I((Si , Xi ) ∈ B(t))
i=1
where B(t) = {(s, x) : s ≤ t, s + x ≥ t}, and so N (t) ∼ Pois(µ(t)) where ∞ t
µ(t) =
λf (x) ds dx 0
t−x
∞
= λ
x f (x) dx 0
= λE(Xi ). Thus the distribution of N (t) is independent of t. Figure 2.3 shows a plot of N (t) versus t for simulated data; the calls arrive as a Poisson process with rate 10 calls/minute while the call lengths are Exponential random variables with mean 5. Based on the calculations √ above the mean of N (t) is 50 while the standard deviation is 50 ≈ 7.1; note that the simulated number of calls N (t) is, for the most part, within 2 standard deviations of the mean. ✸ c 2000 by Chapman & Hall/CRC
70 60 50 40 30
Number of ongoing calls
0
100
200
300
400
500
time
Figure 2.3 Plot of simulated data over a 500 minute period; the call arrival rate is 10 calls/minute while the call lengths are Exponential with mean 5.
2.6 Generating random variables In complex problems in probability and statistics (as well as in other disciplines), computer simulation is necessary because exact analytic properties of certain random processes may be diﬃcult or even impossible to derive exactly. Some examples of problems where computer simulation is used are • evaluation or approximation of the probability distributions; • examination of sample path properties of random processes (such as a Poisson process); • evaluation of integrals. In order to simulate random processes on a computer, we must be able to generate (by computer) random variables having a speciﬁed joint distribution. However, for simplicity, we will consider generating independent random variables with a common distribution F . c 2000 by Chapman & Hall/CRC
It is important to note that computergenerated random variables are not really random as they are typically produced by deterministic algorithms (and so cannot be truly random). For this reason, the “random variables” produced by such algorithms are referred to as pseudorandom. The key lies in ﬁnding an algorithm which produces (pseudo) random variables possessing all the analytic properties of the random variables to an acceptable tolerance. In general, we produce a stream of random variables via a recursive algorithm; that is, given xn , we produce xn+1 by xn+1 = g(xn ) for some function g. For a given distribution function F , we would like to choose g so that for any a < b n 1 I(a < xk ≤ b) → F (b) − F (a) N k=1
as N → ∞
independently of the starting value x0 . In fact, this condition is quite easy to satisfy. A much more stringent requirement is that the pseudorandom variables {x1 , x2 , x3 , · · ·} behave like outcomes of i.i.d. random variables X1 , X2 , X3 , · · · (again independently of x0 ). Clearly this is impossible since xn+1 depends explicitly on xn ; nonetheless, if g is chosen appropriately, it may be possible to achieve “pseudoindependence”. We will ﬁrst consider generating independent Uniform random variables on the interval [0, 1]. A Uniform random variable is a continuous random variable and so any real number between 0 and 1 is a possible outcome of the random variable. However, real numbers cannot be represented exactly on a computer but instead are represented as ﬂoating point numbers. Since ﬂoating point numbers are countable, it is impossible to generate random variables having exactly a Uniform distribution. What is typically done is to generate a random variable R that is uniformly distributed on the set {0, 1, 2, · · · , N − 1} and deﬁne U = R/N . If N is large then U will be approximately Uniform (since P (R = k) = 1/N for k = 0, · · · , N − 1). There are a number of methods for generating the integer random variable r1 , · · · , rn so that the resulting pseudorandom variables are approximately independent. Perhaps the most commonly used generator is the linear congruential generator. We deﬁne our c 2000 by Chapman & Hall/CRC
sequence r1 , · · · , rn via the recursive relationship ri = mod(λri−1 + α, N ) where mod(a, b) is the remainder when a is divided by b (for example, mod(5, 3) = 2) and λ, α are integers. (The number r0 used to start the generator is sometimes called the seed.) It is easy to see that the sequence of ri ’s will repeat after P ≤ N steps (that is, ri = ri+P ) with P called the period of the generator. Clearly, we would like the period of the generator to be as long as possible. However, if λ and α are chosen appropriately then the period P = N and the integers r1 , r2 , · · · , rn will behave more orless like outcomes of independent random variables R1 , · · · , Rn when n is much smaller than N . (Hence u1 , · · · , un (where ui = ri /N ) will behave like outcomes of independent Uniform random variables.) Finding good linear congruential generators is much more complicated than ﬁnding a generator with period P = N (which is quite easy to achieve); in fact, linear congruential and other Uniform generators should be subjected to a battery of tests to ensure that the dependence in U1 , · · · , Un is minimal. EXAMPLE 2.23: One reliable linear congruential generator uses λ = 25211, α = 0 and N = 215 = 32768. The period of this generator is N . Using the seed R0 = 29559, we obtain U1 U2 U3 U4 U5
= = = = =
0.06387329 0.30953979 0.80776978 0.68380737 0.46768188
and so on. An example of a bad linear congruential generator is the socalled “randu” generator that was ﬁrst introduced by IBM in 1963 and was widely used in the 1960s and 1970s. This generator uses λ = 65539, α = 0 and N = 231 = 2147483648. This generator has period N but produces U1 , U2 , · · · that are fairly strongly dependent even though the dependence is not obvious. Figure 2.4 shows a plot of U3i−1 versus U3i−2 + 0.1 U3i (for i = 1, · · · , 1000); note that the points fall along diagonal lines, which would not happen if the Ui ’s were truly independent. ✸ c 2000 by Chapman & Hall/CRC
1.0 0.8 0.6 0.4 0.2 0.0
•
••••• ••••
•• •
• •••• •••
•• • •• •• • • •••• •••• ••• •••••• •• • ••••• • • • •••• • ••• • ••• ••• • • • • • • • • • • • •• • •• •• • •••• •••• • •• •• •• •• ••• •••• • ••• ••• •• ••• •••• •• • •• ••• ••••• • •• •••• • • • • • • • • •• • • • • •• • • • • •• • • • • • • • • • • • • • • • • • • •• •• ••• ••••• ••• •• ••• •••• ••• • •• •••• •• •••• •• •••• •• ••••• •••• • • •••••• •• • • •••• •••• •••• • • • • • • • • • • • •• •• •• • •• •••• •• ••• • ••• •• ••• •••• ••••• •••• • ••• •••• •• ••• • ••• ••••• • • • • • • • • • • • • • • • • • • • •• • • ••• • ••• •• ••• ••• ••• • ••••• •• •• •••• ••• ••• • • ••• ••••• •• ••• •• •••• ••• ••• • • • • • • • •• • • • • • • • • • • • • • • • •• •• • •• • • •• ••••• ••••• •• • • • • • • • • • • • • • • • • • • • • •• •• •• • •• •• ••• •••• ••• ••• •••• ••• •• •••• • •• •• • ••• •••• •• ••••• •••• ••• • • • • • • • • • • • • • • • • • • • • • • • •• • ••• •• •••• ••• ••• • •• •• •••••• •••• •• ••• ••• •••• •• ••• •• •••• • ••• ••• •• ••••• •• • • •• • •• • •• • • • • • • • • • • • • • • • • • •• •••• •• •• •• •••• •• •• ••• •• •• • •• •• •••• • •• •• • ••• ••
0.0
• ••••• •••
•• ••••
••• •••• •••••
0.2
0.4
0.6
0.8
1.0
Figure 2.4 Plot of U3i−1 versus U3i−2 + 0.1U3i for the “randu” generator; the initial seed r0 = 12345.
We now move to the more general problem. Suppose we want to generate outcomes of independent random variables X1 , · · · , Xn with common distribution function F . Several techniques exist for generating random variables. These techniques generally exploit known properties of the distributions. However, generation of Uniform random variables is inevitably involved (to some degree) in virtually all methods as these random variables are typically manipulated to produce X1 , · · · , Xn . Thus given a stream of independent Uniform random variables U1 , U2 , · · ·, Xi and Xj (i = j) will be independent provided they do not depend on the same Uniform random variables. Perhaps the most obvious method for generating random variables with a distribution function F is the inverse method. Given a random variable U that has a Uniform distribution on [0, 1], we can transform U to obtain a random variable X with an arbitrary distribution function F . This can be done by using the inverse distribution function F −1 (t) = inf{x : F (x) ≥ t}. c 2000 by Chapman & Hall/CRC
We showed in Chapter 1 that if X = F −1 (U ) then the distribution function of X is F . (Since U and 1 − U have the same distribution, F −1 (1 − U ) will also have distribution function F .) The inverse method is easy to implement provided the F −1 (t) is easily computable as in the following example. EXAMPLE 2.24: Suppose we want to generate Exponential random variables with parameter λ; for this distribution, we have F (x) = λ exp(−λx) for x ≥ 0. Solving the equation F (F −1 (t)) = t, we get 1 F −1 (t) = − ln(1 − t). λ Thus X = −λ−1 ln(1 − U ) has an Exponential distribution. Alternatively, we could also deﬁne X = −λ−1 ln(U ). ✸ The inverse method becomes less feasible when F −1 is diﬃcult to compute. In such cases, there are a variety of other methods that may be used; some of these will be investigated in the exercises. 2.7 Problems and complements 2.1: Suppose that X and Y are independent Geometric random variables with frequency function f (x) = θ(1 − θ)x
for x = 0, 1, 2, · · ·
(a) Show that Z = X + Y has a Negative Binomial distribution and identify the parameters of Z. (b) Extend the result of part (a): If X1 , · · · , Xn are i.i.d. Geometric random variables, show that S = X1 + X2 + · · · + Xn has a Negative Binomial distribution and identify the parameters of S. (Hint: Use mathematical induction or, alternatively, the result of Problem 2.5 below.) 2.2: (a) Suppose that Y is a nonnegative random variable. Show that X + Y is always stochastically greater than X (for any random variable X). (Hint: See Problem 1.10; you need to show that P (X + Y > x) ≥ P (X > x) for all x.) (b) Suppose that X is stochastically greater than Y and suppose that X and Y are deﬁned on the same sample space. Show that X −Y is not necessarily stochastically greater than 0. c 2000 by Chapman & Hall/CRC
2.3: If f1 (x), · · · , fk (x) are density (frequency) functions then g(x) = p1 f1 (x) + · · · + pk fk (x) is also a density (frequency) function provided that pi ≥ 0 (i = 1, · · · , k) and p1 + · · · + pk = 1. We can think of sampling from g(x) as ﬁrst sampling a discrete random variable Y taking values 1 through k with probabilities p1 , · · · , pk and then, conditional on Y = i, sampling from fi (x). The distribution whose density or frequency function is g(x) is called a mixture distribution. (a) Suppose that X has frequency function g(x). Show that P (Y = iX = x) =
pi fi (x) g(x)
provided that g(x) > 0. (b) Suppose that X has density function g(x). Show that we can reasonably deﬁne P (Y = iX = x) =
pi fi (x) g(x)
in the sense that P (Yi = i) = E[P (Y = iX)]. 2.4: Suppose that X1 , · · · , X5 are i.i.d. random variables such that P (Xi = 0) = 0.5 and P (Xi > x) = 0.5 exp(−x)
for x ≥ 0.
Deﬁne S = X1 + · · · + X5 . (a) Find the distribution function of S. (Hint: Note that the distribution of S can be written as Binomial mixture of Gamma distributions.) (b) Evaluate P (Y ≤ 5). 2.5: Mixture distributions can be extended in the following way. Suppose that f (x; θ) is a density or frequency function where θ lies in some set Θ ⊂ R. Let p(θ) be a density function on Θ and deﬁne g(x) = f (x; θ)p(θ) dθ. Θ
Then g(x) is itself a density or frequency function. As before, we can view sampling from g(x) as ﬁrst sampling from p(θ) and then given θ, sampling from f (x; θ). c 2000 by Chapman & Hall/CRC
(a) Suppose that X has the mixture density or frequency function g(x). Show that E(X) = E[E(Xθ)] and Var(X) = Var[E(Xθ)] + E[Var(Xθ)] where E(Xθ) and Var(Xθ) are the mean and variance of a random variable with density or frequency function f (x; θ). (b) The Negative Binomial distribution introduced in Example 1.12 can be obtained as a Gamma mixture of Poisson distributions. Let f (x; λ) be a Poisson frequency function with mean λ and p(λ) be a Gamma distribution with mean µ and variance µ2 /α. Show that the mixture distribution has frequency function
α
x α µ Γ(x + α) g(x) = x!Γ(α) α+µ α+µ for x = 0, 1, 2, · · ·. Note that this form of the Negative Binomial is richer than the form given in Example 1.12. (c) Suppose that X has a Negative Binomial distribution as given in part (b). Find the mean and variance of X. (Hint: Use the approach outlined in part (a).) (d) Show that the moment generating function of the Negative Binomial distribution in (b) is
m(t) =
α α + µ(1 − exp(t))
α
for t < ln(1 + α/µ). 2.6: A distribution F is said to be inﬁnitely divisible if for each n = 1, 2, · · · there exist i.i.d. random variables X1 , · · · , Xn such that X1 + · · · + Xn ∼ F . For example, a Normal distribution with mean µ and variance σ 2 is inﬁnitely divisible; to see this, we take X1 , · · · , Xn to be i.i.d. Normal random variables with mean µ/n and variance σ 2 /n. (a) Suppose that the distribution F has moment generating function m(t). Show that F is inﬁnitely divisible if, and only if, [m(t)]1/n is a moment generating function for each n. (b) Show that all Poisson distributions are inﬁnitely divisible. (c) Show that all Negative Binomial distributions are inﬁnitely divisible. c 2000 by Chapman & Hall/CRC
2.7: Suppose that X1 , X2 , · · · are i.i.d. random variables with moment generating function m(t) = E[exp(tXi )]. Let N be a Poisson random variable (independent of the Xi ’s) with parameter λ and deﬁne the compound Poisson random variable S=
N
Xi
i=1
where S = 0 if N = 0. (a) Show that the moment generating function of S is E[exp(tS)] = exp[λ(m(t) − 1)]. (b) Suppose that the Xi ’s are Exponential with E(Xi ) = 1 and λ = 5. Evaluate P (S > 5). (Hint: First evaluate P (S > 5N ) using the result of Problem 1.14.) 2.8: Suppose that X1 , · · · , Xn are independent nonnegative integervalued random variables with probability generating functions p1 (t), · · · , pn (t). Show that the probability generating function of S = X1 + · · · + Xn is p(t) = p1 (t) × · · · × pn (t). 2.9: Consider the experiment in Problem 1.3 where a coin is tossed an inﬁnite number of times where the probability of heads on the kth toss is (1/2)k . Deﬁne X to be the number of heads observed in the experiment. (a) Show that the probability generating function of X is p(t) =
∞
k=1
1−t 1− k . 2
(Hint: Think of X as a sum of independent Bernoulli random variables.) (b) Use the result of part (a) to evaluate P (X = x) for x = 0, · · · , 5. 2.10: Consider the following method (known as the rejection method) for generating random variables with a density f (x). Suppose that γ(x) be a function such that γ(x) ≥ f (x) for all x, and ∞
−∞
γ(x) dx = α < ∞.
Then g(x) = γ(x)/α is a probability density function. Suppose we generate a random variable X by the following algorithm: c 2000 by Chapman & Hall/CRC
• Generate a random variable T with density function g(x). • Generate a random variable U ∼ Unif(0, 1), independent of T . If U ≤ f (T )/γ(T ) then set X = T ; if U > f (T )/γ(T ) then repeat steps I and II. (a) Show that the generated random variable X has density f (x). (Hint: you need to evaluate P (T ≤ xU ≤ f (T )/γ(T )).) (b) Show that the number of “rejections” before X is generated has a Geometric distribution. Give an expression for the parameter of this distribution. (c) Show that the rejection method also works if we want to generate from a joint density f (x). (In this case, U ∼ Unif(0, 1) as before but now T is a random vector with density g(x).) 2.11: Suppose we want to generate random variables with a Cauchy distribution. As an alternative to the method described in Problem 1.24, we can generate independent random variables V and W where P (V = 1) = P (V = −1) = 1/2 and W has density 2 g(x) = for x ≤ 1. π(1 + x2 ) (W can be generated by using the rejection method in Problem 2.10) Then we deﬁne X = W V ; show that X has a Cauchy distribution. 2.12: Suppose that X and Y are independent Uniform random variables on [0, 1]. (a) Find the density function of X + Y . (b) Find the density function of XY . 2.13: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on [0, 1]. Deﬁne Sn = (X1 + · · · + Xn ) mod 1; Sn is simply the “decimal” part of X1 + · · · + Xn . (a) Show that Sn = (Sn−1 + Xn ) mod 1 for all n ≥ 2. (b) Show that Sn ∼ Unif(0, 1) for all n ≥ 1. (Hint: Prove the result for n = 2 and apply the result of part (a).) 2.14: Suppose that X and Y are independent Exponential random variables with parameter λ. Deﬁne Z = X − Y . (a) Show that the density function of Z is fZ (x) = c 2000 by Chapman & Hall/CRC
λ exp(−λx). 2
(Hint: Evaluate P (Z ≤ x) for x < 0 and x > 0.) (b) Find the moment generating function of Z. (Hint: use the fact that Z = X − Y for independent Exponential random variables. 2.15: Suppose that X1 , · · · , Xn are independent, nonnegative continuous random variables where Xi has hazard function λi (x) (i = 1, · · · , n). (a) If U = min(X1 , · · · , Xn ), show that the hazard function of U is λU (x) = λ1 (x) + · · · + λn (x). (b) If V = max(X1 , · · · , Xn ), show that the hazard function of V satisﬁes λV (x) ≤ min(λ1 (x), · · · , λn (x)). (c) Show that the result of (b) holds even if the Xi ’s are not independent. 2.16: Jensen’s inequality (see Problem 1.11) can be extended to convex functions in higher dimensions. g(x) is a convex function if g(tx + (1 − t)y) ≤ t g(x) + (1 − t)g(y) for 0 ≤ t ≤ 1. (a) Let X be a random vector with welldeﬁned expected value E(X). Show that E[g(X)] ≥ g(E(X)) for any convex function g. (Hint: Repeat the approach used in Problem 1.11 making appropriate changes.) (b) Let g(x) = max(x1 , · · · , xk ). Show that g is a convex function and so E[max(X1 , · · · , Xk )] ≥ max(E(X1 ), · · · , E(Xk )) for any random variables X1 , · · · , Xk . 2.17: Suppose that X and Y are random variables such that both E(X 2 ) and E(Y 2 ) are ﬁnite. Deﬁne g(t) = E[(Y + tX)2 ]. (a) Show that g(t) is minimized at t = −E(XY )/E(X 2 ). (b) Show that [E(XY )]2 ≤ E(X 2 )E(Y 2 ); this is called the CauchySchwarz inequality. (Hint: Note that g(t) ≥ 0 for all t.) (c) Use part (b) to show that Corr(X, Y ) ≤ 1. 2.18: Suppose that R and U are independent continuous random variables where U has a Uniform distribution on [0, 1] and R has the density function fR (x) = x exp(−x2 /2) c 2000 by Chapman & Hall/CRC
for x ≥ 0.
(a) Show that R2 has an Exponential distribution. (b) Deﬁne X = R cos(2πU ) and Y = R sin(2πU ). Show that X and Y are independent standard Normal random variables. (c) Suggest a method for generating Normal random variables based on the results in part (a) and (b). 2.19: Suppose that X and Y are independent random variables with X discrete and Y continuous. Deﬁne Z = X + Y . (a) Show that Z is a continuous random variable with P (Z ≤ z) =
P (Y ≤ z − x)P (X = x).
x
(b) If Y has a density function fY (y), show that the density of Z is fZ (z) = fY (z − x)fX (x) x
where fX (x) is the frequency function of X. 2.20: Suppose that X and Y are independent Normal random variables each with mean µ and variance σ 2 and let U = Y and V = X(1 + Y ). Evaluate the following: (a) E(V U = u). (b) Var(V U = u). (c) Var(V ). (d) Cov(U, V ) and Corr(U, V ). (Hint: You do not need to evaluate any integrals.) 2.21: (a) Show that Cov(X, Y ) = E[Cov(X, Y Z)] + Cov(E(XZ), E(Y Z)). (Hint: Follow the proof for the similar result involving variances.) (b) Suppose that X1 , X2 , · · · be i.i.d. Exponential random variables with parameter 1 and take N1 , N2 to be independent Poisson random variables with parameters λ1 , λ2 that are independent of the Xi ’s. Deﬁne compound Poisson random variables S1 =
N1 i=1
c 2000 by Chapman & Hall/CRC
Xi
S2 =
N2
Xi
i=1
and evaluate Cov(S1 , S2 ) and Corr(S1 , S2 ). When is this correlation maximized? 2.22: Suppose that X = (X1 , · · · , Xk ) is a random vector and deﬁne the (joint) moment generating function of X; m(t) = E[exp(t1 X1 + · · · + tk Xk )]; we say that this exists if m(t) < ∞ for t < b where b > 0. (a) Show that +
E(Xik Xj ) =
+ ∂ k+ m(t)++ k ∂ti ∂tj t=0
for k, : = 0, 1, 2, · · · (b) Show that
+
+ ∂2 ln m(t)++ . Cov(Xi , Xj ) = ∂ti ∂tj t=0
(c) Suppose that X 1 , · · · , X n are independent random vectors with moment generating functions m1 (t), · · · , mn (t), respectively. Show that the moment generating function of S = X 1 + · · · + X n is mS (t) = m1 (t) × · · · × mn (t). 2.23: The mean residual life function r(t) of a nonnegative random variable X is deﬁned to be r(t) = E(X − tX ≥ t). (r(t) would be of interest, for example, to a life insurance company.) (a) Suppose that F is the distribution function of X. Show that ∞ 1 r(t) = (1 − F (x)) dx. 1 − F (t) t (b) Show that r(t) is constant if, and only if, X has an Exponential distribution. (c) Show that 2
∞
E(X ) = 2 0 c 2000 by Chapman & Hall/CRC
r(t)(1 − F (t)) dt.
(Hint: Show that E(X 2 ) can be written as ∞ t
2 0
0
(1 − F (t)) ds dt
and change the order of integration.) (d) Suppose that X has a density function f (x) that is diﬀerentiable and f (x) > 0 for x > 0. Show that
lim r(t) = lim
t→∞
t→∞
f (t) − . f (t)
(e) Suppose that X has a Gamma distribution: f (x) =
1 α α−1 exp(−λx) λ x Γ(α)
for x > 0.
Evaluate the limit in part (c) for this distribution. Give an interpretation of this result. 2.24: Suppose that X is a nonnegative random variable with mean µ > 0 and variance σ 2 < ∞. The coeﬃcient of variation of X is deﬁned to be CV(X) = σ/µ. (a) Suppose that X and Y are independent nonnegative random variables with CV(X) and CV(Y ) ﬁnite. Show that CV(X + Y ) ≤ CV(X) + CV(Y ). (b) Deﬁne r(t) = E(X − tX ≥ t) to be the mean residual life function of X. Show that CV(X) ≤ 1 if r(t) ≤ r(0) = E(X) and CV(X) ≥ 1 if r(t) ≥ r(0). (Hint: Note that CV(X) ≤ 1 if, and only if, E(X 2 )/[E(X)]2 ≤ 2 and use the result of Problem 2.23(c).) 2.25: Suppose that X1 , · · · , Xn are i.i.d. continuous random variables with distribution function F (x) and density function f (x); let X(1) < X(2) < · · · < X(n) be the order statistics. (a) Show that the distribution function of X(k) is Gk (x) =
n j=k
n F (x)j (1 − F (x))n−j . j
(Hint: Let Yi = I(Xi ≤ x) and deﬁne S = Y1 + · · · + Yn ; S has a Binomial distribution and P (X(k) ≤ x) = P (S ≥ k).) c 2000 by Chapman & Hall/CRC
(b) Show that the density function of X(k) is gk (x) =
n! F (x)k−1 (1 − F (x))n−k f (x). (n − k)!(k − 1)!
(Hint: Assume that F (x) = f (x) and diﬀerentiate Gk (x).) 2.26: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. Let X(1) < · · · < X(n) be the order statistics and deﬁne Y1 Y2 Y3 .. . Yn
= = = .. . =
nX(1) (n − 1)(X(2) − X(1) ) (n − 2)(X(3) − X(2) ) .. . X(n) − X(n−1) .
Show that Y1 , · · · , Yn are i.i.d. Exponential random variables with parameter λ. (Note that the “Jacobian” matrix here is triangular and so the Jacobian itself can be computed as the product of the diagonal elements.) 2.27: Suppose that X1 , · · · , Xn+1 be i.i.d. Exponential random variables with parameter λ and deﬁne Uk =
k 1 Xi T i=1
for k = 1, · · · , n
where T = X1 + · · · + Xn+1 . (a) Find the joint density of (U1 , · · · , Un , T ). (Note that 0 < U1 < U2 < · · · < Un < 1.) (b) Show that the joint distribution of (U1 , · · · , Un ) is exactly the same as the joint distribution of the order statistics of an i.i.d. sample of n observations from a Uniform distribution on [0, 1]. 2.28: A discrete random vector X is said to have a Multinomial distribution if its joint frequency function is f (x) =
n! θx1 × · · · × θkxk x1 ! × · · · × xk ! 1
for nonnegative integers x1 , · · · , xk with x1 + · · · + xk = n and nonnegative parameters θ1 , · · · , θk with θ1 + · · · + θk = 1. (We will write X ∼ Mult(n, θ) where θ = (θ1 , · · · , θk ).) c 2000 by Chapman & Hall/CRC
(a) Show that the marginal distribution of Xi is Binomial with parameters n and θi . (b) Show that E(X) = nθ and
Cov(X) = n
θ1 (1 − θ1 ) −θ1 θ2 ··· −θ1 θk −θ1 θ2 θ2 (1 − θ2 ) · · · −θ2 θk .. .. .. .. . . . . −θ1 θk −θ2 θk · · · θk (1 − θk )
.
(c) Suppose that X 1 , · · · , X n are independent Multinomial random vectors with X i ∼ Mult(ni , θ) (i = 1, · · · , n). Show n that n
X i ∼ Mult
i=1
ni , θ .
i=1
(Hint: For parts (b) and (c), evaluate the moment generating function of a Multinomial random vector using the Multinomial Theorem.) 2.29: Suppose that X and Y are independent Exponential random variables with parameters λ and µ respectively. Deﬁne random variables
T = min(X, Y )
and ∆ =
1 if X < Y 0 otherwise.
Note that T has a continuous distribution while ∆ is discrete. (This is an example of type I censoring in reliability or survival analysis.) (a) Find the density of T and the frequency function of ∆. (b) Find the joint distribution function of (T, ∆). 2.30: Suppose that X has a Beta distribution with parameters α and β (see Example 2.5). (a) Show that for r > 0, E(X r ) =
Γ(α + r)Γ(α + β) . Γ(α)Γ(α + β + r)
(b) Use part (a) and properties of Γ(x) to evaluate E(X) and Var(X). 2.31: Suppose that X has a Beta distribution with parameters α and β. c 2000 by Chapman & Hall/CRC
(a) Find the density function of Y = (1 − X)−1 . (b) Suppose that α = m/2 and β = n/2 and deﬁne Y as in part (a). Using the deﬁnition of the F distribution, show that nY /m ∼ F(m, n). (Hint: Take U ∼ χ2 (m) and V ∼ χ2 (n) to be independent random variables. Show that U/(U + V ) has a Beta distribution and then apply part (a).) 2.32: Suppose that X ∼ Np (0, C) where C −1 exists. Show that X T C −1 X ∼ χ2 (p). (Hint: Write C = OT ΛO where O is an orthogonal matrix and Λ is a diagonal matrix whose entries are the eigenvalues of C; then deﬁne C 1/2 = OT Λ1/2 O to be a symmetric root of C.) 2.33: Suppose that X ∼ χ2 (n). (a) Show that E(X r ) = 2−r Γ(r + n/2)/Γ(n/2) if r > −n/2. (b) Using part (a), show that E(X) = n and Var(X) = 2n. 2.34: Suppose that T ∼ T (n). Show that (a) E(T r ) < ∞ if, and only if, n > r ≥ 0. (b) E(T ) = 0 if n > 1; Var(T ) = n/(n − 2) if n > 2. (Hint: You don’t need to do any integration here. Write T = Z/ V /n where Z ∼ N (0, 1) and V ∼ χ2 (n) are independent.) 2.35: Suppose that W ∼ F(m, n). Show that
r
E(W ) =
n m
r
Γ(r + m/2)Γ(−r + n/2) Γ(m/2)Γ(n/2)
if −m/2 < r < n/2. 2.36: Suppose that X is a Normal random variable with mean θ and variance 1 and deﬁne Y = X 2 . (a) Show that the density of Y is
1 1 √ √ fY (y) = √ exp (y + θ2 ) (exp(θ y) + exp(−θ y)) 2 2πy 2 for y > 0. (Y is said to have a noncentral χ2 distribution with 1 degree of freedom and noncentrality parameter θ2 .) (b) Show that the density of Y can be written as fY (y) =
∞ exp(−θ2 /2)(θ2 /2)k k=0
c 2000 by Chapman & Hall/CRC
k!
f2k+1 (y)
where f2k+1 (y) is the density function of a χ2 random variable √ with 2k + 1 degrees of freedom. (Hint: Expand exp(θ y) and √ exp(−θ y) as power series; note that the odd terms in the two expansions will cancel each other out.) 2.37: Suppose that X ∼ Nn (µ, I); the elements of X are independent Normal random variables with variances equal to 1. (a) Suppose that O is an orthogonal matrix whose ﬁrst row is µT /µ and let Y = OX. Show that E(Y1 ) = µ and E(Yk ) = 0 for k ≥ 2. (b) Using part (a), show that the distribution of X2 is the same as that of Y 2 and hence depends on µ only through its norm µ. (c) Let θ2 = µ2 . Show that the density of V = X2 is fV (x) =
∞ exp(−θ2 /2)(θ2 /2)k
k!
k=0
f2k+n (x)
where f2k+n (x) is the density function of a χ2 random variable with 2k + n degrees of freedom. (V has a noncentral χ2 distribution with n degrees of freedom and noncentrality parameter θ2 .) 2.38: Consider a marked Poisson process similar to that given in Example 2.22 such that the call starting times arrive as a homogeneous Poisson process (with rate λ calls/minute) on the positive real line. Assume that the call lengths are continuous random variables with density function f (x) and deﬁne N (t) to be the number of calls active at time t for t ≥ 0. Show that N (t) ∼ Pois(µ(t)) where t
µ(t) = λ 0
(1 − F (s)) ds
and F is the distribution function of the call lengths. 2.39: Consider the marked Poisson process in Example 2.22 where the call starting times arrive as a homogeneous Poisson process (with rate λ calls/minute) on the entire real line and the call lengths are continuous random variables with density function f (x). In Example 2.22, we showed that the distribution of N (t) is independent of t. (a) Show that for any r, Cov[N (t), N (t + r)] c 2000 by Chapman & Hall/CRC
∞
= λ
x f (x) dx
r
,
= λ r(1 − F (r) +
∞ r

(1 − F (x)) dx
and hence is independent of t and depends only on r. (Hint: Assume that r > 0. Then N (t) =
∞
I(Si ≤ t, t ≤ Si + Xi < t + r)
i=1
+ N (t + r) =
∞
∞
I(Si ≤ t, Si + Xi ≥ t + r)
i=1
I(Si ≤ t, Si + Xi ≥ t + r)
i=1
+
∞
I(t < Si ≤ t + r, Si + Xi ≥ t + r)
i=1
and use the independence of Poisson processes on disjoint sets.) (b) Suppose that the call lengths are Exponential random variables with mean µ. Evaluate Cov[N (t), N (t + r)]. (This is called the autocovariance function of N (t).) (c) Suppose that the call lengths have a density function f (x) = αx−α−1
for x ≥ 1.
Show that E(Xi ) < ∞ if, and only if, α > 1 and evaluate Cov[N (t), N (t + r)] in this case. (d) Compare the autocovariance functions obtained in parts (b) and (c). For which distribution does Cov[N (t), N (t + r)] decay to 0 more slowly as r → ∞?
c 2000 by Chapman & Hall/CRC
CHAPTER 3
Convergence of Random Variables 3.1 Introduction In probability and statistics, it is often necessary to consider the distribution of a random variable that is itself a function of several random variables, for example, Y = g(X1 , · · · , Xn ); a simple example is the sample mean of random variables X1 , · · · , Xn . Unfortunately, ﬁnding the distribution exactly is often very diﬃcult or very timeconsuming even if the joint distribution of the random variables is known exactly. In other cases, we may have only partial information about the joint distribution of X1 , · · · , Xn in which case it is impossible to determine the distribution of Y . However, when n is large, it may be possible to obtain approximations to the distribution of Y even when only partial information about X1 , · · · , Xn is available; in many cases, these approximations can be remarkably accurate. The standard approach to approximating a distribution function is to consider the distribution function as part of an inﬁnite sequence of distribution functions; we then try to ﬁnd a “limiting” distribution for the sequence and use that limiting distribution to approximate the distribution of the random variable in question. This approach, of course, is very common in mathematics. For example, if n is large compared to x, one might approximate (1 + x/n)n by exp(x) since
lim
n→∞
x 1+ n
n
= exp(x).
(However, this approximation may be very poor if x/n is not close to 0.) A more interesting example is Stirling’s approximation, which is used to approximate n! for large values of n: √ n! ≈ 2π exp(−n)nn+1/2 = s(n) where the approximation holds in the sense that n!/s(n) → 1 as c 2000 by Chapman & Hall/CRC
Table 3.1 Comparison of n! and its Stirling approximation s(n).
n
n!
s(n)
1 2 3 4 5 6
1 2 6 24 120 720
0.92 1.92 5.84 23.51 118.02 710.08
n → ∞. In fact, Stirling’s approximation is not too bad even for small n as Table 3.1 indicates. In a sense, Stirling’s approximation shows that asymptotic approximations can be useful in a more general context. In statistical practice, asymptotic approximations (typically justiﬁed for large sample sizes) are very commonly used even in situations where the sample size is small. Of course, it is not always clear that the use of such approximations is warranted but nonetheless there is a suﬃciently rich set of examples where it is warranted to make the study of convergence of random variables worthwhile. To motivate the notion of convergence of random variables, consider the following example. Suppose that X1 , · · · , Xn are i.i.d. random variables with mean µ and variance σ 2 and deﬁne n ¯n = 1 Xi X n i=1
to be their sample mean; we would like to look at the behaviour ¯ n when n is large. First of all, it seems of the distribution of X ¯ reasonable that Xn will be close to µ if n is suﬃciently large; that ¯ n − µ should have a distribution that, for is, the random variable X large n, is concentrated around 0 or, more precisely, ¯ n − µ ≤ ] ≈ 1 P [X ¯ n ) = σ 2 /n → 0 as n → ∞.) when is small. (Note that Var(X This latter observation is, however, not terribly informative about ¯ n . However, it is also possible to look at the the distribution of X ¯ n and µ on a “magniﬁed” scale; we do this diﬀerence between X c 2000 by Chapman & Hall/CRC
¯ n − µ by √n so that the mean and by multiplying the diﬀerence X variance are constant. Thus deﬁne √ ¯ n − µ) Zn = n(X and note that E(Zn ) = 0 and Var(Zn ) = σ 2 . We can now consider the behaviour of the distribution function of Zn as n increases. If this sequence of distribution functions has a limit (in some sense) then we can use the limiting distribution function to approximate ¯ n ). For example, if the distribution function of Zn (and hence of X we have √ ¯ n − µ) ≤ x ≈ F0 (x) P (Zn ≤ x) = P n(X then
√ √ ¯ n ≤ y) = P ¯ n − µ) ≤ n(y − µ) P (X n(X √ ≈ F0 n(y − µ)
provided that n is suﬃciently large to make the approximation valid. 3.2 Convergence in probability and distribution In this section, we will consider two diﬀerent types of convergence for sequences of random variables, convergence in probability and convergence in distribution. DEFINITION. Let {Xn }, X be random variables. Then {Xn } converges in probability to X as n → ∞ (Xn →p X) if for each > 0, lim P (Xn − X > ) = 0. n→∞
If Xn →p X then for large n we have that Xn ≈ X with probability close to 1. Frequently, the limiting random variable X is a constant; Xn →p θ (a constant) means that for large n there is almost no variation in the random variable Xn . (A stronger form of convergence, convergence with probability 1, is discussed in section 3.7.) DEFINITION. Let {Xn }, X be random variables. Then {Xn } converges in distribution to X as n → ∞ (Xn →d X) if lim P (Xn ≤ x) = P (X ≤ x) = F (x)
n→∞
for each continuity point of the distribution function F (x). c 2000 by Chapman & Hall/CRC
It is important to remember that Xn →d X implies convergence of distribution functions and not of the random variables themselves. For this reason, it is often convenient to replace Xn →d X by Xn →d F where F is the distribution function of X, that is, the limiting distribution; for example, Xn →d N (0, σ 2 ) means that {Xn } converges in distribution to a random variable that has a Normal distribution (with mean 0 and variance σ 2 ). If Xn →d X then for suﬃciently large n we can approximate the distribution function of Xn by that of X; thus, convergence in distribution is potentially useful for approximating the distribution function of a random variable. However, the statement Xn →d X does not say how large n must be in order for the approximation to be practically useful. To answer this question, we typically need a further result dealing explicitly with the approximation error as a function of n. EXAMPLE 3.1: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on the interval [0, 1] and deﬁne Mn = max(X1 , · · · , Xn ). Intuitively, Mn should be approximately 1 for large n. We will ﬁrst show that Mn →p 1 and then ﬁnd the limiting distribution of n(1 − Mn ). The distribution function of Mn is Fn (x) = xn
for
0 ≤ x ≤ 1.
Thus for 0 < < 1, P (Mn − 1 > ) = P (Mn < 1 − ) = (1 − )n → 0 as n → ∞ since 1 −  < 1. To ﬁnd the limiting distribution of n(1 − Mn ), note that P (n(1 − Mn ) ≤ x)
P (Mn ≥ 1 − x/n)
x n = 1− 1− n → 1 − exp(−x) =
as n → ∞ for x ≥ 0. Thus n(1 − Mn ) has a limiting Exponential distribution with parameter 1. In this example, of course, there is no real advantage in knowing the limiting distribution of n(1 − Mn ) as its exact distribution is known. ✸ c 2000 by Chapman & Hall/CRC
EXAMPLE 3.2: Suppose that X1 , · · · , Xn are i.i.d. random variables with 1 P (Xi = j) = for j = 0, 1, 2, · · · , 9 10 and deﬁne n Xk . Un = 10k k=1 Un can be thought of as the ﬁrst n digits of a decimal representation of a number between 0 and 1 (Un = 0.X1 X2 X3 X4 · · · Xn ). It turns out that Un tends in distribution to a Uniform on the interval [0, 1]. To see this, note that each outcome of (X1 , · · · , Xn ) produces a unique value of Un ; these possible values are j/10n for j = 0, 1, 2, · · · , 10n − 1, and so it follows that 1 for j = 0, 1, 2, · · · , 10n − 1. 10n If j/10n ≤ x < (j + 1)/10n then P (Un = j/10n ) =
P (Un ≤ x) = and so
j+1 10n
P (Un ≤ x) − x ≤ 10−n → 0
as n → ∞
and so P (Un ≤ x) → x for each x between 0 and 1.
✸
Some important results We noted above that convergence in probability deals with convergence of the random variables themselves while convergence in distribution deals with convergence of the distribution functions. The following result shows that convergence in probability is stronger than convergence in distribution unless the limiting random variable is a constant in which case the two are equivalent. THEOREM 3.1 Let {Xn }, X be random variables. (a) If Xn →p X then Xn →d X. (b) If Xn →d θ (a constant) then Xn →p θ. Proof. (a) Let x be a continuity point of the distribution function of X. Then for any > 0, P (Xn ≤ x) = P (Xn ≤ x, Xn − X ≤ ) +P (Xn ≤ x, Xn − X > ) ≤ P (X ≤ x + ) + P (Xn − X > ) c 2000 by Chapman & Hall/CRC
where the latter inequality follows since [Xn ≤ x, Xn − X ≤ ] implies [X ≤ x + ]. Similarly, P (X ≤ x − ) ≤ P (Xn ≤ x) + P (Xn − X > ) and so P (Xn ≤ x) ≥ P (X ≤ x − ) − P (Xn − X > ). Thus putting the two inequalities for P (Xn ≤ x) together, we have P (X ≤ x − ) − P (Xn − X > ) ≤ P (Xn ≤ x) ≤ P (X ≤ x + ) + P (Xn − X > ). By hypothesis, P (Xn − X > ) → 0 as n → ∞ for any > 0. Moreover, since x is a continuity point of the distribution function of X, P (X ≤ x ± ) can be made arbitrarily close to P (X ≤ x) by taking close to 0. Hence, lim P (Xn ≤ x) = P (X ≤ x).
n→∞
(b) Deﬁne F (x) to be the distribution function of the degenerate random variable taking the single value θ; thus, F (x) = 0 for x < θ and F (x) = 1 for x ≥ θ. Note that F is continuous at all but one point. Then P (Xn − θ > ) = P (Xn > θ + ) + P (Xn < θ − ) ≤ 1 − P (Xn ≤ θ + ) + P (Xn ≤ θ − ). However, since Xn →d θ, it follows that P (Xn ≤ θ + ) → 1 and P (Xn ≤ θ − ) → 0 as n → ∞ and so P (Xn − θ > ) → 0. It is often diﬃcult (if not impossible) to verify the convergence of a sequence of random variables using simply its deﬁnition. Theorems 3.2, 3.3 and 3.4 are sometimes useful for showing convergence in such cases. THEOREM 3.2 (Continuous Mapping Theorem) Suppose that g(x) is a continuous realvalued function. (a) If Xn →p X then g(Xn ) →p g(X). (b) If Xn →d X then g(Xn ) →d g(X). The proofs will not be given here. The proof of (a) is sketched as an exercise. The proof of (b) is somewhat more technical; however, if we further assume g to be strictly increasing or decreasing (so that g has an inverse function), a simple proof of (b) can c 2000 by Chapman & Hall/CRC
be given. (Also see Example 3.16 for a simple proof assuming more technical machinery.) The assumption of continuity can also be relaxed somewhat. For example, Theorem 3.2 will hold if g has a ﬁnite or countable number of discontinuities provided that these discontinuity points are continuity points of the distribution function of X. For example, if Xn →d θ (a constant) and g(x) is continuous at x = θ then g(Xn ) →d g(θ). THEOREM 3.3 (Slutsky’s Theorem) Suppose that Xn →d X and Yn →p θ (a constant). Then (a) Xn + Yn →d X + θ. (b) Xn Yn →d θX. Proof. (a) Without loss of generality, let θ = 0. (If θ = 0 then Xn + Yn = (Xn + θ) + (Yn − θ) and Yn − θ →p 0.) Let x be a continuity point of the distribution function of X. Then P (Xn + Yn ≤ x) = P (Xn + Yn ≤ x, Yn  ≤ ) +P (Xn + Yn ≤ x, Yn  > ) ≤ P (Xn ≤ x + ) + P (Yn  > ). Also, P (Xn ≤ x − ) = P (Xn ≤ x − , Yn  ≤ ) +P (Xn ≤ x − , Yn  > ) ≤ P (Xn + Yn ≤ x) + P (Yn  > ) (since [Xn ≤ x − , Yn  ≤ ] implies [Xn + Yn ≤ x]). Hence, P (Xn ≤ x − ) − P (Yn  > ) ≤ P (Xn + Yn ≤ x) ≤ P (Xn ≤ x + ) + P (Yn  > ). Now take x ± to be continuity points of the distribution function of X. Then lim P (Xn ≤ x ± ) = P (X ≤ x ± )
n→∞
and the limit can be made arbitrarily close to P (X ≤ x) by taking to 0. Since P (Yn  > ) → 0 as n → ∞ the conclusion follows. (b) Again we will assume that θ = 0. (To see that it suﬃces to consider this single case, note that Xn Yn = Xn (Yn − θ) + θXn . Since θXn →d θX the conclusion will follow from part (a) if we show that Xn (Yn − θ) →p 0.) We need to show that Xn Yn →p 0. c 2000 by Chapman & Hall/CRC
Taking > 0 and M > 0, we have P (Xn Yn  > ) = P (Xn Yn  > , Yn  ≤ 1/M ) +P (Xn Yn  > , Yn  > 1/M ) ≤ P (Xn Yn  > , Yn  ≤ 1/M ) + P (Yn  > 1/M ) ≤ P (Xn  > M ) + P (Yn  > 1/M ). Since Yn →p 0, P (Yn  > 1/M ) → 0 as n → ∞ for any ﬁxed M > 0. Now take and M such that ±M are continuity points of the distribution function of X; then P (Xn  > M ) → P (X > M ) and the limit can be made arbitrarily close to 0 by making M suﬃciently large. Since Yn →p θ is equivalent to Yn →d θ when θ is a constant, we could replace “Yn →p θ” by “Yn →d θ” in the statement of Slutsky’s Theorem. We can also generalize this result as follows. Suppose that g(x, y) is a continuous function and that Xn →d X and Yn →p θ for some constant θ. Then it can be shown that g(Xn , Yn ) →d g(X, θ). In fact, this result is sometimes referred to as Slutsky’s Theorem with Theorem 3.3 a special case for g(x, y) = x+y and g(x, y) = xy. THEOREM 3.4 (The Delta Method) Suppose that an (Xn − θ) →d Z where θ is a constant and {an } is a sequence of constants with an ↑ ∞. If g(x) is a function with derivative g (θ) at x = θ then an (g(Xn ) − g(θ)) →d g (θ)Z. Proof. We’ll start by assuming that g is continuously diﬀerentiable at θ. First, note that Xn →p θ. (This follows from Slutsky’s Theorem.) By a Taylor series expansion of g(x) around x = θ, we have g(Xn ) = g(θ) + g (θn∗ )(Xn − θ) where θn∗ lies between Xn and θ; thus θn∗ − θ ≤ Xn − θ and so θn∗ →p θ. Since g (x) is continuous at x = θ, it follows that g (θn∗ ) →p g (θ). Now, an (g(Xn ) − g(θ))
= g (θn∗ )an (Xn − θ) →d g (θ)Z
by Slutsky’s Theorem. For the more general case (where g is not necessarily continuously diﬀerentiable at θ), note that g(Xn ) − g(θ) = g (θ)(Xn − θ) + Rn c 2000 by Chapman & Hall/CRC
where Rn /(Xn − θ) →p 0. Thus an Rn = an (Xn − θ)
Rn →p 0 an (Xn − θ)
and so the conclusion follows by Slutsky’s Theorem. A neater proof of the Delta Method is given in Example 3.17. Also note that if g (θ) = 0, we would have that an (g(Xn ) − g(θ)) →p 0. In this case, we may have akn (g(Xn ) − g(θ)) →d some V for some k ≥ 2; see Problem 3.10 for details. If Xn →d X (or Xn →p X), it is tempting to say that E(Xn ) → E(X); however, this statement is not true in general. For example, suppose that P (Xn = 0) = 1 − n−1 and P (Xn = n) = n−1 . Then Xn →p 0 but E(Xn ) = 1 for all n (and so converges to 1). To ensure convergence of moments, additional conditions are needed; these conditions eﬀectively bound the amount of probability mass in the distribution of Xn concentrated near ±∞ for large n. The following result deals with the simple case where the random variables {Xn } are uniformly bounded; that is, there exists a constant M such that P (Xn  ≤ M ) = 1 for all n. THEOREM 3.5 If Xn →d X and Xn  ≤ M (ﬁnite) then E(X) exists and E(Xn ) → E(X). Proof. For simplicity, assume that Xn is nonnegative for all n; the general result will follow by considering the positive and negative parts of Xn . From Chapter 1, we have that E(Xn ) − E(X)
= =
+ ∞ + + + + (P (Xn > x) − P (X > x)) dx++ + +0 + + M + + + (P (Xn > x) − P (X > x)) dx+ + + 0 +
(since P (Xn > M ) = P (X > M ) = 0)
≤
M
→ 0
0
P (Xn > x) − P (X > x) dx
since P (Xn > x) → P (X > x) for all but a countable number of x’s and the interval of integration is bounded. c 2000 by Chapman & Hall/CRC
3.3 Weak Law of Large Numbers An important result in probability theory is the Weak Law of Large Numbers (WLLN), which deals with the convergence of the sample mean to the population mean as the sample size increases. We start by considering the simple case where X1 , · · · , Xn are i.i.d. Bernoulli random variables with P (Xi = 1) = θ and P (Xi = 0) = 1 − θ so that E(Xi ) = θ. Deﬁne Sn = X1 + · · · + Xn , which has a Binomial distribution with parameters n and θ. We now consider the behaviour of Sn /n as n → ∞; Sn /n represents the proportion of 1’s in the n Bernoulli trials. Our intuition tells us that for large n, this proportion should be approximately equal to θ, the probability that any Xi = 1. Indeed, since the distribution of Sn /n is known, it is possible to show the following law of large numbers: Sn →p θ n as n → ∞. In general, the WLLN applies to any sequence of independent, identical distributed random variables whose mean exists. The result can be stated as follows: THEOREM 3.6 (Weak Law of Large Numbers) Suppose that X1 , X2 , · · · are i.i.d. random variables with E(Xi ) = µ (where E(Xi ) < ∞). Then n 1 ¯ Xi →p µ Xn = n i=1
as n → ∞. While this result certainly agrees with intuition, a rigorous proof of the result is certainly not obvious. However, before proving the WLLN, we will give a nontrivial application of it by proving that the sample median of i.i.d. random variables X1 , · · · , Xn converges in probability to the population median. EXAMPLE 3.3: Suppose that X1 , · · · , Xn are i.i.d. random variables with a distribution function F (x). Assume that the Xi ’s have a unique median µ (F (µ) = 1/2); in particular, this implies that for any > 0, F (µ + ) > 1/2 and F (µ − ) < 1/2. Let X(1) , · · · , X(n) be the order statistics of the Xi ’s and deﬁne Zn = X(mn ) where {mn } is a sequence of positive integers with mn /n → 1/2 as n → ∞. For example, we could take mn = n/2 if n c 2000 by Chapman & Hall/CRC
is even and mn = (n + 1)/2 if n is odd; in this case, Zn is essentially the sample median of the Xi ’s. We will show that Zn →p µ as n → ∞. Take > 0. Then we have
P (Zn > µ + ) = P and
P (Zn < µ − ) = P
n 1 mn I(Xi > µ + ) ≥ n i=1 n
n 1 n − mn I(Xi ≥ µ − ) ≤ n i=1 n
.
By the WLLN, we have n 1 I(Xi > µ + ) →p 1 − F (µ + ) < 1/2 n i=1
and
n 1 I(Xi > µ − ) →p 1 − F (µ − ) > 1/2. n i=1
Since mn /n →p 1/2, it follows that P (Zn > µ + ) → 0 and P (Zn < µ − ) → 0 as n → ∞ and so Zn →p µ. ✸ Proving the WLLN The key to proving the WLLN lies in ﬁnding a good bound for ¯ n − µ > ]; one such bound is Chebyshev’s inequality. P [X THEOREM 3.7 (Chebyshev’s inequality) Suppose that X is a random variable with E(X 2 ) < ∞. Then for any > 0, P [X > ] ≤
E(X 2 ) . 2
Proof. The key is to write X 2 = X 2 I(X ≤ ) + X 2 I(X > ). Then E(X 2 ) = E[X 2 I(X ≤ )] + E[X 2 I(X > )] ≥ E[X 2 I(X > )] ≥ 2 P (X > ) where the last inequality holds since X 2 ≥ 2 when X > and E[I(X > )] = P (X > ). c 2000 by Chapman & Hall/CRC
From the proof, it is quite easy to see that Chebyshev’s inequality remains valid if P [X > ] is replaced by P [X ≥ ]. Chebyshev’s inequality is primarily used as a tool for proving various convergence results for sequences of random variables; for example, if {Xn } is a sequence of random variables with E(Xn2 ) → 0 then Chebyshev’s inequality implies that Xn →p 0. However, Chebyshev’s inequality can also be used to give probability bounds for random variables. For example, let X be a random variable with mean µ and variance σ 2 . Then by Chebyshev’s inequality, we have E[(X − µ)2 ] 1 = 1 − 2. k2 σ2 k However, the bounds given by Chebyshev’s inequality are typically very crude and are seldom of any practical use. Chebyshev’s inequality can be also generalized in a number of ways; these generalizations are examined in Problem 3.8. We will now sketch the proof of the WLLN. First of all, we will assume that E(Xi2 ) < ∞. In this case, the WLLN follows trivially since (by Chebyshev’s inequality) P [X − µ ≤ kσ] ≥ 1 −
¯ ¯ n − µ > ] ≤ Var(Xn ) = Var(X1 ) P [X 2 n2 and the latter quantity tends to 0 as n → ∞ for each > 0. How can the weak law of large numbers be proved if we assume only that E(Xi ) < ∞? The answer is to write Xk = Unk + Vnk where Unk = Xk if Xk  ≤ δn (for some 0 < δ < 1) and Unk = 0 otherwise; it follows that Vnk = Xk if Xk  > δn and 0 otherwise. Then n n 1 ¯n + V¯n ¯n = 1 Uni + Vni = U X n i=1 n i=1 ¯n →p µ and V¯n →p 0. First, we and so it suﬃces to show that U have ¯n − µ)2 ] = Var(Un1 ) + (E(Un1 ) − µ)2 E[(U n 2 ) E(Un1 ≤ + (E(Un1 ) − µ)2 n ≤ E[Un1 ]δ + E 2 (Un1 − X1 ) ≤ E[X1 ]δ + E 2 [X1 I(X1  > δn)] c 2000 by Chapman & Hall/CRC
and so by Chebyshev’s inequality 2 ¯n − µ > ] ≤ E[X1 ]δ + E [X1 I(X1  > δn)] , P [U 2 which can be made close to 0 by taking n → ∞ and then δ to 0. Second,
P [V¯n  > ] ≤ P
n
[Vni = 0]
i=1
≤
n
P [Vni = 0]
i=1
= nP [X1 > δn] and the latter can be shown to tend to 0 as n → ∞ (for any δ > 0). The details of the proof are left as exercises. The WLLN can be strengthened to a strong law of large numbers (SLLN) by introducing another type of convergence known as convergence with probability 1 (or “almost sure” convergence). This is discussed in section 3.7. The WLLN for Bernoulli random variables was proved by Jacob Bernoulli in 1713 and strengthened to random variables with ﬁnite variance by Chebyshev in 1867 using the inequality that bears his name. Chebyshev’s result was extended by Khinchin in 1928 to sums of i.i.d. random variables with ﬁnite ﬁrst moment. 3.4 Proving convergence in distribution Recall that a sequence of random variables {Xn } converges in distribution to a random variable X if the corresponding sequence of distribution functions {Fn (x)} converges to F (x), the distribution function of X, at each continuity point of F . It is often diﬃcult to verify this condition directly for a number of reasons. For example, it is often diﬃcult to work with the distribution functions {Fn }. Also, in many cases, the distribution function Fn may not be speciﬁed exactly but may belong to a wider class; we may know, for example, the mean and variance corresponding to Fn but little else about Fn . (From a practical point of view, the cases where Fn is not known exactly are most interesting; if Fn is known exactly, there is really no reason to worry about a limiting distribution F unless Fn is diﬃcult to work with computationally.) For these reasons, we would like to have alternative methods for c 2000 by Chapman & Hall/CRC
establishing convergence in distribution. Fortunately, there are several other suﬃcient conditions for convergence in distribution that are useful in practice for verifying that a sequence of random variables converges in distribution and determining the distribution of the limiting random variable (the limiting distribution). • Suppose that Xn has density function fn (for n ≥ 1) and X has density function f . Then fn (x) → f (x) (for all but a countable number of x) implies that Xn →d X. Similarly, if Xn has frequency function fn and X has frequency function f then fn (x) → f (x) (for all x) implies that Xn →d X. (This result is known as Scheﬀ´e’s Theorem.) The converse of this result is not true; in fact, a sequence of discrete random variables can converge in distribution to a continuous variable (see Example 3.2) and a sequence of continuous random variables can converge in distribution to a discrete random variable. • If Xn has moment generating function mn (t) and X has moment generating function m(t) then mn (t) → m(t) (for all t ≤ some b > 0) implies Xn →d X. Convergence of moment generating functions is actually quite strong (in fact, it implies that E(Xnk ) → E(X k ) for integers k ≥ 1); convergence in distribution does not require convergence of moment generating functions. It is also possible to substitute other generating functions for the moment generating function to prove convergence in distribution. For example, if Xn has characteristic function ϕn (t) = E[exp(i tX)] and X has characteristic function ϕ(t) then ϕn (t) → ϕ(t) (for all t) implies Xn →d X; in fact, Xn →d X if, and only if, ϕn (t) → ϕ(t) for all t. In addition to the methods described above, we can also use some of the results given earlier (for example, Slutsky’s Theorem and the Delta Method) to help establish convergence in distribution. EXAMPLE 3.4: Suppose that {Xn } is a sequence of random variables where Xn has Student’s t distribution with n degrees of freedom. The density function of Xn is
Γ((n + 1)/2) x2 fn (x) = √ 1+ πnΓ(n/2) n
−(n+1)/2
Stirling’s approximation, which may be stated as √ yΓ(y) lim √ =1 y→∞ 2π exp(−y)y y c 2000 by Chapman & Hall/CRC
.
allows us to approximate Γ((n + 1)/2) and Γ(n/2) for large n. We then get Γ((n + 1)/2) 1 lim √ =√ . n→∞ πnΓ(n/2) 2π Also −(n+1)/2 x2 x2 = exp − lim 1 + n→∞ n 2 and so
1 x2 lim fn (x) = √ exp − n→∞ 2 2π
where the limit is a standard Normal density function. Thus Xn →d Z where Z has a standard Normal distribution. An alternative (and much simpler) approach is to note that the t distributed random variable Xn has the same distribution as Z/ Vn /n where Z and Vn are independent random variables with Z ∼ N (0, 1) distribution and Vn ∼ χ2 (n). Since Vn can be thought of as a sum of n i.i.d. χ2 random variables with 1 degree of freedom, it follows from the WLLN and the Continuous Mapping Theoremthat Vn /n →p 1 and hence from Slutsky’s Theorem that Z/ Vn /n →d N (0, 1).The conclusion follows since Xn has the same distribution as Z/ Vn /n. ✸ EXAMPLE 3.5: Suppose that U1 , · · · , Un are i.i.d. Uniform random variables on the interval [0, 1] and let U(1) , · · · , U(n) be their order√statistics. Deﬁne Zn = U(mn ) where mn ≈ n/2 in the sense that n(mn /n − 1/2) → 0 as n → ∞; note that we are requiring that mn /n converge to 1/2 at a faster rate than in Example 3.3. (Note that taking mn = n/2 for n even and mn = (n + 1)/2 for n odd will satisfy this condition.) We will consider √ the asymptotic distribution of the sequence of random variables √ { n(Zn − 1/2)} by computing its limiting density. The density of n(Zn − 1/2) is
1 n! x mn −1 1 x n−mn √ fn (x) = +√ −√ (mn − 1)!(n − mn )! n 2 n 2 n √ √ for − n/2 ≤ x ≤ n/2. We will show that fn (x) converges to a Normal density with mean 0 and variance 1/4. First using Stirling’s approximation (as in Example 3.4, noting that n! = nΓ(n)), we obtain n! 2n √ ≈√ n(mn − 1)!(n − mn )! 2π c 2000 by Chapman & Hall/CRC
in the sense that the ratio of the righthand to lefthand side tends to 1 as n → ∞. We also have
1 x +√ 2 n =
We now obtain
mn −1
1 2n−1
4x2 1− n
4x2 1− n
and
1 x −√ 2 n
n−mn
mn −1
2x 1− √ n
n−2mn +1
.
mn −1
→ exp(−2x2 )
2x n−2mn +1 1− √ →1 n where, in both cases, we use the fact that (1 + t/an )cn → exp(kt) if an → ∞ and cn /an → k. Putting the pieces from above together, we get 2 fn (x) → √ exp(−2x2 ) 2π √ for any x. Thus n(Zn − 1/2) →d N (0, 1/4). ✸ EXAMPLE 3.6: We can easily extend Example 3.5 to the case where X1 , · · · , Xn are i.i.d. random variables with distribution function F and unique median µ where F (x) is diﬀerentiable at x = µ with F (µ) > 0; if F has a density f then F (µ) = f (µ) typically. Deﬁning F −1 (t) = inf{x : F (x) ≥ t} to be the inverse of F , we note that the order statistic X(k) has the same distribution as F −1 (U(k) ) where U(k) is an order statistic from an i.i.d. sample of Uniform random variables on [0, 1] and also that F −1 (1/2) = µ. Thus √ √ n(X(mn ) − µ) =d n(F −1 (U(mn ) ) − F −1 (1/2)) and so by the Delta Method, we have √ n(X(mn ) − µ) →d N (0, (F (µ))−2 /4) √ if n(mn /n − 1/2) → 0. The limiting variance follows from the fact that F −1 (t) is diﬀerentiable at t = 1/2 with derivative 1/F (µ). Note that existence of a density is not suﬃcient to imply existence of the derivative of F (x) at x = µ; however, if F is continuous but c 2000 by Chapman & Hall/CRC
1.5 1.0 0.0
0.5
density
0.0
0.5
1.0
1.5
x
Figure 3.1 Density of X(n/2) for n = 10 Exponential random variables; the dotted line is the approximating Normal density.
√ not diﬀerentiable at x = µ then n(X(mn ) − µ) may still converge in distribution but the limiting distribution will be diﬀerent. As an illustration of the convergence of the distribution of the sample median to a Normal distribution, we will consider the density of the order statistic X(n/2) for i.i.d. Exponential random variables X1 , · · · , Xn with density f (x) = exp(−x) for x ≥ 0. Figures 3.1, 3.2, and 3.3 give the densities of X(n/2) for n = 10, 50, and 100 respectively; the corresponding approximating Normal density is indicated with dotted lines. ✸ EXAMPLE 3.7: Suppose that {Xn } is a sequence of Binomial random variables with Xn having parameters n and θn where nθn → λ > 0 as n → ∞. The moment generating function of Xn is mn (t) = (1 + θn (exp(t) − 1))n
nθn (exp(t) − 1) n = 1+ . n c 2000 by Chapman & Hall/CRC
3.0 2.5 2.0 1.5 0.0
0.5
1.0
density
0.4
0.6
0.8
1.0
x
Figure 3.2 Density of X(n/2) for n = 50 Exponential random variables; the dotted line is the approximating Normal density.
Since nθn → λ, it follows that lim mn (t) = exp[λ(exp(t) − 1)]
n→∞
where the limiting moment generating function is that of a Poisson distribution with parameter λ. Thus Xn →d X where X has a Poisson distribution with parameter λ. This result can be used to compute Binomial probabilities when n is large and θ is small so that nθ ≈ nθ(1 − θ). For example, suppose that X has a Binomial distribution with n = 100 and θ = 0.05. Then using the Poisson approximation P [a ≤ X ≤ b] ≈
exp(−nθ)(nθ)x a≤x≤b
x!
we get, for example, P [4 ≤ X ≤ 6] ≈ 0.497 compared to the exact probability P [4 ≤ X ≤ 6] = 0.508. ✸ c 2000 by Chapman & Hall/CRC
4 3 2 0
1
density
0.4
0.5
0.6
0.7
0.8
0.9
1.0
x
Figure 3.3 Density of X(n/2) for n = 100 Exponential random variables; the dotted line is the approximating Normal density.
EXAMPLE 3.8: As in Example 3.2, deﬁne Un by Un =
n Xk k=1
10k
where X1 , Xn , · · · are i.i.d. discrete random variables uniformly distributed on the integers 0, 1, · · · , 9. We showed in Example 3.2 that Un →d Unif(0, 1) by showing convergence of the distribution functions. In this example, we will do the same using moment generating functions. The moment generating function of each Xk is m(t) =
1 exp(10t) − 1 (1 + exp(t) + · · · + exp(9t)) = 10 10(exp(t) − 1)
and so the moment generating function of Un is φn (t) = =
n
m(t/10k )
k=1 n k=1
c 2000 by Chapman & Hall/CRC
exp(t/10k−1 ) − 1 10(exp(t/10k ) − 1
=
exp(t) − 1 . 10n (exp(t/10n ) − 1)
Using the expansion exp(x) = 1 + x + x2 /2 + · · ·, it follows that lim 10n (exp(t/10n ) − 1) = t
n→∞
and so
1 1 lim φn (t) = (exp(t) − 1) = exp(tx) dx, n→∞ t 0 which is the moment generating function of the Uniform distribution on [0, 1]. Thus we have shown (using moment generating functions) that Un →d Unif(0, 1). ✸
3.5 Central Limit Theorems In probability theory, central limit theorems (CLTs) establish conditions under which the distribution of a sum of random variables may be approximated by a Normal distribution. (We have seen already in Examples 3.4, 3.5, and 3.6 cases where the limiting distribution is Normal.) A wide variety of CLTs have been proved; however, we will consider CLTs only for sums and weighted sums of i.i.d. random variables. THEOREM 3.8 (CLT for i.i.d. random variables) Suppose that X1 , X2 , · · · are i.i.d. random variables with mean µ and variance σ 2 < ∞ and deﬁne √ ¯ n n(Xn − µ) 1 (Xi − µ) = . Sn = √ σ n i=1 σ Then Sn →d Z ∼ N (0, 1) as n → ∞. (In practical terms, the CLT implies that for “large” n, the ¯ n is approximately Normal with mean µ and distribution of X 2 variance σ /n.) Before discussing the proof of this CLT, we will give a little of the history behind the result. The Frenchborn mathematician de Moivre is usually credited with proving the ﬁrst CLT (in the 18th century); this CLT dealt with the special case that the Xi ’s were Bernoulli random variables (that is, P [Xi = 1] = θ and P [Xi = 0] = 1 − θ). His work (de Moivre, 1738) was not signiﬁcantly improved until Laplace (1810) extended de Moivre’s c 2000 by Chapman & Hall/CRC
work to sums of independent bounded random variables. The Russian mathematician Chebyshev extended Laplace’s work to sums of random variables with ﬁnite moments E(Xi k ) for all k ≥ 1. However, it was not until the early twentieth century that Markov and Lyapunov (who were students of Chebyshev) removed nearly all unnecessary moment restrictions. Finally, Lindeberg (1922) proved the CLT assuming only ﬁnite variances. It should be noted that most of the work subsequent to Laplace dealt with sums of independent (but not necessarily identically distributed) random variables; it turns out that this added generality does not pose any great technical complications. Proving the CLT We will consider two proofs of the CLT for i.i.d. random variables. In the ﬁrst proof, we will assume the existence of moment generating function of the Xi ’s and show that the moment generating function of Sn converges to the moment generating function of Z. Of course, assuming the existence of moment generating functions implies that E(Xik ) exists for all integers k ≥ 1. The second proof will require only that E(Xi2 ) is ﬁnite and will show directly that P [Sn ≤ x] → P [Z ≤ x]. We can assume (without loss of generality) that E(Xi ) = 0 and Var(Xi ) = 1. Let m(t) = E[exp(tXi )] be the moment generating function of the Xi ’s. Then m(t) = 1 +
t2 t3 E(Xi3 ) + + ··· 2 6
(since E(Xi ) = 0 and E(Xi2 ) = 1) and the moment generating function of Sn is √ mn (t) = [mn (t/ n)]n ,
= ,
= ,
=
n
t2 t3 E(Xi3 ) 1+ + ··· + 2n 6n3/2
n
t2 tE(Xi3 ) t2 E(Xi4 ) 1+ 1+ + + ··· 2n 12n 3n1/2 n
t2 1+ rn (t) 2n
where rn (t) < ∞ and rn (t) → 1 as n → ∞ for each t ≤ some c 2000 by Chapman & Hall/CRC
b > 0. Thus
t2 lim mn (t) = exp n→∞ 2 and the limit is the moment generating function of a standard Normal random variable. It should be noted that a completely rigorous proof of Theorem 3.8 can be obtained by replacing moment generating functions by characteristic functions in the proof above. The second proof we give shows directly that P [Sn ≤ x] →
x −∞
1 t2 √ exp − 2 2π
dt;
we will ﬁrst assume that E(Xi 3 ) is ﬁnite and then indicate what modiﬁcations are necessary if we assume only that E(Xi2 ) is ﬁnite. The method used in this proof may seem at ﬁrst somewhat complicated but, in fact, is extremely elegant and is actually the method used by Lindeberg to prove his 1922 CLT. The key to the proof directly lies in approximating P [Sn ≤ x] by E[fδ+ (Sn )] and E[fδ− (Sn )] where fδ+ and fδ− are two bounded, continuous functions. In particular, we deﬁne fδ+ (y) = 1 for y ≤ x, fδ+ (y) = 0 for y ≥ x + δ and 0 ≤ fδ+ (y) ≤ 1 for x < y < x + δ; we deﬁne fδ− (y) = fδ+ (y + δ). If g(y) = I(y ≤ x), it is easy to see that fδ− (y) ≤ g(y) ≤ fδ+ (y) and
P [Sn ≤ x] = E[g(Sn )]. Then if Z is a standard Normal random variable, we have P [Sn ≤ x] ≤ E[fδ+ (Sn )] ≤ E[fδ+ (Sn )] − E[fδ+ (Z)] + E[fδ+ (Z)] ≤ E[fδ+ (Sn )] − E[fδ+ (Z)] + P [Z ≤ x + δ] and similarly, P [Sn ≤ x] ≥ P [Z ≤ x − δ] − E[fδ− (Sn )] − E[fδ− (Z)]. Thus we have E[fδ+ (Sn )] − E[fδ+ (Z)] + P [Z ≤ x + δ] ≥ P [Sn ≤ x] ≥ P [Z ≤ x − δ] − E[fδ− (Sn )] − E[fδ− (Z)];
c 2000 by Chapman & Hall/CRC
since P [Z ≤ x ± δ] can be made arbitrarily close to P (Z ≤ x) (because Z has a continuous distribution function), it suﬃces to show that E[fδ+ (Sn )] → E[fδ+ (Z)] and E[fδ− (Sn )] → E[fδ− (Z)] for suitable choices of fδ+ and fδ− . In particular, we will assume that fδ+ (and hence fδ− ) has three bounded continuous derivatives. Let f be a bounded function (such as fδ+ or fδ− ) with three bounded continuous derivatives and let Z1 , Z2 , Z3 , · · · be a sequence of i.i.d. standard Normal random variables that are also independent of X1 , X2 , · · ·; note that n−1/2 (Z1 + · · · + Zn ) is also standard Normal. Now deﬁne random variables Tn1 , · · · , Tnn where Tnk
n 1 k−1 1 =√ Zj + √ Xj n j=1 n j=k+1
(where the sum is taken to be 0 if the upper limit is less than the lower limit). Then E[f (Sn )] − E[f (Z)] =
n
E[f (Tnk + n−1/2 Xk ) − f (Tnk + n−1/2 Zk )].
k=1
Expanding f (Tnk + n−1/2 Xk ) in a Taylor series around Tnk , we get Xk X2 f (Tnk + n−1/2 Xk ) = f (Tnk ) + √ f (Tnk ) + k f (Tnk ) + RkX n 2n where RkX is a remainder term (whose value will depend on the third derivative of f ); similarly, Zk Z2 f (Tnk + n−1/2 Zk ) = f (Tnk ) + √ f (Tnk ) + k f (Tnk ) + RkZ . n 2n Taking expected values (and noting that Tnk is independent of both Xk and Zk ), we get E[f (Sn )] − E[f (Z)] =
n
[E(RkX ) − E(RkZ )].
k=1
We now try to ﬁnd bounds for RkX and RkZ ; it follows that RkX  ≤
K Xk 3 6 n3/2
and RkZ  ≤
K Zk 3 6 n3/2
where K is an upper bound on f (y), which (by assumption) is c 2000 by Chapman & Hall/CRC
ﬁnite. Thus E[f (Sn )] − E[f (Z)]
≤ ≤
n k=1 n
E(RkX ) − E(RkZ ) [E(RkX ) + E(RkZ )]
k=1
K √ E[X1 3 ] + E[Z1 3 ] 6 n → 0
≤
as n → ∞ since both E[X1 3 ] and E[Z1 3 ] are ﬁnite. Applying the previous result to fδ+ and fδ− having three bounded continuous derivatives, it follows that E[fδ+ (Sn )] → E[fδ+ (Z)] and E[fδ− (Sn )] → E[fδ− (Z)]. Now since P [Z ≤ x ± δ] → P [Z ≤ x] as δ → 0, it follows from above that for each x, P [Sn ≤ x] → P [Z ≤ x]. We have, of course, assumed that E[Xi 3 ] is ﬁnite; to extend the result to the case where we assume only that E(Xi2 ) is ﬁnite, it is necessary to ﬁnd a more accurate bound on RkX . Such a bound is given by
RkX 
≤K
√ √ Xk 3 Xk2 I(X  ≤ n) + I(Xk  > n) k 3/2 n 6n
where K is an upper bound on both f (y) and f (y). It then can be shown that n E[RkX ] → 0 k=1
and so E[f (Sn )] → E[f (Z)]. Using the CLT as an approximation theorem In mathematics, a distinction is often made between limit theorems and approximation theorems; the former simply speciﬁes the limit of a sequence while the latter provides an estimate or bound on the diﬀerence between an element of the sequence and its limit. For example, it is wellknown that ln(1 + x) can be approximated c 2000 by Chapman & Hall/CRC
by x when x is small; a crude bound on the absolute diﬀerence  ln(1 + x) − x is x2 /2 when x ≥ 0 and x2 /[2(1 + x)2 ] when x < 0. For practical purposes, approximation theorems are more useful as they allow some estimate of the error made in approximating by the limit. The CLT as stated here is not an approximation theorem. That is, it does not tell us how large n should be in order for a Normal distribution to approximate the distribution of Sn . Nonetheless, with additional assumptions, the CLT can be restated as an approximation theorem. Let Fn be the distribution function of Sn and Φ be the standard Normal distribution function. To gain some insight into the factors aﬀecting the speed of convergence of Fn to Φ, we will use Edgeworth expansions. Assume that Fn is a continuous distribution function and that E(Xi4 ) < ∞ and deﬁne γ=
E[(Xi − µ)3 ] σ3
and κ =
E[(Xi − µ)4 ] − 3; σ4
γ and κ are, respectively, the skewness and kurtosis of the distribution of Xi both of which are 0 when Xi is normally distributed. It is now possible to show that
γ κ γ2 Fn (x) = Φ(x) − φ(x) √ p1 (x) + p2 (x) + p3 (x) + rn (x) 6 n 24n 72n where φ(x) is the standard Normal density function and p1 (x), p2 (x), p3 (x) are the polynomials p1 (x) = x2 − 1 p2 (x) = x3 − 3x p3 (x) = x5 − 10x3 + 15x; the remainder term rn (x) satisﬁes nrn (x) → 0. From this expansion, it seems clear that the approximation error Fn (x) − Φ(x) depends on the skewness and kurtosis (that is, γ and κ) of the Xi ’s. The skewness and kurtosis are simple measures of how a particular distribution diﬀers from normality; skewness is a measure of the asymmetry of a distribution (γ = 0 if the distribution is symmetric around its mean) while kurtosis is a measure of the thickness of the tails of a distribution (κ > 0 indicates heavier tails than a Normal distribution while κ < 0 indicates lighter tails). For example, a Uniform distribution has γ = 0 and κ = −1.2 while c 2000 by Chapman & Hall/CRC
0.20 0.15 0.10 0.0
0.05
density
10
5
0
5
10
x
Figure 3.4 Density of the sum of 10 Uniform random variables; the dotted curve is the approximating Normal density.
an Exponential distribution has γ = 2 and κ = 6. Thus we should expect convergence to occur more quickly for sums of Uniform random variables than for sums of Exponential random variables. Indeed, this is true; in fact, the distribution of a sum of as few as ten Uniform random variables is suﬃciently close to a Normal distribution to allow generation of Normal random variables on a computer by summing Uniform random variables. To illustrate the diﬀerence in the accuracy of the Normal approximation, we consider the distribution of X1 + · · · + X10 when the Xi ’s are Uniform and Exponential; Figures 3.4 and 3.5 give the exact densities and their Normal approximations in these two cases. The speed of convergence of the CLT (and hence the goodness of approximation) can often be improved by applying transformations ¯ n . Recall that if √n(X ¯n − to reduce the skewness and kurtosis of X µ) →d Z then √ ¯ n ) − g(µ)) →d g (µ)Z. n(g(X ¯ n ) is more symmetric If g is chosen so that the distribution of g(X ¯ and has lighter tails than that of Xn then the CLT should √ provide ¯n) − a more accurate approximation for the distribution of n(g(X √ ¯ g(µ)) than it does for the distribution of n(Xn − µ). c 2000 by Chapman & Hall/CRC
0.12 0.08 0.0
0.04
density
0
5
10
15
20
x
Figure 3.5 Density of the sum of 10 Exponential random variables; the dotted curve is the approximating Normal density.
Although the Edgeworth expansion above does not always hold when the Xi ’s are discrete, the preceding comments regarding speed of convergence and accuracy of the Normal approximation are still generally true. However, when the Xi ’s are discrete, there is a simple technique that can improve the accuracy of the Normal approximation. We will illustrate this technique for the Binomial distribution. Suppose that X is a Binomial random variable with parameters n and θ; X can be thought of as a sum of n i.i.d. Bernoulli random variables so the distribution of X can be approximated by a Normal distribution if n is suﬃciently large. More speciﬁcally, the distribution of X − nθ nθ(1 − θ)
is approximately standard Normal for large n. Suppose we want to evaluate P [a ≤ X ≤ b] for some integers a and b. Ignoring the fact that X is a discrete random variable and the Normal distribution c 2000 by Chapman & Hall/CRC
0.12 0.08 0.04 0.0
density/frequency
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Figure 3.6 Binomial distribution (n = 40, θ = 0.3) and approximating Normal density
is a continuous distribution, a naive application of the CLT gives P [a ≤ X ≤ b] ,
a − nθ X − nθ b − nθ = P ≤ ≤ nθ(1 − θ) nθ(1 − θ) nθ(1 − θ)
b − nθ ≈ Φ nθ(1 − θ)

a − nθ −Φ . nθ(1 − θ)
How can this approximation be improved? The answer is clear if we compare the exact distribution of X to its Normal approximation. The distribution of X can be conveniently represented as a probability histogram as in Figure 3.6 with the area of each bar representing the probability that X takes a certain value. The naive Normal approximation given above merely integrates the approximating Normal density from a = 8 to b = 17; this probability is represented by the shaded area in Figure 3.7. It seems that the naive Normal approximation will underestimate the true probability and Figures 3.7 and 3.8 suggest that a better approximation may be obtained by integrating from a − 0.5 = 7.5 c 2000 by Chapman & Hall/CRC
0.12 0.08 0.04 0.0
density/frequency
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
Figure 3.7 Naive Normal approximation of P (8 ≤ X ≤ 17)
to b + 0.5 = 17.5. This corrected Normal approximation is P [a ≤ X ≤ b] = P [a − 0.5 ≤ X ≤ b + 0.5]
b + 0.5 − nθ ≈ Φ nθ(1 − θ)
a − 0.5 − nθ −Φ nθ(1 − θ)
.
The correction used here is known as a continuity correction and can be applied generally to improve the accuracy of the Normal approximation for sums of discrete random variables. (In Figures 3.6, 3.7, and 3.8, X has a Binomial distribution with parameters n = 40 and θ = 0.3.) Some other Central Limit Theorems CLTs can be proved under a variety of conditions; neither the assumption of independence nor that of identical distributions are necessary. In this section, we will consider two simple modiﬁcations of the CLT for sums of i.i.d. random variables. The ﬁrst modiﬁcation deals with weighted sums of i.i.d. random variables while c 2000 by Chapman & Hall/CRC
0.12 0.08 0.04 0.0
density/frequency
4
5
6
7
8
9
10
11
12
13
14
15
16
Figure 3.8 Normal approximation of P (8 ≤ X correction
17
18
19
20
21
≤ 17) with continuity
the second deals with sums of independent but not identically distributed random variables with ﬁnite third moment. THEOREM 3.9 (CLT for weighted sums) Suppose that X1 , X2 , · · · are i.i.d. random variables with E(Xi ) = 0 and Var(Xi ) = 1 and let {ci } be a sequence of constants. Deﬁne Sn =
n 1 ci Xi sn i=1
where
s2n =
n
c2i .
i=1
Then Sn →d Z, a standard Normal random variable, provided that c2i →0 1≤i≤n s2 n max
as n → ∞. What is the practical meaning of the “negligibility” condition on the constants {ci } given above? For each n, it is easy to see that c 2000 by Chapman & Hall/CRC
Var(Sn ) = 1. Now writing Sn =
n ci i=1
sn
Xi =
n
Yni
i=1
and noting that Var(Yni ) = c2i /s2n , it follows that this condition implies that no single component of the sum Sn contributes an excessive proportion of the variance of Sn . For example, the condition rules out situations where Sn depends only on a negligible proportion of the Yni ’s. An extreme example of this occurs when c1 = c2 = · · · = ck = 1 (for some ﬁxed k) and all other ci ’s are 0; in this case, c2 1 max 2i = , 1≤i≤n sn k which does not tend to 0 as n → ∞ since k is ﬁxed. On the other hand, if ci = i then s2n =
n
1 i2 = n(2n2 + 3n + 1) 6 i=1
and
c2i 6n2 = →0 1≤i≤n s2 n(2n2 + 3n + 1) n as n → ∞ and so the negligibility condition holds. Thus if the Xi ’s are i.i.d. random variables with E(Xi ) = 0 and Var(Xi ) = 1, it follows that n 1 iXi →d Z sn i=1 max
where Z has a standard Normal distribution. When the negligibility condition of Theorem 3.9 fails, it may still be possible to show that the weighted sum Sn converges in distribution although the limiting distribution will typically be nonNormal. EXAMPLE 3.9: Suppose that X1 , X2 , · · · are i.i.d. random variables with common density function 1 f (x) = exp(−x) 2 (called a Laplace distribution) and deﬁne n Xk 1 Sn = sn k=1 k c 2000 by Chapman & Hall/CRC
n ∞ −2 2 −2 = π 2 /6 as where s2n = i=1 k . Note that sn → k=1 k n → ∞ and the negligibility condition does not hold. √ However, it can be shown that Sn →d S. Since sn → π/ 6, we will consider√the limiting distribution of Vn = sn Sn ; if Vn →d V then Sn →d 6/πV = S. The moment generating function of Xi is
1 1 − t2
m(t) =
(for t < 1)
and so the moment generating function of Vn is mn (t) =
n
m(t/k) =
k=1
n k=1
k2 k 2 − t2
.
As n → ∞, mn (t) → mV (t) where mV (t) =
∞
k=1
k2 k 2 − t2
= Γ(1 + t)Γ(1 − t)
for t < 1. The limiting moment generating function mV (t) is not immediately recognizable. However, note that ∞
exp(x) exp(tx) dx = (1 + exp(x))2 −∞
1 0
ut (1 − u)−t du
Γ(1 + t)Γ(1 − t) Γ(2) = Γ(1 + t)Γ(1 − t)
=
and so the density of V is fV (x) =
exp(x) (1 + exp(x))2
(this distribution is often called the Logistic distribution). The density of S is thus √ exp(πx/ 6) π √ fS (x) = √ . 6 (1 + exp(πx/ 6))2 The limiting density fS (x) is shown in Figure 3.9.
✸
Another useful CLT for sums of independent (but not identically distributed) random variables is the Lyapunov CLT. Like the CLT for weighted sums of i.i.d. random variables, this CLT depends on a condition that can be easily veriﬁed. c 2000 by Chapman & Hall/CRC
0.30 0.20 0.0
0.10
density
4
2
0
2
4
x
Figure 3.9 Density of S; the dotted curve is a Normal density with the same mean and variance as S.
THEOREM 3.10 (Lyapunov CLT) Suppose that X1 , X2 , · · · are independent random variables with E(Xi ) = 0, E(Xi2 ) = σi2 and E(Xi 3 ) = γi and deﬁne
Sn =
where s2n =
n
2 i=1 σi .
n 1 Xi sn i=1
If
lim
n→∞
n 1 3/2
sn
γi = 0
i=1
then Sn →d Z, a standard Normal random variable. It is possible to adapt the proof of the CLT for sums of i.i.d. random variables to the two CLTs given in this section. In the case of the CLT for weighted sums, the key modiﬁcation lies in redeﬁning c 2000 by Chapman & Hall/CRC
Tnk to be Tnk =
n 1 k−1 1 cj Zj + cj Xj sn j=1 sn j=k+1
where Z1 , Z2 , · · · are independent standard Normal random variables independent of the Xi ’s. Then letting f be a bounded function with three bounded derivatives, we have E[f (Sn )] − E[f (Z)] =
n
E[f (Tnk + ck Xk /sn ) − f (Tnk + ck Zk /sn )].
k=1
The remainder of the proof is much the same as before and is left as an exercise. It is also possible to give a proof using moment generating functions assuming, of course, that the moment generating function of Xi exists. Multivariate Central Limit Theorem CLTs for sums of random variables can be generalized to deal with sums of random vectors. For example, suppose that X 1 , X 2 , · · · are i.i.d. random vectors with mean vector µ and variancecovariance matrix C; deﬁne n ¯n= 1 X Xi n i=1 to be the (coordinatewise) sample mean of X 1 , · · · , X n . The logical extension of the CLT for i.i.d. random variables √ ¯ is to consider the limiting behaviour of the distributions of n(X n − µ). Before considering any multivariate CLT, we need to extend the notion of convergence in distribution to sequences of random vectors. This extension is fairly straightforward and involves the joint distribution functions of the random vectors; given random vectors {X n } and X, we say that X n →d X if lim P [X n ≤ x] = P [X ≤ x] = F (x)
n→∞
at each continuity point x of the joint distribution function F of X. (X ≤ x means that each coordinate of X is less than or equal to the corresponding coordinate of x.) This deﬁnition, while simple enough, is diﬃcult to prove analytically. Fortunately, convergence in distribution of random vectors can be cast in terms of their onedimensional projections. c 2000 by Chapman & Hall/CRC
THEOREM 3.11 Suppose that {X n } and X are random vectors. Then X n →d X if, and only if, tT X n →d tT X for all vectors t. Theorem 3.11 is called the Cram´erWold device. The proof of this result will not be given here. The result is extremely useful for proving multivariate CLTs since it essentially reduces multivariate CLTs to special cases of univariate CLTs. We will only consider a multivariate CLT for sums of i.i.d. random vectors but more general multivariate CLTs can also be deduced from appropriate univariate CLTs. THEOREM 3.12 (Multivariate CLT) Suppose that X 1 , X 2 , X 3 , · · · are i.i.d. random vectors with mean vector µ and variancecovariance matrix C and deﬁne n √ 1 ¯ n − µ). Sn = √ (X i − µ) = n(X n i=1
Then S n →d Z where Z has a multivariate Normal distribution with mean 0 and variancecovariance matrix C. Proof. It suﬃces to show that tT S n →d tT Z; note that tT Z is Normally distributed with mean 0 and variance tT Ct. Now tT S n = =
n 1 √ tT (X i − µ) n i=1 n 1 √ Yi n i=1
where the Yi ’s are i.i.d. with E(Yi ) = 0 and Var(Yi ) = tT Ct. Thus by the CLT for i.i.d. random variables, tT S n →d N (0, tT Ct) and the theorem follows. The deﬁnition of convergence in probability can be extended quite easily to random vectors. We will say that X n →p X if each coordinate of X n converges in probability to the corresponding coordinate of X. Equivalently, we can say that X n →p X if lim P [X n − X > ] = 0
n→∞
c 2000 by Chapman & Hall/CRC
where · is the Euclidean norm of a vector. It is possible to generalize many of the results proved above. For example, suppose that X n →d X; then if g is a continuous realvalued function, g(X n ) →d g(X). (The same is true if →d is replaced by →p .) This multivariate version of the Continuous Mapping Theorem can be used to obtain a generalization of Slutsky’s Theorem. Suppose that Xn →d X and Yn →p θ. By using the Cram´erWold device (Theorem 3.11) and Slutsky’s Theorem, it follows that (Xn , Yn ) →d (X, θ). Thus if g(x, y) is a continuous function, we have g(Xn , Yn ) →d g(X, θ).
EXAMPLE 3.10: Suppose that {X n } is a sequence of random vectors with X n →d Z where Z ∼ Np (0, C) and C is nonsingular. Deﬁne the function g(x) = xT C −1 x, which is a continuous function of x. Then we have g(X n ) = X Tn C −1 X n →d Z T C −1 Z = g(Z). It follows from Chapter 2 that the random variable Z T C −1 Z has a χ2 distribution with p degrees of freedom. Thus for large n, X Tn C −1 X n is approximately χ2 with p degrees of freedom. ✸ It is also possible to extend the Delta Method to the multivariate case. Let {an } be a sequence of constants tending to inﬁnity and suppose that an (X n − θ) →d Z. If g(x) = (g1 (x), · · · , gk (x)) is a vectorvalued function that is continuously diﬀerentiable at x = θ, we have an (g(X n ) − g(θ)) →d D(θ)Z where D(θ) is a matrix of partial derivatives of g with respect to x evaluated at x = θ; more precisely, if x = (x1 , · · · , xp ), D(θ) =
· · · ∂x∂ p g1 (θ) · · · ∂x∂ p g2 (θ) ........................ ∂ ∂ ∂x1 gk (θ) · · · ∂xp gk (θ) ∂ ∂x1 g1 (θ) ∂ ∂x1 g2 (θ)
c 2000 by Chapman & Hall/CRC
.
The proof of this result parallels that of the Delta Method given earlier and is left as an exercise. EXAMPLE 3.11: Suppose that (X1 , Y1 ), · · · , (Xn , Yn ) are i.i.d. pairs of random variables with E(Xi ) = µX > 0, E(Yi ) = µY > 0, and E(Xi2 ) and E(Yi2 ) both ﬁnite. By the multivariate CLT, we have ¯n √ µX X n − →d Z ∼ N2 (0, C) ¯ µY Yn where C is the variancecovariance matrix of (Xi , Yi ). We want to ¯ n /Y¯n . Applying the Delta consider the asymptotic distribution of X Method (with g(x, y) = x/y), we have √
¯ n µX X n ¯ − µY Yn
→d D(µX , µY )Z ∼
N (0, D(µX , µY )CD(µX , µY )T )
where D(µX , µY ) =
1 µX ,− 2 µY µY
.
2 , Var(Y ) = σ 2 , and Cov(X , Y ) = σ Letting Var(Xi ) = σX i i i XY , it Y follows that the variance of the limiting Normal distribution is
D(µX , µY )CD(µX , µY )T =
2 − 2µ µ σ 2 2 µ2Y σX X Y XY + µX σY . µ4Y
✸
3.6 Some applications In subsequent chapters, we will use many of the concepts and results developed in this chapter to characterize the large sample properties of statistical estimators. In this section, we will give some applications of the concepts and results given so far in this chapter. Variance stabilizing transformations The CLT states that if X1 , X2 , · · · are i.i.d. random variables with mean µ and variance σ 2 then √ ¯ n − µ) →d Z n(X where Z has a Normal distribution with mean 0 and variance σ 2 . For many distributions, the variance σ 2 depends only on the mean c 2000 by Chapman & Hall/CRC
µ (that is, σ 2 = V (µ)). In statistics, it is often desirable to ﬁnd a function g such that the limit distribution of √ ¯ n ) − g(µ)) n(g(X does not depend on µ. (We could then use this result to ﬁnd an approximate conﬁdence interval for µ; see Chapter 7.) If g is diﬀerentiable, we have √ ¯ n ) − g(µ)) →d g (µ)Z n(g(X and g (µ)Z is Normal with mean 0 and variance [g (µ)]2 V (µ); in order to make the limiting distribution independent of µ, we need to ﬁnd g so that this variance is 1 (or some other constant). Thus, given V (µ), we would like to ﬁnd g such that [g (µ)]2 V (µ) = 1 or
1 . V (µ)1/2 The function g can be the solution of either of the two diﬀerential equations depending on whether one wants g to be an increasing or a decreasing function of µ; g is called a variance stabilizing transformation. g (µ) = ±
EXAMPLE 3.12: Suppose that X1 , · · · , Xn are i.i.d. Bernoulli random variables with parameter θ. Then √ ¯ n − θ) →d Z ∼ N (0, θ(1 − θ)). n(X To ﬁnd g such that √
¯ n ) − g(θ)) →d N (0, 1) n(g(X
we solve the diﬀerential equation 1 . θ(1 − θ)
g (θ) =
The general form of the solutions to this diﬀerential equation is g(θ) = sin−1 (2θ − 1) + c where c is an arbitrary constant that could be taken to be 0. (The solutions√to the diﬀerential equation can also be written ✸ g(θ) = 2 sin−1 ( θ) + c.) c 2000 by Chapman & Hall/CRC
Variance stabilizing transformations often improve the speed of ¯ n ) can be convergence to normality; that is, the distribution of g(X ¯ n if better approximated by a Normal distribution than that of X g is a variance stabilizing transformation. However, there may be other transformations that result in a better approximation by a Normal distribution. A CLT for dependent random variables Suppose that {Ui } is an inﬁnite sequence of i.i.d. random variables with mean 0 and variance σ 2 and deﬁne Xi =
p
cj Ui−j
j=0
where c0 , · · · , cp are constants. Note that X1 , X2 , · · · are not necessarily independent since they can depend on common Ui ’s. (In time series analysis, {Xi } is called a moving average process.) A natural question to ask is whether a CLT holds for sample means ¯ n based on X1 , X2 , · · · X We begin by noting that n 1 √ Xi = n i=1
n 1 √ cj Ui−j n i=1 j=0
=
n 1 √ cj Ui−j n j=0 i=1
p
p
= where
p
n 1 cj √ Ui + cj Rnj n i=1 j=0 j=1
Rnj
p
=
n n 1 √ Ui−j − Ui n i=1 i=1
=
1 √ (U1−j + · · · + U0 − Un−j+1 − · · · − Un ) . n
2 ) = 2jσ 2 /n and so by ChebyNow E(Rnj ) = 0, Var(Rnj ) = E(Rnj shev’s inequality Rnj →p 0 as n → ∞ for j = 1, · · · , p; thus, p j=1 c 2000 by Chapman & Hall/CRC
cj Rnj →p 0.
Finally,
n 1 √ Ui →d N (0, σ 2 ) n i=1
and so applying Slutsky’s Theorem, we get n 1 √ Xi →d Z n i=1
where Z has a Normal distribution with mean 0 and variance
σ2
p
2
cj .
j=0
When pj=0 cj = 0, the variance of the limiting Normal distribution is 0; this suggests that n 1 √ Xi →p 0 n i=1
p
(if that
j=0 cj
= 0). This is, in fact, the case. It follows from above n 1 √ Xi = cj Rnj , n i=1 j=1 p
which tends to 0 in probability. An extension to inﬁnite moving averages is considered in Problem 3.22. In general, what conditions are necessary to obtain a CLT for sums of dependent random variables {Xi }? Loosely speaking, it may be possible to approximate the distribution of X1 + · · · + Xn by a Normal distribution (for suﬃciently large n) if both • the dependence between Xi and Xi+k becomes negligible as k → ∞ (for each i), and • each Xi accounts for a negligible proportion of the variance of X1 + · · · + Xn . However, the conditions above provide only a very rough guideline for the possible existence of a CLT; much more speciﬁc technical conditions are typically needed to establish CLTs for sums of dependent random variables. c 2000 by Chapman & Hall/CRC
Monte Carlo integration. Suppose we want to evaluate the multidimensional integral
···
g(x) dx
where the function g is suﬃciently complicated that the integral cannot be evaluated analytically. A variety of methods exist for evaluating integrals numerically. The most wellknown of these involve deterministic approximations of the form
···
g(x) dx ≈
m
ai g(xi )
i=1
where a1 , · · · , an , x1 , · · · , xn are ﬁxed points that depend on the method used; for a given function g, it is usually possible to give an explicit upper bound on the approximation error + + m + + + + ai g(xi )+ + · · · g(x) dx − + + i=1
and so the points {xi } can be chosen to make this error acceptably small. However, as the dimension of domain of integration B increases, the number of points m needed to obtain a given accuracy increases exponentially. An alternative is to use socalled “Monte Carlo” sampling; that is, we evaluate g at random (as opposed to ﬁxed) points. The resulting approximation is of the form m
Ai g(X i )
i=1
where the X i ’s and possibly the Ai ’s are random. One advantage of using Monte Carlo integration is the fact that the order of the approximation error depends only on m and not the dimension of B. Unfortunately, Monte Carlo integration does not give a guaranteed error bound; hence, for a given value of m, we can never be absolutely certain that the approximation error is suﬃciently small. Why does Monte Carlo integration work? Monte Carlo integration exploits the fact that any integral can be expressed as the expected value of some realvalued function of a random variable or random vector. Since the WLLN says that sample means approximate population means (with high probability) if the sample c 2000 by Chapman & Hall/CRC
size is suﬃciently large, we can use the appropriate sample mean to approximate any given integral. To illustrate, we will consider evaluating the integral I=
1
g(x) dx. 0
Suppose that a random variable U has a Uniform distribution on [0, 1]. Then 1 E[g(U )] = g(x) dx. 0
If U1 , U2 , · · · are i.i.d. Uniform random variables on [0, 1], the WLLN says that n 1 g(Ui ) →p E[g(U )] = n i=1
which suggests that Carlo estimate
.1 0
1
g(x) dx
as
0
n → ∞,
g(x) dx may be approximated by the Monte I/ =
n 1 g(Ui ) n i=1
if n is suﬃciently large. (In practice, U1 , · · · , Un are pseudorandom variables and so are not truly independent.) Generally, it is not possible to obtain a useful absolute bound on the approximation error + + 1 n +1 + + + g(Ui ) − g(x) dx+ + +n + 0 i=1
.
since this error is random. However, if 01 g 2 (x) dx < ∞, it is possible to use the CLT to make a probability statement about the approximation error. Deﬁning σg2
1
= Var[g(Ui )] =
0
g (x) dx − 2
1
2
g(x) dx
,
0
it follows that
+ + 1 n +1 + aσg + + P + g(Ui ) − g(x) dx+ < √ ≈ Φ(a) − Φ(−a) +n + n 0 i=1
where Φ(x) is the standard Normal distribution function. . The simple Monte Carlo estimate of 01 g(x) dx can be improved in c 2000 by Chapman & Hall/CRC
a number of ways. We will mention two such methods: importance sampling and antithetic sampling. Importance sampling exploits the fact that 1
g(x) dx = 0
1 g(x)
f (x)
0
f (x) dx = E
g(X) f (X)
where the random variable X has density f on [0, 1]. Thus if X1 , · · · , Xn are . 1 i.i.d. random variables with density f , we can approximate 0 g(x) dx by I0 =
n g(Xi ) 1 . n i=1 f (Xi )
.
How do I/ and I0 compare as estimates of I = 01 g(x) dx? In simple terms, the estimates can be assessed by comparing their variances or, equivalently, Var[g(Ui )] and Var[g(Xi )/f (Xi )]. It can be shown that Var[g(Xi )/f (Xi )] is minimized by sampling the Xi ’s from the density g(x) ; f (x) = . 1 0 g(x) dx in practice, however, it may be diﬃcult to generate random variables with this density. However, a signiﬁcant reduction in variance can be obtained if the Xi ’s are sampled from a density f that is approximately proportional to g. (More generally, we .∞ could approximate the integral −∞ g(x) dx by n 1 g(Xi ) n i=1 f (Xi )
where X1 , · · · , Xn are i.i.d. random variables with density f .) Antithetic sampling exploits the fact that 1
g(x) dx = 0
1 2
1 0
(g(x) + g(1 − x)) dx.
Hence if U1 , U2 , · · · are i.i.d. Uniform random variables on [0, 1], we . can approximate 01 g(x) dx by n 1 (g(Ui ) + g(1 − Ui )). 2n i=1
Antithetic sampling is eﬀective when g is a monotone function c 2000 by Chapman & Hall/CRC
(either increasing or decreasing); in this case, it can be shown that Cov[g(Ui ), g(1 − Ui )] ≤ 0. Now comparing the variances of 2n 1 / Is = g(Ui )
2n
and
n 1 / Ia = (g(Ui ) + g(1 − Ui )),
i=1
we obtain
Var(I/s ) =
2n
i=1
Var[g(U1 )] 2n
and Var[g(U1 ) + g(1 − U1 )] 4n Var[g(U1 )] + Cov[g(U1 ), g(1 − U1 )] . 2n
Var(I/a ) = =
Since Cov[g(U1 ), g(1 − U1 )] ≤ 0 when g is monotone, it follows that Var(I/a ) ≤ Var(I/s ). EXAMPLE 3.13: Consider evaluating the integral 1
exp(−x) cos(πx/2) dx; 0
the integrand is shown in Figure 3.10. This integral can be evaluated in closedform as 2 (2 + π exp(−1)) ≈ 0.4551. 4 + π2 We will evaluate this integral using four Monte Carlo approaches. Let U1 , · · · , U1000 be i.i.d. Uniform random variables on the interval [0, 1] and deﬁne the following four Monte Carlo estimates of the integral: I/1 =
1 1000 exp(−Ui ) cos(πUi /2) 1000 i=1
I/2 =
500 1 exp(−Ui ) cos(πUi /2) 1000 i=1
+ I/3 =
500 1 exp(Ui − 1) cos(π(1 − Ui )/2) 1000 i=1
exp(−Vi ) cos(πVi /2) 1 1000 1000 i=1 2(1 − Vi )
c 2000 by Chapman & Hall/CRC
1.0 0.8 0.6 0.0
0.2
0.4
g(x)
0.0
0.2
0.4
0.6
0.8
1.0
x
Figure 3.10 Graph of the function g(x) = exp(−x) cos(πx/2).
I/4 =
exp(−Wi ) cos(πWi /2) 1 1000 1000 i=1 2Wi
where Vi = 1 − (1 − Ui )1/2 and Wi = Ui . I/2 is an antithetic sampling estimate of the integral while I/3 and I/4 are both importance sampling estimates; note that the density of Vi is fV (x) = 2(1 − x) while the density of Wi is fW (x) = 2x (for 0 ≤ x ≤ 1 in both cases). Each of the four estimates was evaluated for 10 samples of U1 , · · · , U1000 and the results presented in Table 3.2. A glance at Table 3.2 reveals that the estimates I/2 and I/3 are the best while I/4 is the clear loser; I/2 comes the closet to the true value 4 times while I/3 comes the closest the other 6 times. It is not surprising that I/3 does so well; from Figure 3.10, we can see that the integrand g(x) = exp(−x) cos(πx/2) is approximately 1 − x so that g(x)/fV (x) ≈ 2 and so I/3 should be close to the optimal importance sampling estimate. The fact that g(x) ≈ 1 − x also explains the success of the antithetic sampling estimate I/2 : g(Ui )+g(1−Ui ) ≈ 1, which suggests that Var(g(Ui )+g(1−Ui )) ≈ 0. 1/2
c 2000 by Chapman & Hall/CRC
Table 3.2 Monte Carlo estimates of
.1 0
exp(−x) cos(πx/2) dx.
I/1
I/2
I/3
I/4
0.4596 0.4706 0.4653 0.4600 0.4496 0.4412 0.4601 0.4563 0.4541 0.4565
0.4522 0.4561 0.4549 0.4559 0.4551 0.4570 0.4546 0.4555 0.4538 0.4534
0.4555 0.4563 0.4560 0.4559 0.4546 0.4527 0.4552 0.4549 0.4555 0.4556
0.4433 0.5882 0.6569 0.4259 0.4907 0.4206 0.4289 0.4282 0.4344 0.4849
(In fact, Var(I/2 ) = 0.97 × 10−6 while Var(I/3 ) = 2.25 × 10−6 and Var(I/1 ) = 90.95 × 10−6 ; the variance of I/4 is inﬁnite.) ✸ 3.7 Convergence with probability 1 Earlier in this chapter, we mentioned the existence of another type of convergence for sequences of random variables, namely convergence with probability 1. In the interest of completeness, we will discuss this type of convergence although we will not make use of it subsequently in the text; therefore, this section can be skipped without loss of continuity. DEFINITION. A sequence of random variables {Xn } converges with probability 1 (or almost surely) to a random variable X (Xn →wp1 X) if
P
ω : lim Xn (ω) = X(ω) n→∞
= 1.
What exactly does convergence with probability 1 mean? By the deﬁnition above, if Xn →wp1 X then Xn (ω) → X(ω) for all outcomes ω ∈ A with P (A) = 1. For a given ω ∈ A and > 0, there exists a number n0 (ω) such that Xn (ω) − X(ω) ≤ for all c 2000 by Chapman & Hall/CRC
n ≥ n0 (ω). Now consider the sequence of sets {Bn ()} with ∞
Bn () =
[Xk − X > ];
k=n
{Bn ()} is a decreasing sequence of sets (that is, Bn+1 () ⊂ Bn ()) and its limit will contain all ω’s lying in inﬁnitely many of the Bn ()’s. If ω ∈ A then for n suﬃciently large Xk (ω)−X(ω) ≤ for k ≥ n and so ω ∈ Bn () (for n suﬃciently large). Thus Bn ()∩A ↓ ∅ as n → ∞. Likewise, if ω ∈ A then ω will lie in inﬁnitely many of the Bn ’s and so Bn ∩ Ac ↓ Ac . Thus P (Bn ())
= P (Bn () ∩ A) + P (Bn () ∩ Ac ) → lim P (Bn () ∩ A) + lim P (Bn () ∩ Ac ) n→∞
=
n→∞
0
Thus Xn →wp1 X implies that P (Bn ()) → 0 as n → ∞ for all > 0. Conversely, if P (Bn ()) → 0 then using the argument given above, it follows that Xn →wp1 X. Thus Xn →wp1 X is equivalent to lim P
n→∞
∞
[Xk − X > ]
=0
k=n
for all > 0. Using the condition given above, it is easy to see that if Xn →wp1 X then Xn →p X; this follows since ∞
[Xn − X > ] ⊂
[Xk − X > ]
k=n
and so P (Xn − X > ) ≤ P
∞
[Xk − X > ]
→ 0.
k=n
Note that if [Xn+1 − X > ] ⊂ [Xn − X > ] for all n then P
∞
[Xk − X > ]
= P (Xn − X > )
k=n
in which case Xn →p X implies that Xn →wp1 X. c 2000 by Chapman & Hall/CRC
EXAMPLE 3.14: Suppose that X1 , X2 , · · · are i.i.d. Uniform random variables on the interval [0, 1] and deﬁne Mn = max(X1 , · · · , Xn ). In Example 3.1, we showed that Mn →p 1. However, note that 1 ≥ Mn+1 (ω) ≥ Mn (ω) for all ω and so [Mn+1 − 1 > ] = [Mn+1 < 1 − ] ⊂ [Mn < 1 − ] = [Mn − 1 > ]. Thus P
∞
[Mk − 1 > ]
= P (Mn − 1 > ) → 0
as n → ∞
k=n
as shown in Example 3.1. Thus Mn →wp1 1.
✸
Example 3.14 notwithstanding, it is, in general, much more diﬃcult to prove convergence with probability 1 than it is to prove convergence in probability. However, assuming convergence with probability 1 rather than convergence in probability in theorems (for example, in Theorems 3.2 and 3.3) can sometimes greatly facilitate the proofs of these results. EXAMPLE 3.15: Suppose that Xn →wp1 X and g(x) is a continuous function. Then g(Xn ) →wp1 g(X). To see this, let A be the set of ω’s for which Xn (ω) → X(ω). Since g is a continuous function Xn (ω) → X(ω) implies that g(Xn (ω)) → g(X(ω)); this occurs for all ω’s in the set A with P (A) = 1 and so g(Xn ) →wp1 g(X). ✸ We can also extend the WLLN to the socalled Strong Law of Large Numbers (SLLN). THEOREM 3.13 (Strong Law of Large Numbers) Suppose that X1 , X2 , · · · are i.i.d. random variables with E(Xi ) < ∞ and E(Xi ) = µ. Then n ¯n = 1 X Xi →wp1 µ n i=1
as n → ∞. c 2000 by Chapman & Hall/CRC
The SLLN was proved by Kolmogorov (1930). Its proof is more diﬃcult than that of the WLLN but similar in spirit; see, for example, Billingsley (1995) for details. There is a very interesting connection between convergence with probability 1 and convergence in distribution. It follows that convergence with probability 1 implies convergence in distribution; the converse is not true (since Xn →d X means that the distribution functions converge and so all the Xn ’s can be deﬁned on diﬀerent sample spaces). However, there is a partial converse that is quite useful technically. Suppose that Xn →d X; thus Fn (x) → F (x) for all continuity points of F . Then it is possible to deﬁne random variables {Xn∗ } and X ∗ with Xn∗ ∼ Fn and X ∗ ∼ F such that Xn∗ →wp1 X ∗ . Constructing these random variables is remarkably simple. Let U be a Uniform random variable on the interval [0, 1] and deﬁne Xn∗ = Fn−1 (U )
and X ∗ = F −1 (U )
where Fn−1 and F −1 are the inverses of the distribution functions of Xn and X respectively. It follows now that Xn∗ ∼ Fn and X ∗ ∼ F . Moreover, Xn∗ →wp1 X ∗ ; the proof of this fact is left as an exercise but seems reasonable given that Fn (x) → F (x) for all but (at most) a countable number of x’s. This representation is due to Skorokhod (1956). As mentioned above, the ability to construct these random variables {Xn∗ } and X ∗ is extremely useful from a technical point of view. In the following examples, we give elementary proofs of the Continuous Mapping Theorem (Theorem 3.2) and the Delta Method (Theorem 3.4). EXAMPLE 3.16: Suppose that Xn →d X and g(x) is a continuous function. We can then construct random variables {Xn∗ } and X ∗ such that Xn∗ =d Xn and X ∗ =d X and Xn∗ →wp1 X ∗ . Since g is continuous, we have g(Xn∗ ) →wp1 g(X ∗ ). However, g(Xn∗ ) =d g(Xn ) and g(X ∗ ) =d g(X) and so it follows that g(Xn ) →d g(X) since g(Xn∗ ) →d g(X ∗ ). c 2000 by Chapman & Hall/CRC
✸
EXAMPLE 3.17: Suppose that Zn = an (Xn − θ) →d Z where an ↑ ∞. Let g(x) be a function that is diﬀerentiable at x = θ. We construct the random variables {Zn∗ } and Z ∗ having the same distributions as {Zn } and Z with Zn∗ →wp1 Z ∗ . We can also deﬁne ∗ Xn∗ = a−1 n Zn + θ which will have the same distribution as Xn ; ∗ clearly, Xn →wp1 θ. Thus an (g(Xn∗ )
− g(θ))
= →wp1
since
g(Xn∗ ) − g(θ) an (Xn∗ − θ) Xn∗ − θ g (θ)Z ∗
g(Xn∗ ) − g(θ) →wp1 g (θ). Xn∗ − θ
Now since an (g(Xn∗ ) − g(θ)) and an (g(Xn ) − g(θ)) have the same distribution as do g (θ)Z ∗ and g (θ)Z, it follows that an (g(Xn ) − g(θ)) →d g (θ)Z. Note that we have required only existence (and not continuity) of the derivative of g(x) at x = θ. ✸ 3.8 Problems and complements (1)
(k)
3.1: (a) Suppose that {Xn }, · · · , {Xn } are sequences of random (i) variables with Xn →p 0 as n → ∞ for each i = 1, · · · , k. Show that max Xn(i)  →p 0 1≤i≤k
as n → ∞. (b) Find an example to show that the conclusion of (a) is not necessarily true if the number of sequences k = kn → ∞. 3.2: Suppose that X1 , · · · , Xn are i.i.d. random variables with a distribution function F (x) satisfying lim xα (1 − F (x)) = λ > 0
x→∞
for some α > 0. Let Mn = max(X1 , · · · , Xn ). We want to show that n−1/α Mn has a nondegenerate limiting distribution. (a) Show that n[1−F (n1/α x)] → λx−α as n → ∞ for any x > 0. c 2000 by Chapman & Hall/CRC
(b) Show that
P n−1/α Mn ≤ x
=
[F (n1/α )]n
= [1 − (1 − F (n1/α ))]n → 1 − exp −λx−α as n → ∞ for any x > 0. (c) Show that P (n−1/α Mn ≤ 0) → 0 as n → ∞. (d) Suppose that the Xi ’s have a Cauchy distribution with density function 1 f (x) = . π(1 + x2 ) Find the value of α such that n−1/α Mn has a nondegenerate limiting distribution and give the limiting distribution function. 3.3: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ and let Mn = max(X1 , · · · , Xn ). Show that Mn − ln(n)/λ →d V where P (V ≤ x) = exp[− exp(−λx)] for all x. 3.4: Suppose that X1 , · · · , Xn are i.i.d. nonnegative random variables with distribution function F . Deﬁne Un = min(X1 , · · · , Xn ). (a) Suppose that F (x)/x → λ as x → 0. Show that nUn →d Exp(λ). (b) Suppose that F (x)/xα → λ as x → 0 for some α > 0. Find the limiting distribution of n1/α Un . 3.5: Suppose that XN has a Hypergeometric distribution (see Example 1.13) with the following frequency function
fN (x) =
MN x
N − MN rN − x
N rN
for x = max(0, rN + MN − N ), · · · , min(MN , rN ). When the population size N is large, it becomes somewhat diﬃcult to compute probabilities using fN (x) so that it is desirable to ﬁnd approximations to the distribution of XN as N → ∞. c 2000 by Chapman & Hall/CRC
(a) Suppose that rN → r (ﬁnite) and MN /N → θ for 0 < θ < 1. Show that XN →d Bin(r, θ) as N → ∞ (b) Suppose that rN → ∞ with rN MN /N → λ > 0. Show that XN →d Pois(λ) as N → ∞. 3.6: Suppose that {Xn } and {Yn } are two sequences of random variables such that an (Xn − Yn ) →d Z for a sequence of numbers {an } with an → ∞ (as n → ∞). (a) Suppose that Xn →p θ. Show that Yn →p θ. (b) Suppose that Xn →p θ and g(x) is a function continuously diﬀerentiable at x = θ. Show that an (g(Xn ) − g(Yn )) →d g (θ)Z. 3.7: (a) Let {Xn } be a sequence of random variables. Suppose that E(Xn ) → θ (where θ is ﬁnite) and Var(Xn ) → 0. Show that Xn →p θ. (b) A sequence of random variables {Xn } converges in probability to inﬁnity (Xn →p ∞) if for each M > 0, lim P (Xn ≤ M ) = 0.
n→∞
Suppose that E(Xn ) → ∞ and Var(Xn ) ≤ kE(Xn ) for some k < ∞. Show that Xn →p ∞. (Hint: Use Chebyshev’s inequality to show that P [Xn − E(Xn ) > E(Xn )] → 0 for each > 0.) 3.8: (a) Let g be a nonnegative even function (g(x) = g(−x)) that is increasing on [0, ∞) and suppose that E[g(X)] < ∞. Show that E[g(X)] P [X > ] ≤ g() for any > 0. (Hint: Follow the proof of Chebyshev’s inequality making the appropriate changes.) (b) Suppose that E[Xn r ] → 0 as n → ∞. Show that Xn →p 0. 3.9: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ. By the CLT, √ ¯ n − λ) →d N (0, λ). n(X c 2000 by Chapman & Hall/CRC
(a) Find the limiting distribution of (b) Find a function g such that √
√
¯ n ) − ln(λ)). n(ln(X
¯ n ) − g(λ)) →d N (0, 1). n(g(X
3.10: Let {an } be a sequence of constants with an → ∞ and suppose that an (Xn − θ) →d Z where θ is a constant. (a) Let g be a function that is twice diﬀerentiable at θ and suppose that g (θ) = 0. Show that a2n (g(Xn ) − g(θ)) →d
1 2 Z . 2
(b) Now suppose that g is k times diﬀerentiable at θ with g (θ) = · · · = g (k−1) (θ) = 0. Find the limiting distribution of akn (g(Xn ) − g(θ)). (Hint: Expand g(Xn ) in a Taylor series around θ.) 3.11: The sample median of i.i.d. random variables is asymptotically Normal provided that the distribution function F has a positive derivative at the median; when this condition fails, an asymptotic distribution may still exist but will be nonNormal. To illustrate this, let X1 , · · · , Xn be i.i.d. random variables with density 1 f (x) = x−2/3 for x ≤ 1. 6 (Notice that this density has a singularity at 0.) (a) Evaluate the distribution function of Xi and its inverse (the quantile function). (b) Let Mn be the sample median of X1 , · · · , Xn . Find the limiting distribution of n3/2 Mn . (Hint: use the extension of the Delta Method in Problem 3.10 by applying the inverse transformation from part (a) to the median of n i.i.d. Uniform random variables on [0, 1].) 3.12: Suppose that X1 , · · · , Xn are i.i.d. random variables with common density f (x) = αx−α−1 c 2000 by Chapman & Hall/CRC
for x ≥ 1
where α > 0. Deﬁne Sn =
n
1/n
Xi
.
i=1
(a) Show that ln(Xi ) has an Exponential distribution. (b) Show that Sn →p exp(1/α). (Hint: Consider ln(Sn ).) (c) Suppose α = 10 and n = 100. Evaluate P (Sn > 1.12) using an appropriate approximation. 3.13: Suppose that X1 , · · · , Xn be i.i.d. discrete random variables with frequency function x for x = 1, 2, · · · , 6. f (x) = 21 (a) Let Sn =
n
k=1 kXk .
Show that
(Sn − E(Sn )) →d N (0, 1). Var(Sn ) (b) Suppose n = 20. Use a Normal approximation to evaluate P (S20 ≥ 1000). (c) Suppose n = 5. Compute the exact distribution of Sn using the probability generating function of Sn (see Problems 1.18 and 2.8). 3.14: Suppose that Xn1 , Xn2 , · · · , Xnn are independent random variables with P (Xni = 0) = 1 − pn Suppose that γ(t) = neighbourhood of 0.
and P (Xni ≤ xXni = 0) = F (x).
.∞
−∞ exp(tx) dF (x)
< ∞ for t in a
(a) Show that the moment generating function of Xni is mn (t) = pn γ(t) + (1 − pn )
(b) Let Sn = ni=1 Xni and suppose that npn → λ > 0 as n → ∞. Show that Sn converges in distribution to a random variable S that has a compound Poisson distribution. (Hint: See Problem 2.7 for the moment generating function of a compound Poisson distribution.) 3.15: Suppose that Xn1 , Xn2 , · · · , Xnn are independent Bernoulli c 2000 by Chapman & Hall/CRC
random variables with parameters θn1 , · · · , θnn respectively. Deﬁne Sn = Xn1 + Xn2 + · · · + Xnn . (a) Show that the moment generating function of Sn is n
mn (t) =
(1 − θni + θni exp(t)) .
i=1
(b) Suppose that n
θni → λ > 0
and
i=1
max θni → 0
1≤i≤n
as n → ∞. Show that ln mn (t) = λ[exp(t) − 1] + rn (t) where for each t, rn (t) → 0 as n → ∞. (Hint: Use the fact that ln(1 + x) = x − x2 /2 + x3 /3 + · · · for x < 1.) (c) Deduce from part (b) that Sn →d Pois(λ). 3.16: Suppose that {Xn } is a sequence of nonnegative continuous random variables and suppose that Xn has hazard function λn (x). .Suppose that for each x, λn (x) → λ0 (x) as n → ∞ where 0∞ λ0 (x) dx = ∞. Show that Xn →d X where
P (X > x) = exp −
x 0
λ0 (t) dt
3.17: Suppose that X1 , · · · , Xn are independent nonnegative random variables with hazard functions λ1 (x), · · · , λn (x) respectively. Deﬁne Un = min(X1 , · · · , Xn ). (a) Suppose that for some α > 0, n 1 λi (t/nα ) = λ0 (t) n→∞ nα i=1
lim
.
for all t > 0 where 0∞ λ0 (t) dt = ∞. Show that nα Un →d V
x where P (V > x) = exp − λ0 (t) dt . 0
(b) Suppose that X1 , · · · , Xn are i.i.d. Weibull random variables (see Example 1.19) with density function
f (x) = λβxβ−1 exp −λxβ c 2000 by Chapman & Hall/CRC
(x > 0)
where λ, α > 0. Let Un = min(X1 , · · · , Xn ) and ﬁnd α such that nα Un →d V . 3.18: Suppose that Xn ∼ χ2 (n). √ √ (a) Show that Xn − n →d N (0, 1/2) as n → ∞. (Hint: Recall that Xn can be represented as a sum of n i.i.d. random variables.) (b) Suppose that n = 100. Use the result in part (a) to approximate P (Xn > 110). 3.19: Suppose that {Xn } is a sequence of random variables such that Xn →d X where E(X) is ﬁnite. We would like to investigate suﬃcient conditions under which E(Xn ) → E(X) (assuming that E(Xn ) is welldeﬁned). Note that in Theorem 3.5, we indicated that this convergence holds if the Xn ’s are uniformly bounded. (a) Let δ > 0. Show that E(Xn 1+δ ) = (1 + δ)
∞ 0
xδ P (Xn  > x) dx.
(b) Show that for any M > 0 and δ > 0, M 0
P (Xn  > x) dx ≤ E(Xn ) ≤
M 0
P (Xn  > x) dx
1 + δ M
∞ M
xδ P (Xn  > x) dx.
(c) Again let δ > 0 and suppose that E(Xn 1+δ ) ≤ K < ∞ for all n. Assuming that Xn →d X, use the results of parts (a) and (b) to show that E(Xn ) → E(X) and E(Xn ) → E(X) as n → ∞. (Hint: Use the fact that M 0
P (Xn  > x) − P (X > x) dx → 0
as n → ∞ for each ﬁnite M .) 3.20: A sequence of random variables {Xn } is said to be bounded in probability if for every > 0, there exists M0 < ∞ such that P (Xn  > M0 ) < for all n. (a) If Xn →d X, show that {Xn } is bounded in probability. c 2000 by Chapman & Hall/CRC
(b) If E(Xn r ) ≤ M < ∞ for some r > 0, show that {Xn } is bounded in probability. (c) Suppose that Yn →p 0 and {Xn } is bounded in probability. Show that Xn Yn →p 0. 3.21: If {Xn } is bounded in probability, we often write Xn = Op (1). Likewise, if Xn →p 0 then Xn = op (1). This useful shorthand notation generalizes the bigoh and littleoh notation that is commonly used for sequences of numbers to sequences of random variables. If Xn = Op (Yn ) (Xn = op (Yn )) then Xn /Yn = Op (1) (Xn /Yn = op (1)). (a) Suppose that Xn = Op (1) and Yn = op (1). Show that Xn + Yn = Op (1). (b) Let {an } and {bn } be sequences of constants where an /bn → 0 as n → ∞ (that is, an = o(bn )) and suppose that Xn = Op (an ). Show that Xn = op (bn ). 3.22: Suppose that {Ui } is an inﬁnite sequence of i.i.d. random variables with mean 0 and variance 1, and deﬁne {Xi } by Xi =
∞
cj Ui−j
j=0
where we assume that ∞ j=0 cj  < ∞ to guarantee that the inﬁnite summation is welldeﬁned.
(a) Deﬁne c0j = ∞ k=j+1 ck and deﬁne Zi =
∞
c0j Ui−j
j=0
and assume that that
∞
02j j=0 c
Xi =
< ∞ (so that Zi is welldeﬁned). Show
∞
cj Ui + Zi − Zi−1 .
j=0
This decomposition is due to Beveridge and Nelson (1981). (b) Using the decomposition in part (a), show that n 1 √ Xi →d N (0, σ 2 ) n i=1 c 2000 by Chapman & Hall/CRC
where
σ2 =
∞
2
cj .
j=0
3.23: Suppose that A1 , A2 , · · · is a sequence of events. We are sometimes interested in determining the probability that inﬁnitely many of the Ak ’s occur. Deﬁne the event ∞ ∞
B=
Ak .
n=1 k=n
It is possible to show that an outcome lies in B if, and only if, it belongs to inﬁnitely many of the Ak ’s. (To see this, ﬁrst suppose that an outcome ωlies in inﬁnitely many of the Ak ’s. Then it belongs to Bn = ∞ k=n Ak for each n ≥ 1 and hence ∞ in B = n=1 Bn . On the other hand, suppose that ω lies in B; then it belongs to Bn for all n ≥ 1. If ω were in only a ﬁnite number of Ak ’s, there would exist a number m such that Ak did not contain ω for k ≥ m. Hence, ω would not lie in Bn for n ≥ m and so ω would not lie in B. This is a contradiction, so ω must lie in inﬁnitely many of the Ak ’s.)
(a) Prove the ﬁrst BorelCantelli Lemma: If ∞ k=1 P (Ak ) < ∞ then P (Ak inﬁnitely often) = P (B) = 0. (Hint: note that B ⊂ Bn for any n and so P (B) ≤ P (Bn ).) (b) When the Ak ’s are mutually independent, we can strengthen the ﬁrst BorelCantelli Lemma. Suppose that ∞
P (Ak ) = ∞
k=1
for mutually independent events {Ak }. Show that P (Ak inﬁnitely often) = P (B) = 1; this result is called the second BorelCantelli Lemma. (Hint: Note that c
B =
∞ ∞
n=1 k=n c 2000 by Chapman & Hall/CRC
Ack
and so P (B ) ≤ P c
∞
Ack
k=n
=
∞
(1 − P (Ak )).
k=n
Now use the fact that ln(1 − P (Ak )) ≤ −P (Ak ).) 3.24: Suppose that {Xk } is an inﬁnite sequence of identically distributed random variables with E(Xk ) < ∞. (a) Show that for > 0,
+ +
+ +
Xk + P ++ > inﬁnitely often = 0. k + (From this, it follows that Xn /n → 0 with probability 1 as n → ∞.) (b) Suppose that the Xk ’s are i.i.d. Show that Xn /n → 0 with probability 1 if, and only if, E(Xk ) < ∞. 3.25: Suppose that X1 , X2 , · · · are i.i.d. random variables with E(Xi ) = 0 and E(Xi4 ) < ∞. Deﬁne n ¯n = 1 X Xi . n i=1
¯ n 4 ] ≤ k/n2 for some constant k. (Hint: To (a) Show that E[X evaluate n n n n
E [Xi Xj Xk X ]
i=1 j=1 k=1 =1
note that most of the n4 terms in the fourfold summation are exactly 0.) (b) Using the ﬁrst BorelCantelli Lemma, show that ¯ n →wp1 0. X (This gives a reasonably straightforward proof of the SLLN albeit under much stronger than necessary conditions.)
c 2000 by Chapman & Hall/CRC
CHAPTER 4
Principles of Point Estimation 4.1 Introduction To this point, we have assumed (implicitly or explicitly) that all the parameters necessary to make probability calculations for a particular probability model are available to us. Thus, for example, we are able to calculate the probability that a given event occurs either exactly or approximately (with the help of limit theorems). In statistics, however, the roles of parameters (of the probability model) and outcomes (of the experiment) are somewhat reversed; the outcome of the experiment is observed by the experimenter while the true value of the parameter (or more generally, the true probability distribution) is unknown to the experimenter. In very broad terms, the goal of statistics is to use the outcome of the experiment (that is, the data from the experiment) to make inference about the values of the unknown parameters of the assumed underlying probability distribution. The previous paragraph suggests that no ambiguity exists regarding the probability model for a given experiment. However, in “real life” statistical problems, there may be considerable uncertainty as to the choice of the appropriate probability model and the model is only chosen after the data have been observed. Moreover, in many (perhaps almost all) problems, it must be recognized that any model is, at best, an approximation to reality; it is important for a statistician to verify that any assumed model is more or less close to reality and to be aware of the consequences of misspecifying a model. A widely recognized philosophy in statistics (and in science more generally) is that a model should be as simple as possible. This philosophy is often expressed by the maxim known as Occam’s razor (due to the philosopher William of Occam): “explanations should not be multiplied beyond necessity”. In terms of statistical modeling, Occam’s razor typically means that we should prefer a c 2000 by Chapman & Hall/CRC
model with few parameters to one with many parameters if the data are explained equally well by both. There are several philosophies of statistical inference; we will crudely classify these into two schools, the Frequentist school and the Bayesian school. The Frequentist approach to inference is perhaps the most commonly used in practice but is, by no means, superior (or inferior) to the Bayesian approach. Frequentist methods assume (implicitly) that any experiment is inﬁnitely repeatable and that we must consider all possible (but unobserved) outcomes of the experiment in order to carry out statistical inference. In other words, the uncertainty in the outcome of the experiment is used to describe the uncertainty about the parameters of the model. In contrast, Bayesian methods depend only on the observed data; uncertainty about the parameters is described via probability distributions that depend on these data. However, there are Frequentist methods that have a Bayesian ﬂavour and vice versa. In this book, we will concentrate on Frequentist methods although some exposure will be given to Bayesian methods. 4.2 Statistical models Let X1 , · · · , Xn be random variables (or random vectors) and suppose that we observe x1 , · · · , xn , which can be thought of as outcomes of the random variables X1 , · · · , Xn . Suppose that the joint distribution of X = (X1 , · · · , Xn ) is unknown but belongs to some particular family of distributions. Such a family of distributions is called a statistical model. Although we usually assume that X is observed, it is also possible to talk about a model for X even if some or all of the Xi ’s are not observable. It is convenient to index the distributions belonging to a statistical model by a parameter θ; θ typically represents the unknown or unspeciﬁed part of the model. We can then write X = (X1 , · · · , Xn ) ∼ Fθ
for θ ∈ Θ
where Fθ is the joint distribution function of X and Θ is the set of possible values for the parameter θ; we will call the set Θ the parameter space. In general, θ can be either a single realvalued parameter or a vector of parameters; in this latter case, we will often write θ to denote a vector of parameters (θ1 , · · · , θp ) to emphasize that we have a vectorvalued parameter. Whenever it is not notationally cumbersome to do so, we c 2000 by Chapman & Hall/CRC
will write (for example) Pθ (A), Eθ (X), and Varθ (X) to denote (respectively) probability, expected value, and variance with respect to a distribution with unknown parameter θ. The reasons for doing this are purely stylistic and mainly serve to emphasize the dependence of these quantities on the parameter θ. We usually assume that Θ is a subset of some Euclidean space so that the parameter θ is either real or vectorvalued (in the vector case, we will write θ = (θ1 , · · · , θk )); such a model is often called a parametric model in the sense that the distributions belonging to the model can be indexed by a ﬁnite dimensional parameter. Models whose distributions cannot be indexed by a ﬁnite dimensional parameter are often (somewhat misleadingly) called nonparametric models; the parameter space for such models is typically inﬁnite dimensional. However, for some nonparametric models, we can express the parameter space Θ = Θ1 × Θ2 where Θ1 is a subset of a Euclidean space. (Such models are sometimes called semiparametric models.) For a given statistical model, a given parameter θ corresponds to a single distribution Fθ . However, this does not rule out the possibility that there may exist distinct parameter values θ1 and θ2 such that Fθ1 = Fθ2 . To rule out this possibility, we often require that a given model, or more precisely, its parametrization be identiﬁable; a model is said to have an identiﬁable parametrization (or to be an identiﬁable model) if Fθ1 = Fθ2 implies that θ1 = θ2 . A nonidentiﬁable parametrization can lead to problems in estimation of the parameters in the model; for this reason, the parameters of an identiﬁable model are often called estimable. Henceforth unless stated otherwise, we will assume implicitly that any statistical model with which we deal is identiﬁable. EXAMPLE 4.1: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ. The joint frequency function of X = (X1 , · · · , Xn ) is f (x; λ) =
n exp(−λ)λxi i=1
xi !
for x1 , · · · , xn = 0, 1, 2, · · ·. The parameter space for this parametric model is {λ : λ > 0}. ✸ EXAMPLE 4.2: Suppose that X1 , · · · , Xn are i.i.d. random variables with a continuous distribution function F that is unknown. c 2000 by Chapman & Hall/CRC
The parameter space for this model consists of all possible continuous distributions. These distributions cannot be indexed by a ﬁnite dimensional parameter and so this model is nonparametric. We may also assume that F (x) has a density f (x − θ) where θ is an unknown parameter and f is an unknown density function satisfying f (x) = f (−x). This model is also nonparametric but depends on the realvalued parameter θ. (This might be considered a semiparametric model because of the presence of θ.) ✸ EXAMPLE 4.3: Suppose that X1 , · · · , Xn are independent Normal random variables with E(Xi ) = β0 + β1 ti + β2 si (where t1 , · · · , tn and s1 , · · · , sn are known constants) and Var(Xi ) = σ 2 ; the parameter space is {(β0 , β1 , β2 , σ) : −∞ < β0 , β1 , β2 < ∞, σ > 0}. We will see that the parametrization for this model is identiﬁable if, and only if, the vectors
1 t1 .. .. z0 = . , z1 = . , tn 1
s1 .. and z 2 = . sn
are linearly independent, that is, a0 z 0 + a1 z 1 + a2 z 2 = 0 implies that a0 = a1 = a2 = 0. To see why this is true, let
E(X1 ) .. µ= . E(Xn ) and note that the parametrization is identiﬁable if there is a onetoone correspondence between the possible values of µ and the parameters β0 , β1 , β2 . Suppose that z 0 , z 1 , and z 2 are linearly dependent; then a0 z 0 + a1 z 1 + a2 z 2 = 0 where at least one of a0 , a1 , or a2 is nonzero. In this case, we would have µ = β0 z 0 + β1 z 1 + β2 z 2 = (β0 + a0 )z 0 + (β1 + a1 )z 1 + (β2 + a2 )z 2 and thus there is not a onetoone correspondence between µ and (β0 , β1 , β2 ). However, when z 0 , z 1 , and z 2 are linearly dependent, it is possible to obtain an identiﬁable parametrization by restricting c 2000 by Chapman & Hall/CRC
the parameter space; this is usually achieved by putting constraints on the parameters β0 , β1 , and β2 . ✸ Exponential families One important class of statistical models is exponential family models. Suppose that X1 , · · · , Xn have a joint distribution Fθ where θ = (θ1 , · · · , θp ) is an unknown parameter. We say that the family of distributions {Fθ } is a kparameter exponential family if the joint density or joint frequency function of (X1 , · · · , Xn ) is of the form f (x; θ) = exp
, k

ci (θ)Ti (x) − d(θ) + S(x)
i=1
for x = (x1 , · · · , xn ) ∈ A where A does not depend on the parameter θ. It is important to note that k need not equal p, the dimension of θ, although, in many cases, they are equal. EXAMPLE 4.4: Suppose that X has a Binomial distribution with parameters n and θ where θ is unknown. Then the frequency function of X is
n x θ (1 − θ)n−x x
f (x; θ) =
,

θ n = exp ln x + n ln(1 − θ) + ln 1−θ x
for xinA = {0, 1, · · · , n} and so the distribution of X has a oneparameter exponential family. ✸ EXAMPLE 4.5: Suppose that X1 , · · · , Xn are i.i.d. Gamma random variables with unknown shape parameter α and unknown scale parameter λ. Then the joint density function of X = (X1 , · · · , Xn ) is f (x; α, λ) =
, n λα xα−1 exp(−λxi ) i
i=1
,
Γ(α)
= exp (α − 1)
n
ln(xi ) − λ
i=1 c 2000 by Chapman & Hall/CRC
n i=1

xi + nα ln(λ) − n ln(Γ(α))
(for x1 , · · · , xn > 0) and so the distribution of X is a twoparameter exponential family. ✸ EXAMPLE 4.6: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean θ and variance θ2 where θ > 0. The joint density function of (X1 , · · · , Xn ) is f (x; θ) =
n )
i=1
1 1 √ exp − 2 (xi − θ)2 2θ θ 2π
*
,

n n 1 1 n = exp − 2 x2i + xi − (1 + ln(θ2 ) + ln(2π)) 2θ i=1 θ i=1 2
and so A = Rn . Note that this is a twoparameter exponential family despite the fact that the parameter space is onedimensional. ✸ EXAMPLE 4.7: Suppose that X1 , · · · , Xn are independent Poisson random variables with E(Xi ) = exp(α + βti ) where t1 , · · · , tn are known constants. Setting X = (X1 , · · · , Xn ), the joint frequency function of X is f (x; α, β) =
* n ) exp(− exp(α + βti )) exp(αxi + βxi ti )
i=1
,
= exp α
x! n
xi + β
i=1
n i=1
xi ti +
n
exp(α + βti ) −
i=1
n

ln(xi !) .
i=1
This is a twoparameter exponential family model; the set A is ✸ simply {0, 1, 2, 3, · · ·}n . EXAMPLE 4.8: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on the interval [0, θ]. The joint density function of X = (X1 , · · · , Xn ) is f (x; θ) =
1 θn
for 0 ≤ x1 , · · · , xn ≤ θ.
The region on which f (x; θ) is positive clearly depends on θ and so this model is not an exponential family model. ✸ The following result will prove to be useful in the sequel. c 2000 by Chapman & Hall/CRC
PROPOSITION 4.1 Suppose that X = (X1 , · · · , Xn ) has a oneparameter exponential family distribution with density or frequency function f (x; θ) = exp [c(θ)T (x) − d(θ) + S(x)] for x ∈ A where (a) the parameter space Θ is open, (b) c(θ) is a onetoone function on Θ, (c) c(θ), d(θ) are twice diﬀerentiable functions on Θ. Then Eθ [T (X)] = and
Varθ [T (X)] =
d (θ) c (θ) d (θ)c (θ) − d (θ)c (θ) . [c (θ)]3
Proof. Deﬁne φ = c(θ); φ is called the natural parameter of the exponential family. Let d0 (φ) = d(c−1 (φ)) where c−1 is welldeﬁned since c is a onetoone continuous function on Θ. Then for s suﬃciently small (so that φ+s lies in the natural parameter space), we have (Problem 4.1) Eφ [exp(sT (X))] = exp[d0 (φ + s) − d0 (φ)], which is the moment generating function of T (X). Diﬀerentiating and setting s = 0, we get Eφ [T (X)] = d0 (φ)
and Varφ [T (X)] = d0 (φ).
Now note that d0 (φ) = and d0 (φ) =
d (θ) c (θ) d (θ)c (θ) − d (θ)c (θ) [c (θ)]3
and so the conclusion follows. Proposition 4.1 can be extended to ﬁnd the means, variances and covariances of the random variables T1 (X), · · · , Tk (X) in kparameter exponential family models; see Problem 4.2. Statistics Suppose that the model for X = (X1 , · · · , Xn ) has a parameter space Θ. Since the true value of the parameter θ (or, equivalently, c 2000 by Chapman & Hall/CRC
the true distribution of X) is unknown, we would like to summarize the available information in X without losing too much information about the unknown parameter θ. At this point, we are not interested in estimating θ per se but rather in determining how to best use the information in X. We will start by attempting to summarize the information in X. Deﬁne a statistic T = T (X) to be a function of X that does not depend on any unknown parameter; that is, the statistic T depends only on observable random variables and known constants. A statistic can be real or vectorvalued. ¯ = n−1 n Xi . Since n (the sample EXAMPLE 4.9: T (X) = X i=1 size) is known, T is a statistic. ✸ EXAMPLE 4.10: T (X) = (X(1) , · · · , X(n) ) where X(1) ≤ X(2) ≤ · · · ≤ X(n) are the order statistics of X. Since T depends only on the values of X, T is a statistic. ✸ It is important to note that any statistic is itself a random variable and so has its own probability distribution; this distribution may or may not depend on the parameter θ. Ideally, a statistic T = T (X) should contain as much information about θ as X does. However, this raises several questions. For example, how does one determine if T and X contain the same information about θ? How do we ﬁnd such statistics? Before attempting to answer these questions, we will deﬁne the concept of ancillarity. DEFINITION. A statistic T is an ancillary statistic (for θ) if its distribution is independent of θ; that is, for all θ ∈ Θ, T has the same distribution. EXAMPLE 4.11: Suppose that X1 and X2 are independent Normal random variables each with mean µ and variance σ 2 (where σ 2 is known). Let T = X1 − X2 ; then T has a Normal distribution with mean 0 and variance 2σ 2 . Thus T is ancillary for the unknown parameter µ. However, if both µ and σ 2 were unknown, T would not be ancillary for θ = (µ, σ 2 ). (The distribution of T depends on σ 2 so T contains some information about σ 2 .) ✸ EXAMPLE 4.12: Suppose that X1 , · · · , Xn are i.i.d. random variables with density function 1 f (x; µ, θ) = for µ − θ ≤ x ≤ µ + θ. 2θ c 2000 by Chapman & Hall/CRC
Deﬁne a statistic R = X(n) − X(1) , which is the sample range of X1 , · · · , Xn . The density function of R is
n(n − 1)xn−2 x 1− fR (x) = n−1 (2θ) 2θ
for 0 ≤ x ≤ 2θ,
which depends on θ but not µ. Thus R is ancillary for µ.
✸
Clearly, if T is ancillary for θ then T contains no information about θ. In other words, if T is to contain any useful information about θ, its distribution must depend explicitly on θ. Moreover, intuition also tells us that the amount of information contained will increase as the dependence of the distribution on θ increases. EXAMPLE 4.13: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on the interval [0, θ] where θ > 0 is an unknown parameter. Deﬁne two statistics, S = min(X1 , · · · , Xn ) and T = max(X1 , · · · , Xn ). The density of S is
n x 1− fS (x; θ) = θ θ while the density of T is n fT (x; θ) = θ
n−1
n−1 x
θ
for
0≤x≤θ
for 0 ≤ x ≤ θ.
Note that the densities of both S and T depend on θ and so neither is ancillary for θ. However, as n increases, it becomes clear that the density of S is concentrated around 0 for all possible values of θ while the density of T is concentrated around θ. This seems to indicate that T provides more information about θ than does S. ✸ Example 4.13 suggests that not all nonancillary statistics are created equal. In the next section, we will elaborate on this observation. 4.3 Suﬃciency The notion of suﬃciency was developed by R.A. Fisher in the early 1920s. The ﬁrst mention of suﬃciency was made by Fisher (1920) in which he considered the estimation of the variance σ 2 of a Normal distribution based on i.i.d. observations X1 , · · · , Xn . (This is formalized in Fisher (1922).) In particular, he considered c 2000 by Chapman & Hall/CRC
estimating σ 2 based on the statistics T1 =
n
¯ and T2 = Xi − X
i=1
n
¯ 2 (Xi − X)
i=1
¯ is the average of X1 , · · · , Xn ). Fisher showed that the (where X distribution of T1 conditional on T2 = t does not depend on the parameter σ while the distribution of T2 conditional on T1 = t does depend on σ. He concluded that all the information about σ 2 in the sample was contained in the statistic T2 and that any estimate of σ 2 should be based on T2 ; that is, any estimate of σ 2 based on T1 could be improved by using the information in T2 while T2 could not be improved by using T1 . We will now try to elaborate on Fisher’s argument in a more general context. Suppose that X = (X1 , · · · , Xn ) ∼ Fθ for some θ ∈ Θ and let T = T (X) be a statistic. For each t in the range of T , deﬁne the level sets of T At = {x : T (x) = t}. Now look at the distribution of X on the set At , that is, the conditional distribution of X given T = t. If this conditional distribution is independent of θ then X contains no information about θ on the set At ; that is, X is an ancillary statistic on At . If this is true for each t in the range of the statistic T , it follows that T contains the same information about θ as X does; in this case, T is called a suﬃcient statistic for θ. The precise deﬁnition of suﬃciency follows. DEFINITION. A statistic T = T (X) is a suﬃcient statistic for a parameter θ if for all sets A, P [X ∈ AT = t] is independent of θ for all t in the range of T . Suﬃcient statistics are not unique; from the deﬁnition of suﬃciency, it follows that if g is a onetoone function over the range of the statistic T then g(T ) is also suﬃcient. This emphasizes the point that it is not the suﬃcient statistic itself that is important but rather the partition of the sample space induced by the statistic (that is, the level sets of the statistic). It also follows that if T is suﬃcient for θ then the distribution of any other statistic S = S(X) conditional on T = t is independent of θ. How can we check if a given statistic is suﬃcient? In some cases, suﬃciency can be veriﬁed directly from the deﬁnition. EXAMPLE 4.14: Suppose that X1 , · · · , Xk are independent c 2000 by Chapman & Hall/CRC
random variables where Xi has a Binomial distribution with parameters ni (known) and θ (unknown). Let T = X1 + · · · + Xk ; T will also have a Binomial distribution with parameters m = n1 + · · · + nk and θ. To show that T is suﬃcient, we need to show that Pθ [X = xT = t] is independent of θ (for all x1 , · · · , xk and t). First note that if t = x1 + · · · + xk then this conditional probability is 0 (and hence independent of θ). If t = x1 + · · · + xk then Pθ [X = xT = t] =
Pθ [X = x] Pθ [T = t]
ni x ni −xi i i=1 xi θ (1 − θ) m t m−t t θ (1 − θ) 1k ni i=1 xi m , t 1k
= =
which is independent of θ. Thus T is a suﬃcient statistic for θ. ✸ Unfortunately, there are two major problems with using the deﬁnition to verify that a given statistic is suﬃcient. First, the condition given in the deﬁnition of suﬃciency is sometimes very diﬃcult to verify; this is especially true when X has a continuous distribution. Second, the deﬁnition of suﬃciency does not allow us to identify suﬃcient statistics easily. Fortunately, there is a simple criterion due to Jerzy Neyman that gives a necessary and suﬃcient condition for T to be a suﬃcient statistic when X has a joint density or frequency function. THEOREM 4.2 (Neyman Factorization Criterion) Suppose that X = (X1 , · · · , Xn ) has a joint density or frequency function f (x; θ) (θ ∈ Θ). Then T = T (X) is suﬃcient for θ if, and only if, f (x; θ) = g(T (x); θ)h(x). (Both T and θ can be vectorvalued.) A rigorous proof of the Factorization Criterion in its full generality is quite technical and will not be pursued here; see Billingsley (1995) or Lehmann (1991) for complete details. However, the proof when X is discrete is quite simple and will be sketched here. c 2000 by Chapman & Hall/CRC
Suppose ﬁrst that T is suﬃcient. Then f (x; θ) = Pθ [X = x] =
Pθ [X = x, T = t]
t
= Pθ [X = x, T = T (x)] = Pθ [T = T (x)]P [X = xT = T (x)]. Since T is suﬃcient, P [X = xT = T (x)] is independent of θ and so f (x; θ) = g(T (x); θ)h(x). Now suppose that f (x; θ) = g(T (x); θ)h(x). Then if T (x) = t, Pθ [X = xT = t] = = =
Pθ [X = x] Pθ [T = t] g(T (x); θ)h(x)
T (y)=t g(T (y); θ)h(y) h(x) , T (y)=t h(y)
which does not depend on θ. If T (x) = t then Pθ [X = xT = t] = 0. In both cases, Pθ [X = xT = t] is independent of θ and so T is suﬃcient. EXAMPLE 4.15: Suppose that X1 , · · · , Xn are i.i.d. random variables with density function 1 f (x; θ) = for 0 ≤ x ≤ θ θ where θ > 0. The joint density function of X = (X1 , · · · , Xn ) is 1 for 0 ≤ x1 , · · · , xn ≤ θ θn 1 I(0 ≤ x1 , · · · , xn ≤ θ) = θn
1 I max x ≤ θ I min x ≥ 0 = i i 1≤i≤n 1≤i≤n θn = g(max(x1 , · · · , xn ); θ)h(x)
f (x; θ) =
and so X(n) = max(X1 , · · · , Xn ) is suﬃcient for θ.
✸
EXAMPLE 4.16: Suppose that X = (X1 , · · · , Xn ) have a distribution belonging to a kparameter exponential family with c 2000 by Chapman & Hall/CRC
joint density or frequency function satisfying f (x; θ) = exp
, k

ci (θ)Ti (x) − d(θ) + S(x) I(x ∈ A).
i=1
Then (taking h(x) = exp[S(x)]I(x ∈ A)), it follows from the Factorization Criterion that the statistic T = (T1 (X), · · · , Tk (X)) ✸
is suﬃcient for θ.
From the deﬁnition of suﬃciency, it is easy to see that the data X is itself always suﬃcient. Thus suﬃciency would not be a particularly useful concept unless we could ﬁnd suﬃcient statistics that truly represent a reduction of the data; however, from the examples given above, we can see that this is indeed possible. Thus, the real problem lies in determining whether a suﬃcient statistic represents the best possible reduction of the data. There are two notions of what is meant by the “best possible” reduction of the data. The ﬁrst of these is minimal suﬃciency; a suﬃcient statistic T is minimal suﬃcient if for any other suﬃcient statistic S, there exists a function g such that T = g(S). Thus a minimal suﬃcient statistic is the suﬃcient statistic that represents the maximal reduction of the data that contains as much information about the unknown parameter as the data itself. A second (and stronger) notion is completeness which will be discussed in more depth in Chapter 6. If X ∼ Fθ then a statistic T = T (X) is complete if Eθ (g(T )) = 0 for all θ ∈ Θ implies that Pθ (g(T ) = 0) = 1 for all θ ∈ Θ. In particular, if T is complete then g(T ) is ancillary for θ only if g(T ) is constant; thus a complete statistic T contains no ancillary information. It can be shown that if a statistic T is suﬃcient and complete then T is also minimal suﬃcient; however, the converse is not true. For example, suppose that X1 , · · · , Xn are i.i.d. random variables whose density function is f (x; θ) =
exp(x − θ) . [1 + exp(x − θ)]2
For this model, a onedimensional suﬃcient statistic for θ does not exist and, in fact, the order statistics (X(1) , · · · , X(n) ) can be shown to be minimal suﬃcient. However, the statistic T = X(n) − X(1) is ancillary and so the order statistics are not complete. Thus despite c 2000 by Chapman & Hall/CRC
the fact that (X(1) , · · · , X(n) ) is a minimal suﬃcient statistic, it still contains “redundant” information about θ. How important is suﬃciency in practice? The preceding discussion suggests that any statistical procedure should depend only on the minimal suﬃcient statistic. In fact, we will see in succeeding chapters that optimal statistical procedures (point estimators, hypothesis tests and so on discussed in these chapters) almost invariably depend on minimal suﬃcient statistics. Nonetheless, statistical models really serve only as approximations to reality and so procedures that are nominally optimal can fail miserably in practice. For example, suppose X1 , · · · , Xn are i.i.d. random variables with mean µ and variance σ 2 . It is common to assume that the Xi ’s have a Normal distribution in which case ( ni=1 Xi , ni=1 Xi2 ) is a minimal suﬃcient statistic for (µ, σ 2 ). However, optimal procedures for the Normal distribution can fail miserably if the Xi ’s are not Normal. For this reason, it is important to be ﬂexible in developing statistical methods. 4.4 Point estimation A point estimator or estimator is a statistic whose primary purpose is to estimate the value of a parameter. That is, if X ∼ Fθ for θ ∈ Θ, then an estimator θ/ is equal to some statistic T (X). Assume that θ is a realvalued parameter and that θ/ is an estimator of θ. The probability distribution of an estimator θ/ is / Ideally, we often referred to as the sampling distribution of θ. / would like the sampling distribution of θ to be concentrated closely around the true value of the parameter, θ. There are several simple measures of the quality of an estimator based on its sampling distribution. DEFINITION. The bias of an estimator θ/ is deﬁned to be / = E (θ) / − θ. bθ (θ) θ / = 0, that is, An estimator is said to be unbiased if bθ (θ) / = θ. Eθ (θ)
DEFINITION. The mean absolute error (MAE) of θ/ is deﬁned to be / = E [θ/ − θ]. MAEθ (θ) θ DEFINITION. The mean square error (MSE) of θ/ is deﬁned to be / = E [(θ/ − θ)2 ]; MSEθ (θ) θ c 2000 by Chapman & Hall/CRC
/ = Var (θ) / + [b (θ)] / 2. it is easy to show that MSEθ (θ) θ θ
The bias of θ/ gives some indication of whether the sampling / and MSE (θ) / are distribution is centered around θ while MAEθ (θ) θ / measures of the dispersion of the sampling distribution of θ around θ. MAE and MSE are convenient measures for comparing diﬀerent estimators of a parameter θ; since we would like θ/ to be close to θ, it is natural to prefer estimators with small MAE or MSE. Although MAE may seem to be a better measure for assessing the accuracy of an estimator, MSE is usually preferred to MAE. There are several reasons for preferring MSE; most of these derive from / into variance and bias components: the decomposition of MSEθ (θ) / = Var (θ) / + [b (θ)] / 2. MSEθ (θ) θ θ
This decomposition makes MSE much easier to work with than MAE. For example, when θ/ is a linear function of X1 , · · · , Xn , the mean and variance of θ/ (and hence its MSE) are easily computed; computation of the MAE is much more diﬃcult. Frequently, the sampling distribution of an estimator is approximately Normal; for example, it is often true that the distribution of θ/ is approximately Normal with mean θ and variance σ 2 (θ)/n. In such cases, the / variance σ 2 (θ)/n is often approximated reasonably well by MSEθ (θ) and so the MSE essentially characterizes the dispersion of the / (Typically, the variance component of sampling distribution of θ. / ≈ the MSE is much larger than the bias component and so MSEθ (θ) / However, it is also important to note that the MSE of an Varθ (θ).) estimator can be inﬁnite even when its sampling distribution is approximately Normal. Unbiasedness is a very controversial issue. The use of the word “biased” to describe an estimator is very loaded; it suggests that a biased estimator is somehow misleading or prejudiced. Thus, at ﬁrst glance, it may seem reasonable to require an estimator to be unbiased. However, in many estimation problems, unbiased estimators do not exist; moreover, there are situations where all unbiased estimators are ridiculous. A further diﬃculty with unbiasedness is the fact that unbiasedness is not generally preserved by transformation; that is, if θ/ is an unbiased estimator of θ then / is typically not an unbiased estimator of g(θ) unless g is a g(θ) linear function. Thus unbiasedness is not an intrinsically desirable quality of an estimator; we should not prefer an unbiased estimator to a biased estimator based only on unbiasedness. However, this is c 2000 by Chapman & Hall/CRC
not to say that bias should be ignored. For example, if an estimator θ/ systematically over or underestimates θ (in the sense that the sampling distribution of θ/ lies predominantly to the right or left of θ), steps should be taken to remove the resulting bias. EXAMPLE 4.17: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance σ 2 . An unbiased estimator of σ 2 is n 1 ¯ 2. S2 = (Xi − X) n − 1 i=1 √ However, S = S 2 is not an unbiased estimator of σ; using the fact that (n − 1)S 2 ∼ χ2 (n − 1), Y = σ2 it follows that √ σ E( Y ) Eσ (S) = √ n−1 √ 2Γ(n/2) σ = √ n − 1 Γ((n − 1)/2) = σ. However, as n → ∞, it can be show that E(S) → σ.
✸
EXAMPLE 4.18: Suppose that X1 , · · · , Xn are i.i.d. random variables with a Uniform distribution on [0, θ]. Let θ/ = X(n) , the sample maximum; the density of θ/ is n f (x; θ) = n xn−1 for 0 ≤ x ≤ θ. θ / < θ; in fact, it is easy to show Note that θ/ ≤ θ and hence Eθ (θ) that / = n θ. Eθ (θ) n+1 / The form of Eθ (θ) makes it easy to construct an unbiased estimator / of θ. If we deﬁne θ0 = (n + 1)θ/n then clearly θ0 is an unbiased estimator of θ. ✸
Suppose that θ/n is an estimator of some parameter θ based on n random variables X1 , · · · , Xn . As n increases, it seems reasonable to expect that the sampling distribution of θ/n should become increasingly concentrated around the true parameter value c 2000 by Chapman & Hall/CRC
θ. This property of the sequence of estimators {θ/n } is known as consistency. DEFINITION. A sequence of estimators {θ/n } is said to be consistent for θ if {θ/n } converges in probability to θ, that is, if lim Pθ [θ/n − θ > ] = 0
n→∞
for each > 0 and each θ. Although, strictly speaking, consistency refers to a sequence of estimators, we often say that θ/n is a consistent estimator of θ if it is clear that θ/n belongs to a welldeﬁned sequence of estimators; an example of this occurs when θ/n is based on n i.i.d. random variables. EXAMPLE 4.19: Suppose that X1 , · · · , Xn are i.i.d. random variables with E(Xi ) = µ and Var(Xi ) = σ 2 . Deﬁne Sn2 =
n 1 ¯ n )2 , (Xi − X n − 1 i=1
which is an unbiased estimator of σ 2 . To show that Sn2 is a consistent estimator (or more correctly {Sn2 } is a consistent sequence of estimators), note that Sn2
=
n 1 ¯ n )2 (Xi − µ + µ − X n − 1 i=1
=
n 1 n ¯ n − µ)2 (Xi − µ)2 + (X n − 1 i=1 n−1
=
n n−1
n 1 n ¯ n − µ)2 . (Xi − µ)2 + (X n i=1 n−1
By the WLLN, we have n 1 (Xi − µ)2 →p σ 2 n i=1
¯ n →p µ and X
and so by Slutsky’s Theorem, it follows that Sn2 →p σ 2 . Note that Sn2 will be a consistent estimator of σ 2 = Var(Xi ) for any distribution with ﬁnite variance. ✸ c 2000 by Chapman & Hall/CRC
EXAMPLE 4.20: Suppose that X1 , · · · , Xn are independent random variables with Eβ (Xi ) = βti
and Varβ (Xi ) = σ 2
where t1 , · · · , tn are known constants and β, σ 2 unknown parameters. A possible estimator of β is
n i=1 ti Xi / βn = n 2 . i=1 ti
It is easy to show that β/n is an unbiased estimator of β for each n and hence to show that β/n is consistent, it suﬃces to show that Varβ (β/n ) → 0. Because of the independence of the Xi ’s, it follows that σ2 Varβ (β/n ) = n 2 . i=1 ti
Thus β/n is consistent provided that ni=1 t2i → ∞ as n → ∞.
✸
4.5 The substitution principle In statistics, we are frequently interested in estimating parameters that depend on the underlying distribution function of the data; we will call such parameters functional parameters (although the term statistical functional is commonly used in the statistical literature). For example, the mean of a random variable with distribution function F may be written as ∞
µ(F ) =
−∞
x dF (x).
The value of µ(F ) clearly depends on the distribution function F ; thus we can think of µ(·) as a realvalued function on the space of distribution functions much in the same way that g(x) = x2 is a realvalued function on the realline. Some other examples of functional parameters include .∞ • the variance: σ 2 (F ) = −∞ (x − µ(F ))2 dF (x). • the median: med(F ) = F −1 (1/2) = inf{x : F (x) ≥ 1/2}. • the density at x0 : θ(F ) = F (x0 ) (θ(F ) is deﬁned only for those distributions with a density). • a measure of location θ(F ) deﬁned by the equation ∞
−∞
ψ(x − θ(F )) dF (x) = 0
c 2000 by Chapman & Hall/CRC
1.0 0.8 0.6 0.0
0.2
0.4
q(x)
0.0
0.2
0.4
0.6
0.8
1.0
x
Figure 4.1 Lorenz curves for the Gamma distribution with α = 0.5 (solid curve) and α = 5 (dotted curve).
(where ψ is typically a nondecreasing, odd function). The following example introduces a somewhat more complicated functional parameter that is of interest in economics. EXAMPLE 4.21: Economists are often interested in the distribution of personal income in a population. More speciﬁcally, they are interested in measuring the “inequality” of this distribution. One way to do so is to consider the socalled Lorenz curve that gives the percentage of income held by the poorest 100t% as a function of t. Let F be a distribution function (with F (x) = 0 for x < 0) whose expected value is µ(F ). For t between 0 and 1, we deﬁne .t
qF (t) =
F −1 (s) ds . 01 −1 (s) ds 0 F
.t
=
0
F −1 (s) ds . µ(F )
(Note that the denominator in the deﬁnition of qF (t) is simply the expected value of the distribution F .) It is easy to verify that qF (t) ≤ t with qF (t) = t (for 0 < t < 1) if, and only if, F is concentrated at a single point (that is, all members of the c 2000 by Chapman & Hall/CRC
population have the same income). The Lorenz curves for Gamma distributions with shape parameters 0.5 and 5 are given in Figure 4.1. (It can be shown that the Lorenz curve will not depend on the scale parameter.) One measure of inequality based on the Lorenz curve is the Gini index deﬁned by 1
θ(F ) = 2 0
(t − qF (t)) dt = 1 − 2
1 0
qF (t) dt.
The Gini index θ(F ) is simply twice the area between the functions t and qF (t) and so 0 ≤ θ(F ) ≤ 1; when perfect equality exists (qF (t) = t) then θ(F ) = 0 while as the income gap between the richest and poorest members of the population widens, θ(F ) increases. (For example, according to the World Bank (1999), estimated Gini indices for various countries range from 0.195 (Slovakia) to 0.629 (Sierra Leone); the Gini indices reported for Canada and the United States are 0.315 and 0.401, respectively.) The Gini index for the Gamma distribution with shape parameter 0.5 is 0.64 while for shape parameter 5, the Gini index is 0.25. (It can be shown that as the shape parameter tends to inﬁnity, the Gini index tends to 0.) ✸ The substitution principle Suppose that X1 , · · · , Xn are i.i.d. random variables with distribution function F ; F may be completely unknown or may depend on a ﬁnite number of parameters. (Hence the model can be parametric or nonparametric). In this section, we will consider the problem of estimating a parameter θ that can be expressed as a functional parameter of F , that is, θ = θ(F ). The dependence on θ of the distribution function F suggests that it may be possible to estimate θ by ﬁnding a good estimate of F and then substituting this estimate, F/ , for F in θ(F ) to get an estimate of θ, θ/ = θ(F/ ). Thus we have changed the problem from estimating θ to estimating the distribution function F . Substituting an estimator F/ for F in θ(F ) is known as the substitution principle. However, the substitution principle raises two basic questions: ﬁrst, how do we estimate F and second, does the substitution principle always lead to good estimates of the parameter in question? We will ﬁrst discuss estimation of F . If F is the distribution function of X1 , · · · , Xn then for a given value x, F (x) is the c 2000 by Chapman & Hall/CRC
probability that any Xi is no greater than x or (according to the WLLN) the longrun proportion of Xi ’s that are not greater than x. Thus it seems reasonable to estimate F (x) by F/ (x) =
n 1 I(Xi ≤ x) n i=1
(where I(A) is 1 if condition A is true and 0 otherwise), which is simply the proportion of Xi ’s in the sample less than or equal to x; this estimator is called the empirical distribution function of X1 , · · · , Xn . From the WLLN, it follows that F/ (x) = F/n (x) →p F (x) for each value of x (as n → ∞) so that F/n (x) is a consistent estimator of F (x). (In fact, the consistency of F/n holds uniformly over the real line: sup
−∞<x 0, we have Γ(r + 1) . λr Thus for a given r > 0, a method of moments estimator of λ is Eλ (Xir ) =
/= λ
n 1 Xr nΓ(r + 1) i=1 i
−1/r
.
/ = 1/X.) ¯ Since r is moreorless arbitrary (If we take r = 1 then λ here, it is natural to ask what value of r gives the best estimator of λ; a partial answer to this question is given in Example 4.39. ✸
EXAMPLE 4.26: Suppose that X1 , · · · , Xn are i.i.d. Gamma random variables with unknown parameters α and λ. It is easy to show that α α η1 (F ) = E(Xi ) = and η2 (F ) = Var(Xi ) = 2 . λ λ / satisfy the equations / and λ Thus α n ¯= 1 X Xi = n i=1 c 2000 by Chapman & Hall/CRC
/ α / λ
/2 = σ
n 1 ¯ 2 = (Xi − X) n i=1
/ α / λ2
/ = X/ ¯ 2 /σ ¯ σ /=X / 2 and λ /2. and so α
✸
EXAMPLE 4.27: Suppose that X1 , · · · , Xn are i.i.d. random variables with a “zeroinﬂated” Poisson distribution; the frequency function of Xi is
f (x; θ, λ) =
θ + (1 − θ) exp(−λ) for x = 0 (1 − θ) exp(−λ)λx /x! for x = 1, 2, 3, · · ·
where 0 ≤ θ ≤ 1 and λ > 0. To estimate θ and λ via the method of moments, we will use η1 (F ) = P (Xi = 0)
and η2 (F ) = E(Xi );
it is easy to show that P (Xi = 0) = θ + (1 − θ) exp(−λ) and E(Xi ) = (1 − θ)λ. / satisfy the equations Thus θ/ and λ /λ / ¯ = (1 − θ) X n 1 / exp(−λ); / I(Xi = 0) = θ/ + (1 − θ) n i=1
/ do not exist although however, closed form expressions for θ/ and λ they may be computed for any given sample. (For this model, the
statistic ( ni=1 Xi , ni=1 I(Xi = 0)) is suﬃcient for (θ, λ).) ✸
It is easy to generalize the method of moments to noni.i.d. settings. Suppose that (X1 , · · · , Xn ) has a joint distribution depending on realvalued parameters θ = (θ1 , · · · , θp ). Suppose that T1 , · · · , Tp are statistics with Eθ (Tk ) = gk (θ)
for k = 1, · · · , p.
If, for all possible values of Eθ (T1 ), · · · , Eθ (Tp ), this system of / equations has a unique solution then we can deﬁne the estimator θ such that / for k = 1, · · · , p. Tk = gk (θ) However, in the general (that is, noni.i.d.) setting, greater care c 2000 by Chapman & Hall/CRC
must be taken in choosing the statistics T1 , · · · , Tp ; in particular, it is important that Tk be a reasonable estimator of its mean Eθ (Tk ) (for k = 1, · · · , p). 4.6 Inﬂuence curves Suppose that h is a realvalued function on the real line and that {xn } is a sequence of real numbers whose limit (as n → ∞) is x0 . If h is continuous at x0 , then h(xn ) → h(x0 ) as n → ∞; thus for large enough n, h(xn ) ≈ h(x0 ). If we assume that h is diﬀerentiable, it is possible to obtain a more accurate approximation of h(xn ) by making a one term Taylor series expansion: h(xn ) ≈ h(x0 ) + h (x0 )(xn − x0 ). This approximation can be written more precisely as h(xn ) = h(x0 ) + h (x0 )(xn − x0 ) + rn where the remainder term rn goes to 0 with n faster than xn − x0 : rn lim = 0. n→∞ xn − x0 An interesting question to ask is whether notions of continuity and diﬀerentiability can be extended to functional parameters and whether similar approximations can be made for substitution principle estimators of functional parameters. Let θ(F ) be a functional parameter and F/n be the empirical distribution function of i.i.d. random variables X1 , · · · , Xn . Since F/n converges in probability to F uniformly over the real line, it is tempting to say that θ(F/n ) converges in probability to θ(F ) given the right kind of continuity of θ(F ). However, continuity and diﬀerentiability of functional parameters are very diﬃcult and abstract topics from a mathematical point of view and will not be dealt with here in any depth. In principle, deﬁning continuity of the realvalued functional parameter θ(F ) at F is not diﬃcult; we could say that θ(F ) is continuous at F if θ(Fn ) → θ(F ) whenever a sequence of distribution functions {Fn } converges to F . However, there are several ways in which convergence of {Fn } to F can be deﬁned and the continuity of θ(F ) may depend on which deﬁnition is chosen. Diﬀerentiability of θ(F ) is an even more diﬃcult concept. Even if we agree on the deﬁnition of convergence of {Fn } to F , there are several diﬀerent concepts of diﬀerentiability. Thus we will not c 2000 by Chapman & Hall/CRC
touch on diﬀerentiability in any depth. We will, however, deﬁne a type of directional derivative for θ(F ) whose properties are quite useful for heuristic calculations; this derivative is commonly called the inﬂuence curve of θ(F ). The idea behind deﬁning the inﬂuence curve is to look at the behaviour of θ(F ) for distributions that are close to F in some sense. More speciﬁcally, we look at the diﬀerence between θ(F ) evaluated at F and at (1 − t)F + t∆x where ∆x is a degenerate distribution function putting all its probability at x so that ∆x (y) = 0 for y < x and ∆x (y) = 1 for y ≥ x; for 0 ≤ t ≤ 1, (1 − t)F + t∆x is a distribution function and can be thought of as F contaminated by probability mass at x. Note that as t ↓ 0, we typically have θ((1 − t)F + t∆x ) → θ(F ) for any x where this convergence is linear in t, that is, θ((1 − t)F + t∆x ) − θ(F ) ≈ φ(x; F )t for t close to 0. DEFINITION. The inﬂuence curve of θ(F ) at F is the function φ(x; F ) = lim t↓0
θ((1 − t)F + t∆x ) − θ(F ) t
provided that the limit exists. The inﬂuence curve can also be evaluated as + + d φ(x; F ) = θ((1 − t)F + t∆x )++ dt t=0 whenever this limit exists. The inﬂuence curve allows for a “linear approximation” of the diﬀerence θ(F/n ) − θ(F ) much in the same way that the derivative of a function h allows for a linear approximation of h(xn ) − h(x0 ); in particular, it is often possible to write θ(F/n ) − θ(F ) =
∞
−∞
φ(x; F )d(F/n (x) − F (x)) + Rn
where Rn tends in probability to 0 at a faster rate than F/n converges to F . This representation provides a useful heuristic method for √ determining the limiting distribution of n(θ(F/n )−θ(F )). In many cases, it is possible to show that ∞
−∞
φ(x; F ) dF (x) = E[φ(Xi ; F )] = 0
c 2000 by Chapman & Hall/CRC
and so
n 1 φ(Xi ; F ) + Rn n i=1 √ where the remainder term Rn satisﬁes nRn →p 0. Thus by the Central Limit Theorem and Slutsky’s Theorem, it follows that √ n(θ(F/n ) − θ(F )) →d N (0, σ 2 (F ))
θ(F/n ) − θ(F ) =
∞
where 2
σ (F ) =
−∞
φ2 (x; F ) dF (x),
provided that σ 2 (F ) is ﬁnite. This socalled “inﬂuence curve heuristic” turns out to be very useful in practice. However, despite the fact that this heuristic approach works in many examples, we actually require a stronger notion of diﬀerentiability to make this approach rigorous; fortunately, the inﬂuence curve heuristic can typically be made rigorous using other approaches. The inﬂuence curve is a key concept in theory of robustness, which essentially studies the sensitivity (or robustness) of estimation procedures subject to violations of the nominal model assumptions. For more information on the theory of robust estimation, see Hampel et al (1986). We will now discuss some simple results that are useful for computing inﬂuence curves. To make the notation more compact, we will set Ft,x = (1 − t)F + t∆x . • (Moments) Deﬁne θ(F ) = θ(F ) = E[h(X)]. Then
.∞
−∞ h(x) dF (x);
θ(Ft,x ) = (1 − t)
if X ∼ F then
∞ −∞
h(u) dF (u)
∞
+t
h(u)d∆x (u)
−∞ ∞
= (1 − t)
−∞
h(u) dF (u) + th(x)
and so
1 (θ(Ft,x ) − θ(F )) = h(x) − θ(F ). t Thus the inﬂuence curve is φ(x; F ) = h(x) − θ(F ). • (Sums and integrals) Suppose that θ(F ) = θ1 (F ) + θ2 (F ) where c 2000 by Chapman & Hall/CRC
φi (x; F ) is the inﬂuence curve of θi (F ) (for i = 1, 2). Then φ(x; F ), the inﬂuence curve of θ(F ), is simply φ(x; F ) = φ1 (x; F ) + φ2 (x; F ). This result can be extend to any ﬁnite sum of functional parameters. Often we need to consider functional parameters of the form θ(F ) = g(s)θs (F ) ds A
where θs (F ) is a functional parameter for each s ∈ A and g is a function deﬁned on A. Then we have 1 (θ(Ft,x ) − θ(F )) t 1 = g(s) (θs (Ft,x ) − θs (F )) ds. t A Thus, if φs (x; F ) is the inﬂuence curve of θs (F ) and we can take the limit as t ↓ 0 inside the integral sign, the inﬂuence curve of θ(F ) is deﬁned by
φ(x; F ) = A
g(s)φs (x; F ) ds.
(The trimmed mean considered in Example 4.30 is an example of such a functional parameter.) • (The chain rule) Suppose that θ(F ) has inﬂuence curve φ(x; F ). What is the inﬂuence curve of g(θ(F )) if g is a diﬀerentiable function? First of all, we have 1 (g(θ(Ft,x )) − g(θ(F ))) t
g(θ(Ft,x )) − g(θ(F )) θ(Ft,x ) − θ(F ) = . θ(Ft,x ) − θ(F ) t As t → 0, θ(Ft,x ) → θ(F ) (for each x) and so g(θ(Ft,x )) − g(θ(F )) → g (θ(F )) θ(Ft,x ) − θ(F )
as t → 0
and by deﬁnition 1 (θ(Ft,x ) − θ(F )) → φ(x; F ). t Therefore the inﬂuence curve of g(θ(F )) is g (θ(F ))φ(x; F ); this is a natural extension of the chain rule. For a given distribution c 2000 by Chapman & Hall/CRC
function F , the inﬂuence curve of g(θ(F )) is simply a constant multiple of the inﬂuence curve of θ(F ). • (Implicitly deﬁned functional parameters) Functional parameters are frequently deﬁned implicitly. For example, θ(F ) may satisfy the equation h(F, θ(F )) = 0 where for a ﬁxed number u, h(F, u) has inﬂuence curve λ(x; F, u) and for a ﬁxed distribution function F , h(F, u) has derivative (with respect to u), h (u; F ). We then have 0
1 (h(Ft,x , θ(Ft,x )) − h(F, θ(F ))) t 1 = (h(Ft,x , θ(Ft,x )) − h(Ft,x , θ(F ))) t 1 + (h(Ft,x , θ(F )) − h(F, θ(F ))) t → h (θ(F ); F )φ(x; F ) + λ(x; F, θ(F )) =
as t → 0 where φ(x; F ) is the inﬂuence curve of θ(F ). Thus h (θ(F ); F )φ(x; F ) + λ(x; F, θ(F )) = 0 and so
φ(x; F ) = −
λ(x; F, θ(F )) . h (θ(F ); F )
EXAMPLE 4.28: One example of an implicitly deﬁned functional parameter is the median of a continuous distribution F , θ(F ), deﬁned by the equation F (θ(F )) =
1 2
or equivalently θ(F ) = F −1 (1/2) where F −1 is the inverse of F . ∞ Since I(x ≤ u) dF (x), F (u) = −∞
it follows that the inﬂuence curve of F (u) is λ(x; F, u) = I(x ≤ u) − F (u). Thus if F (u) is diﬀerentiable at u = θ(F ) with F (θ(F )) > 0, it follows that the inﬂuence curve of θ(F ) = F −1 (1/2) is φ(x; F ) = − c 2000 by Chapman & Hall/CRC
I(x ≤ θ(F )) − F (θ(F )) F (θ(F ))
=
sgn(x − θ(F )) 2F (θ(F ))
where sgn(u) is the “sign” of u (sgn(u) is 1 if u > 0, −1 if u < 0 and 0 if u = 0). Note that we require F (u) to be diﬀerentiable at u = θ(F ) so φ(x; F ) is not deﬁned for all F (although F does not have to be a continuous distribution function). Using the heuristic n 1 / med(Fn ) − θ(F ) = φ(Xi ; F ) + Rn
n
i=1
it follows that √ n(θ(F/n ) − θ(F )) →d N (0, [2F (θ(F ))]−2 ) since Var(sgn(Xi − θ(F ))) = 1. Indeed, the convergence indicated above can be shown to hold when the distribution function F is diﬀerentiable at its median; see Example 3.6 for a rigorous proof of the asymptotic normality of the sample median. ✸ EXAMPLE 4.29: Let σ(F ) be the standard deviation of a random variable X with distribution function F ; that is,
1/2
σ(F ) = θ2 (F ) − θ12 (F ) .
∞ where θ1 (F ) = −∞ y dF (y) and θ2 (F ) = inﬂuence curve of θ2 (F ) is
.∞
−∞ y
2 dF (y).
The
φ2 (x; F ) = x2 − θ2 (F ) while the inﬂuence curve of θ12 (F ) is φ1 (x; F ) = 2θ1 (F )(x − θ1 (F )) by applying the chain rule for inﬂuence curves. Thus the inﬂuence curve of θ2 (F ) − θ12 (F ) is φ3 (x; F ) = x2 − θ2 (F ) − 2θ1 (F )(x − θ1 (F )). Since σ(F ) = (θ2 (F ) − θ12 (F ))1/2 , it follows that the inﬂuence curve of σ(F ) is φ(x; F ) =
x2 − θ2 (F ) − 2θ1 (F )(x − θ1 (F )) 2σ(F )
by applying the chain rule. Note that φ(x; F ) → ∞ as x → ±∞ and that φ(x; F ) = 0 when x = θ1 (F ) ± σ(F ). ✸ EXAMPLE 4.30: A functional parameter that includes the mean c 2000 by Chapman & Hall/CRC
and median as limiting cases is the αtrimmed mean deﬁned for continuous distribution functions F by µα (F ) =
1 1 − 2α
1−α
F −1 (t) dt
α
where 0 < α < 0.5. If f (x) = F (x) is continuous and strictly positive over the interval
F −1 (α), F −1 (1 − α)
as well as symmetric around some point µ then µα (F ) = µ and the inﬂuence curve of µα (F ) is −1 (F (α) − µ)/(1 − 2α)
for x < F −1 (α) − α) − µ)/(1 − 2α) for x > F −1 (1 − α) φα (x; F ) = (x − µ)/(1 − 2α) otherwise. (F −1 (1
To ﬁnd a substitution principle estimator for µα (F ) based on i.i.d. observations X1 , · · · , Xn , we ﬁrst ﬁnd a substitution principle estimator of F −1 (t) for 0 < t < 1 based on the inverse of the empirical distribution function F/n : F/n−1 (t) = X(i)
if (i − 1)/n < t ≤ i/n
(where X(i) is the ith order statistic) and substitute this into the deﬁnition of µα (F ) yielding µα (F/n ) =
1 1 − 2α
1−α α
F/n−1 (t) dt.
Applying the inﬂuence curve heuristic, we have √ n(µα (F/n ) − µα (F )) →d N (0, σ 2 (F )) where
∞
2
σ (F ) = =
−∞
φ2α (x; F ) dF (x) ,
−1

F (1−α) 2 (x − µ)2 dF (x) (1 − 2α)2 F −1 (α) 2 2α −1 + F (1 − α) − µ . (1 − 2α)2
c 2000 by Chapman & Hall/CRC
A somewhat simpler alternative that approximates µα (F/n ) is 0n = µ
1 n − 2gn
n−g n
X(i)
i=gn +1
0n = where gn is chosen so that gn /n ≈ α. (If gn /n = α then µ √ / 0n − µα (F )) is the same µα (Fn .) The limiting distribution of n(µ √ ✸ as that of n(µα (F/n ) − µα (F )).
EXAMPLE 4.31: Consider the Gini index θ(F ) deﬁned in Example 4.21. To determine the substitution principle estimator of θ(F ) based on i.i.d. observations X1 , · · · , Xn , we use the substitution principle estimator of F −1 (t) from Example 4.30: 1 t 0
0
1 1
F/n−1 (s) ds dt =
0
1
= 0
s
(1 − s)F/n−1 (s) ds
n
=
i/n
X(i)
i=1
θ(F/n ) =
n
Xi
(i−1)/n
−1 n
2i − 1
i=1
(1 − s) ds
n 1 2i − 1 1− X(i) n i=1 2n
= and so
F/n−1 (s) dt ds
n
i=1
− 1 X(i) .
As with the trimmed mean, the inﬂuence curve of the Gini index is complicated to derive. With some work, it can be shown that ) 1
φ(x; F ) = 2 0
*
qF (t) dt − qF (F (x))
x +2 µ(F )
1 0
qF (t) dt − 1 − F (x)
where ∞
1
x dF (x) =
µ(F ) = 0
and qF (t) = c 2000 by Chapman & Hall/CRC
1 µ(F )
0
t
F 0
F −1 (t) dt
−1
(t) dt.
The inﬂuence curve heuristic suggests that √ n(θ(F/n ) − θ(F )) →d N (0, σ 2 (F )) ∞
with 2
σ (F ) =
φ2 (x; F ) dF (x).
0
Unfortunately, σ 2 (F ) is diﬃcult to evaluate (at least as a closedform expression) for most distributions F . ✸ The inﬂuence curve has a nice ﬁnite sample interpretation. Suppose that we estimate θ(F ) based on observations x1 , · · · , xn and set θ/n = θ(F/n ). Now suppose that we obtain another observation xn+1 and reestimate θ(F ) by θ/n+1 = θ(F/n+1 ) where F/n+1 (x) =
n / 1 Fn (x) + ∆x (x). n+1 n + 1 n+1
Letting t = 1/(n + 1) and assuming that n is suﬃciently large to make t close to 0, the deﬁnition of the inﬂuence curve suggests that we can approximate θ/n+1 by θ/n+1 ≈ θ/n +
1 φ(xn+1 ; F/n ). n+1
(This approximation assumes that φ(x; F/n ) is well deﬁned; it need not be. For example, the inﬂuence curve of the median is not deﬁned for discrete distributions such as F/n .) From this, we can see that the inﬂuence curve gives an approximation for the inﬂuence that a single observation exerts on a given estimator. For example, consider the inﬂuence curve of the standard deviation σ(F ) given in Example 4.29; based on x1 , · · · , xn , the substitution principle estimate is 1/2 n 1 2 /n = σ (xi − x ¯n ) n i=1 where x ¯n is the sample mean. The approximation given above /n and suggests that if the observation xn+1 lies between x ¯n − σ /n then σ /n+1 < σ /n and otherwise σ /n+1 ≥ σ /n . Moreover, σ /n+1 x ¯n + σ can be made arbitrarily large by taking xn+1  suﬃciently large. Suppose that X1 , · · · , Xn are i.i.d. random variables with distribution function Fθ where θ is a realvalued parameter and let G = {φ : φ(Fθ ) = θ}; the functional parameters φ in G are called c 2000 by Chapman & Hall/CRC
Fisher consistent for θ. Many statisticians consider it desirable for a functional parameter to have a bounded inﬂuence curve as this will limit the eﬀect that a single observation can have on the value of an estimator. This would lead us to consider only those Fisher consistent φ’s with bounded inﬂuence curves. For example, suppose that X1 , · · · , Xn are i.i.d. random variables with a distribution symmetric around θ; if E[Xi ] < ∞ then we could estimate θ by the sample mean with the substitution principle estimator .∞ of µ(F ) = −∞ x dF (x). However, the inﬂuence curve of µ(F ) is φ(x; F ) = x − µ(F ), which is unbounded (in x) for any given F . As an alternative, we might instead estimate θ by the sample median or some trimmed mean as these are substitution principle estimators of functional parameters with bounded inﬂuence curves. 4.7 Standard errors and their estimation The standard error of an estimator is deﬁned to be the standard deviation of the estimator’s sampling distribution. Its purpose is to convey some information about the uncertainty of the estimator. Unfortunately, it is often very diﬃcult to calculate the standard error of an estimator exactly. In fact, there are really only two situations where the standard error of an estimator θ/ can be computed exactly: • the sampling distribution of θ/ is known. • θ/ is a linear function of random variables X1 , · · · , Xn where the variances and covariances of the Xi ’s are known. However, if the sampling distribution of θ/ can be approximated by a distribution whose standard deviation is known, this standard / deviation can be used to give an approximate standard error for θ. The most common example of such an approximation occurs when the√sampling distribution is approximately Normal; for example, if n(θ/ − θ) is approximately Normal with√mean 0 and variance σ 2 (where σ 2 may depend on θ) then σ/ n can be viewed as / In fact, it is not uncommon an approximate standard√error of θ. in such cases to see σ/ n referred to as the standard error of θ/ despite the fact that it is merely an approximation. Moreover, approximate standard errors can be more useful than their exact / can be inﬁnite despite the fact counterparts. For example, Varθ (θ) that the distribution of θ/ is approximately Normal; in this case, the approximate standard error is more informative about the c 2000 by Chapman & Hall/CRC
/ (The variance can be distorted by small amounts uncertainty of θ. of probability in the tails of the distribution; thus the variance of the approximating Normal distribution gives a better indication of the true variability.) “Delta Method” type arguments are useful for ﬁnding approximate standard errors, especially for method of moments estimators. For example, suppose that X1 , · · · , Xn are independent random variables with E(Xi ) = µi and Var(Xi ) = σi2 and
θ/ = g
n
Xi
i=1
where θ=g
n
µi .
i=1
Then a Taylor series expansion gives θ/ − θ
= g ≈ g
n
Xi − g
i=1 n i=1
µi
n
n
µi
i=1
(Xi − µi )
i=1
and taking the variance of the last expression, we obtain the following approximate standard error: + + n 1/2 n + + + + 2 / µi + σi . se(θ) ≈ +g + + i=1
i=1
The accuracy of this approximation depends on the closeness the distribution of θ/ to normality. When X1 , · · · , Xn are i.i.d. it is usually possible to prove directly that θ/ is approximately Normal (provided n is suﬃciently large). EXAMPLE 4.32: Suppose that X1 , · · · , Xn are i.i.d. random variables with mean µ and variance σ 2 . The substitution principle ¯ whose variance is σ 2 /n. Thus the standard error estimator of µ is X √ ¯ is σ/ n. of X ✸ EXAMPLE 4.33: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. Since Eλ (Xi ) = 1/λ, a method / = 1/X. ¯ If n is suﬃciently large of moments estimator of λ is λ c 2000 by Chapman & Hall/CRC
√ ¯ then n(X − λ−1 ) is approximately Normal with mean 0 and √ / variance λ−2 ; applying the Delta Method, we have n(λ − λ) is approximately Normal with mean √ 0 and variance λ2 . Thus an / is λ/ n. approximate standard error of λ ✸ EXAMPLE 4.34: Suppose that X1 , · · · , Xn are independent Poisson random variables with Eβ (Xi ) = exp(βti ) where β is an unknown parameter and t1 , · · · , tn are known constants. Deﬁne β/ to satisfy the equation n
Xi =
i=1
n
/ i ) = g(β). / exp(βt
i=1
/ we will use a To compute an approximate standard error for β, “Delta Method” type argument. Expanding g in a Taylor series, we get / − g(β) = g(β)
n
(Xi − exp(βti ))
i=1
≈ g (β)(β/ − β) and so β/ − β ≈ =
/ − g(β) g(β) g (β)
n i=1 (Xi − exp(βti ))
n . i=1 ti exp(βti )
Since Varβ (Xi ) = Eβ (Xi ) = exp(βti ), it follows that an approximate standard error of β/ is
( ni=1 exp(βti ))1/2 / se(β) ≈ n .

i=1 ti exp(βti )
This approximation assumes that the distribution of β/ is approximately Normal. The standard error of β/ can be estimated by substituting β/ for β in the expression given above. ✸ EXAMPLE 4.35: Suppose that X1 , · · · , Xn are i.i.d. random variables with density f (x − θ) where f (x) = f (−x); that is, the Xi ’s have distribution that is symmetric around 0. Let ψ(x) be c 2000 by Chapman & Hall/CRC
a nondecreasing odd function (ψ(x) = −ψ(−x)) with derivative ψ (x) and deﬁne θ/ to be the solution to the equation n 1 / = 0. ψ(x − θ) n i=1
Note that θ/ is a substitution principle estimator of the functional parameter θ(F ) deﬁned by ∞ −∞
ψ(x − θ(F )) dF (x) = 0;
the inﬂuence curve of θ(F ) is ψ(x − θ(F )) . −∞ ψ (x − θ(F )) dF (x)
φ(x; F ) = . ∞
Hence for n suﬃciently large, with mean 0 and variance σ
∞
2
= =
=
√
n(θ/ − θ) is approximately Normal
φ2 (x; F ) dF (x)
−∞ .∞ 2 −∞ ψ (x
.
− θ)f (x − θ) dx
∞ −∞ ψ (x − θ)f (x − .∞ 2 −∞ ψ (x)f (x) dx . 2 ∞ −∞ ψ (x)f (x) dx
2
θ) dx
√ and so an approximate standard error of θ/ is σ/ n.
✸
As we noted above, standard errors (and their approximations) can and typically do depend on unknown parameters. These standard errors can themselves be estimated by substituting estimates for the unknown parameters in the expression for the standard error. EXAMPLE 4.36: In Example √ 4.35, we showed that the approxi/ mate standard error of θ is σ/ n where .∞
σ = . −∞ 2
ψ 2 (x − θ)f (x − θ) dx
∞ −∞ ψ (x
c 2000 by Chapman & Hall/CRC
− θ)f (x − θ) dx
2 .
Substituting θ/ for θ, we can obtain the following substitution principle estimator of σ 2 :
−2
n 1 / ψ (Xi − θ) n i=1
n 1 2 / . / = σ ψ(Xi − θ) n i=1 √ / / n. The estimated standard error of θ/ is simply σ
✸
Another method of estimating standard errors is given in Section 4.9. 4.8 Asymptotic relative eﬃciency Suppose that θ/n and θ0n are two possible estimators (based on X1 , · · · , Xn ) of a realvalued parameter θ. There are a variety of approaches to comparing two estimators. For example, we can compare the MSEs or MAEs (if they are computable) and choose the estimator whose MSE (or MAE) is smaller (although this choice may depend on the unknown value of θ). If both estimators are approximately Normal, we can use a measure called the asymptotic relative eﬃciency (ARE). DEFINITION. Let X1 , X2 , · · · be a sequence of random variables and suppose that θ/n and θ0n are estimators of θ (based on X1 , · · · , Xn ) such that θ/n − θ →d N (0, 1) σ1n (θ)
θ0n − θ →d N (0, 1) σ2n (θ)
and
for some sequences {σ1n (θ)} and {σ2n (θ)}. Then the asymptotic relative eﬃciency of θ/n to θ0n is deﬁned to be 2 (θ) σ2n n→∞ σ 2 (θ) 1n
AREθ (θ/n , θ0n ) = lim provided this limit exists.
What is the interpretation of asymptotic relative eﬃciency? In many applications (for example, if the Xi ’s are i.i.d.), we have σ1 (θ) σ1n (θ) = √ n
σ2 (θ) and σ2n (θ) = √ n
and so AREθ (θ/n , θ0n ) = c 2000 by Chapman & Hall/CRC
σ22 (θ) . σ12 (θ)
Suppose we can estimate θ using either θ/n or θ0m where n and m are the sample sizes on which the two estimators are based. Suppose we want to choose m and n such that
Pθ θ/n − θ < ∆ ≈ Pθ θ0m − θ < ∆
for any ∆. Since for m and n suﬃciently large both estimators are approximately Normal, m and n satisfy √ √ P Z < ∆σ1 (θ)/ n ≈ P Z < ∆σ2 (θ)/ m (where Z ∼ N (0, 1)), which implies that σ1 (θ) σ2 (θ) √ ≈ √ n m or
m σ22 (θ) ≈ . 2 n σ1 (θ) Thus the ratio of sample sizes needed to achieve the same precision is approximately equal to the asymptotic relative eﬃciency; for example, if AREθ (θ/n , θ0n ) = k, we would need m ≈ kn so that θ0m has the same precision as θ/n (when θ is the true value of the parameter). In applying ARE to compare two estimators, we should keep in mind that it is a large sample measure and therefore may be misleading in small sample situations. If measures such as MSE and MAE cannot be accurately evaluated, simulation is useful for comparing estimators. EXAMPLE 4.37: Suppose that X1 , · · · , Xn are i.i.d. Normal /n be the random variables with mean µ and variance σ 2 . Let µ 0n be the sample median of X1 , · · · , Xn . Then sample mean and µ we have √ √ /n − µ) →d N (0, σ 2 ) and 0n − µ) →d N (0, πσ 2 /2). n(µ n(µ Hence
π = 1.571. 2 /n is more eﬃcient than µ 0n . We say that µ /n , µ 0n ) = AREµ (µ
✸
EXAMPLE 4.38: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ. Suppose we want to estimate θ = exp(−λ) = Pλ (Xi = 0). Consider the two estimators ¯n) θ/n = exp(−X c 2000 by Chapman & Hall/CRC
and θ0n =
n 1 I(Xi = 0). n i=1
It is easy to show (using the CLT and the Delta Method) that √ and
√
n(θ/n − θ) →d N (0, λ exp(−2λ)) n(θ0n − θ) →d N (0, exp(−λ) − exp(−2λ)).
Hence
exp(λ) − 1 . λ Using the expansion exp(λ) = 1 + λ + λ2 /2 + · · ·, it is easy to see that this ARE is always greater than 1; however, for small values of λ, the ARE is close to 1 indicating that there is little to choose between the two estimators when λ is small. ✸ AREλ (θ/n , θ0n ) =
EXAMPLE 4.39: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. In Example 4.25, we gave a family of method of moments estimators of λ using the fact that Eλ (Xi ) = Γ(r + 1)/λr for r > 0. Deﬁne / (r) = λ n
n 1 Xr nΓ(r + 1) i=1 i
−1/r
.
Using the fact that Varλ (Xir ) = (Γ(2r+1)−Γ2 (r+1))/λ2r , it follows from the Central Limit Theorem and the Delta Method that √ / (r) n(λn − λ) →d N (0, σ 2 (r)) where
λ2 σ (r) = 2 r
2
Γ(2r + 1) −1 . Γ2 (r + 1)
The graph of σ 2 (r)/λ2 is given in Figure 4.2; it is easy to see that ¯ is the most eﬃcient σ 2 (r) is minimized for r = 1 so that 1/X (asymptotically) estimator of λ of this form. ✸ EXAMPLE 4.40: Suppose that X1 , · · · , Xn are i.i.d. Cauchy random variables with density function f (x; θ, σ) =
σ 1 . 2 π σ + (x − θ)2
This density function is symmetric around θ; however, since E(Xi ) ¯ n is not a is not deﬁned for this distribution, the sample mean X good estimator of θ. A possible estimator of θ is the αtrimmed c 2000 by Chapman & Hall/CRC
10 8 6 4 2
scaled variance
0
1
2
3
4
5
r
Figure 4.2 σ 2 (r)/λ2 in Example 4.39 as a function of r.
mean θ/n (α) =
1 n − 2gn
n−g n
X(i)
i=gn +1
where the X(i) ’s are the order statistics and gn /n → α as n → ∞ where 0 < α < 0.5. It can be shown (for example, by using the inﬂuence curve of the trimmed mean functional parameter given in Example 4.30) that √
n(θ/n (α) − θ) →d N (0, γ 2 (α))
where γ 2 (α) σ2 =
2π −1 tan (π(0.5 − α)) + 2α − 1 + 2α tan2 (π(0.5 − α)) . (1 − 2α)2
If θ0n is the sample median of X1 , · · · , Xn , we have √
n(θ0n − θ) →d N 0, σ 2 π 2 /4 .
c 2000 by Chapman & Hall/CRC
1.0 0.8 0.6 0.0
0.2
0.4
ARE
0.0
0.1
0.2
0.3
0.4
0.5
trimming fraction
Figure 4.3 ARE of αtrimmed means (for 0 ≤ α ≤ 0.5) with respect to the sample median in Example 4.40.
The ARE of θ/n (α) with respect to θ0n is thus given by the formula AREθ (θ/n (α), θ0n ) =
π 2 (1 − 2α)2 4 [2π −1 tan (π(0.5 − α)) + 2α − 1 + 2α tan2 (π(0.5 − α))]
A plot of AREθ (θ/n (α), θ0n ) for α between 0 and 0.5 is given in Figure 4.3. The trimmed mean θ/n (α) is more eﬃcient than θ0n for α > 0.269 and the ARE is maximized at α = 0.380. We will see in Chapter 5 that we can ﬁnd even more eﬃcient estimators of θ for this model. ✸ 4.9 The jackknife The jackknife provides a generalpurpose approach to estimating the bias and variance (or standard error) of an estimator. Suppose that θ/ is an estimator of θ based on i.i.d. random variables X1 , · · · , Xn ; θ could be an unknown parameter from some parametric model or θ could be functional parameter of the common c 2000 by Chapman & Hall/CRC
distribution function F of the Xi ’s (in which case θ = θ(F )). The jackknife is particularly useful when standard methods for computing bias and variance cannot be applied or are diﬃcult to apply. Two such examples are given below. EXAMPLE 4.41: Suppose that X1 , · · · , Xn are i.i.d. random variables with density f (x − θ) that is symmetric around θ (f (x) = f (−x)). One possible estimator of θ is the trimmed mean θ/ =
n−g 1 X , n − 2g i=g+1 (i)
which averages X(g+1) , · · · , X(n−g) , the middle n − 2 g order statistics. The trimmed mean is less susceptible to extreme values than the sample mean of the Xi ’s, and is often a useful estimator of θ. However, unless the density function f is known precisely, it is dif/ (If f is known, it is possible ﬁcult to approximate the variance of θ. / to approximate the variance of θ using the inﬂuence curve given in Example 4.30; see also Example 4.40.) ✸ EXAMPLE 4.42: In survey sampling, it is necessary to estimate the ratio of two means. For example, we may be interested in estimating the unemployment rate for males aged 18 to 25. If we take a random sample of households, we can obtain both the number of males between 18 and 25 and the number of these males who are unemployed in each of the sampled households. Our estimate of the unemployment rate for males aged 18 to 25 would then be r/ =
number of unemployed males aged 18  25 in sample . number of males aged 18  25 in sample
The general problem may be expressed as follows. Suppose that (X1 , Y1 ), · · · , (Xn , Yn ) are independent random vectors from the same joint distribution with E(Xi ) = µX and E(Yi ) = µY ; we want to estimate r = µX /µY . A method of moments estimator of r is
n ¯ Xi X r/ = i=1 = ¯. n Y i=1 Yi ¯ Y¯ ) Unfortunately, there is no easy way to evaluate either E(X/ ¯ ¯ or Var(X/Y ) (although the Delta Method provides a reasonable approximation). ✸ c 2000 by Chapman & Hall/CRC
The name “jackknife” was originally used by Tukey (1958) to suggest the broad usefulness of the technique as a substitute to more specialized techniques much in the way that a jackknife can be used as a substitute for a variety of more specialized tools (although, in reality, a jackknife is not a particularly versatile tool!). More complete references on the jackknife are the monographs by Efron (1982), and Efron and Tibshirani (1993). The jackknife estimator of bias The jackknife estimator of bias was developed by Quenouille (1949) although he did not refer to it as the jackknife. The basic idea behind the jackknife estimators of bias and variance lies in recomputing the parameter estimator using all but one of the observations. Suppose that θ/ is an estimator of a parameter θ based on sample / of i.i.d. random variables X1 , · · · , Xn : θ/ = θ(X). (For example, / / θ = θ(F ) if θ = θ(F ).) Quenouille’s method for estimating the bias of θ/ is based on sequentially deleting a single observation Xi and recomputing θ/ based on n − 1 observations. Suppose that / = θ + b (θ) / Eθ (θ) θ / is the bias of θ. / Let θ/−i be the estimator of θ evaluated where bθ (θ) after deleting Xi from the sample: / 1 , · · · , Xi−1 , Xi+1 , · · · , Xn ). θ/−i = θ(X
Now deﬁne θ/• to be the average of θ/−1 , · · · , θ/−n : θ/• =
n 1 θ/−i . n i=1
The jackknife estimator of bias is then / / = (n − 1)(θ/• − θ). / b(θ) / A biascorrected version of θ/ can be constructed by subtracting /b(θ) / we will show below that this procedure reduces the bias of θ. / from θ; / / / The theoretical rationale behind b(θ) assumes that Eθ (θ) can be expressed as a series involving powers of 1/n; for simplicity, we will ﬁrst assume that for any n / =θ+ Eθ (θ) c 2000 by Chapman & Hall/CRC
a1 (θ) n
where a1 (θ) can depend on θ or the distribution of the Xi ’s but not / = a1 (θ)/n. Since θ−i is based the sample size n; in this case, bθ (θ) on n − 1 observations (for each i), it follows that Eθ (θ/• ) =
n 1 a1 (θ) Eθ (θ/−i ) = θ + . n i=1 n−1
Thus Eθ (θ/ − θ/• ) =
a1 (θ) a1 (θ) a1 (θ) − = n n−1 n(n − 1)
/ and so (n − 1)(θ/ − θ/• ) is an unbiased estimator of bθ (θ). In the general case, we will have / =θ+ Eθ (θ)
a1 (θ) a2 (θ) a3 (θ) + + ··· + n n2 n3
or
a1 (θ) a2 (θ) a3 (θ) + + ··· + n n2 n3 where a1 (θ), a2 (θ), a3 (θ), · · · can depend on θ or the distribution of the Xi ’s but not on n. Again, it follows that / = bθ (θ)
Eθ (θ/• ) =
n 1 Eθ (θ/−i ) n i=1
= θ+
a1 (θ) a3 (θ) a2 (θ) + + ··· + 2 n − 1 (n − 1) (n − 1)3
(since each θ/−i is based on n − 1 observations). Thus the expected value of the jackknife estimator of bias is
/ / Eθ (/b(θ)) = (n − 1) Eθ (θ/• ) − Eθ (θ)
=
a1 (θ) (2n − 1)a2 (θ) + n n2 (n − 1) (3n2 − 3n + 1)a3 (θ) + ···. + n3 (n − 1)2
/ is not an unbiased estimator of We can see from above that /b(θ) / as it was in the simple case considered earlier. However, note bθ (θ) / (namely a1 (θ)/n) agrees with that that the ﬁrst term of Eθ (/b(θ)) / Thus if we deﬁne of bθ (θ). / = nθ/ − (n − 1)θ/• θ/jack = θ/ − /b(θ) c 2000 by Chapman & Hall/CRC
/ it follows that to be the biascorrected (or jackknifed) version of θ,
a2 (θ) (2n − 1)a3 (θ) + ··· − n(n − 1) n2 (n − 1)2 a2 (θ) 2a3 (θ) − + ··· ≈ θ− n2 n3
Eθ (θ/jack ) = θ −
for large n. Since 1/n2 , 1/n3 , · · · go to 0 faster than 1/n goes to 0 (as n gets large), it follows that the bias of θ/jack is smaller than the bias of θ/ for n suﬃciently large. In the case where / =θ+ Eθ (θ)
a1 (θ) n
(so that a2 (θ) = a3 (θ) = · · · = 0), θ/jack will be unbiased. EXAMPLE 4.43: Suppose that X1 , · · · , Xn are i.i.d. random variables from a distribution with mean µ and variance σ 2 , both unknown. The estimator /2 = σ
n 1 ¯ 2 (Xi − X) n i=1
/ 2 ) = −σ 2 /n. Thus the bias in σ /2 is a biased estimator with b(σ can be removed by using the jackknife. An educated guess for the resulting unbiased estimator is
S2 =
n 1 ¯ 2. (Xi − X) n − 1 i=1
To ﬁnd the unbiased estimator using the jackknife, we ﬁrst note that 1 ¯ −i = 1 ¯ − Xi ) X Xj = (nX n − 1 j=i n−1 and so 2 /−i σ =
1 ¯ −i )2 (Xj − X n − 1 j=i
=
1 n ¯ Xi Xj − X+ n − 1 j=i n−1 n−1
=
n 1 ¯ + 1 (Xi − X) ¯ Xj − X n − 1 j=1 n−1
c 2000 by Chapman & Hall/CRC
− =
n 1 n ¯ 2+ ¯ 2 (Xj − X) (Xi − X) n − 1 j=1 (n − 1)3
− =
n2 ¯ 2 (Xi − X) (n − 1)3
n2 ¯ 2 (Xi − X) (n − 1)3
n 1 n ¯ 2− ¯ 2. (Xj − X) (Xi − X) n − 1 j=1 (n − 1)2
2 ’s so that /•2 is just the average of the σ /−i Now σ
/•2 = σ
n n 1 1 ¯ 2− ¯ 2 (Xi − X) (Xi − X) n − 1 i=1 (n − 1)2 i=1
and the unbiased estimator of σ 2 is / 2 − (n − 1)σ /•2 = nσ
n 1 ¯ 2 = S2 (Xi − X) n − 1 i=1
as was guessed above.
✸
EXAMPLE 4.44: Suppose that X1 , · · · , Xn are i.i.d. random variables with probability density function 1 for 0 ≤ x ≤ θ θ where θ is an unknown parameter. Since θ is the maximum possible value of the Xi ’s, a natural estimator of θ is f (x; θ) =
θ/ = X(n) = max(X1 , · · · , Xn ). However, since the Xi ’s cannot exceed θ, it follows that their maximum cannot exceed θ and so θ/ is biased; in fact, n θ n+1 1 = θ 1 + 1/n
1 1 1 = θ 1 − + 2 − 3 + ··· . n n n
/ = E(θ)
Since
θ/−i = max(X1 , · · · , Xi−1 , Xi+1 . · · · , Xn ),
c 2000 by Chapman & Hall/CRC
it follows that θ/−i = X(n) for n − 1 values of i and θ/−i = X(n−1) for the other value of i. Thus, we obtain θ/• =
n−1 1 X(n) + X(n−1) n n
and so the jackknifed estimator of θ is θ/jack = X(n) +
n−1 (X(n) − X(n−1) ). n
/ nonetheless, we The bias of θ/jack will be smaller than that of θ; / can easily modify θ to make it unbiased without resorting to the jackknife by simply multiplying it by (n + 1)/n. ✸
The latter example points out one of the drawbacks in using any general purpose method (such as the jackknife), namely that in speciﬁc situations, it is often possible to improve upon that method with one that is tailored speciﬁcally to the situation at hand. Removing the bias in θ/ = X(n) by multiplying X(n) by (n + 1)/n relies on the fact that the form of the density is known. Suppose instead that the range of the Xi ’s was still [0, θ] but that the density f (x) was unknown for 0 ≤ x ≤ θ. Then X(n) is still a reasonable estimator of θ and still always underestimates θ. However, (n+1)X(n) /n need not be unbiased and, in fact, may be more severely biased than X(n) . However, the jackknifed estimator θ/jack = X(n) +
n−1 (X(n) − X(n−1) ) n
will have a smaller bias than X(n) and may be preferable to X(n) in this situation. The jackknife estimator of variance The jackknife estimator of bias uses the estimators θ/−1 , · · · , θ/−n (which use all the observations but one in their computation) to / Tukey (1958) construct an estimator of bias of an estimator θ. / that uses θ/−1 , · · · , θ/−n . suggested a method of estimating Var(θ) / Tukey’s jackknife estimator of Var(θ) is / = 2 θ) Var( c 2000 by Chapman & Hall/CRC
n n−1 (θ/−i − θ/• )2 n i=1
where as before θ/−i is the estimator evaluated using all the observations except Xi and θ/• =
n 1 θ/−i . n i=1
The formula for the jackknife estimator of variance is somewhat unintuitive. In deriving the formula, Tukey assumed that the estimator θ/ can be approximated well by an average of independent random variables; this assumption is valid for a wide variety of estimators but is not true for some estimators (for example, sample maxima or minima). More precisely, Tukey assumed that θ/ ≈
n 1 φ(Xi ), n i=1
which suggests that / ≈ Var(θ)
Var(φ(X1 )) . n
(In the case where the parameter of interest θ is a functional parameter of the distribution function F (that is, θ = θ(F )), the function φ(·) − θ(F ) is typically the inﬂuence curve of θ(F ).) In general, we do not know the function φ(x) so we cannot use the above formula directly. However, it is possible to ﬁnd reasonable surrogates for φ(X1 ), · · · , φ(Xn ). Using the estimators / we deﬁne pseudovalues θ/−i (i = 1, · · · , n) and θ, Φi = θ/ + (n − 1)(θ/ − θ/−i ) (for i = 1, · · · , n) that essentially play the same role as the φ(Xi )’s above; in the case where θ = θ(F ), (n − 1)(θ/ − θ/−i ) is an attempt to estimate the inﬂuence curve of θ(F ) at x = Xi . (In the case where θ/ is exactly a sample mean θ/ =
n 1 φ(Xi ), n i=1
it easy to show that Φi = φ(Xi ) and so the connection between Φi and φ(Xi ) is clear in this simple case.) We can then take the sample variance of the pseudovalues Φi to be an estimate of the variance c 2000 by Chapman & Hall/CRC
Table 4.1 Pretax incomes for Example 4.45.
3841 22588 32528 39464 54339
7084 23972 32921 40506 57935
7254 25694 33724 44516 75137
15228 27592 36887 46538 82612
18042 27927 37776 51088 83381
19089 31576 37992 51955 84741
/ Note that of φ(X1 ) and use it to estimate the variance of θ. n n n−1 1 Φi = nθ/ − θ/−i n i=1 n i=1
= nθ/ − (n − 1)θ/• = θ/jack / The sample variance where θ/jack is the biascorrected version of θ. of the Φi ’s is n 1 ¯ 2 = (Φi − Φ) n − 1 i=1
n 1 [(n − 1)(θ/• − θ/−i )]2 n − 1 i=1
= (n − 1)
n
(θ/−i − θ/• )2 .
i=1
We now get the jackknife estimator of variance by dividing the sample variance of the Φi ’s by n: / = 2 θ) Var(
n n−1 (θ/−i − θ/• )2 . n i=1
It should be noted that the jackknife estimator of variance does not work in all situations. One such situation is the sample median; the problem here seems to be the fact that the inﬂuence curve of the median is deﬁned only for continuous distributions and so is diﬃcult to approximate adequately from ﬁnite samples. EXAMPLE 4.45: The data in Table 4.1 represent a sample of 30 pretax incomes. We will assume that these data are outcomes of i.i.d. random variables X1 , · · · , X30 from a distribution function F ; c 2000 by Chapman & Hall/CRC
Table 4.2 Values of θ/−i obtained by leaving out the corresponding entry in Table 4.1.
0.2912 0.3092 0.3153 0.3170 0.3152
0.2948 0.3103 0.3154 0.3170 0.3140
0.2950 0.3115 0.3157 0.3169 0.3069
0.3028 0.3127 0.3166 0.3167 0.3033
0.3055 0.3129 0.3168 0.3161 0.3028
0.3064 0.3148 0.3168 0.3159 0.3020
we will use the data to estimate the Gini index θ(F ) = 1 − 2 where
.t
qF (t) = . 01 0
1 0
qF (t) dt
F −1 (s) ds F −1 (s) ds
is the Lorenz curve. The substitution principle estimator of θ(F ) is θ/ = θ(F/ ) =
30 i=1
Xi
−1 30
2i − 1 i=1
30
− 1 X(i)
where X(1) ≤ X(2) ≤ · · · ≤ X(30) are the order statistics of X1 , · · · , X30 . For these data, the estimate of θ(F ) is 0.311. The standard error of this estimate can be estimated using the jackknife. The leaveoutestimates θ/−i of θ(F ) are given in Table 4.2. The jackknife estimate of the standard error of θ/ is / = / θ) se(
30 29 (θ/−i − θ/• )2 = 0.0398 30 i=1
where θ/• = 0.310 is the average of θ/−1 , · · · , θ/−30 .
✸
Comparing the jackknife and Delta Method estimators How does the jackknife estimator of variance compare to the Delta Method estimator? We will consider the simple case of estimating ¯ where X ¯ is the sample mean of i.i.d. random the variance of g(X) c 2000 by Chapman & Hall/CRC
variables X1 , · · · , Xn . The Delta Method estimator is 2 d (g(X)) ¯ = [g (X)] ¯ 2 Var
n 1 ¯ 2 (Xi − X) n(n − 1) i=1
while the jackknife estimator is 2 j (g(X)) ¯ = Var
n n−1 ¯ −i ) − g• )2 (g(X n i=1
where g• =
n 1 ¯ −i ). g(X n i=1
Recalling that 1 ¯ − Xi ) (nX n−1 ¯ − 1 (Xi − X), ¯ = X n−1
¯ −i = X
it follows from a Taylor series expansion that ¯ + (X ¯ −i − X)g ¯ (X) ¯ ¯ −i ) ≈ g(X) g(X ¯ (X) ¯ ¯ − 1 (Xi − X)g = g(X) n−1 and hence n 1 ¯ −i ) g(X n i=1 ¯ ≈ g(X).
g• =
2 j (g(X)), ¯ we get Substituting these approximations into Var 2 j (g(X)) ¯ ¯ 2 Var ≈ [g (X)]
n 1 ¯ 2 (Xi − X) n(n − 1) i=1
2 d (g(X)). ¯ = Var
Thus the jackknife and Delta Method estimators are approximately ¯ equal when θ/ = g(X). 4.10 Problems and complements 4.1: Suppose that X = (X1 , · · · , Xn ) has a oneparameter exponential family distribution with joint density or frequency funcc 2000 by Chapman & Hall/CRC
tion
f (x; θ) = exp [θT (x) − d(θ) + S(x)] where the parameter space Θ is an open subset of R. Show that
Eθ [exp(sT (X))] = d(θ + s) − d(θ) if s is suﬃciently small. (Hint: Since Θ is open, f (x; θ + s) is a density or frequency function for s suﬃciently small and hence integrates or sums to 1.) 4.2: Suppose that X = (X1 , · · · , Xn ) has a kparameter exponential family distribution with joint density or frequency function f (x; θ) = exp
, p

θi Ti (x) − d(θ) + S(x)
i=1
where the parameter space Θ is an open subset of Rk . (a) Show that ∂ d(θ) Eθ [Ti (X)] = ∂θi for i = 1, · · · , k. (b) Show that Covθ [Ti (X), Tj (X)] =
∂2 d(θ) ∂θi ∂θj
for i, j = 1, · · · , k. 4.3: Suppose that X1 , · · · , Xn are i.i.d. random variables with density
f (x; θ1 , θ2 ) =
a(θ1 , θ2 )h(x) for θ1 ≤ x ≤ θ2 0 otherwise
where h(x) is a known function deﬁned on the real line. (a) Show that θ2
a(θ1 , θ2 ) =
−1
h(x) dx
.
θ1
(b) Show that (X(1) , X(n) ) is suﬃcient for (θ1 , θ2 ). 4.4: Suppose that X = (X1 , · · · , Xn ) has joint density or frequency function f (x; θ1 , θ2 ) where θ1 and θ2 vary independently (that is, Θ = Θ1 × Θ2 ) and the set S = {x : f (x; θ1 , θ2 ) > 0} c 2000 by Chapman & Hall/CRC
does not depend on (θ1 , θ2 ). Suppose that T1 is suﬃcient for θ1 when θ2 is known and T2 is suﬃcient for θ2 when θ1 is known. Show that (T1 , T2 ) is suﬃcient for (θ1 , θ2 ) if T1 and T2 do not depend on θ2 and θ1 respectively. (Hint: Use the Factorization Criterion.) 4.5: Suppose that the lifetime of an electrical component is known to depend on some stress variable that varies over time; speciﬁcally, if U is the lifetime of the component, we have 1 P (x ≤ U ≤ x + ∆U ≥ x) = λ exp(βφ(x)) ∆↓0 ∆ lim
where φ(x) is the stress at time x. Assuming that we can measure φ(x) over time, we can conduct an experiment to estimate λ and β by replacing the component when it fails and observing the failure times of the components. Because φ(x) is not constant, the interfailure times will not be i.i.d. random variables. Deﬁne nonnegative random variables X1 < · · · < Xn such that X1 has hazard function λ1 (x) = λ exp(βφ(x)) and conditional on Xi = xi , Xi+1 has hazard function
λi+1 (x) =
0 if x < xi λ exp(βφ(x)) if x ≥ xi
where λ, β are unknown parameters and φ(x) is a known function. (a) Find the joint density of (X1 , · · · , Xn ). (b) Find suﬃcient statistics for (λ, β). 4.6: Let X1 , · · · , Xn be i.i.d. Exponential random variables with parameter λ. Suppose that we observe only the smallest r values of X1 , · · · , Xn , that is, the order statistics X(1) , · · · , X(r) . (This is called type II censoring in reliability.) (a) Find the joint density of X(1) , · · · , X(r) . (b) Show that V = X(1) + · · · + X(r−1) + (n − r + 1)X(r) is suﬃcient for λ. c 2000 by Chapman & Hall/CRC
4.7: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on [0, θ]: 1 f (x; θ) = for 0 ≤ x ≤ θ. θ Let X(1) = min(X1 , · · · , Xn ) and X(n) = max(X1 , · · · , Xn ). (a) Deﬁne T = X(n) /X(1) . Is T ancillary for θ? (b) Find the joint distribution of T and X(n) . Are T and X(n) independent? 4.8: Suppose that X1 , · · · , Xn are i.i.d. random variables with density function f (x; θ) = θ(1 + x)−(θ+1)
for x ≥ 0
where θ > 0 is an unknown parameter.
(a) Show that T = ni=1 ln(1 + Xi ) is suﬃcient for θ. (b) Find the mean and variance of T . 4.9: Consider the Gini index θ(F ) as deﬁned in Example 4.21. (a) Suppose that X ∼ F and let G be the distribution function of Y = aX for some a > 0. Show that θ(G) = θ(F ). (b) Suppose that Fp is a discrete distribution with probability p at 0 and probability 1 − p at x > 0. Show that θ(Fp ) → 0 as p → 0 and θ(Fp ) → 1 as p → 1. (c) Suppose that F is a Pareto distribution whose density is α f (x; α) = x0
x x0
−α−1
for x > x0 > 0
α > 0. (This is sometimes used as a model for incomes exceeding a threshold x0 .) Show that θ(F ) = (2α − 1)−1 for α > 1. (f (x; α) is a density for α > 0 but for α ≤ 1, the expected value is inﬁnite.) 4.10: An alternative to the Gini index as a measure of inequality is the Theil index. Given a distribution function F whose probability is concentrated on nonnegative values, the Theil index is deﬁned to be the functional parameter ∞
θ(F ) = .
0
x x ln µ(F ) µ(F )
dF (x)
where µ(F ) = 0∞ x dF (x). (a) Suppose that X ∼ F and let G be the distribution function of Y = aX for some a > 0. Show that θ(G) = θ(F ). c 2000 by Chapman & Hall/CRC
(b) Find the inﬂuence curve of θ(F ). (c) Suppose that X1 , · · · , Xn are i.i.d. random variables with distribution function F . Show that
n Xi Xi 1 / θn = ¯ ln ¯
n
i=1
Xn
Xn
is the substitution principle estimator of θ(F ). √ (d) Find the limiting distribution of n(θ/n − θ(F )). 4.11: The inﬂuence curve heuristic can be used to obtain the joint limiting distribution of a ﬁnite number of substitution principle estimators. Suppose that θ1 (F ), · · · , θk (F ) are functional parameters with inﬂuence curves φ1 (x; F ), · · · , φk (x; F ). Then if X1 , · · · , Xn is an i.i.d. sample from F , we typically have √
√
n(θ1 (F/n ) − θ1 (F )) = .. .
.. .
n(θk (F/n ) − θk (F )) =
n 1 √ φ1 (Xi ; F ) + Rn1 n i=1
.. .
n 1 √ φk (Xi ; F ) + Rnk n i=1
where Rn1 , · · · , Rnk →p 0. (a) Suppose that X1 , · · · , Xn are i.i.d. random variables from a distribution F with mean µ and median θ; assume that /n is the sample mean and Var(Xi ) = σ 2 and F (θ) > 0. If µ / θn is the sample median, use the inﬂuence curve heuristic to show that √ /n − µ µ n →d N2 (0, C) θ/n − θ and give the elements of the variancecovariance matrix C. (b) Now assume that the Xi ’s are i.i.d. with density p f (x; θ) = exp(−x − θp ) 2Γ(1/p) where θ is the mean and median of the distribution and p > 0 is another parameter (that may be known or unknown). Show that the matrix C in part (a) is
C=
Γ(3/p)/Γ(1/p) Γ(2/p)/p Γ(2/p)/p [Γ(1/p)/p]2 .
c 2000 by Chapman & Hall/CRC
/ + (1 − s)θ/n . (c) Consider estimators of θ of the form θ0n = sµ √n 0 For a given s, ﬁnd the limiting distribution of n(θn − θ). (d) For a given value of p > 0, ﬁnd the value of s that minimizes the variance of this limiting distribution. For which value(s) of p is this optimal value equal to 0; for which value(s) is it equal to 1? 4.12: (a) Suppose that F is a continuous distribution function with density f = F . Find the inﬂuence curve of the functional parameter θp (F ) deﬁned by F (θp (F )) = p for some p ∈ (0, 1). (θp (F ) is the pquantile of F .)
(b) Let F/n (x) be the empirical distribution function of i.i.d. random variables X1 , · · · , Xn (with continuous distribution F and density f = F ) and for 0 < t < 1 deﬁne F/n−1 (t) = inf{x : F/n (x) ≥ t}. Deﬁne τ/n = F/n−1 (0.75)− F/n−1 (0.25) to be the interquartile range of X1 , · · · , Xn . Find the limiting distribution of √ n(τ/n − τ (F )) where τ (F ) = θ3/4 (F )−θ1/4 (F ). (Hint: Find the inﬂuence curve of τ (F ); a rigorous derivation of the limiting distribution can be obtained by mimicking Examples 3.5 and 3.6.) 4.13: Suppose that X1 , X2 , · · · are i.i.d. nonnegative random variables with distribution function F and deﬁne the functional parameter . ( 0∞ x dF (x))2 θ(F ) = . ∞ 2 . 0 x dF (x) (Note that θ(F ) = (E(X))2 /E(X 2 ) where X ∼ F .) (a) Find the inﬂuence curve of θ(F ). (b) Using X1 , · · · , Xn , ﬁnd a substitution principle estimator, √ θ/n , of θ(F ) and ﬁnd the limiting distribution of n(θ/n − θ). (You can use either the inﬂuence curve or the Delta Method to do this.) 4.14: Sizebiased (or lengthbiased) sampling occurs when the size or length of a certain object aﬀects its probability of being sampled. For example, suppose we are interested in estimating the mean number of people in a household. We could take a random sample of households, in which case the natural c 2000 by Chapman & Hall/CRC
estimate would be the sample mean (which is an unbiased estimator). Alternatively, we could take a random sample of individuals and record the number of people in each individual’s household; in this case, the sample mean is typically not a good estimator since the sampling scheme is more likely to include individuals from large households than would be the case if households were sampled. In many cases, it is possible to correct for the biased sampling if the nature of the biased sampling is known. (Another example of biased sampling is given in Example 2.21.) (a) Suppose we observe i.i.d. random variables X1 , · · · , Xn from the distribution
∞
G(x) =
−1 x
w(t) dF (t)
w(t) dF (t)
0
0
where w(t) is a known (nonnegative) function and F is an unknown distribution function. Deﬁne F/n (x) =
n
[w(Xi )]−1
i=1
−1 n
[w(Xi )]−1 I(Xi ≤ x).
i=1
Show that for each x, F/n (x) is a consistent estimator of F (x) provided that E[1/w(Xi )] < ∞. (b) Using the estimator. in part (a), give a substitution principle estimator of θ(F ) = g(x) dF (x). What is the estimator of . x dF (x) when w(x) = x? Find the limiting distribution of this estimator when E[1/w2 (Xi )] < ∞. (c) Suppose that we have the option of sampling from F or from the. biased version G where w(x) = x. Show that the estimator of x dF (x) based on the biased sample is asymptotically more eﬃcient than that based on the sample from F if
3
x dF (x)
x−1 dF (x)
0
so that E(Λi ) = µ. Given Λi , let Xi and Yi be independent Poisson random variables with E(Xi Λi ) = Λi and E(Yi Λi ) = θΛi . We will observe i.i.d. pairs of (dependent) random variables (X1 , Y1 ), · · · , (Xn , Yn ) (that is, the Λi ’s are unobservable). (See Lee (1996) for an application of such a model.) (a) Show that the joint frequency function of (Xi , Yi ) is
f (x, y) =
α θy Γ(x + y + α) 1+θ+ α x!y! Γ(α)(µ/α) µ
−(x+y+α)
for x, y = 0, 1, 2, 3, · · ·. (Hint: P (Xi = x, Yi = y) = E[P (Xi = x, Yi = yΛi )].) c 2000 by Chapman & Hall/CRC
(b) Find the expected values and variances of Xi and Yi as well as Cov(Xi , Yi ). ¯ n is a consistent estimator of θ. (c) Show that θ/n = Y¯n /X √ (d) Find the asymptotic distribution of n(θ/n − θ). 4.19: Suppose that X1 , · · · , Xn are i.i.d. random variables with a continuous distribution function F . It can be shown that g(t) = E(Xi − t) (or g(t) = E[Xi − t − Xi ]) is minimized at t = θ where F (θ) = 1/2 (see Problem 1.25). This suggests that the median θ can be estimated by choosing θ/n to minimize gn (t) =
n
Xi − t.
i=1
(a) Let X(1) ≤ X(2) ≤ · · · ≤ X(n) be the order statistics. Show that if n is even then gn (t) is minimized for X(n/2) ≤ t ≤ X(1+n/2) while if n is odd then gn (t) is minimized at t = X((n+1)/2) . (Hint: Evaluate the derivative of gn (t) for X(i) < t < X(i+1) (i = 1, · · · , n − 1); determine for which values of t gn (t) is decreasing and for which it is increasing.) (b) Let F/n (x) be the empirical distribution function. Show that F/n−1 (1/2) = X(n/2) if n is even and F/n−1 (1/2) = X((n+1)/2) if n is odd. 4.20: Suppose that X1 , X2 , · · · are i.i.d. random variables with distribution function
x−µ F (x) = (1 − θ)Φ σ
x−µ + θΦ 5σ
where 0 < θ < 1, µ and σ are unknown parameters. (Φ is /n to be the sample the N (0, 1) distribution function.) Deﬁne µ 0n to be the sample median of X1 , · · · , Xn . (This is mean and µ an example of a contaminated Normal model that is sometimes used to study the robustness of estimators.) √ √ /n − µ) and n(µ 0n − (a) Find the limiting distributions of n(µ 2 µ). (These will depend on θ and σ .) (b) For which values (if any) of θ is the sample median more eﬃcient than the sample mean? 4.21: Suppose that X1 , · · · , Xn are i.i.d. random variables with distribution function. The substitution principle can be extended c 2000 by Chapman & Hall/CRC
to estimating functional parameters of the form θ(F ) = E[h(X1 , · · · , Xk )] where h is some speciﬁed function. (We assume that this expected value is ﬁnite.) If n ≥ k, a substitution principle estimator of θ(F ) is θ/ =
−1
n k
h(Xi1 , · · · , Xik )
i1 0. The likelihood function is L(θ) = θ−n I(0 ≤ x1 , · · · , xn ≤ θ) = θ−n I(θ ≥ max(x1 , · · · , xn )). Thus if θ < max(x1 , · · · , xn ), L(θ) = 0 while L(θ) is a decreasing function of θ for θ ≥ max(x1 , · · · , xn ). Hence, L(θ) attains its maximum at θ = max(x1 , · · · , xn ) and so θ/ = X(n) = max(X1 , · · · , Xn ) ✸
is the MLE of θ.
Note that in Example 5.1, we could have deﬁned the likelihood function to be L(θ) = θ−n I(θ < max(x1 , · · · , xn )) (by deﬁning the density of Xi to be f (x; θ) = 1/θ for 0 < x < θ). In this case, the MLE does not exist. We have sup L(θ) = [max(x1 , · · · , xn )]−n ; θ>0
however, there exists no S(x) such that L(S(x)) = sup L(θ). θ>0
c 2000 by Chapman & Hall/CRC
EXAMPLE 5.2: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ > 0. The likelihood function is L(λ) =
3 n exp(−λ)λxi
xi !
i=1
and the loglikelihood is ln L(λ) = −nλ + ln(λ)
n
xi −
i=1
n
ln(xi !).
i=1
Assuming that ni=1 xi > 0 and taking the derivative with respect to λ, we ﬁnd that n d 1 xi ln L(λ) = −n + dλ λ i=1
/=x and setting the derivative to 0 suggests that λ ¯. To verify that this is indeed a maximum, note that n d2 1 ln L(λ) = − xi , dλ2 λ2 i=1
which is always negative. Thus x ¯ maximizes the likelihood function / = X ¯ (for a given sample x , · · · , x ) and the MLE of λ is λ 1 n
n n provided that i=1 Xi > 0. If i=1 Xi = 0 then strictly speaking no MLE exists since the loglikelihood function ln L(λ) = −nλ has no maximum on the interval (0, ∞). ✸ EXAMPLE 5.3: Suppose that the joint density or frequency function of X = (X1 , · · · , Xn ) is a oneparameter exponential family; the loglikelihood function is then ln L(θ) = c(θ)T (x) − d(θ) + S(x). Diﬀerentiating with respect to θ and setting this derivative to 0, we / get the following equation for the MLE θ: / c (θ) / d (θ)
= T (X).
However, since T (X) is the suﬃcient statistic, we know that Eθ [T (X)] = c 2000 by Chapman & Hall/CRC
d (θ) c (θ)
and so it follows that the MLE is simply a method of moments estimator for oneparameter exponential families. Also note that θ/ does maximize the likelihood function; the second derivative of the loglikelihood function is d2 ln L(θ) = c (θ)T (x) − d (θ) dθ2 and substituting θ/ for θ, we get / d2 / = c (θ) / d (θ) − d (θ) / 0. c (θ) A similar result also holds for kparameter exponential families. ✸ Varθ [T (X)] = d (θ) − c (θ)
Why does maximum likelihood estimation make sense? For simplicity, assume that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f0 (x) and distribution function F0 (x). For any other density (frequency) function f (x), we can deﬁne the KullbackLeibler information number K(f : f0 ) = E0 [ln (f0 (Xi )/f (Xi ))] where the expected value E0 is computed assuming that f0 (x) is the true density (frequency) function of Xi . K(f : f0 ) can be interpreted as measuring the distance to the “true” density (frequency) function f0 of some other density (frequency) function f ; it is easy to see that K(f0 : f0 ) = 0 and since − ln(x) is a convex function, it follows (from Jensen’s inequality) that K(f : f0 ) = E0 [− ln (f (Xi )/f0 (Xi ))] ) * f (Xi ) ≥ − ln E0 f0 (Xi ) = 0. Thus over all density (frequency) functions f , K(f : f0 ) is minimized (for a given f0 ) at f = f0 . Moreover, unless f (x) = f0 (x) for all x, K(f : f0 ) > 0. K(f : f0 ) can also be interpreted as the c 2000 by Chapman & Hall/CRC
ineﬃciency in assuming that the density (frequency) function is f when the true density (frequency) function is f0 . We will now use these facts to rationalize maximum likelihood estimation in the case of i.i.d. observations. Since K(f : f0 ) = −E0 [ln(f (Xi ))] + E0 [ln(f0 (Xi ))] , we can also see that for ﬁxed f0 , L(f : f0 ) = E0 [ln(f (Xi ))] is maximized over all f at f = f0 . This suggests the following method for estimating the density (frequency) function f0 of i.i.d. observations X1 , · · · , Xn : • For each f in some family F, estimate L(f : f0 ) using the substitution principle n 1 / L(f : f0 ) = ln f (Xi );
n
i=1
/ : f0 ). • Find f to maximize L(f / : f0 ) is simply the loglikelihood If F = {f (x; θ) : θ ∈ Θ}, then nL(f function. Thus maximum likelihood estimation can be viewed as a sort of substitution principle estimation. Maximum likelihood estimation has been described as the original jackknife in the sense that it is an estimation procedure that is applicable in a wide variety of problems. As will be shown in Chapter 6, maximum likelihood estimation has some very attractive optimality properties and for this reason, it is often viewed as the “gold standard” of estimation procedures. However, as with any methodology, maximum likelihood estimation should be viewed with the appropriate degree of scepticism as there may be estimation procedures that are better behaved in terms of robustness or eﬃciency when the model is slightly (or grossly) misspeciﬁed.
5.3 The likelihood principle The likelihood function has a much greater signiﬁcance in statistical inference; we will discuss this brieﬂy in this section. Suppose we are given a choice between two experiments for estimating a parameter θ. From the ﬁrst experiment, we obtain data x while x∗ is obtained from the second experiment. The likelihood principle provides a c 2000 by Chapman & Hall/CRC
simple criterion for when identical inferences for θ should be drawn from both x and x∗ ; it is stated as follows: Likelihood principle. Let L(θ) be the likelihood function for θ based on observing X = x and L∗ (θ) be the likelihood function for θ based on observations X ∗ = x∗ . If L(θ) = kL∗ (θ) (for some k that does not depend on θ) then the same inference for θ should be drawn from both samples. In point estimation, the likelihood principle implies that if x and x∗ are two samples with likelihood functions L(θ) and L∗ (θ) where L(θ) = kL∗ (θ) (for all θ) then the two point estimates T (x) and T ∗ (x∗ ) should be equal. Clearly, maximum likelihood estimation satisﬁes this condition but many of the substitution principle and other estimators discussed in Chapter 4 do not. For example, the biasedreduced jackknife estimator typically violates the likelihood principle. However, certain inferential procedures based on maximum likelihood estimation may violate the likelihood principle. It is also worth noting at this point that the likelihood principle refers only to the information about θ contained in the sample; we may also have additional information about θ available to us from outside the sample. For example, we can express our beliefs about the true value of θ (prior to observing the sample) by specifying a probability distribution over the parameter space Θ; this distribution is called a prior distribution. This approach (called the Bayesian approach) is often cited as the “correct” approach to implementing the likelihood principle. More details on the Bayesian approach are given in section 5.7. Does the likelihood principle make sense? This is a very controversial issue in theoretical statistics and there is really no clear answer. The likelihood principle can be shown to follow by assuming two other somewhat less controversial principles, the (weak) suﬃciency principle and the (weak) conditionality principle. The suﬃciency principle states that the suﬃcient statistic contains as much information about the value of a parameter as the data themselves while the conditionality principle essentially states that no information is lost by conditioning on an ancillary statistic. For some compelling arguments in favour of the likelihood principle, see the monograph by Berger and Wolpert (1988). The book by Lee (1989) also contains some interesting discussion and examples on the likelihood principle. c 2000 by Chapman & Hall/CRC
The likelihood principle is not universally accepted; in fact, much of classical statistical practice violates it. Of course, the likelihood principle assumes that the true parametric model for the data is known; we have stated previously that statistical models are almost inevitably used to approximate reality and are seldom exactly true. This, of course, does not immediately invalidate the likelihood principle but rather emphasizes the point that any principle or philosophy should not be accepted without closely examining its tenets. 5.4 Asymptotic theory for MLEs Under what conditions is the MLE a consistent and asymptotically Normal estimator of a parameter? We will show in this section that, under fairly mild regularity conditions, it is possible to prove consistency and asymptotic normality for the MLE of a realvalued parameter based on i.i.d. observations. However, in many cases, it is possible to ﬁnd the asymptotic distribution of a sequence of MLEs using standard techniques. This is common when the MLE corresponds to a method of moments estimator. EXAMPLE 5.4: Suppose that X1 , · · · , Xn are i.i.d. Geometric random variables with frequency function f (x; θ) = θ(1 − θ)x
for x = 0, 1, 2, · · ·
The MLE of θ based on X1 , · · · , Xn is 1 . θ/n = ¯ Xn + 1 By the Central Limit Theorem, we have that √ ¯ n − (θ−1 − 1)) →d N (0, θ−2 (1 − θ)). n(X Thus we obtain √ / n(θn − θ)
√ ¯ n ) − g(θ−1 − 1)) = n(g(X →d N (0, θ2 (1 − θ))
by applying the Delta Method with g(x) = 1/(1 + x) and g (x) = −1/(1 + x)2 . ✸ EXAMPLE 5.5: Suppose that X1 , · · · , Xn are i.i.d. Uniform c 2000 by Chapman & Hall/CRC
random variables on [0, θ]. The MLE of θ is θ/n = X(n) = max(X1 , · · · , Xn ) whose distribution function is given by P (θ/n ≤ x) =
n
x θ
for 0 ≤ x ≤ θ.
Thus for > 0, P (θ/n − θ > )
P (θ/n < θ − ) θ− n = θ → 0 as n → ∞ =
since (θ − )/θ < 1. We also have that P n(θ − θ/n ) ≤ x
)
=
* x / P θn ≥ θ − n
n
x θn → 1 − exp(−x/θ) =
1− 1−
for x ≥ 0
and so n(θ− θ/n ) converges in distribution to an Exponential random variable. ✸ For the remainder of this section, we will assume that X1 , X2 , · · · , Xn are i.i.d. random variables with common density or frequency function f (x; θ) where θ is a realvalued parameter. Deﬁne :(x; θ) = ln f (x; θ) and let : (x; θ), : (x; θ), and : (x; θ) be the ﬁrst three partial derivatives of :(x; θ) with respect to θ. We will make the following assumptions about f (x; θ): (A1) The parameter space Θ is an open subset of the realline. (A2) The set A = {x : f (x; θ) > 0} does not depend on θ. (A3) f (x; θ) is three times continuously diﬀerentiable with respect to θ for all x in A. (A4) Eθ [: (Xi ; θ)] = 0 for all θ and Varθ [: (Xi ; θ)] = I(θ) where 0 < I(θ) < ∞ for all θ. (A5) Eθ [: (Xi ; θ)] = −J(θ) where 0 < J(θ) < ∞ for all θ. (A6) For each θ and δ > 0, : (x; t) ≤ M (x) for θ − t ≤ δ where Eθ [M (Xi )] < ∞. c 2000 by Chapman & Hall/CRC
Suppose that f (x; θ) is a density function. If condition (A2) holds then f (x; θ) dx = 1 for all θ ∈ Θ A
and so
d f (x; θ) dx = 0. dθ A If the derivative can be taken inside the integral we then have
∂ f (x; θ) dx ∂θ A : (x; θ)f (x; θ) dx =
0 =
A
= Eθ [: (Xi ; θ)] and so the assumption that Eθ [: (Xi ; θ)] = 0 is, in many cases, a natural consequence of condition (A2). Moreover, if we can . diﬀerentiate A f (x; θ) dx twice inside the integral sign, we have
∂ : (x; θ)f (x; θ) dx ∂θ A 2 : (x; θ)f (x; θ) dx + : (x; θ) f (x; θ) dx =
0 =
A
= −J(θ) + I(θ)
A
and so I(θ) = J(θ). (Similar results apply if f (x; θ) is a frequency function.) I(θ) is called the Fisher information. EXAMPLE 5.6: Suppose that X1 , X2 , · · · are i.i.d. random variables with the oneparameter exponential family density or frequency function f (x; θ) = exp[c(θ)T (x) − d(θ) + S(x)]
for x ∈ A.
In this case, : (x; θ) = c (θ)T (x) − d (θ) and : (x; θ) = c (θ)T (x) − d (θ). Since we have Eθ [T (Xi )] = and
Varθ [T (Xi )] =
c 2000 by Chapman & Hall/CRC
d (θ) c (θ)
1 d (θ) d (θ) − c (θ) , [c (θ)]2 c (θ)
it follows that Eθ [: (Xi ; θ)] = c (θ)Eθ [T (Xi )] − d (θ) = 0, I(θ) = [c (θ)]2 Varθ [T (Xi )] d (θ) = d (θ) − c (θ) c (θ) and J(θ) = d (θ) − c (θ)Eθ [T (Xi )] d (θ) = d (θ) − c (θ) c (θ) ✸
and so I(θ) = J(θ).
EXAMPLE 5.7: Suppose that X1 , · · · , Xn are i.i.d. random variables with a Logistic distribution whose density function is f (x; θ) =
exp(x − θ) . [1 + exp(x − θ)]2
The derivatives of :(x; θ) = ln f (x; θ) with respect to θ are : (x; θ) =
exp(x − θ) − 1 1 + exp(x − θ)
and : (x; θ) = −2 It follows then that
Eθ [: (Xi ; θ)] =
exp(x − θ) . [1 + exp(x − θ)]2
∞ exp(2(x − θ)) − exp(x − θ) −∞
[1 + exp(x − θ)]3
dx = 0,
I(θ) = Varθ [: (Xi ; θ)] ∞
= = and J(θ) = = =
−∞
(exp(x − θ) − 1)2
exp(x − θ) dx [1 + exp(x − θ)]4
1 3 −Eθ [: (Xi ; θ)] ∞ exp(2(x − θ)) 2 dx 4 −∞ [1 + exp(x − θ)] 1 3
c 2000 by Chapman & Hall/CRC
✸
and so I(θ) = J(θ).
Under conditions (A1) to (A3), if θ/n maximizes the likelihood, we have n : (Xi ; θ/n ) = 0 i=1
and expanding this equation in a Taylor series expansion, we get 0=
n
: (Xi ; θ/n ) =
i=1
n
: (Xi ; θ)
i=1
+(θ/n − θ)
n
: (Xi ; θ)
i=1 n 1 + (θ/n − θ)2 : (Xi ; θn∗ ) 2 i=1
√ where θn∗ lies between θ and θ/n . Dividing both sides by n, we get 0 =
n n √ 1 1 √ : (Xi ; θ) + n(θ/n − θ) : (Xi ; θ) n i=1 n i=1 n 1√ / 21 + n(θn − θ) : (Xi ; θn∗ ), 2 n i=1
which suggests that √ / n(θn − θ) =
n
n −1
−n−1/2
i=1 : (Xi ; θ)
+
n
i=1 : (Xi ; θ) .
(θ/n − θ)(2n)−1 ni=1 : (Xi ; θn∗ )
From the Central Limit Theorem (and condition (A4)), it follows that n 1 √ : (Xi ; θ) →d Z ∼ N (0, I(θ)) n i=1 and from the WLLN (and condition (A5)), we have that n 1 : (Xi ; θ) →p −J(θ). n i=1
Thus it follows from Slutsky’s Theorem that √ / Z n(θn − θ) →d ∼ N (0, I(θ)/J 2 (θ)) J(θ) c 2000 by Chapman & Hall/CRC
provided that (θ/n − θ)
n 1 ψ (Xi ; θn∗ ) →p 0; n i=1
we will show later that this latter statement holds provided that condition (A6) holds and θ/n →p θ. Proving consistency of the MLE θ/n is somewhat subtle. Consider the (random) function φn (t) =
n 1 [ln f (Xi ; t) − ln f (Xi ; θ)] , n i=1
which is maximized at t = θ/n . By the WLLN, for each ﬁxed t ∈ Θ, )
φn (t) →p φ(t) = Eθ
f (Xi ; t) ln f (Xi ; θ)
*
.
Now note that −φ(t) is simply a KullbackLeibler information number, which is minimized when t = θ and so φ(t) is maximized at t = θ (φ(θ) = 0). Moreover, unless f (x; t) = f (x; θ) for all x ∈ A then φ(t) < 0; since we are assuming (implicitly) identiﬁability of the model, it follows that φ(t) is uniquely maximized at t = θ. Does the fact that φn (t) →p φ(t) for each t (where φ(t) is maximized at t = θ) imply that θ/n →p θ? Unfortunately, the answer to this question is, in general, “no” (unless we make more assumptions about the φn ’s). To keep things simple, we will consider some examples where {φn (t)} and φ(t) are nonrandom functions. EXAMPLE 5.8: Let {φn (t)} be a sequence of functions with + + 1 − n +t −
φn (t) =
1 2
− t − 2
0
+
1+ n+
for 0 ≤ t ≤ 2/n for 3/2 ≤ t ≤ 5/2 otherwise.
Note that φn (t) is maximized at tn = 1/n and φn (1/n) = 1. It is easy to see that for each t, φn (t) → φ(t) where φ(t) =
1 2 − t − 2 for 3/2 ≤ t ≤ 5/2
0
otherwise.
Thus φ(t) is maximized at t0 = 2; clearly, tn → 0 = t0 .
✸
What goes wrong in the previous example? The main problem c 2000 by Chapman & Hall/CRC
seems to be that although φn (t) → φ(t) for each t, the convergence is not uniform; that is, for any M > 0, sup φn (t) − φ(t) = 1
t≤M
(for n suﬃciently large)
and so φn (t) does not converge to φ(t) uniformly for t ≤ M . However, this uniform convergence is not by itself suﬃcient to guarantee the convergence of the maximizers of φn (t) to the maximizer of φ(t) as evidenced by the following example. EXAMPLE 5.9: Let {φn (t)} be a sequence of functions with φn (t) =
1 2 (1 − 2t)
for t ≤ 1/2 1 − 2t − n for n − 1/2 ≤ t ≤ t + 1/2 0 otherwise.
It is easy to see that for any M > 0, sup φn (t) − φ(t) → 0
t≤M
where φ(t) =
1 2 (1 − 2t)
0
for t ≤ 1/2 otherwise.
However, φn (t) is maximized at tn = n while φ(t) is maximized at t0 = 0; again, tn → ∞ = t0 . ✸ Even though uniform convergence of φn (t) to φ(t) holds over closed and bounded sets in this example, the sequence of maximizers {tn } cannot be contained within a closed and bounded set and so the sequence does not converge to the maximizer of φ(t). The following result shows that adding the condition that the sequence of maximizers is bounded is suﬃcient for convergence of tn to t0 where t0 maximizes φ(t); moreover, this result covers the case where {φn (t)} and φ(t) are random. THEOREM 5.1 Suppose that {φn (t)} and φ(t) are realvalued random functions deﬁned on the real line. Suppose that (a) for each M > 0, sup φn (t) − φ(t) →p 0;
t≤M
(b) Tn maximizes φn (t) and T0 is the unique maximizer of φ(t); (c) for each > 0, there exists M0 such that P (Tn  > M0 ) < for c 2000 by Chapman & Hall/CRC
all n. Then Tn →p T0 . As appealing as Theorem 5.1 seems to be, it is very diﬃcult to use in practice. For example, although it is not too diﬃcult to establish the uniform convergence, it can be extremely diﬃcult to show that P (Tn  > M0 ) < . However, if we assume that φn (t) is a concave function for each n, we can weaken the conditions of the previous theorem considerably. THEOREM 5.2 Suppose that {φn (t)} and φ(t) are random concave functions. If (a) for each t, φn (t) →p φ(t), and (b) Tn maximizes φn (t) and T0 is the unique maximizer of φ(t) then Tn →p T0 . In most applications of the preceding theorem, φn (t) will be an average of n random variables (depending on t) and so we can use the WLLN (or a similar result) to show that φn (t) →p φ(t); in such cases, the limiting function φ(t) will be nonrandom and so its maximizer T0 will also be nonrandom. Theorem 5.2 also holds if the functions {φn } are deﬁned on Rp ; the same is true for Theorem 5.1. Theorem 5.2 can also be used if the functions φn (t) are convex. Since the negative of a convex function is a concave function, we can use this theorem to establish convergence in probability of minimizers of convex functions. EXAMPLE 5.10: Suppose that X1 , X2 , · · · , Xn are i.i.d. random variables with continuous distribution function F (x) and assume that µ is the unique median of the distribution so that F (µ) = 0.5. /n of X1 , · · · , Xn can be deﬁned as a minimizer The sample median µ of n
Xi − t
i=1
or equivalently as the minimizer of n 1 φn (t) = [Xi − t − Xi ] . n i=1
It is easy to see that φn (t) is a convex function since a−t is convex in t for any a. By the WLLN, φn (t) →p E [X1 − t − X1 ] = φ(t) c 2000 by Chapman & Hall/CRC
and it can be easily shown that φ(t) is minimized at t = µ. Thus /n →p µ. (Note that this result is valid even if E[Xi ] = ∞ since µ E[Xi − t − Xi ] < ∞ for all t.) It is interesting to compare this proof of consistency to the proof given in Example 3.3. ✸ EXAMPLE 5.11: Suppose that X1 , X2 , · · · , Xn are i.i.d. random variables from a oneparameter exponential family density or frequency function f (x; θ) = exp [c(θ)T (x) − d(θ) + S(x)]
for x ∈ A.
The MLE of θ maximizes φn (t) =
n 1 [c(t)T (Xi ) − d(t)] ; n i=1
however, φn (t) is not necessarily a concave function. Nonetheless, if c(·) is a onetoone continuous function with inverse c−1 (·), we can deﬁne u = c(t) and consider φ∗n (u) =
n 1 [uT (Xi ) − d0 (u)] = φn (c(t)) n i=1
where d0 (u) = d(c−1 (u)). It follows that φ∗n (u) is a concave function since its second derivative is −d0 (u), which is negative. By the WLLN, for each u, we have φ∗n (u) →p uEθ [T (X1 )] − d0 (u) = φ∗ (u) and φ∗ (u) is maximized when d0 (u) = Eθ [T (X1 )]. Since Eθ [T (X1 )] = d0 (c(θ)), it follows that φ∗ (u) is maximized at c(θ). Since u = c(t), it follows that c(θ/n ) →p c(θ) and since c(·) is onetoone and continuous, it ✸ follows that θ/n →p θ. We will now state a result concerning the asymptotic normality of MLEs for i.i.d. sequences. We will assume that consistency of the estimators has been proved. THEOREM 5.3 (Asymptotic normality of MLEs) Suppose that X1 , X2 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) that satisﬁes conditions (A1)(A6) and supc 2000 by Chapman & Hall/CRC
pose that the MLEs satisfy θ/n →p θ where n
: (Xi ; θ/n ) = 0.
i=1
Then
√
n(θ/n − θ) →d N (0, I(θ)/J 2 (θ)). √ When I(θ) = J(θ), we have n(θ/n − θ) →d N (0, 1/I(θ)). Proof. From above we have that √ / n(θn − θ) =
n−1
n
−n−1/2
i=1 : (Xi ; θ)
+
n
i=1 : (Xi ; θ) .
(θ/n − θ)(2n)−1 ni=1 : (Xi ; θn∗ )
Given our previous development, we need only show that Rn = (θ/n − θ)
n 1 : (Xi ; θn∗ ) →p 0. 2n i=1
We have that for any > 0, P (Rn  > ) = P (Rn  > , θ/n − θ > δ) + P (Rn  > , θ/n − θ ≤ δ) and
P (Rn  > , θ/n − θ > δ) ≤ P (θ/n − θ > δ) → 0
as n → ∞. If θ/n − θ ≤ δ , we have (by condition (A6)), Rn  ≤ and since
n δ M (Xi ) 2n i=1
n 1 M (Xi ) →p Eθ [M (X1 )] < ∞ n i=1
(by the WLLN), it follows that P (Rn  > , θ/n − θ ≤ δ) can be made arbitrarily small (for large n) by taking δ suﬃciently small. Thus Rn →p 0 and so √ / n(θn − θ) →d N (0, I(θ)/J 2 (θ)) applying Slutsky’s Theorem. c 2000 by Chapman & Hall/CRC
The regularity conditions (A1) through (A6) are by no means minimal conditions. In particular, it is possible to weaken the diﬀerentiability conditions at the cost of increasing the technical diﬃculty of the proof. For example, we could replace conditions (A5) and (A6) by a somewhat weaker condition that n 1 : (Xi ; Tn ) →p −J(θ) n i=1
for any sequence of random variables {Tn } with Tn →p θ; it is easy to see that this latter condition is implied by (A5) and (A6). However, conditions (A1) through (A6) are satisﬁed for a wide variety of oneparameter models and are relatively simple to verify. An alternative condition to (A6) is considered in Problem 5.11. Estimating standard errors In all but rare cases, we have I(θ) = J(θ) and so the result of Theorem 5.3 suggests that for suﬃciently large n, the MLE θ/n is approximately Normal with mean θ and variance 1/(nI(θ)). This result can be used to approximate the standard error of θ/n by [nI(θ)]−1/2 . Since I(θ) typically depends on θ, it is necessary to estimate I(θ) to estimate the standard error of θ/n . There are two approaches to estimating I(θ) and hence the standard error of θ/n . • If I(θ) has a closedform, we can substitute θ/n for θ; our standard error estimator becomes 1 / θ/n ) = . se( nI(θ/n ) nI(θ/n ) is called the expected Fisher information for θ. • Since I(θ) = −Eθ [: (Xi ; θ)], we can estimate I(θ) by 2 = −1 I(θ)
n
n
: (Xi ; θ/n ),
i=1
which leads to the standard error estimator 1 / θ/n ) = se( = 2 nI(θ)
−
n
−1/2 / : (Xi ; θn ) .
i=1
2 is called the observed Fisher information for θ. nI(θ) c 2000 by Chapman & Hall/CRC
EXAMPLE 5.12: Suppose X1 , · · · , Xn have a oneparameter exponential family density or frequency function f (x; θ) = exp [c(θ)T (x) − d(θ) + S(x)]
for x ∈ A
Using the facts that n 1 d (θ/n ) T (Xi ) = n i=1 c (θ/n )
and
d (θ) , c (θ) it follows that the expected and observed Fisher information for θ are the same. ✸ I(θ) = d (θ) − c (θ)
EXAMPLE 5.13: Suppose that X1 , · · · , Xn are i.i.d. Cauchy random variables with density function 1 1 f (x; θ) = . π 1 + (x − θ)2 Then
: (x; θ) =
and : (x; θ) = −
2(x − θ) 1 + (x − θ)2 2(1 − (x − θ)2 ) . (1 + (x − θ)2 )2
It possible to show that
1 2 and so the expected Fisher information for θ is n/2 and the corresponding standard error estimate is I(θ) =
4
2 . n On the other hand, the observed Fisher information for θ is / θ/n ) = se(
n 2(1 − (Xi − θ/n )2 ) (1 + (Xi − θ/n )2 )2 i=1
(which is not equal to the expected Fisher information) and so the corresponding standard error estimator is n −1/2 2(1 − (Xi − θ/n )2 ) / θ/n ) = se( . (1 + (Xi − θ/n )2 )2 i=1
c 2000 by Chapman & Hall/CRC
It is interesting to compare the limiting variance of the MLE with that of the trimmed means in Example 4.40; we can compute the minimum limiting variance for all trimmed means to 2.278, which implies that the MLE is more eﬃcient since its limiting variance is 2. ✸ In Example 5.13, we see that diﬀerent estimates of the standard error can result by using the observed and expected Fisher information. This raises the question of whether either of the two estimators can be shown to be superior to the other. While there is some debate, results of Efron and Hinkley (1978) indicate that there are reasons to prefer the estimator based on the observed Fisher information. Their rationale is that this estimator actually estimates 1/2 Varθ (θ/n S = s) where S = S(X) is an ancillary statistic for θ and s = S(x). Multiparameter models Extending the consistency and asymptotic normality results from the single to multiparameter cases is simple, if somewhat notationally messy. Assume that X1 , X2 , · · · are i.i.d. random variables with density or frequency function f (x; θ) where θ = (θ1 , · · · , θp ). The MLE θ/n based on X1 , · · · , Xn satisﬁes the likelihood equations n
/n) = 0 : (Xi ; θ
i=1
: (x; θ)
where now is the vector of partial derivatives of :(x; θ) = ln f (x; θ) (with respect to the components of θ). The idea behind proving asymptotic normality is really exactly the same as that used in the single parameter case: we make a Taylor series expansion of the likelihood equations around the true parameter value. Doing this we get n 1 0= √ : (Xi ; θ) + n i=1
n √ / 1 : (Xi ; θ ∗n ) n(θ n − θ) n i=1
where (a) : (x; θ) is the matrix of second partial derivatives (with respect to θ1 , · · · , θp ); the (j, k) element of this matrix is given by :jk (x; θ) = c 2000 by Chapman & Hall/CRC
∂2 :(x; θ). ∂θj ∂θk
/n. (b) θ ∗n is on the line segment joining θ and θ √ / Now solving for n(θ n − θ), we get
√
/ n − θ) = n(θ
−1
n 1 − : (Xi ; θ ∗n ) n i=1
n 1 √ : (Xi ; θ) n i=1
and so, under appropriate regularity conditions, we should have n 1 √ : (Xi ; θ) →d Np (0, I(θ)) n i=1
(with I(θ) = Covθ [: (Xi ; θ)]) and n ' ( 1 : (Xi ; θ ∗n ) →p −J(θ) = Eθ : (X1 ; θ) n i=1
where I(θ) and J(θ) are p × p matrices. Now provided J(θ) is invertible, we have √ / n(θ n − θ) →d Np (0, J(θ)−1 I(θ)J(θ)−1 ). As in the single parameter case, for many models (including exponential families) we have I(θ) = J(θ) in which case the limiting variancecovariance matrix J(θ)−1 I(θ)J(θ)−1 above becomes I(θ)−1 . I(θ) = J(θ) if we are able to diﬀerentiate twice inside the integral (or summation) sign with respect to all p components of θ. (I(θ) is called the Fisher information matrix.) We now state the regularity conditions that are suﬃcient to “rigorize” the previous argument. The conditions are simply analogues of conditions (A1)(A6) used previously. (B1) The parameter space Θ is an open subset of Rp . (B2) The set A = {x : f (x; θ) > 0} does not depend on θ. (B3) f (x; θ) is three times continuously diﬀerentiable with respect to θ for all x in A. (B4) Eθ [: (Xi ; θ)] = 0 for all θ and Covθ [: (Xi ; θ)] = I(θ) where I(θ) is positive deﬁnite for all θ. (B5) Eθ [: (Xi ; θ)] = −J(θ) where J(θ) is positive deﬁnite for all θ. (B6) Let : jkl (x; θ) be the mixed partial derivative of :(x; θ) with respect to θj , θk , θl . For each θ, δ > 0 and 1 ≤ j, k, l ≤ p, : jkl (x; t) ≤ Mjkl (x) for θ − t ≤ δ where Eθ [Mjkl (Xi )] < ∞. c 2000 by Chapman & Hall/CRC
THEOREM 5.4 Suppose that X1 , X2 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) that satisﬁes / n →p θ conditions (B1)(B6) and suppose that the MLEs satisfy θ where n / n ) = 0. : (Xi ; θ i=1
Then
√
/ n − θ) → Np (0, J(θ)−1 I(θ)J(θ)−1 ). n(θ d √ / −1 When I(θ) = J(θ), we have n(θ n − θ) →d Np (0, I(θ) ).
The proof of Theorem 5.4 parallels that of Theorem 5.3 and will /n. not be given. Note that Theorem 5.4 assumes consistency of θ This can be proved in a variety of ways; for example, in the case of pparameter exponential families, Theorem 5.2 can be used as in Example 5.11. It should also be noted that both Theorems 5.3 and 5.4 hold even if X1 , X2 , · · · , Xn are i.i.d. random vectors. EXAMPLE 5.14: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance σ 2 . The MLEs of µ and σ 2 are n 1 /n )2 . (Xi − µ n i=1 √ √ /n − µ) and n(σ / The joint limiting distribution of n(µ √n − σ) can /n − µ) be derived quite easily using Theorem 5.4. (Of course, n(µ is exactly Normal for any n.) Writing
/n = µ
n 1 Xi n i=1
:(x; µ, σ) = −
/n2 = and σ
1 1 (x − µ)2 − ln(σ) − ln(2π) 2 2σ 2
and taking partial derivatives of : with respect to µ and σ, it is easy to show that
I(µ, σ) = J(µ, σ) =
σ −2 0 0 2σ −2
.
Thus (after verifying conditions (B1)(B6) of Theorem 5.4 hold), we have
√ /n − µ µ n →d N2 (0, I −1 (µ, σ)); /n − σ σ note that the variancecovariance matrix I −1 (µ, σ) is diagonal with diagonal entries σ 2 and σ 2 /2. It is also worth noting that we could c 2000 by Chapman & Hall/CRC
have determined the joint limiting distribution via the Multivariate CLT more or less directly. ✸ EXAMPLE 5.15: Suppose that X1 , · · · , Xn are i.i.d. Gamma random variables with shape parameter α and scale parameter λ. The MLEs of α and λ satisfy the equations n 1 Xi = n i=1
/n α /n λ
n 1 /n) / n ) − ln(λ ln(Xi ) = ψ(α n i=1
where ψ(α) is the derivative of ln Γ(α). (ψ(α) is called the digamma function and its derivative ψ (α) the trigamma function.) Since this is a twoparameter exponential family model, we can compute the information matrix by looking at the second partial derivatives of the logdensity; we have
and
∂2 ln f (x; α, λ) = −ψ (α) ∂α2 α ∂2 ln f (x; α, λ) = − 2 ∂λ2 λ 1 ∂2 ln f (x; α, λ) = . ∂α∂λ λ
Since none of these derivatives depends on x, we can easily compute
I(α, λ) =
ψ (α) −1/λ −1/λ α/λ2
.
Inverting I(α, λ), we obtain the limiting variancecovariance matrix I −1 (α, λ) =
λ2 αψ (α) − 1
α/λ2 1/λ 1/λ ψ (α)
.
In particular, this implies that √ / n − α) →d N 0, α/(αψ (α) − 1) n(α and
√
/ n − λ) → N 0, λ2 ψ (α)/(αψ (α) − 1) . n(λ d √ / It is interesting to compare the limiting distribution of n(λ n − λ) (assuming α is unknown) to the limiting distribution of the MLE c 2000 by Chapman & Hall/CRC
Table 5.1 Comparison of α−1 and g(α) = ψ (α)/(αψ (α) − 1) in Example 5.15 for various α.
α
α−1
g(α)
0.1 0.5 1 2 5 10 100
10 2 1 0.5 0.2 0.1 0.01
11.09 3.36 2.55 2.22 2.08 2.04 2.00
of λ when α is known. In this latter case, the MLE is 0n = α λ ¯n X
and we have
√
0 n − λ) → N (0, λ2 /α). n(λ d
0 n to be the more eﬃcient estimator of λ since We should expect λ we are able to incorporate our knowledge of α into the estimation of λ. In fact, it can be shown that α−1 < ψ (α)/(αψ (α)−1) for any α > 0; Table 5.1 compares α−1 and g(α) = ψ (α)/(αψ (α) − 1) for several values of α. Note that as α becomes larger, the diﬀerence in eﬃciency is more substantial. In practice, of course, one would /n. ✸ rarely “know” α and so we would have no choice but to use λ
Example 5.15 illustrates a more general point. Given two equally valid statistical models, we can obtain more eﬃcient estimators in the model with fewer parameters (or more precisely, we will not do worse with the lower parameter model). However, the penalties we pay in assuming a “too small” model can be quite severe. In practice, model selection is often the most diﬃcult and important step in the analysis of data. Some of the consequences of misspecifying a model are examined in the next section. 5.5 Misspeciﬁed models It is important to remember that statistical models are typically merely approximations to reality and so the wrong model is, more c 2000 by Chapman & Hall/CRC
often than not, ﬁt to the observed data. As troubling as this observation may seem, it may not be a problem from a practical point of view. First, the assumed model may be “close enough” to the true model so that very little is lost by assuming the wrong model. Second, the parameters estimated for a given model can often be interpreted usefully even if the assumed model is wrong. The following two examples illustrate these points. EXAMPLE 5.16: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. However, suppose that we decide that the appropriate model for the data is given by the following twoparameter family of densities for the Xi ’s:
x −(α+1) α 1+ for x > 0. f (x; α, θ) = θ θ α > 0 and θ > 0 are the unknown parameters. Even though the Exponential distribution is not a member of this family of distributions, it is easy to see that by letting α and θ tend to inﬁnity such that α/θ tends to λ, we have (for x > 0), f (x; α, θ) → λ exp(−λx)
(as α, θ → ∞ and α/θ → λ).
/ the estimated density / and θ, Thus given reasonable estimators α / / θ) will be close to true density of the Xi ’s. f (x; α, ✸
EXAMPLE 5.17: Suppose that X1 , · · · , Xn are independent random variables with E(Xi ) = α + βti
and Var(Xi ) = σ 2
where α, β, and σ 2 are unknown, and t1 , · · · , tn are known constants. In estimating α, β and σ 2 , it is often assumed that the Xi ’s are Normal and the parameters estimated by maximum likelihood; we will see in Chapter 8 that these estimators remain valid even when the Xi ’s are nonnormal. ✸ Suppose that X1 , X2 , · · · , Xn are i.i.d. random variables with distribution function F . We assume, however, that the Xi ’s have a common density or frequency function f (x; θ) for some θ ∈ Θ where the true distribution function F does not necessarily correspond to any f (x; θ). Suppose that an estimator θ/n satisﬁes the likelihood equation n i=1 c 2000 by Chapman & Hall/CRC
: (Xi ; θ/n ) = 0
where : (x; θ) = : (x; θ). (In general, this type of relation that deﬁnes an estimator is called an estimating equation and includes the likelihood equation as a special case.) What exactly is θ/n estimating? What is the behaviour of the sequence of estimators {θ/n } for large n? Consider the functional parameter θ(F ) deﬁned by ∞
−∞
: (x; θ(F )) dF (x) = 0.
The substitution principle estimator of θ(F ) is simply the solution to the likelihood equation given above. The inﬂuence curve of θ(F ) is : (x; θ(F )) φ(x; F ) = − . ∞ , −∞ : (x; θ(F )) which suggests that √ where
n(θ/n − θ(F )) →d N (0, σ 2 ) .∞
σ = .−∞ ∞ 2
[: (x; θ(F ))]2 dF (x)
2 .
−∞ : (x; θ(F )) dF (x)
The following theorem gives precise conditions under which the preceding statement is true; these conditions parallel the diﬀerentiability conditions (A4) through (A6). THEOREM 5.5 Suppose that X1 , X2 , · · · , Xn are i.i.d. random variables with distribution function F and that the estimator θ/n satisﬁes the estimating equation n
: (Xi ; θ/n ) = 0
i=1
for some θ/n in an open set Θ. If (a) : (x; θ) is a strictly decreasing (or strictly decreasing) function of θ .(over the open set Θ) for each x, ∞ (b) −∞ : (x; θ) dF (x) = 0 has a unique solution θ = θ(F ) where θ(F ) ∈ Θ, . ∞ (c) I(F ) = −∞ [: (x; θ(F ))]2 dF (x) < ∞, .∞ (d) J(F ) = − −∞ : (x; θ(F )) dF (x) < ∞, (e) :.(x; t) ≤ M (x) for θ(F ) − δ ≤ t ≤ θ(F ) + δ and some δ > 0 ∞ where −∞ M (x) dF (x) < ∞, √ / then θn →p θ(F ) and n(θ/n − θ(F )) →d N (0, I(F )/J 2 (F )). c 2000 by Chapman & Hall/CRC
Proof. Since : (x; θ) is strictly decreasing in θ and θ(F ) is the unique solution of the equation ∞
−∞
: (x; θ) dF (x) = 0,
we have for any > 0, ,

,

n 1 P : (Xi ; θ(F ) + ) > 0 → 0 n i=1
and
n 1 : (Xi ; θ(F ) − ) < 0 → 0 P n i=1
as n → ∞. From this it follows that
P θ/n − θ(F ) > → 0 and so θ/n →p θ(F ). Now expanding the estimating equation in a Taylor series, we have 0 =
n n √ 1 1 √ : (Xi ; θ(F )) + n(θ/n − θ(F )) : (Xi ; θ(F )) n i=1 n i=1 n √ 1 : (Xi ; θn∗ ) + n(θ/n − θ(F ))2 2n i=1
where θn∗ lies between θ/n and θ(F ). The remainder of the proof is identical to the proof of Theorem 5.3. We can remove the assumption that : (x; θ) is strictly monotone in θ for each x by adding the assumption that θ/n →p θ(F ). In certain cases, we will not have
: (x; θ) dF (x) = 0
for any θ ∈ Θ but may instead have
lim
θ→a
: (x; θ) dF (x) = 0
for some a lying at the boundary of the set Θ (but not in Θ); for example, a can be ±∞. In this case, it is usually possible to show that the sequence of estimators {θ/n } converges in probability to a. It is also possible to extend the result to the multiparameter case. c 2000 by Chapman & Hall/CRC
Suppose X1 , · · · , Xn are i.i.d. random variables with distribution F / n satisfy and let θ n
/n) = 0 : (Xi ; θ
i=1
where : (x; θ) is the vector of partial derivatives (with respect to the components of θ) of ln f (x; θ). Then under appropriate regularity conditions, we have √ / n(θ n − θ(F )) →d Np (0, J(F )−1 I(F )J(F )−1 ) where θ(F ) = (θ1 (F ), · · · , θp (F )) satisﬁes ∞
−∞
: (x; θ(F )) dF (x) = 0
and the matrices I(F ), J(F ) are deﬁned by ∞
I(F ) =
[: (x; θ(F ))][: (x; θ)]T dF (x)
−∞ ∞
J(F ) = −
−∞
: (x; θ(F )) dF (x).
(As before, : (x; θ) is the matrix of second partial derivatives of :(x; θ) with respect to θ1 , · · · , θp ; the integrals deﬁning I(F ) and J(F ) are deﬁned element by element.) EXAMPLE 5.18: Suppose the X1 , X2 , · · · , Xn are i.i.d. Normal random variables with (true) mean θ0 and variance σ 2 . However, we (erroneously) assume that the density of the Xi ’s is f (x; θ) =
exp(x − θ) [1 + exp(x − θ)]2
for some θ. The MLE of θ for this model based on X1 , · · · , Xn is the solution to the equation n exp(Xi − θ/n ) − 1 = 0. exp(Xi − θ/n ) + 1 i=1
It is possible to show that θ/n →p θ0 since ∞
exp(x − θ0 ) − 1
1 (x − θ0 )2 √ exp − 2σ 2 σ 2π
dx = 0. exp(x − θ0 ) + 1 √ We also have that n(θ/n − θ0 ) →d N (0, γ 2 (σ 2 )). The values of −∞
c 2000 by Chapman & Hall/CRC
Table 5.2 Asymptotic variances of
√
n(θ/n − θ) in Example 5.18.
σ
0.5
1
2
5
10
20
50
γ 2 (σ 2 )/σ 2
1.00
1.02
1.08
1.24
1.37
1.46
1.53
γ 2 (σ 2 )/σ 2 (as a function of σ) are given in Table 5.2. (The quantity γ 2 /σ 2 is the asymptotic relative eﬃciency of the sample mean to ✸ the estimator θ/n .) EXAMPLE 5.19: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on the interval [0, b]. We erroneously assume that the Xi ’s have the Gamma density f (x; α, λ) =
λα α−1 exp(−λx) for x > 0. x Γ(α)
/ n satisfy the equations / n and λ The estimators α /n) + n ln(λ
n
/n) = 0 ln(Xi ) − nψ(α
i=1
n
/n α /n λ
−
n
Xi = 0
i=1
/ n →p α(b) where ψ(α) is the derivative of ln Γ(α). It follows that α / and λn →p λ(b) where α(b) and λ(b) satisfy
ln(λ(b)) +
1 b
b 0
ln(x) dx − ψ(α(b)) = 0 α(b) 1 − λ(b) b
b
x dx = 0. 0
From the second equation above, it is easy to see that α(b) = bλ(b)/2 and so λ(b) satisﬁes ln(λ(b)) + ln(b) − 1 − ψ (bλ(b)/2) = 0. Numerically solving this equation, we get λ(b) =
3.55585 b
and so α(b) = bλ(b)/2 = 1.77793. Now using α(b) and λ(b), we can c 2000 by Chapman & Hall/CRC
compute the matrices I(F ) and J(F ):
I(F ) =
J(F ) =
1 −b/4 −b/4 b2 /12
0.74872 −b/3.55585 −b/3.55585 b2 /7.11171
.
From this it follows that √ / n − α(b) α n / →d N2 (0, J(F )−1 I(F )J(F )−1 ) λn − λ(b) with J(F )−1 I(F )J(F )−1 =
9.61 16.03/b 16.03/b 29.92/b2
.
✸
EXAMPLE 5.20: Suppose that X1 , · · · , Xn are i.i.d. Gamma random variables with shape parameter α and scale parameter λ that represent incomes sampled from some population. Using X1 , · · · , Xn , we wish to estimate the Gini index (see Examples 4.21 and 4.31) of the distribution; for a Gamma distribution, this depends only on α and is given by the expression Γ(α + 1/2) . π −1/2 √ πΓ(α + 1) However, we erroneously assume that the observations are logNormal with density
1 (ln(x) − µ)2 f (x; µ, σ) = √ exp − 2σ 2 2πσx
for x > 0
(ln(Xi ) will have a Normal distribution with mean µ and variance σ 2 ). The Gini index for√the logNormal depends only on σ and is given by γ(σ) = 2Φ(σ/ 2) − 1 where Φ(x) is the standard Normal distribution function. The MLE of σ is
/n = σ
n 1 (Yi − Y¯n )2 n i=1
1/2
where Yi = ln(Xi )
and the corresponding estimator of the Gini index (assuming the /n ). By the WLLN, logNormal model) is γ(σ n 1 (Yi − Y¯n )2 →p Var(ln(Xi )) = ψ (α) n i=1 c 2000 by Chapman & Hall/CRC
Table 5.3 Comparison of true Gini index (from a Gamma model) with its misspeciﬁed version based on a logNormal model in Example 5.20.
α
True
Misspeciﬁed
0.1 0.2 0.5 1 2 5 10 20 50
0.883 0.798 0.637 0.500 0.375 0.246 0.176 0.125 0.080
1.000 1.000 0.884 0.636 0.430 0.261 0.181 0.127 0.080
where ψ is the derivative of the function ψ in Example 5.19 (that is, the second derivative of the logarithm of the Gamma function). Thus applying the Continuous Mapping Theorem, it follows that
/n ) →p 2Φ (ψ (α)/2)1/2 − 1. γ(σ
The limit can now be compared to the true value of the Gini index in order to assess the asymptotic bias involved in assuming the wrong model; these values are given in Table 5.3 for various α. /n ) is quite badly biased for Table 5.3 shows that the estimator γ(σ small α but also that the bias gradually disappears as α increases. Note that we could also use the substitution principle estimator discussed in Example 4.31; this estimator does not require us to know the form of the distribution of the Xi ’s but could be extremely ineﬃcient compared to an estimator that assumes the data comes from a particular parametric model. ✸ The result of Theorem 5.5 suggests that an estimator of the standard error of θ/n (when a single parameter is estimated) is given by
/ θ/n ) = se(
−
n
−1 n 1/2 2 / / : (Xi ; θn ) [: (Xi ; θn )] .
i=1
i=1
/n In the multiparameter case, the variancecovariance matrix of θ c 2000 by Chapman & Hall/CRC
can be estimated by / n ) = J( / n )−1 I( / n )J( / n )−1 2 θ /θ /θ /θ Cov(
where /n) = − 1 /θ J(
n
/n) = /θ I(
n
/n) : (Xi ; θ
i=1
n 1 / n )][: (Xi ; θ / n )]T . [: (Xi ; θ n i=1
This estimator has come to be known as the sandwich estimator. 5.6 Nonparametric maximum likelihood estimation In discussing maximum likelihood estimation, we have assumed that the observations have a joint density or frequency function (depending on a real or vectorvalued parameter), from which we obtain the likelihood function. This formulation eﬀectively rules out maximum likelihood estimation for nonparametric models as for these models we typically do not make suﬃciently strong assumptions to deﬁne a likelihood function in the usual sense. However, it is possible to deﬁne a notion of nonparametric maximum likelihood estimation although its formulation is somewhat tenuous. Suppose that X1 , · · · , Xn are i.i.d. random variables with unknown distribution function F ; we want to deﬁne a (nonparametric) MLE of F . In order to make the estimation problem welldeﬁned, we will consider only distributions putting positive probability mass only at the points X1 , · · · , Xn . For simplicity, we will assume here that the Xi ’s are distinct (as would be the case if the Xi ’s were sampled from a continuous distribution). If pi is the probability mass at Xi then the nonparametric loglikelihood function is n ln L(p1 , · · · , pn ) =
ln(pi )
i=1
where pi ≥ 0 (i = 1, · · · , n) and p1 + · · · + pn = 1. Maximizing the nonparametric loglikelihood, we obtain p/i = 1/n (i = 1, · · · , n); thus the nonparametric MLE of F is the empirical distribution function F/ with n 1 F/ (x) = I(Xi ≤ x). n i=1 c 2000 by Chapman & Hall/CRC
See Scholz (1980) and Vardi (1985) for more discussion of nonparametric maximum likelihood estimation. Given our nonparametric MLE of F , we can determine the nonparametric MLE of an arbitrary functional parameter θ(F ) to be θ/ = θ(F/ ), provided that this latter estimator is welldeﬁned. Thus the nonparametric MLE of θ(F ) is simply the substitution principle estimator as described in section 4.5. More discussion of nonparametric maximum likelihood estimation will be given in section 7.5 where we discuss using the nonparametric likelihood function to ﬁnd conﬁdence intervals for certain functional parameters θ(F ). 5.7 Numerical computation of MLEs In many estimation problems, it is diﬃcult to obtain closedform analytical expressions for maximum likelihood estimates. In these situations, it is usually necessary to calculate maximum likelihood estimates numerically. Many numerical methods are available for calculating maximum likelihood (or other) estimates; a good survey of such methods is available in the monograph by Thisted (1988). We will describe just two methods here: the NewtonRaphson and EM algorithms. The NewtonRaphson algorithm The NewtonRaphson algorithm is a general purpose algorithm for ﬁnding the solution of a nonlinear equation; it can also be generalized to ﬁnding the solution of a system of nonlinear equations. The NewtonRaphson algorithm is natural in the context of computing maximum likelihood estimates as these estimates are often the solution of a system of equations (the likelihood equations). Suppose that we want to ﬁnd the solution to the equation g(x0 ) = 0 where g is a diﬀerentiable function. Given a number x that is close to x0 , it follows from a Taylor series expansion around x that 0 = g(x0 ) ≈ g(x) + g (x)(x0 − x) and solving for x0 , we get x0 ≈ x − c 2000 by Chapman & Hall/CRC
g(x) . g (x)
Thus given an estimate xk , we can obtain a new estimate xk+1 by xk+1 = xk −
g(xk ) g (xk )
and this procedure can be iterated for k = 1, 2, 3, · · · until g(xk ) (or g(xk )/g (xk )) is suﬃciently small. The NewtonRaphson algorithm can be applied to the computation of maximum likelihood estimates (as well as other estimates deﬁned as the solution to a single equation or system of equations). Suppose that the joint density or joint frequency function of X = (X1 , · · · , Xn ) is f (x; θ) and let L(θ) = f (x; θ) be the likelihood function based on X = x. Suppose that the maximum like/ = 0 where S(θ) is the derivative of lihood estimate θ/ satisﬁes S(θ) the loglikelihood function, ln L(θ). (S(θ) is often called the score function.) Let θ/(k) be the estimate of θ after k iterations of the algorithm; then S(θ/(k) ) θ/(k+1) = θ/(k) + H(θ/(k) ) where
∂2 ln L(θ). ∂θ2 The procedure is then iterated until convergence (that is, when S(θ/(k) ) or the absolute diﬀerence between θ/(k) and θ/(k+1) is suﬃciently small). In order to use the NewtonRaphson algorithm, one needs an initial estimate of θ, θ/(0) . In fact, in some cases, this initial estimate is critical as the algorithm will not always converge for a given θ/( 0). It is also possible that there may be several solutions to the / = 0, each solution corresponding to either a local equation S(θ) maximum, local minimum, or “saddlepoint” of the loglikelihood function; thus it is possible that the sequence of estimates {θ/(k) } / = 0. (The convergence will converge to the “wrong” solution of S(θ) question is much more important when estimating three or more parameters; with one or two unknown parameters, it is possible to plot the loglikelihood function and determine appropriate initial estimates from this plot.) If it is not clear that the algorithm will converge to the maximizer of the likelihood function then several diﬀerent initial estimates can be tried. Alternatively, one can use another estimate (for example, a method of moments estimate) as H(θ) = −
c 2000 by Chapman & Hall/CRC
an initial estimate for the algorithm. In fact, if θ/(0) is a suﬃciently good estimator of θ then the “onestep” estimator θ/(1) = θ/(0) +
S(θ/(0) ) H(θ/(0) )
has virtually the same large sample properties as the MLE. More √ (0) (0) precisely, if θ/n is based on n observations and n(θ/n − θ) √ (1) converges in distribution then typically n(θ/n − θ/n ) →p 0 where (1) (0) θ/n is the MLE and θ/n is the “onestep” estimator using θ/n as a (0) starting value. Thus we could take θ/n to be a method of moments estimator (or some other substitution principle estimator) of θ. EXAMPLE 5.21: Suppose that X1 , · · · , Xn are i.i.d. Cauchy random variables with density function f (x; θ) =
1 . π(1 + (x − θ)2
Given outcomes x1 , · · · , xn , the loglikelihood function is ln L(θ) = −
n
ln[1 + (xi − θ)2 ] − n ln(π).
i=1
The maximum likelihood estimate θ/ satisﬁes the equation / = S(θ)
n i=1
/ 2(xi − θ)
/2 1 + (xi − θ)
= 0.
Note that S(θ) is not monotone in θ and hence the equation / = 0 may have more than one solution for a given sample S(θ) x1 , · · · , xn . To illustrate the NewtonRaphson algorithm, we take a sample of 100 observations with θ = 10. Plots of the loglikelihood ln L(θ) and the score function S(θ) are given in Figures 5.1 and 5.2. To ﬁnd θ/(0) , we must ﬁrst ﬁnd a reasonable initial estimate of θ. Since the density of the Xi ’s is symmetric around θ, it makes sense to consider either the sample mean or sample median as an initial estimate; however, E(Xi ) is not welldeﬁned so that the sample mean need not be a good estimate of θ. Thus we will use the sample median as an initial estimate. Successive values of θ/(k) are deﬁned by S(θ/(k) ) θ/(k+1) = θ/(k) + H(θ/(k) ) c 2000 by Chapman & Hall/CRC
300 400 500
Loglikelihood
5
1 0
15
θ
Figure 5.1 Loglikelihood function for the Cauchy data in Example 5.21.
Table 5.4 Iterates of the NewtonRaphson algorithm for Cauchy data in Example 5.21.
k
θ/(k)
ln L(θ/(k) )
0 1 2 3
10.04490 10.06934 10.06947 10.06947
239.6569 239.6433 239.6433 239.6433
where H(θ) = 2
n 1 − (xi − θ)2 i=1
(1 + (xi − θ)2 )2
.
The values of θ/(k) and ln L(θ/(k) ) are given in Table 5.4. The choice of θ/(0) is crucial here; for example, if θ/(0) is taken to be less than 8.74 or greater than 11.86 then the sequence {θ/(k) } will not converge (at c 2000 by Chapman & Hall/CRC
40 20 0 40
20
Score
5
1 0
15
θ
Figure 5.2 Score function for the Cauchy data in Example 5.21.
✸
least to the maximum likelihood estimate).
The NewtonRaphson algorithm can also be extended to the multiparameter case. Let θ = (θ1 , · · · , θp )T and suppose that the / =0 maximum likelihood estimate of θ is given by the equation S(θ) where
S(θ) =
T
∂ ∂ ln L(θ), · · · , ln L(θ) ∂θ1 ∂θp
.
/ (k) , we deﬁne θ / (k+1) by Then given θ / θ
(k+1)
/ =θ
(k)
)
/ (k) + H θ
*−1
/ (k) S θ
where H(θ) is the matrix of negative second partial derivatives of ln L(θ); the (i, j) element of H is given by Hij (θ) = − c 2000 by Chapman & Hall/CRC
∂2 ln L(θ). ∂θi ∂θj
/ is sometimes called the observed Fisher information matrix.) (H(θ) Estimates of standard error for the parameter estimates can be obtained from the NewtonRaphson algorithm. In the single / −1 parameter case, the variance of θ/ can be estimated by [H(θ)] while in the multiparameter case, the variancecovariance matrix / −1 . These variance of θ/ can be estimated by the matrix [H(θ)] estimates assume that the model is correctly speciﬁed (that is, the data belong to the family of distributions being estimated). In the case where the model is incorrectly speciﬁed, these variance estimates can be modiﬁed to make them valid as in Section 4.5.
The Fisher scoring algorithm A simple modiﬁcation of the NewtonRaphson algorithm is the Fisher scoring algorithm. This algorithm replaces H by H ∗ where the (i, j) element of H ∗ (θ) is ,
Hij∗ (θ)
= Eθ [Hij (θ)] = −Eθ

∂2 ln f (X; θ) ; ∂θi ∂θj
the expected value above is computed assuming that θ is the true value of the parameter. (H is the observed Fisher information matrix while H ∗ is the expected Fisher information matrix.) Now / (k) is the estimate of θ after k iterations, we deﬁne θ / (k+1) by if θ )
*−1
(k+1) (k) (k) (k) / / / / θ =θ + H θ S θ .
The important distinction between the NewtonRaphson and Fisher scoring algorithms is the fact that H ∗ (θ) depends on the observed value of X, x, only through the value of θ while H(θ) depends, in general, on both θ and x. EXAMPLE 5.22: As in Example 5.21, let X1 , · · · , Xn be i.i.d. Cauchy random variables. From before, we have H(θ) = 2
n 1 − (xi − θ)2 i=1
and so H ∗ (θ) =
n π
∞ −∞
c 2000 by Chapman & Hall/CRC
(1 + (xi − θ)2 )2
1 − (x − θ)2 n dx = . (1 + (x − θ)2 )3 2
Table 5.5 Iterates of the Fisher scoring algorithm for the Cauchy data in Example 5.21.
k
θ/(k)
ln L(θ/(k) )
0 1 2 3 4
10.04490 10.06710 10.06923 10.06945 10.06947
239.6569 239.6434 239.6433 239.6433 239.6433
Hence the Fisher scoring algorithm is θ/(k+1) = θ/(k) +
n xi − θ/(k) 4 . n i=1 1 + (xi − θ/(k) )2
Values for θ/(k) and ln L(θ/(k) ) are given in Table 5.5; as before, θ/(0) is taken to be the sample median. One advantage of the Fisher scoring algorithm in this particular example is the fact that θ/(k) converges for a much wider interval of starting values θ/(0) . ✸ The diﬀerences between the NewtonRaphson and Fisher scoring algorithms are subtle but still important. Although it is diﬃcult to generalize too much, we can make the following observations: • The convergence of the NewtonRaphson algorithm is often faster when both algorithms converge. • The radius of convergence for the Fisher scoring algorithm is often larger; this suggests that the choice of an initial estimate is less important for the Fisher scoring algorithm. In the case of exponential family models, the NewtonRaphson and Fisher scoring algorithms are almost equivalent; if (θ1 , · · · , θp ) are the natural parameters then H(θ) = H ∗ (θ) and so the two algorithms are identical. The EM algorithm The EM algorithm provides a useful framework for computing maximum likelihood estimates in socalled incomplete data situations, for example, when data are missing or are not observed exactly. It was formalized in a paper by Dempster, Laird and Rubin (1977) c 2000 by Chapman & Hall/CRC
although special cases of the algorithm had been used in speciﬁc problems for many years before 1977. Today, the EM algorithm is widely used in a variety of problems. The following example illustrates an incomplete data problem where a continuous random variable is not observed exactly but rather in a given interval. EXAMPLE 5.23: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. Given outcomes x1 , · · · , xn of X1 , · · · , Xn , the maximum likelihood estimate of θ is 1/¯ x, where x ¯ is the average of the xi ’s. However, suppose that rather than observing Xi exactly, we observe lower and upper bounds, ui and vi , for Xi so that ui < Xi < vi . Given (u1 , v1 ), · · · , (un , vn ), the likelihood function of λ is L(λ) = =
n i=1 n
Pλ (ui ≤ Xi ≤ vi ) (exp(−λui ) − exp(−λvi )) .
i=1
This is simple example of interval censored data.
✸
The main idea behind the EM algorithm is to construct an estimate of the complete data likelihood, which can then be maximized using traditional numerical methods (such as the NewtonRaphson algorithm). Thus the EM algorithm is not really a numerical algorithm but instead a general purpose procedure for computing parameter estimates from incomplete data by iteratively computing parameter maximum likelihood estimates based on the complete data likelihood function; implicit in the use of the EM algorithm is the assumption that complete data maximum likelihood estimates can be readily computed and the incomplete data likelihood function is diﬃcult to work with. The EM algorithm iterates 2 steps (called the E and M steps) to convergence. Before discussing the algorithm in any depth, we will illustrate it using the example given above. EXAMPLE 5.24: In Example 5.23, the loglikelihood function of λ based on the exact outcomes x1 , · · · , xn is ln Lc (λ) = n ln(λ) − λ
n i=1
c 2000 by Chapman & Hall/CRC
xi .
The ﬁrst step of the EM algorithm involves ﬁnding an estimate of ln Lc (λ) given the incomplete data (u1 , v1 ), · · · , (un , vn ) and / (k) where λ / (k) is the assuming that the true parameter value is λ estimate of λ after k iterations of the algorithm; this is done by ﬁnding the expected value of n ln(λ) − λ
n
Xi
i=1
given u1 < X1 < v1 , · · · , un < Xn < vn and assuming that the / (k) . By simple integration, it follows that value of the parameter is λ E(Xi ui < Xi < vi ; λ) =
1 ui exp(−λui ) − vi exp(−λvi ) + λ exp(−λui ) − exp(−λvi )
and so to complete the E step of the algorithm, we substitute (k)
/i x
=
1 / (k) λ
+
/ (k) ui ) − vi exp(−λ / (k) vi ) ui exp(−λ / (k) ui ) − exp(−λ / (k) vi ) exp(−λ
for xi in ln Lc . The M step of the algorithm involves maximizing this new loglikelihood; it is maximized at
n 1 (k) / x n i=1 i
/ (k+1) = λ
−1
.
The E and M steps are then iterated until convergence.
✸
We can now discuss the EM algorithm in more or less complete generality. Suppose that X = (X1 , · · · , Xn ) are continuous random variables with joint density function fX (x; θ) (where θ can be realor vectorvalued) and let Y = (Y1 , · · · , Ym ) be random variables such that Yi = gi (X) where g1 , · · · , gm are known functions. The mapping that produces the Yi ’s from the Xi ’s is typically not onetoone; this means that any given outcome of the Yi ’s, y1 , · · · , ym can be produced by more than one outcome of the Xi ’s. We thus think of the Xi ’s as being the complete data and the Yi ’s as being the incomplete data. The joint density of Y is
fY (y; θ) = where
A(y)
fX (x; θ) dx
A(y) = {x : yi = gi (x) for i = 1, · · · m}.
c 2000 by Chapman & Hall/CRC
If X1 , · · · , Xn are discrete, the integral sign is replaced by a summation over A(y). If we observed X = x, we could estimate θ by maximizing the complete data likelihood function Lc (θ) = fX (x; θ). Instead we observe Y = y. In general, the joint density function of Y (as given above) will be quite diﬃcult to compute directly; hence, it will be diﬃcult to compute maximum likelihood estimates from incomplete data using standard algorithms. However, it is often possible to impute (or “estimate”) the values of the complete data using the observed (incomplete) data. Given a preliminary estimate of the parameter, the EM algorithm ﬁrst constructs an estimate of the complete data likelihood function and then maximizes this likelihood to obtain a new parameter estimate; this two step procedure is then iterated until convergence. • (E step) The E (for expectation) step of the EM algorithm involves ﬁnding an estimate of the likelihood function of θ for the complete data given the observed (or incomplete) data. Given an estimate of θ, θ/(k) , after k iterations of the algorithm, we deﬁne ln L(k) (θ) = E[ln fX (X; θ)Y = y; θ/(k) ] where the expectation is taken assuming that the true parameter value is θ/(k) . • (M step) An updated estimate of θ is obtained in the M (maximization) step of the EM algorithm. The updated estimate, θ/(k+1) , is chosen to maximize the estimate of the complete data loglikelihood function, ln L(k) (θ), that was obtained in the E step. Computation of θ/(k+1) will often involve the use of some numerical method such as the NewtonRaphson algorithm; it is often useful to use θ/(k) as an initial estimate of θ/(k+1) in this case. Once θ/(k+1) is computed, we return to the E step. The E and M steps are repeated until convergence. In many situations (for example, exponential family models), fX (x; θ) is a product of two nonconstant factors, one of which does not depend on θ. We can then write ln fX (x; θ) = h1 (x; θ) + h2 (x). In such cases, we can redeﬁne ln L(k) (θ) in the E step by ln L(k) (θ) = E[h1 (X; θ)Y = y; θ/(k) ]. c 2000 by Chapman & Hall/CRC
This modiﬁcation of the algorithm can considerably simplify computations. The M step of the algorithm remains the same; as before, θ/(k+1) maximizes ln L(k) (θ). What can be said about the convergence properties of the EM algorithm? If L(θ) is the likelihood function based on y1 , · · · , ym , it can be shown that ln L(θ/(k+1) ) ≥ ln L(θ/(k) ) for any initial estimate θ/(0) . Unfortunately, this does not necessarily imply that the sequence {θ/(k) } converges to the maximum likelihood estimate (namely the maximizer of L(θ)). This suggests that the choice of initial estimate can be very important as it is for the NewtonRaphson algorithm. In practice, it is a good idea to use a variety of initial estimates. If θ/(k+1) is computed numerically in the M step of the algorithm, it is important to ensure that an appropriate initial estimate is chosen to guarantee that θ/(k+1) will maximize L(k) (θ). A rigorous treatment of the convergence properties of the EM algorithm is given by Wu (1983). EXAMPLE 5.25: Suppose that X1 , · · · , Xn are i.i.d. random variables whose distribution is a mixture of two Poisson distributions: fX (x; λ, θ) = θ
exp(−λ)λx exp(−µ)µx + (1 − θ) x! x!
for x = 1, 2, · · ·. The frequency function given above arises if we observe a Poisson random variable with mean λ with probability θ and a Poisson distribution with mean µ with probability 1 − θ. (Note that when µ = λ, the model becomes a i.i.d. Poisson model and the parametrization is not identiﬁable. The loglikelihood function for θ, λ, and µ is given by n
exp(−µ)µxi exp(−λ)λxi ln L(θ, λ, µ) = ln θ + (1 − θ) xi ! xi ! i=1
.
Given x1 , · · · , xn , it is possible to estimate the parameters using the NewtonRaphson algorithm; however, the implementation of the NewtonRaphson algorithm in this example is somewhat diﬃcult and so the EM algorithm is a natural alternative. To use the EM algorithm, we must ﬁnd a suitable “complete” data problem. Assume that (X1 , Y1 ), · · · , (Xn , Yn ) are i.i.d. pairs of c 2000 by Chapman & Hall/CRC
random variables with P (Yi = y) = θy (1 − θ)1−y for y = 0, 1 exp(−µ)µx P (Xi = xY = 0) = for x = 0, 1, 2, · · · x! exp(−λ)λx P (Xi = xY = 1) = for x = 0, 1, 2, · · ·. x! The complete data likelihood based on (x1 , y1 ), · · · , (xn , yn ) is then ln Lc (θ, λ, µ) =
n
yi [ln(θ) + xi ln(λ) − λ]
i=1
+
n
(1 − yi ) [ln(1 − θ) + xi ln(µ) − µ]
i=1
−
n
ln(xi !).
i=1
We must now ﬁnd the expected value of the complete data loglikelihood given the observations x1 , · · · , xn ; to do this, it is suﬃcient to ﬁnd the expected value of Yi given Xi = xi for any values of the parameters θ, λ, and µ; an easy computation using Bayes’ Theorem yields E(Yi Xi = x; θ, λ, µ) =
θ exp(−λ)λx . θ exp(−λ)λx + (1 − θ) exp(−µ)µx
/ (k) , and µ / (k+1) , /(k) , we obtain θ/(k+1) , λ Thus given estimates θ/(k) , λ (k+1) / and µ by maximizing
ln L(k) (θ, λ, µ) =
n (k) y/i [ln(θ) + xi ln(λ) − λ] i=1
+
n
(k)
(1 − y/i ) [ln(1 − θ) + xi ln(µ) − µ]
i=1
where
(k)
y/i
/ (k) , µ /(k) . = E Yi Xi = x; θ/(k) , λ
It is easy to see that θ/(k+1) = c 2000 by Chapman & Hall/CRC
n 1 (k) y/i n i=1
Table 5.6 Frequency distribution of goals in First Division matches.
Goals Frequency
0
1
2
3
4
5
6
7
252
344
180
104
28
11
2
3
/ (k+1) λ
n
=
/(k+1) = µ
(k) /i i=1 xi y
n (k) /i i=1 y
n (k) /i ) i=1 xi (1 − y .
n (k) /i ) i=1 (1 − y
To illustrate the convergence EM algorithm in this example, we consider the number of goals scored during a game by teams in the First Division of the English Football League during the 197879 season; at that time, there were 22 teams with each team playing 42 games for a total of 924 observations. The data are summarized in Table 5.6. The EM algorithm is very slow to converge in this example even when the initial estimates are taken to be close to the maximizing values. For example, starting from initial estimates / (0) = 1.23, and µ /(k) = 3.04, the EM algorithm θ/(0) = 0.95, λ described above takes about 300 iterations to converge (to three decimal places) to the maximum likelihood estimates θ/ = 0.954, / = 1.232, and µ / = 3.043. The NewtonRaphson algorithm can, λ of course, also be applied; its performance in this example is very erratic. For example, using the same starting values as above, the NewtonRaphson algorithm actually diverges. ✸ Comparing the NewtonRaphson and EM algorithms Frequently, it is feasible to use both the NewtonRaphson and EM algorithms to compute maximum likelihood estimates. This raises the question of which algorithm is best to use. There are two compelling theoretical reasons to prefer the NewtonRaphson algorithm: • standard error estimates are a byproduct of the NewtonRaphson algorithm and not of the EM algorithm; • the NewtonRaphson algorithm has a faster rate of convergence c 2000 by Chapman & Hall/CRC
in the sense that θ/(k) → θ/ more quickly when θ/(k) is in a / neighbourhood of θ. Nonetheless, there are also many reasons to prefer the EM algorithm, perhaps the most important being ease of implementation. For example, the EM algorithm often leads to a considerable savings in programming time that far outweigh the savings in computer time incurred by using the NewtonRaphson or other more sophisticated algorithm. Moreover, in many problems, the number of iterations to convergence for the EM algorithm can be comparable to the NewtonRaphson algorithm; much depends on how close the complete data problem is to the incomplete data problem. For example, in Example 5.24, there is virtually no diﬀerence in the convergence rate between the EM and NewtonRaphson algorithm. 5.8 Bayesian estimation To this point, we have regarded a parameter θ as a ﬁxed quantity whose value is unknown and used the data to estimate (or make inference about) its value. In essence, we treat the observable quantities (the data) as outcomes of random variables and the unobservable quantity (the parameter) as a constant. The Bayesian approach describes the uncertainty in θ with a probability distribution on the parameter space Θ. This approach works by specifying a distribution for θ prior to observing the data and considering f (x; θ) as a “conditional” density or frequency function given θ; the observed data is then used to update this distribution. The probabilities placed on diﬀerent subsets of Θ are generally not interpreted as “longrun frequencies” but instead are usually subjective probabilities that reﬂect “degrees of belief” about θ lying in the various subsets of Θ. Before we go into the speciﬁcs of the Bayesian approach, we will discuss some of its pros and cons. First, if we accept the notion of describing uncertainty in the parameter by means of a probability distribution, the Bayesian approach yields a uniﬁed procedure for solving practically any problem in statistical inference. Second, mathematical arguments can be used to show that the only coherent systems for describing uncertainty are those based on probability distributions. A common complaint with Bayesian inference is that it is not “objective” in the sense that estimates depend on the prior distribution that is speciﬁed by the scientist or investigator (perhaps in consultation with someone else). Therefore, c 2000 by Chapman & Hall/CRC
the possibility exists for the unscrupulous scientist to inﬂuence the results by choosing the appropriate prior. The counterargument to this is that most scientiﬁc inference is very subjective; for example, scientists constantly make subjective judgements about how to carry out experiments. Moreover, much of nonBayesian inference is also to some extent subjective. If a scientist or investigator is in a good position to specify a prior distribution then the Bayesian approach to estimation (and inference) can be extremely useful. However, in many cases, there is no obvious, realistic prior distribution available; in such cases, the usefulness of the Bayesian approach is questionable although some remedies do exist. We will now elaborate on the implementation of the Bayesian approach. Suppose that X = (X1 , · · · , Xn ) are random variables with joint density or frequency function f (x; θ) where θ is a realvalued parameter lying in Θ, which we will assume to an open set. Let Π(θ) be a distribution function deﬁned on Θ and suppose that π(θ) is the corresponding density function; π(θ) is called the prior density for θ and describes the uncertainty in the value of the parameter θ prior to observing X. We deﬁne the posterior density of θ given X = x as follows: π(θx) = .
f (x; θ)π(θ) . Θ f (x; t)π(t) dt
Note that the denominator of the righthandside above does not depend on θ and so π(θx) ∝ f (x; θ)π(θ) = L(θ)π(θ) where L(θ) is the likelihood function based on x. It is very easy to generalize posterior distributions to the cases where θ is vectorvalued or is discretevalued. If θ = (θ1 , · · · , θp ) is vectorvalued, the (joint) posterior density of θ is π(θx) = .
···
f (x; θ)π(θ) . Θ f (x; t)π(t) dt
where π(θ) is the joint prior density of θ. The marginal posterior density of any single parameter (or the joint posterior density of a collection of parameters) in θ can be obtained by integrating the joint posterior density over all the other parameters. If θ is discretevalued and π(θ) is the prior frequency function for θ then c 2000 by Chapman & Hall/CRC
the posterior frequency function is given by f (x; θ)π(θ) π(θx) = . t f (x; t)π(t) Again in both these cases, we have that π(θx) ∝ π(θ)L(θ) where L(θ) is the likelihood function based on x. The posterior distribution describes the uncertainty in the value of θ after observing the data, taking into account the prior distribution. This brings up an important distinction between Bayesian and Frequentist statistics. In the Bayesian approach, all of the information about the uncertainty about θ is contained in the posterior distribution. More importantly, inference for θ depends only on the observed data (through the likelihood function) as well as the prior distribution. In contrast, the Frequentist approach often measures uncertainty (for example, by the standard error of a point estimator) by averaging over all possible (but unobserved) samples. Since the posterior distribution depends on the data only through the likelihood function, it follows that all Bayesian procedures satisfy the likelihood principle. EXAMPLE 5.26: Suppose that X1 , · · · , Xn are i.i.d. Bernoulli random variables with unknown parameter θ and assume the following Beta prior density for θ π(θ) =
Γ(α + β) α−1 (1 − θ)β−1 θ Γ(α)Γ(β)
for 0 < θ < 1
where α and β are speciﬁed. Then given X1 = x1 , · · · , Xn = xn , the posterior density for θ is π(θx) = =
=
f (x1 , · · · , xn ; θ)π(θ) .1 0 f (x; t)π(t) dt θy+α−1 (1 − θ)n−y+β−1 y+α−1 (1 − t)n−y+β−1 dt 0 t (where y = x1 + · · · + xn ) Γ(n + α + β) θy+α−1 (1 − θ)n−y+β−1 . Γ(y + α)Γ(n − y + β) .1
Note that the posterior distribution is also a Beta distribution with data dependent parameters. ✸ c 2000 by Chapman & Hall/CRC
EXAMPLE 5.27: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ where the prior density for λ is also an Exponential distribution with known parameter α. To complicate matters somewhat, we will assume that only the r smallest of X1 , · · · , Xn are observed. Letting Yi = X(i) for i = 1, · · · , r, the joint density of Y1 , · · · , Yr is ,
f (y; λ) = λ exp −λ r
r

yi + (n − r)yr
.
i=1
Given Y1 = y1 , · · · , Yr = yr , the posterior density of λ is ,
−1
π(λy) = [K(y, α)]
αλ exp −λ r
r

yi + (n − r)yr + α
i=1
where ∞
K(y, α) = 0
=
,
αλ exp −λ r
r

yi + (n − r)yr + α
dλ
i=1
αr! . ( i=1 yi + (n − r)yr + α)r+1
r
It is easy to see that the posterior distribution is a Gamma distribution with shape parameter r + 1 and scale parameter
r y + (n − r)yr + α. ✸ i=1 i In the previous two examples, the prior distributions themselves depend on parameters (α and β in Example 5.26, α in Example 5.27); these parameters are often called hyperparameters. Orthodox Bayesian statisticians maintain that the values of hyperparameters must be speciﬁed independently of the observed data so the prior distribution reﬂects true a priori beliefs. However, if we are not willing a priori to specify the values of the hyperparameters, it is possible to use the socalled empirical Bayes approach in which hyperparameter values are estimated from the observed data and then substituted into the posterior distribution. More precisely, if π(θ; α) is a prior density for θ depending on a hyperparameter α (possibly vectorvalued), we can deﬁne the “marginal” joint density or frequency function
g(x; α) =
f (x; θ)π(θ; α) dθ. Θ
The hyperparameter α can then be estimated by maximum likelic 2000 by Chapman & Hall/CRC
hood or substitution principle estimation assuming that the joint density or frequency function of the data is g(x; α). Empirical Bayes methods are often very useful although despite their name, they are essentially nonBayesian since the prior distribution is estimated from the data. EXAMPLE 5.28: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ where the prior distribution for λ is given by π(λ; α, β) =
αβ β−1 exp(−αλ) λ Γ(β)
for λ > 0
with α and β the hyperparameters. Setting y = x1 + · · · + xn , the “marginal” joint frequency function is given by g(x; α, β) =
∞ exp(−nλ)λy
π(θ; α, β) dλ x1 ! · · · xn !
β
y α Γ(y + β) 1 . = n+α Γ(α)x1 ! · · · xn ! n + α 0
Estimates for α and β can be obtained by maximum likelihood using g(x; α, β) as the likelihood function. Alternatively, method of moments estimates can be used. For example, if we think of λ as a random variable (with density π(λ; α, β), we have that E(Xi λ) = λ and Var(Xi λ) = λ. Then α β Var(Xi ) = E[Var(Xi λ)] + Var[E(Xi λ)] α α = + 2 β β E(Xi ) = E[E(Xi λ)] =
and
using the fact that the mean and variance of the Gamma prior are / and β/ can be obtained by α/β and α/β 2 , respectively. Estimates α / β/ and setting the sample mean and variance of the xi ’s equal to α/ / β/ + α/ / β/2 respectively. α/ ✸ How do we go from the posterior density for θ to a point estimator for θ? A common practice is to take the mean of the posterior distribution (or the median) to be the Bayesian point estimator. Another commonly used estimator is the posterior mode, that is, the value of θ that maximizes π(θx). In some sense, Bayesian c 2000 by Chapman & Hall/CRC
point estimators are really just descriptive measures of the posterior distribution as it is the posterior distribution itself that is of primary importance. However, one can also study Bayesian point estimators from a Frequentist point of view; in fact, Bayesian point estimators have very attractive Frequentist properties and are often superior to estimators obtained by Frequentist principles. EXAMPLE 5.29: In Example 5.26, the posterior distribution for θ is π(θx) =
Γ(n + α + β) θy+α−1 (1 − θ)n−y+β−1 Γ(y + α)Γ(n − y + β)
where y = x1 + · · · + xn . The posterior mean is θ/ =
1
θ π(θx) dθ = 0
y+α n+α+β
while the posterior mode is θ0 =
y+α−1 . n+α+β−2
It is easy to see that when both y and n are large compared to α and β then both θ/ and θ0 are approximately y/n (which is the maximum likelihood estimate of θ). ✸ EXAMPLE 5.30: In Example 5.27, the posterior distribution for λ is a Gamma distribution with shape parameter r + 1 and scale
parameter ri=1 yi + (n − r)yr + α. The posterior mean is r+1 i=1 yi + (n − r)yr + α
θ/ = r
while the posterior mode is r . i=1 yi + (n − r)yr + α
θ0 = r
Note that the diﬀerence between the posterior mean and mode becomes smaller as n increases. ✸ Conjugate and ignorance priors Although prior distributions are essentially arbitrary, it is often convenient to choose the prior distribution so that the posterior distribution is easily derivable. The classical examples of such prior c 2000 by Chapman & Hall/CRC
distributions are conjugate families, which are typically parametric families of priors such that posterior distribution belongs the same parametric family. For example, suppose that the model for (X1 , · · · , Xn ) is a oneparameter exponential family so that f (x; θ) = exp [c(θ)T (x) − d(θ) + S(x)] and suppose that the prior density for θ is of the form π(θ) = K(α, β) exp [αc(θ) − βd(θ)] for some α and β (where the K(α, β) is chosen so that π(θ) is a density). Then it is easy to see that π(θx) ∝ π(θ)f (x; θ) ∝ exp [(T (x) + α)c(θ) − (β + 1)d(θ)] . From this, we can generally deduce that π(θx) = K(T (x) + α, β + 1) exp [(T (x) + α)c(θ) − (β + 1)d(θ)] , which has exactly the same form as π(θ). The prior distributions {π(θ)} indexed by the hyperparameters α and β is a conjugate family of priors for the oneparameter exponential family. EXAMPLE 5.31: Suppose that X1 , · · · , Xn are i.i.d. Geometric random variables with frequency function f (x; θ) = θ(1 − θ)x for x = 0, 1, 2, · · · where 0 < θ < 1. The joint frequency function of X1 , · · · , Xn is a oneparameter exponential family: ,
f (x; θ) = exp ln(1 − θ)
n

xi + n ln(θ) .
i=1
This suggests that a conjugate family of priors for θ is given by π(θ) = K(α, β) exp (α ln(1 − θ) + β ln(θ)) = K(α, β)(1 − θ)α θβ for 0 < θ < 1. For α > 0 and β > 0, π(θ) is simply a Beta density with Γ(α, β) K(α, β) = . Γ(α)Γ(β)) Note that the same family is conjugate in Example 5.26. ✸ EXAMPLE 5.32: Suppose that X1 , · · · , Xn are i.i.d. Exponential c 2000 by Chapman & Hall/CRC
random variables with parameter λ. Again we have a oneparameter exponential family whose joint density function is ,
f (x; θ) = exp −λ
n

xi + n ln(λ) .
i=1
This suggests that a conjugate family of prior densities for λ is given by π(λ) = K(α, β) exp (−αλ + β ln(λ)) = K(α, β)λβ exp (−αλ) for λ > 0. For α > 0 and β > −1, this is a Gamma density and K(α, β) = αβ+1 /Γ(β + 1). Note that the Exponential prior density given in Example 5.27 is a special case. ✸ Conjugate prior distributions are often used simply for convenience rather than for their ability to accurately describe a priori beliefs. Basically, assuming a conjugate prior for θ greatly facilitates evaluation of the integral
π(θ)f (x; θ) dθ. Θ
However, with the rapid increase in computational power over the past 20 years, the use of conjugate priors to simplify computation has become less important; numerical integration techniques can be used to evaluate posterior distributions with minimal diﬃculty. These numerical techniques are particularly important in multiparameter problems where useful conjugate families are not generally available. Monte Carlo integration (see Chapter 3) is often useful in this context; in recent years, techniques such as Gibbs sampling and related Markov chain Monte Carlo methods have been used eﬀectively in problems with large numbers of parameters. A good reference is the monograph by Gilks et al (1996). EXAMPLE 5.33: Suppose we want to evaluate the integral
K(x) =
π(θ)f (x; θ) dθ. Θ
Let g be a density function on Θ; then
K(x) = Θ
π(θ)f (x; θ) g(θ) dθ. g(θ)
To estimate K(x), let T1 , · · · , Tm be i.i.d. random variables with c 2000 by Chapman & Hall/CRC
density g; for m suﬃciently large, we will have (by the WLLN) 5 K(x) =
m π(Ti )f (x; Ti ) 1 ≈ K(x) m i=1 g(Ti )
5 and K(x) serves as estimate of K(x). This is simply the importance sampling estimate of the integral considered in section 3.6. ✸
As stated above, Bayesian inference is often perceived as not objective. While this perception is perhaps undeserved, it is nonetheless desirable to ﬁnd prior distributions that express an indiﬀerence among all values of the parameter space. In the case where the parameter space is discrete and ﬁnite, the solution is clear; if there are m elements in the parameter space, we can specify a prior distribution putting probability 1/m on each element. However, when the parameter space is not ﬁnite then some diﬃculties arise. As above, we will concentrate on the case where the parameter space Θ is a subset of the realline. Suppose ﬁrst of all that Θ is the interval (a, b). Then a natural prior distribution is a Uniform distribution with density π(θ) = (b − a)−1
for a < θ < b.
However, expression of indiﬀerence by Uniform distributions is not invariant under transformations: the prior density of g(θ) will not itself be Uniform if g(·) is a nonlinear function. If Θ is a inﬁnite interval (for example, the realline) then . “proper” Uniform prior densities do not exist in the sense that Θ k dθ = ∞ for any k > 0. Nonetheless, these “improper” prior densities will often yield valid posterior densities; the posterior density is π(θx) = .
f (x; θ) . Θ f (x; t) dt
Uniform improper priors share the same problem as Uniform proper priors, namely, the lack of invariance under nonlinear transformation. EXAMPLE 5.34: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance 1. We will assume a uniform prior on the realline for µ. It follows that the posterior for µ is , n 1 π(µx) = k(x) exp − (xi − µ)2 2 i=1 c 2000 by Chapman & Hall/CRC
where ∞
k(x) = 4
= Thus
,

n 1 exp − (xi − t)2 dt 2 −∞ i=1
,
−1

n n 1 (xi − x ¯)2 . exp 2π 2 i=1
4
)
*
n n π(θx) = exp − (µ − x ¯)2 2π 2 and so the posterior distribution is Normal with mean x ¯ and variance 1/n. ✸ One approach to deﬁning ignorance priors has been proposed by Jeﬀreys (1961). Let g be a monotone function on Θ and deﬁne ν = g(θ). The information for ν can be deﬁned in the usual way: )
*
∂ ln f (X; g −1 (ν)) ∂ν ) * ∂ 1 Var . ln f (X; θ) θ (g (θ))2 ∂θ
I(ν) = Varν =
Now choose the function g so that I(ν) is constant and put a uniform prior distribution on g(Θ); this implies that the prior distribution on Θ is )
π(θ) = kg (θ) ∝ Varθ
*1/2
∂ ln f (X; θ) ∂θ
= I 1/2 (θ)
where k is some positive constant. Note that this prior distribution could be improper. These prior distributions are called Jeﬀreys priors. EXAMPLE 5.35: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ. The information for λ is I(λ) = n/λ. It is easy to verify that the transformation g must satisfy λ(g (λ))2 = constant. Thus the Jeﬀreys prior for λ is k π(λ) = √ λ c 2000 by Chapman & Hall/CRC
for λ > 0.
.
Since 0∞ λ−1/2 dλ = ∞, the prior is improper and so k can be chosen arbitrarily. Using the Jeﬀreys prior, the posterior density of λ given x1 , · · · , xn is the Gamma density π(λx) =
ny+1/2 λy−1/2 exp(−nλ) Γ(y + 1/2)
where y = x1 + · · · + xn . This posterior density is always proper (even though the prior density is improper). ✸ 5.9 Problems and complements 5.1: Suppose that X1 , · · · , Xn are i.i.d. random variables with density
f (x; θ1 , θ2 ) =
a(θ1 , θ2 )h(x) for θ1 ≤ x ≤ θ2 0 otherwise
where h(x) > 0 is a known continuous function deﬁned on the real line. (a) Show that the MLEs of θ1 and θ2 are X(1) and X(n) respectively. (b) Let θ/1n and θ/2n be the MLEs of θ1 and θ2 and suppose that h(θ1 ) = λ1 > 0 and h(θ2 ) = λ2 > 0. Show that
n
θ/1n − θ1 θ2 − θ/2n
→d
Z1 Z2
where Z1 and Z2 are independent Exponential random variables with parameters λ1 a(θ1 , θ2 ) and λ2 a(θ1 , θ2 ) respectively. 5.2: Suppose that (X1 , Y1 ), · · · , (Xn , Yn ) are i.i.d. pairs of Normal random variables where Xi and Yi are independent N (µi , σ 2 ) random variables. (a) Find the MLEs of µ1 , · · · , µn and σ 2 . (b) Show that the MLE of σ 2 is not consistent. Does this result contradict the theory we have established regarding the consistency of MLEs? Why or why not? (c) Suppose we observe only Z1 , · · · , Zn where Zi = Xi − Yi . Find the MLE of σ 2 based on Z1 , · · · , Zn and show that it is consistent. 5.3: Suppose that X1 , · · · , Xn , Y1 , · · · Yn are independent Exponenc 2000 by Chapman & Hall/CRC
Table 5.7 Data for Problem 5.3.
xi
yi
xi
yi
xi
yi
xi
yi
0.7 11.3 2.1 30.7 4.6
3.8 4.6 2.1 5.6 10.3
20.2 0.3 0.9 0.7 2.3
2.8 1.9 1.4 0.4 0.9
1.1 1.9 0.5 0.8 1.2
2.8 3.2 8.5 14.5 14.4
15.2 0.2 0.7 0.4 2.3
8.8 7.6 1.3 2.2 4.0
tial random variables where the density of Xi is fi (x) = λi θ exp(−λi θx)
for x ≥ 0
and the density of Yi is gi (x) = λi exp(−λi x) for x ≥ 0 where λ1 , · · · , λn and θ are unknown parameters. (a) Show that the MLE of θ (based on X1 , · · · , Xn , Y1 , · · · Yn ) satisﬁes the equation n θ/
−2
n
Ri
i=1
/ i 1 + θR
=0
where Ri = Xi /Yi . (b) Show that the density of Ri is fR (x; θ) = θ (1 + θx)−2
for x ≥ 0.
and show that the MLE for θ based on R1 , · · · , Rn is the same as that given in part (a). (c) Let θ/n be the MLE in part (c). Find the limiting distribution √ of n(θ/n − θ). (d) Use the data for (Xi , Yi ), i = 1, · · · , 20 given in Table 5.7 to compute the maximum likelihood estimate of θ using either the NewtonRaphson or Fisher scoring algorithm. Find an appropriate starting value for the iterations and justify your choice. (e) Give an estimate of the standard error for the maximum likelihood estimate computed in part (c). c 2000 by Chapman & Hall/CRC
5.4: Suppose that X1 , · · · , Xn are i.i.d. nonnegative random variables whose hazard function is given by
λ(x) =
λ1 for x ≤ x0 λ2 for x > x0
where λ1 , λ2 are unknown parameters and x0 is a known constant. (a) Show that the density of Xi is
f (x; λ1 , λ2 ) =
λ1 exp(−λ1 x) for x ≤ x0 λ2 exp(−λ2 (x − x0 ) − λ1 x0 ) for x > x0
(b) Find the MLEs of λ1 and λ2 as well as their joint limiting distribution. 5.5: Suppose that X1 , · · · , Xn are i.i.d. discrete random variables with frequency function
f (x; θ) =
θ for x = −1 (1 − θ)2 θx for x = 0, 1, 2, · · ·
where 0 < θ < 1. (a) Show that the MLE of θ based on X1 , · · · , Xn is 2 θ/n =
n
n
= −1) +
2n + ni=1 Xi
i=1 I(Xi
i=1 Xi
and show that {θ/n } is consistent for θ. √ (b) Show that n(θ/n − θ) →d N (0, σ 2 (θ) and ﬁnd the value of σ 2 (θ). 5.6: Suppose that U1 , · · · , Un are i.i.d. Uniform random variables on [0, θ]. Suppose that only the smallest r values are actually observed, that is, the order statistics U(1) < U(2) < · · · < U(r) . (a) Find the MLE of θ based on U(1) , U(2) , · · · , U(r) . (b) If r = rn = n − k where k is ﬁxed, ﬁnd the limiting distribution of n(θ − θ/n ) as n → ∞ where θ/n is the MLE. 5.7: Suppose that X = (X1 , · · · , Xn ) has a kparameter exponential family distribution with joint density or frequency function f (x; θ) = exp
, k i=1
c 2000 by Chapman & Hall/CRC

ci (θ)Ti (x) − d(θ) + S(x)
where the parameter space Θ is an open subset of Rk and the function c = (c1 , · · · , ck ) is onetoone on Θ. (a) Suppose that Eθ [Ti (X)] = bi (θ) (i = 1, · · · , k). Show that / satisﬁes the equations the MLE θ / Ti (X) = bi (θ) (i = 1, · · · , k).
(b) Suppose that the Xi ’s are also i.i.d. so that Ti (X) can be / taken to be an average of i.i.d. random variables. √ / If θ n is the MLE, use the Delta Method to show that n(θ n − θ) has the limiting distribution given in Theorem 5.4. 5.8: Suppose that X1 , · · · , Xn are i.i.d. continuous random variables with density f (x; θ) where θ is realvalued. (a) We are often not able to observe the Xi ’s exactly rather only if they belong to some region Bk (k = 1, · · · , m); an example of this is interval censoring in survival analysis. Intuitively, we should be able to estimate θ more eﬃciently with the actual values of the Xi ’s; in this problem, we will show that this is true (at least) for MLEs. Assume that B1 , · · · , Bm are disjoint sets such that
Pθ Xi ∈
m
Bk
= 1.
k=1
Deﬁne i.i.d. discrete random variables Y1 , · · · , Yn where Yi = k if Xi ∈ Bk ; the frequency function of Yi is p(k; θ) = Pθ (Xi ∈ Bk ) =
f (x; θ) dx Bk
for k = 1, · · · , m.
Also deﬁne )
*
∂ IX (θ) = Varθ ln f (Xi ; θ) ∂θ ) * ∂ and IY (θ) = Varθ ln p(Yi ; θ) . ∂θ Assume the usual regularity conditions for f (x; θ), in particular, that f (x; θ) can be diﬀerentiated with respect to θ inside integral signs with impunity! Show that IX (θ) ≥ IY (θ) and indicate under what conditions there will be strict inequality. c 2000 by Chapman & Hall/CRC
(Hint: Note that (i) f (x; θ)/p(k; θ) is a density function on Bk , and (ii) for any random variable U , E(U 2 ) ≥ [E(U )]2 with strict inequality unless U is constant.) (b) Suppose that we observe Yi = g(Xi ) (i = 1, · · · , n) for some diﬀerentiable function g. Show that IX (θ) ≥ IY (θ) with equality if g is a monotone function. 5.9: Let X1 , · · · , Xn be i.i.d. Exponential random variables with parameter λ. Suppose that the Xi ’s are not observed exactly but rather we observe random variables Y1 , · · · , Yn where Yi = kδ
if kδ ≤ Xi < (k + 1)δ
for k = 0, 1, 2, · · · where δ > 0 is known. (a) Give the joint frequency function of Y = (Y1 , · · · , Yn ) and
n show that i=1 Yi is suﬃcient for λ. (b) Find the MLE of λ based on Y1 , · · · , Yn . / n be the MLE of λ in part (b). Show that (c) Let λ √ / n(λn − λ) →d N (0, σ 2 (λ, δ))
where σ 2 (λ, δ) → λ2 as δ → 0. 5.10: Let X1 , · · · , Xn be i.i.d. random variables with density or frequency function f (x; θ) satisfying conditions (A1)(A6) with √ I(θ) = J(θ). Suppose that θ0n is such that n(θ0n − θ) →d Z for 0 some √ random variable Z (not necessarily Normal); θn is said to be nconsistent. Deﬁne θ/n = θ0n −
n i=1
−1 n 0 : (Xi ; θn ) : (Xi ; θ0n )
i=1
√
(a) Show that n(θ/n − θ) →d N (0, 1/I(θ)). (b) Suppose that X1 , · · · , Xn are i.i.d. random variables with density function θ f (x; θ) = (1 + θx)2 /n to be the sample median of X1 , · · · , Xn . for x ≥ 0 and deﬁne µ √ /n to construct a nconsistent estimator of θ, θ0n . What Use µ √ is the asymptotic distribution of n(θ0n − θ)?
(c) Using θ0n from part (b), show how to construct an estimator with the same limiting distribution as the MLE in Problem 5.3(b).
c 2000 by Chapman & Hall/CRC
5.11: The key condition in Theorem 5.3 is (A6) as this allows us to the likelihood equation by a linear equation in √ approximate n(θ/n − θ). However, condition (A6) can be replaced by other similar conditions, some of which may be weaker than (A6). Assume that θ/n →p θ and that conditions (A1)(A5) hold. Suppose that for some δ > 0, there exists a function Kδ (x) and a constant α > 0 such that + + +: (x; t) − : (x; θ)+ ≤ K(x)t − θα
for t−θ ≤ δ where Eθ [Kδ (Xi )] < ∞. Show that the conclusion of Theorem 5.3 holds. 5.12: In section 5.4, we noted that the consistency of the MLE is straightforward when the loglikelihood function is concave. Similarly, it is possible to exploit the concavity or convexity of objective functions to derive the limiting distributions of estimators. In this problem, we will derive the limiting distribution of the sample median using the fact that the median minimizes a convex objective function; see Problem 4.19. Suppose that {Zn (u)} is a sequence of random convex functions and for any (u1 , u2 , · · · , uk ), we have (Zn (u1 ), · · · , Zn (uk )) →d (Z(u1 ), · · · , Z(uk )) where Z(u) is a random convex function that is uniquely minimized at U (which will be a random variable). Then it can be shown that if Un minimizes Zn then Un →d U (Davis et al, 1992). This result can be used to rederive the limiting distribution of the sample median (see Examples 3.5 and 3.6). Suppose that X1 , · · · , Xn are i.i.d. random variables with distribution /n is the function F where F (µ) = 1/2 and F (µ) = λ > 0. If µ sample median of X1 , · · · , Xn then we know that √ /n − µ) →d N (0, 1/(4λ2 )). n(µ √ /n − µ) minimizes the objective (a) Show that Un = n(µ function Zn (u) =
n '
√ ( Xi − µ − u/ n − Xi − µ
i=1
and that Zn is a convex function of u. c 2000 by Chapman & Hall/CRC
(b) Show that x − y − x = −y sgn(x) + 2
y 0
[I(x ≤ s) − I(x ≤ 0)] ds
for x = 0 where sgn(x) = I(x > 0) − I(x < 0) is the “sign” of x. (c) Show that n u sgn(Xi − µ) Zn (u) = − √ n i=1 n 2 +√ n i=1
u)
0
s I Xi ≤ µ + √ n
*
− I(Xi ≤ µ) ds
using the formula in part (b). (d) Show that (Zn (u1 ), · · · , Zn (uk )) →d (Z(u1 ), · · · , Z(uk )) where Z(u) = −uW + 2λu2 and W ∼ N (0, 1). (Hint: Note that √ √ E[I(Xi ≤ µ + s/ n) − I(Xi ≤ µ)] = F (µ + s/ n) − F (µ) and
√ √ Var[I(Xi ≤ µ + s/ n) − I(Xi ≤ µ)] ≤ F (µ + s/ n) − F (µ) √ √ for each s, with F (µ + s/ n) − F (µ) ≈ λs/ n.) √ /n − µ) →d W/(2λ). (e) Show that n(µ 5.13: The same approach used in Problem 5.12 can be used to determine the limiting distribution of the sample median under more general conditions. Again let X1 , · · · , Xn be i.i.d. with distribution function F and median µ where now √ lim n[F (µ + s/an ) − F (µ)] = ψ(s) n→∞
for some increasing function ψ and sequence of constants an → /n −µ) will be determined ∞. The asymptotic distribution of an (µ by considering the objective function n an Zn (u) = √ [Xi − µ − u/an  − Xi − µ] . n i=1
/n − µ) minimizes Zn . (a) Show that Un = an (µ c 2000 by Chapman & Hall/CRC
(b) Repeat the steps used in Problem 5.12 to show that (Zn (u1 ), · · · , Zn (uk )) →d (Z(u1 ), · · · , Z(uk )) .
where Z(u) = −uW + 2 0u ψ(s) ds and W ∼ N (0, 1). /n − µ) →d ψ −1 (W/2). (c) Show that an (µ 5.14: The conditions assumed in Theorems 5.3 and 5.4 eﬀectively imply that the loglikelihood function is approximately quadratic in a neighbourhood of the true parameter value. These quadratic approximations can be used to give “heuristic” proofs of Theorems 5.3 and 5.4. Suppose that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) satisfying conditions (B1)(B6). Deﬁne Zn (u) =
n
√ ' ( ln f (Xi ; θ + u/ n)/f (Xi ; θ)
i=1
and note that Zn is maximized at u = the MLE of θ. (a) Show that
√
/ n − θ) where θ / n is n(θ
1 Zn (u) = uT W n − uT J(θ)u + Rn (u) 2 where supK Rn (u) →p 0 for any compact set K and W n →d Np (0, I(θ)). (b) Part (a) suggests that the limit of Zn is 1 Z(u) = uT W − uT J(θ)u 2 where W ∼ Np (0, I(θ)). Show that Z is maximized at u = √ / −1 (θ)W . (If J −1 (θ)W , which suggests that n(θ n − θ) →d J Zn (u) is a concave function of u for each n then this argument is a rigorous proof of Theorem 5.4.) 5.15: In Theorems 5.3 and 5.4, we assume that the parameter space Θ is an open subset of Rp . However, in many situations, this assumption is not valid; for example, the model may impose constraints on the parameter θ, which eﬀectively makes Θ a closed set. If Θ is not an open set then the MLE of θ need not / n may lie on the satisfy the likelihood equations as the MLE θ c 2000 by Chapman & Hall/CRC
boundary of Θ. In determining the asymptotic distribution of / n the main concern is whether or not the true value of the θ parameter lies on the boundary of the parameter space. If θ lies in the interior of Θ then eventually (for suﬃciently large / n will satisfy the likelihood equations and so Theorems 5.3 n) θ and 5.4 will still hold; however, the situation becomes more complicated if θ lies on the boundary of Θ. Suppose that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) (satisfying conditions (B2)(B6)) where θ lies on the boundary of Θ. Deﬁne (as in Problem 5.14) the function Zn (u) =
n
√ ' ( ln f (Xi ; θ + u/ n)/f (Xi ; θ)
i=1
and the set
√ Cn = {u : θ + u/ n ∈ Θ}.
The limiting distribution of the MLE can be determined by the limiting behaviour of Zn and Cn ; see Geyer (1994) for details. √ / (a) Show that n(θ n − θ) maximizes Zn (u) subject to the constraint u ∈ Cn . (b) Suppose that {Cn } is a decreasing sequence of sets whose limit is C. Show that C is nonempty. (c) Parts (a) and (b) (together with Problem 5.14) suggest that √ / n(θ n − θ) converges in distribution to the minimizer of 1 Z(u) = uT W − uT J(θ)u 2 (where W ∼ Np (0, I(θ))) subject to u ∈ C. Suppose that X1 , · · · , Xn are i.i.d. Gamma random variables with shape parameter α and scale parameter λ where the parameter space is restricted so that α ≥ λ > 0 (that is, E(Xi ) ≥ 1). If α = λ, describe the limiting distribution of the MLEs. (Hint: Show that C = {(u1 , u2 ) : u1 ≥ u2 }.) 5.16: Suppose that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) where θ is realvalued. In many cases, the MLE of θ satisﬁes the likelihood equation n i=1 c 2000 by Chapman & Hall/CRC
: (Xi ; θ/n ) = 0.
If :(x; θ) is an unbounded function in x then a single observation can potentially have an arbitrarily large inﬂuence on the value of the MLE. For this reason, it is often desirable to robustify maximum likelihood estimation by replacing : (x; θ) by a function ψc (x; θ) that is bounded in x. In order to have Fisher consistency, we choose ψc so that Eθ [ψc (Xi ; θ)] = 0 where the expected value is taken with respect to the density or frequency function f (x; θ). Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. The MLE of λ satisﬁes the likelihood equation n 1 − Xi = 0, / i=1 λn / n = 1/X ¯ n . Consider replacing which can be solved to yield λ : (x; λ) = 1/λ − x (which is unbounded in x) by
ψc (x; λ) =
1/λ + gc (λ) − x c/λ
if x ≤ c/λ + 1/λ + gc (λ) otherwise
where c > 0 is a tuning parameter and gc (λ) is determined so that Eλ [ψc (Xi ; λ)] = 0. (a) Show that gc (λ) satisﬁes the equation exp(−c − 1 − gc (λ)λ) + gc (λ) = 0 0 n to satisfy the equation (b) Deﬁne λ n
0 n ) = 0. ψc (Xi ; λ
i=1
Assuming that the standard √ 0regularity conditions hold, ﬁnd the limiting distribution of n(λ n − λ). 0 n with respect (c) Find the asymptotic relative eﬃciency of λ / n . For what value of c is the asymptotic relative to the MLE λ eﬃciency equal to 0.95? 5.17: Let X1 , · · · , Xn be i.i.d. random variables with density or frequency function f (x; θ) where θ is a realvalued parameter. / satisﬁes the likelihood equation Suppose that MLE of θ, θ, n i=1 c 2000 by Chapman & Hall/CRC
/ =0 : (Xi ; θ)
where : (x; θ) is the derivative with respect to θ of ln f (x; θ). (a) Let θ/−j be MLE of θ based on all the Xi ’s except Xj . Show that / : (Xj ; θ) θ/−j ≈ θ/ + n / i=1 : (Xi ; θ) (if n is reasonably large). (b) Show that the jackknife estimator of θ/ satisﬁes
n / 2 n−1 j=1 [: (Xj ; θ)] / 2 Var(θ) ≈ 2 . n n / : (Xi ; θ) i=1
(c) The result of part (b) suggests that the jackknife estimator / is essentially the “sandwich” estimator; the latter of Var(θ) estimator is valid when the model is misspeciﬁed. Explain the / apparent equivalence between these two estimators of Var(θ). (Hint: Think of the MLE as a substitution principle estimator of some functional parameter.) 5.18: Millar (1987) considers a statistical model for determining the composition of a mixed stock ﬁshery. The statistical model can be described as follows: We have a sample of N ﬁsh that can be classiﬁed into one of G genotypes (where typically G >> N ). If Yi is the number of ﬁsh in the sample with genotype i then Y = (Y1 , · · · , YG ) is a Multinomial random vector with P (Y = y) =
G N! λi y1 ! × · · · × yG ! i=1
where λi =
S
xij θj
j=1
where θ1 , · · · , θS (unknown) are the proportion of ﬁsh belonging to the S subpopulations or stocks, and xij (known) is the (conditional) probability of belonging to genotype i given membership in stock j. To estimate θ = (θ1 , · · · , θS ), we maximize the loglikelihood function ln L(θ) =
G i=1
c 2000 by Chapman & Hall/CRC
yi ln
S j=1
xij θj + k(y)
where k(y) does not depend on θ. (We also must assume that θj ≥ 0 for all j and θ1 + · · · + θS = 1.) (a) Show that ln L(θ) can be written as ln L(θ) =
N
ln
r=1
S
xir j θj
j=1
where ir is the genotype of the rth ﬁsh in the sample. (b) Show that the MLEs of θ1 , · · · , θS satisfy the equations θ/j =
N xir j θ/j 1
S N r=1 k=1 xir k θ/k
for j = 1, · · · , S.
(c) Assume that S = 2. In this case, we need only estimate a single parameter θ (equal, say, to the proportion belonging to stock 1). Assuming that the standard regularity conditions for maximum likelihood estimators hold, ﬁnd an estimator of the / standard error of θ. 5.19: Suppose that X = (X1 , · · · , Xn ) has a joint density or frequency function f (x; θ) where θ has prior density π(θ). If T = T (X) is suﬃcient for θ, show that the posterior density of θ given X = x is the same as the posterior density of θ given T = T (x). 5.20: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on [0, θ] where θ has a Pareto prior density function: π(θ) =
α θ0
θ θ0
−α−1
for θ > θ0 > 0 and α > 0 where θ0 and α are hyperparameters. (a) Show that the posterior distribution of θ is also Pareto. (b) Suppose that θ∗ is the true value of θ. Under i.i.d. sampling, what happens to the posterior density of θ as n → ∞? (Hint: There are two cases to consider: θ∗ ≥ θ0 and θ∗ < θ0 .) 5.21: The Zeta distribution is sometimes used in insurance as a model for the number of policies held by a single person in an insurance portfolio. The frequency function for this distribution is x−(α+1) f (x; α) = ζ(α + 1) c 2000 by Chapman & Hall/CRC
Table 5.8 Data for Problem 5.21.
Observation
1
2
3
4
5
Frequency
63
14
5
1
2
for x = 1, 2, 3, · · · where α > 0 and ζ(p) =
∞
k −p .
k=1
(The function ζ(p) is called the Riemann zeta function.) (a) Suppose that X1 , · · · , Xn are i.i.d. Zeta random variables. Show that the MLE of α satisﬁes the equation n / n + 1) 1 ζ (α ln(Xi ) = − / n + 1) n i=1 ζ(α
√ / n − α). and ﬁnd the limiting distribution of n(α (b) Assume the following prior density for α: 1 π(α) = α2 exp(−α) for α > 0 2 A sample of 85 observations is collected; its frequency distribution is given in Table 5.8. Find the posterior distribution of α. What is the mode (approximately) of this posterior distribution? (c) Repeat part (b) using the improper prior density π(α) =
1 α
for α > 0.
Compare the posterior densities in part (b) and (c). 5.22: Suppose that X has a Binomial distribution with parameters n and θ where θ is unknown. (a) Find the Jeﬀreys prior for θ. Is this prior density proper? (b) Find the posterior density for θ given X = x using the Jeﬀreys prior. 5.23: The concept of Jeﬀreys priors can be extended to derive “noninformative” priors for multiple parameters. Suppose that c 2000 by Chapman & Hall/CRC
X has joint density or frequency function f (x; θ) and deﬁne the matrix I(θ) = Eθ [S(X; θ)S T (X; θ)] where S(x; θ) is the gradient (vector of partial derivatives) of ln f (x; θ) with respect to θ. The Jeﬀreys prior for θ is proportional to det(I(θ))1/2 . (a) Show that the Jeﬀreys prior can be derived using the same considerations made in the single parameter case. That is, if φ = g(θ) for some onetoone function g such that I(φ) is constant then the Jeﬀreys prior for θ corresponds to a uniform prior for φ. (b) Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance σ 2 . Find the Jeﬀreys prior for (µ, σ).
c 2000 by Chapman & Hall/CRC
CHAPTER 6
Optimality in Estimation 6.1 Introduction To this point, we have discussed a number of approaches for obtaining point estimators. We have also discussed approaches for determining various properties of these estimators, such as limiting distributions and approximate standard errors. However, we have not (so far) attempted to determine if a given estimator is optimal in any sense. As it turns out, “optimal” is a poorly deﬁned term in the context of estimation as there are a number of criteria for optimality. For example, we can approach optimality from a large sample (or asymptotic) point of view. Suppose that X1 , X2 , · · · is a sequence of random variables and θ/n is an estimator of some parameter θ based on X = (X1 , · · · , Xn ) such that √ / n(θn − θ) →d N (0, σ 2 (θ)). A natural question to ask is whether a lower bound for σ 2 (θ) exists; if such a lower bound exists and is attained by some sequence of estimators {θ/n } then we can say that this sequence is optimal. In this chapter, we will discuss optimal estimation from three perspectives. First, we will look at estimation from a decision theoretic point of view, comparing estimators based on their risk for a given loss function. Next, we will narrow our focus to unbiased estimation and attempt to ﬁnd unbiased estimators with uniformly minimum variance over the parameter space. Finally, we will take an asymptotic point of view and consider estimators with asymptotic Normal distributions. 6.2 Decision theory While we have discussed a number of diﬀerent approaches to point estimation, we have so far avoided the issue of which estimator is “best” or optimal in a given situation. One approach to ﬁnding and c 2000 by Chapman & Hall/CRC
evaluating estimators is based on decision theory. We will discuss only the most basic elements of decision theory here; a more detailed treatment is given in Ferguson (1967). In decision theory, in addition to specifying a model for X (depending on some unknown parameter θ), we also specify a loss / θ) that describes the “loss” incurred by making an function L(θ, / estimate θ when the true value of the parameter is θ. Typically, the loss function is chosen so that for ﬁxed θ, L(a, θ) increases as a moves away from θ and L(θ, θ) = 0. DEFINITION. A loss function L(·, ·) is a nonnegativevalued / θ) indicates the loss function deﬁned on Θ × Θ so that L(θ, / incurred in estimating a parameter θ by θ. Perhaps the most commonly used loss functions for realvalued parameters are squared error loss (L(a, b) = (a − b)2 ) and absolute error loss (L(a, b) = a−b). However, many other loss functions are possible. Strictly speaking, the loss function used should be dictated by the particular problem; however, almost invariably squared error loss is used in the vast majority of estimation problems. / the DEFINITION. Given a loss function L and an estimator θ, / / θ): risk function Rθ (θ) is deﬁned to be the expected value of L(θ, / = E [L(θ, / θ)]. Rθ (θ) θ / = MSE (θ), / the mean If squared error loss is used, then Rθ (θ) θ / Likewise, if absolute error loss is used then square error of θ. / / Rθ (θ) = MAEθ (θ), the mean absolute error. Given a particular loss function L, we can try to ﬁnd an estimator / is minimized for all θ. Unfortunately, this is θ/ such that Rθ (θ) impossible to do except in trivial cases; typically, the estimator that minimizes Rθ (·) for a ﬁxed θ will not minimize Rθ (·) uniformly over all θ. However, if we accept risk (for example, mean square error) as a measure of the quality of an estimator, we can rule out estimators whose risk is uniformly higher than another estimator.
DEFINITION. For a given loss function L, an estimator θ/ is an inadmissible estimator of θ if there exists an estimator θ0 such that 0 ≤ R (θ) / for all θ ∈ Θ Rθ (θ) θ and
0 < R (θ) / Rθ0 (θ) θ0
c 2000 by Chapman & Hall/CRC
for some θ0 ∈ Θ.
If no such estimator θ0 exists then θ/ is admissible. EXAMPLE 6.1: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. The MLE of λ is /= 1 λ ¯ X / = nλ/(n− ¯ is the sample mean. It can be shown that Eλ (λ) where X 0 = (n − 1)λ/n. / 1) when n ≥ 2. An unbiased estimator of λ is λ Clearly, 0 < MSE (λ) / MSEλ (λ) λ 0 is unbiased and Var (λ) 0 < Var (λ). / Since MSE (λ) / = R (λ) / since λ λ λ λ λ / is an inadmissible when the loss function is squared error loss, λ estimator of λ under squared error loss. However, in the case of estimating positive parameters, mean square error tends to penalize overestimation more heavily than underestimation (since the maximum possible underestimation is bounded) and thus squared error loss may not be the best loss function to consider in this example. Instead, we might deﬁne the loss function
L(a, b) =
b − 1 − ln(b/a) a
for which we have for each ﬁxed b > 0, lim L(a, b) = lim L(a, b) = ∞.
a→0
a→∞
0 we get Evaluating the risk function of λ, ,
0 Rλ (λ)

¯ ¯ nλX nλX = Eλ − 1 − ln n−1 n−1 ' ( ¯ − 1 − ln λX ¯ = Eλ λ X
n 1 ¯ + Eλ (λX) − ln n−1 n−1
' ( n 1 ¯ ¯ = Eλ λX − 1 − ln λX + − ln n−1 n−1 ' ( ¯ ¯ > Eλ λX − 1 + ln λX / = Rλ (λ)
c 2000 by Chapman & Hall/CRC
0 is (since 1/(n − 1) − ln(n/(n − 1)) > 0 for n ≥ 2) and so λ inadmissible under this loss function. ✸
Note that we cannot make any claims about the admissibility of the estimators in Example 6.1. More precisely, to prove admissibility, we need to show that there exists no other estimator with uniformly lower risk. At ﬁrst glance, this may seem a formidable task; however, there are technical devices available that facilitate it. One of these devices arises by putting a prior distribution on the parameter space. Suppose that we put a (proper) prior density function π(θ) on the parameter space Θ. Then given a loss function L(·, θ) and risk function Rθ (·), we can deﬁne the Bayes risk of an estimator θ/ by / = RB (θ)
/ π(θ) dθ. Rθ (θ)
Θ
/ depends on the loss function, the distribution of (Note that RB (θ)
X as well as the prior distribution on Θ.) Then a Bayes estimator of θ is an estimator that minimizes the Bayes risk. Bayes estimators can usually be determined from the posterior distribution of θ; more precisely, if the expected posterior loss function L(T (x), θ)π(θx) dθ Θ
is minimized at T (x) = T ∗ (x) (where π(θx) is the posterior density function of θ) then T ∗ (X) is a Bayes estimator. For example, for squared error loss, the Bayes estimator is simply the mean of the posterior distribution while for absolute error loss, the Bayes estimator is any median of the posterior distribution. Bayes estimators are also admissible estimators provided that they are unique. For example, suppose that θ/ is a unique Bayes 0 θ) ≤ R (θ, / θ) for some other estimator and suppose that Rθ (θ, θ 0 Then estimator θ. 0 = RB (θ)
≤
Θ Θ
0 π(θ) dθ Rθ (θ) / π(θ) dθ Rθ (θ)
/ = RB (θ),
which contradicts the uniqueness of θ/ as a Bayes estimator; thus θ/ must be admissible. c 2000 by Chapman & Hall/CRC
An important optimality concept in decision theory is minimax estimation. DEFINITION. An estimator θ/ is said to be minimax (with respect to a loss function L) if / ≤ sup R (θ) 0 sup Rθ (θ) θ θ∈Θ
θ∈Θ
0 for any estimator θ.
It is easy to see that a minimax estimator is admissible. In general, proving that an estimator is minimax is not easy although this can be facilitated if it can be shown that the estimator is a Bayes estimator (or a limit of Bayes estimators) and has constant risk over the parameter space; see Problem 6.5 for more details. It should be noted that there are pitfalls involved in using any criterion for evaluating estimators based on risk functions (such as admissibility or minimaxity), particularly when the loss function is not appropriate for the problem. In some problems, estimators are often inadmissible only because they are dominated (in terms of risk) by estimators whose practical use is somewhat dubious; in such cases, it may be the loss function rather than the estimator that is suspect. However, even if we are conﬁdent that our loss function is appropriate, it may be worthwhile to limit the class of estimators under consideration. For example, if θ is a parameter describing the center of a distribution, it may be desirable to consider only those estimators θ/ = S(X1 , · · · , Xn ) for which S(X1 + c, · · · , Xn + c) = S(X1 , · · · , Xn ) + c. More generally, we might consider only those estimators that satisfy a certain invariance or equivariance property, or only unbiased estimators. In the next section, we will consider unbiased estimation under squared error loss. 6.3 Minimum variance unbiased estimation As mentioned above, one approach to ﬁnding estimators is to ﬁnd an estimator having uniformly smallest risk over some restricted class of estimators. If we take the loss function to be squared error loss and consider only unbiased estimators of a parameter then we reduce the problem to ﬁnding the unbiased estimator with the minimum variance (since the mean square error and variance are the c 2000 by Chapman & Hall/CRC
same for unbiased estimators); if such an estimator has minimum variance over the parameter space then this estimator is called a uniformly minimum variance unbiased (UMVU) estimator. Suppose that X1 , · · · , Xn are random variables whose joint distribution depends on some parameter θ, which may be real or vectorvalued. In this section, we will consider unbiased estimation of g(θ) where g is a realvalued function deﬁned on the parameter space Θ. As discussed previously in Chapter 4, it is not clear that unbiased estimators are always desirable. For example, if we believe in the likelihood principle then it is easily shown that the requirement of unbiasedness is a violation of this principle. For the time being, however, we will ignore the possible shortcomings of unbiased estimators and focus our energy on the theory of optimal unbiased estimator under squared error loss. Nonetheless, it should be noted that unbiased estimators do not always exist or, if they do exist, can be nonsensical estimators; the following examples illustrate these points. EXAMPLE 6.2: Suppose that X is a Binomial random variable with parameters n and θ, where 0 < θ < 1 is unknown. We wish to ﬁnd an unbiased estimator of g(θ) = 1/θ; that is, we need to ﬁnd a statistic T (X) such that Eθ [T (X)] = 1/θ for all 0 < θ < 1. Thus we need to ﬁnd T (0), T (1), · · · , T (n) such that n
T (x)
x=0
n x 1 θ (1 − θ)n−x = . θ x
Multiplying both sides by θ, we get n x=0
T (x)
n+1 n x+1 θ (1 − θ)n−x = a(k)θk x k=1
= 1 where a(1), · · · , a(n + 1) depend on T (0), T (1), · · · , T (n). It follows that whatever the choice of a(1), · · · , a(n + 1), the equality a(1)θ + · · · + a(n + 1)θn+1 = 1 is satisﬁed for at most n + 1 values of θ between 0 and 1 and cannot be satisﬁed for all θ. Thus there exists no unbiased estimator of 1/θ. ✸ EXAMPLE 6.3: Suppose that X is a Poisson random variable c 2000 by Chapman & Hall/CRC
with mean λ. We wish to ﬁnd an unbiased estimator of exp(−2λ) based on X. If T (X) is this estimator, we must have Eλ [T (X)] =
∞
T (x)
x=0
exp(−λ)λx = exp(−2λ) x!
for all λ > 0. Multiplying both sides of the equation by exp(λ), we get ∞
T (x)
x=0
λx x!
= exp(−λ) ∞
=
(−1)x
x=0
λx , x!
which implies that T (x) = (−1)x . Thus the only unbiased estimator of exp(−2λ) is (−1)X ; since 0 < exp(−2λ) < 1 for λ > 0, this is clearly a ridiculous estimator. It should be noted, however, that if X1 , · · · , Xn are i.i.d. Poisson random variables with parameter λ
n T then (1 − 2/n) (with T = i=1 Xi ) is unbiased; this estimator is somewhat more sensible especially for larger values of n. ✸ Examples 6.2 and 6.3 notwithstanding, in many cases the class of unbiased estimators is nontrivial. EXAMPLE 6.4: Suppose that X1 , · · · , Xn are i.i.d. random variables with density f (x; µ) = exp (−(x − µ))
for x ≥ µ
(where −∞ < µ < ∞). Two possible unbiased estimators of µ are /1 = X(1) − µ
1 n
¯ − 1; /2 = X and µ
/1 + (1 − t)µ /2 is also unbiased for any t. Simple note also that tµ calculations reveal that /1 ) = Var(µ
1 n2
and
/2 ) = Var(µ
1 n
/1 has the smaller variance. It is interesting to note that µ /1 so that µ depends only on the onedimensional suﬃcient statistic X(1) . ✸ /1 have the minimum variance among all Does the estimator µ c 2000 by Chapman & Hall/CRC
unbiased estimators of µ in the previous example? The following result indicates that any candidate for the minimum variance unbiased estimator should be a function of the suﬃcient statistic. THEOREM 6.1 (RaoBlackwell Theorem) Suppose that X has a joint distribution depending on some unknown parameter θ and that T = T (X) is a suﬃcient statistic for θ. (Both θ and T can be vectorvalued.) Let S = S(X) be a statistic with Eθ (S) = g(θ) and Varθ (S) < ∞ for all θ. If S ∗ = E(ST ) then (a) S ∗ is an unbiased estimator of g(θ), and (b) Varθ (S ∗ ) ≤ Varθ (S) for all θ. Moreover, Varθ (S ∗ ) < Varθ (S) unless Pθ (S = S ∗ ) = 1. Proof. Since T is suﬃcient for θ, h(t) = E(ST = t) does not depend on θ and so S ∗ = h(T ) is a statistic with Eθ (S ∗ ) = Eθ [E(ST )] = Eθ (S) = g(θ). Also Varθ (S) = Varθ [E(ST )] + Eθ [Var(ST )] ≥ Varθ [E(ST )] = Varθ (S ∗ ) and so Varθ (S ∗ ) ≤ Varθ (S). Also Var(ST ) = E[(S − S ∗ )2 T ] > 0 unless P (S = S ∗ T ) = 1. Thus Eθ [Var(ST )] > 0 unless S = S ∗ with probability 1. The RaoBlackwell Theorem says that any unbiased estimator should be a function of a suﬃcient statistic; if not, we can construct an estimator with smaller variance merely by taking the conditional expectation given a suﬃcient statistic. However, this raises the question of which suﬃcient statistic to use to compute the conditional expectation. For example, suppose that S is an unbiased estimator of g(θ) (with ﬁnite variance) and T1 and T2 are both suﬃcient statistics for θ with T2 = h(T1 ) for some function h. (Both T1 and T2 can be vectorvalued.) We deﬁne S1∗ = E(ST1 ) and S2∗ = E(ST2 ). By the RaoBlackwell Theorem, the variances of S1∗ and S2∗ cannot exceed Var(S); however, it is not obvious which “RaoBlackwellized” estimator will have the smaller variance although intuition suggests that Varθ (S2∗ ) should be smaller since T2 is the “simpler” statistic (as T2 = h(T1 )). c 2000 by Chapman & Hall/CRC
PROPOSITION 6.2 Suppose that S is an unbiased estimator of g(θ) with ﬁnite variance and deﬁne S1∗ = E(ST1 ) and S2∗ = E(ST2 ) for suﬃcient statistics T1 and T2 . If T2 = h(T1 ), Varθ (S2∗ ) ≤ Varθ (S1∗ ). Proof. We will use the fact that Eθ [(S − φ(T1 ))2 ] is minimized over all functions φ by E(ST1 ). Since T2 is a function of T1 , it follows that E(S1∗ T2 ) = = = =
E [E(ST1 )T2 ] E [E(ST2 )T1 ] E(S2∗ T1 ) S2∗
where the last equality holds since S2∗ is a function of T1 . The conclusion now follows from the RaoBlackwell Theorem. Proposition 6.2 essentially says that for any unbiased estimator S of g(θ) the best “RaoBlackwellization” of S is achieved by conditioning on a minimal suﬃcient statistic. (Recall that a statistic T is minimal suﬃcient if for any other suﬃcient statistic T ∗ , T = h(T ∗ ) for some function h.) However, even if T is minimal suﬃcient, the estimator S ∗ = E(ST ) will not necessarily have the minimum variance among all unbiased estimators since there may exist another unbiased estimator S1 such that S1∗ = E(S1 T ) has a smaller variance than S ∗ . In the next section, we will show that in certain problems, it is possible to ﬁnd a suﬃcient statistic T such that S ∗ = E(ST ) is independent of S. The following simple (and rather silly) example illustrates the potential problem. EXAMPLE 6.5: Suppose X is a discrete random variable with frequency function
f (x; θ) =
θ for x = −1 (1 − θ)2 θx for x = 0, 1, 2, · · ·
where 0 < θ < 1. The statistic X is suﬃcient for θ and can be shown to be minimal suﬃcient. Two unbiased estimators of θ are S1 = I(X = −1) and S2 = I(X = −1) + X (since Eθ (X) = 0 for all θ). Since both S1 and S2 are functions of the minimal suﬃcient statistic X, we have S1∗ = E(S1 X) = S1 and S2∗ = E(S2 X) = S2 . But Varθ (S1 ) = θ(1 − θ) c 2000 by Chapman & Hall/CRC
while Varθ (S2 ) = Var(S1 ) + Var(X) + 2Cov(S1 , X) 2θ = θ(1 − θ) + − 2θ 1−θ θ(θ2 + 1) = 1−θ > Varθ (S1 ) thus illustrating that dependence on the minimal suﬃcient statistic does not guarantee that an unbiased estimator will have minimum variance. ✸ Complete and Suﬃcient Statistics Suppose that S1 and S2 are two unbiased estimators of g(θ) (with ﬁnite variance) and suppose that T is suﬃcient for θ. We can deﬁne S1∗ = E(S1 T ) and S2∗ = E(S2 T ). Although Varθ (Si∗ ) ≤ Varθ (Si ) (for i = 1, 2), there is no way of knowing a priori whether S1∗ or S2∗ will have a smaller variance. However, for an appropriate choice of T , we would like to have • Varθ (S1∗ ) = Varθ (S2∗ ) for all θ, or • Pθ (S1∗ = S2∗ ) = 1 for all θ. More precisely, for this particular choice of T , there will be only one unbiased estimator that is a function of T and, if T is minimal suﬃcient, this unbiased estimator will have minimum variance. EXAMPLE 6.6: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ and consider unbiased estimators of g(λ) = exp(−λ) = Pλ (Xi = 0). Two unbiased estimators of g(λ) are n 1 I(Xi = 0). S1 = I(X1 = 0) and S2 = n i=1
n A minimal suﬃcient statistic for λ is T = i=1 Xi . To ﬁnd the “RaoBlackwellized” estimators, we must ﬁnd the conditional distributions of S1 and S2 given T = t. It is easy to see that
Pλ (S1 = s, T = t) Pλ (T = t)
1 t 1− = n
P (S1 = 1T = t) =
c 2000 by Chapman & Hall/CRC
and hence
1 n
For S2 , deﬁne U = ni=1 I(Xi = 0); we get S1∗ = E(S1 T ) = 1 −
n u
P (U = uT = t) =
T
.
u 1− n
t
for u = max(n − t, 0), · · · , max(n − 1, n − t). It follows then that E(U T = t) =
u
n u u
u 1− n
t
1 =n 1− n
t
and so S2∗ = (1 − 1/n)T = S1∗ . Thus, in this example, “RaoBlackwellizing” S1 and S2 leads to the same estimator. ✸ DEFINITION. Suppose that X1 , · · · , Xn are random variables whose joint distribution depends on some unknown parameter θ. A statistic T = T (X1 , · · · , Xn ) (possibly vectorvalued) is said to be complete for θ if for any function g, Eθ [g(T )] = 0 for all θ implies that Pθ [g(T ) = 0] = 1 for all θ. Completeness of a statistic T essentially means that T contains no “ancillary” (meaningless) information about θ. For example, if T is complete for θ then g(T ) is an ancillary statistic for θ if, and only if, g(T ) is constant over the range of T . (More precisely, for some constant k, Pθ [g(T ) = k] = 1 for all θ.) It can be shown that if a statistic T is suﬃcient and complete then T is also minimal suﬃcient. However, a minimal suﬃcient statistic need not be complete. EXAMPLE 6.7: Suppose that X1 , · · · , Xn are i.i.d. Poisson n random variables with mean λ and deﬁne T = i=1 Xi . T , of course, is suﬃcient for λ; to see if T is complete, we need to see if there exists a function g such that Eλ [g(T )] = 0 for all λ > 0. Since T has a Poisson distribution with mean nλ, we have Eλ [g(T )] =
∞
g(x)
x=0
exp(−nλ)(nλ)x x!
and so Eλ [g(T )] = 0 for all λ if, and only if, ∞ x=0
g(x)
(nλ)x =0 x!
c 2000 by Chapman & Hall/CRC
for all λ > 0.
k Since ∞ k=0 ck λ = 0 for all a < λ < b if, and only if, ck = 0 for all k ≥ 0, it follows that Eλ [g(T )] = 0 for all λ > 0 if, and only if, g(x)nx /x! = 0 for x = 0, 1, · · ·. Hence Eλ [g(T )] = 0 for all λ > 0 implies that Pλ (g(T ) = 0) = 1 for all λ > 0. ✸
EXAMPLE 6.8: Suppose that X1 , · · · , Xn are i.i.d. discrete random variables with frequency function
f (x; θ) =
θ for x = −1 2 x (1 − θ) θ for x = 0, 1, 2, · · ·
where 0 < θ < 1. The joint frequency function of X = (X1 , · · · , Xn ) can be written as a twoparameter exponential family: f (x; θ) = exp [c1 (θ)T1 (x) + c2 (θ)T2 (x) + 2 ln(1 − θ)] where c1 (θ) = ln(θ) − 2 ln(1 − θ), c2 (θ) = ln(θ) and T1 (x) = T2 (x) =
n i=1 n
I(xi = −1) xi I(xi ≥ 0).
i=1
Thus (T1 (X), T2 (X) is suﬃcient (in fact, minimal suﬃcient) for θ. However, this suﬃcient statistic is not complete. To see this, note that ∞ x(1 − θ)2 θx = 0. Eθ (Xi ) = −θ + x=0
Since n i=1
Xi =
n
Xi I(Xi ≥ 0) −
i=1
n
I(Xi = −1)
i=1
= T2 − T1 , it follows that Eθ (T2 − T1 ) = 0 for all θ. Since Pθ (T2 = T1 ) < 1 for all θ, it follows that the suﬃcient statistic (T1 , T2 ) is not complete. (This explains what happens in Example 6.5.) ✸ The following result gives a condition for a suﬃcient statistic to be complete in an exponential family. THEOREM 6.3 Suppose that X = (X1 , · · · , Xn ) have joint c 2000 by Chapman & Hall/CRC
density or joint frequency function that is a kparameter exponential family: f (x; θ) = exp
, k

ci (θ)Ti (x) − d(θ) + S(x)
for x ∈ A
i=1
Deﬁne C = {(c1 (θ), · · · , ck (θ)) : θ ∈ Θ}. If the set C contains an open set (rectangle) of the form (a1 , b1 ) × · · · × (ak , bk ) then the statistic (T1 (X), · · · , Tk (X)) is complete as well as suﬃcient for θ. The proof of Theorem 6.3 is beyond the scope of this book; a proof can be found in Lehmann (1991). Essentially this result is a consequence of the uniqueness of characteristic functions. It says, roughly speaking, that a kdimensional suﬃcient statistic in a kparameter exponential family will also be complete provided that the dimension of the parameter space is k. Note that in Example 6.8, the parameter space is onedimensional while the model is a twoparameter exponential family. EXAMPLE 6.9: Suppose X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance σ 2 . The joint density can be written as a twoparameter exponential family: ,

n n 1 µ x2i + 2 xi − d(µ, σ) f (x; µ, σ) = exp − 2 2σ i=1 σ i=1
with d(µ, σ) = nµ2 /(2σ 2 ) + n ln(σ) + n ln(2π)/2. Clearly range of the function
1 µ c(µ, σ) = − , 2 2σ σ for −∞ < µ < ∞ and σ > 0 contains an open rectangle in R2 so the suﬃcient statistic n i=1
Xi ,
n
Xi2
i=1
is complete as well as suﬃcient for (µ, σ). Moreover, as with suﬃciency, any onetoone function of complete statistic will a n 2 ¯ ¯ also be complete; thus, for example, X, i=1 (Xi − X) is also complete. ✸ The following result gives a simple criterion for the existence of a UMVU estimator when a complete and suﬃcient statistic exists. c 2000 by Chapman & Hall/CRC
THEOREM 6.4 (LehmannScheﬀ´ e Theorem) Suppose that T is a suﬃcient and complete statistic for θ and that S is a statistic with Eθ (S) = g(θ) and Varθ (S) < ∞ for all θ. If S ∗ = E(ST ) and V is any other unbiased estimator of g(θ) then (a) Varθ (S ∗ ) ≤ Varθ (V ) (b) Varθ (S ∗ ) = Varθ (V ) implies that Pθ (S ∗ = V ) = 1. (Thus S ∗ is the unique UMVU estimator of g(θ).) Proof. Take V to be an arbitrary statistic with Eθ (V ) = g(θ) and Varθ (V ) < ∞. Deﬁne V ∗ = E(V T ) and note that Varθ (V ∗ ) ≤ Varθ (V ) by the RaoBlackwell Theorem. It suﬃces to show that Pθ (S ∗ = V ∗ ) = 1 for all θ. Since both S ∗ and V ∗ are unbiased estimators of g(θ), we have 0 = Eθ (S ∗ − V ∗ ) = Eθ [E(ST ) − E(V T )]
for all θ
or Eθ [h(T )] = 0 for all θ where h(T ) = E(ST )−E(V T ) = S ∗ −V ∗ . Since T is complete, it follows that Pθ (S ∗ − V ∗ = 0) = 1 for all θ, which completes the proof. The LehmannScheﬀ´e Theorem states that if a complete and suﬃcient statistic T exists, then the UMVU estimator of g(θ) (if it exists) must be a function of T ; moreover, if the UMVU estimator exists then it is unique. The LehmannScheﬀ´e Theorem also simpliﬁes the search for unbiased estimators considerably: if a complete and suﬃcient statistic T exists and there exists no function h such that Eθ [h(T )] = g(θ) then no unbiased estimator of g(θ) exists. Taken together, the RaoBlackwell and LehmannScheﬀ´e Theorems also suggest two approaches to ﬁnding UMVU estimators when a complete and suﬃcient statistic T exists. • Find a function h such that Eθ [h(T )] = g(θ). If Varθ [h(T )] < ∞ for all θ then h(T ) is the unique UMVU estimator of g(θ). The function h can be determined by solving the equation Eθ [h(T )] = g(θ) or by making an educated guess. • Given an unbiased estimator S of g(θ), deﬁne the “RaoBlackwellized” estimator S ∗ = E(ST ). Then S ∗ is the unique UMVU estimator of g(θ). EXAMPLE 6.10: Suppose that X1 , · · · , Xn are i.i.d. Bernoulli random variables with parameter θ. By the Neyman Factorization Criterion, T = X1 + · · · + Xn is suﬃcient for θ and since the c 2000 by Chapman & Hall/CRC
distribution of (X1 , · · · , Xn ) is a oneparameter exponential family, T is also complete. Suppose we want to ﬁnd the UMVU estimator of θ2 . First suppose that n = 2. If a UMVU estimator exists, it must be of the form h(T ) where the function h satisﬁes θ2 =
2
h(k)
k=0
2 k θ (1 − θ)2−k . k
It is easy to see that h(0) = h(1) = 0 while h(2) = 1. Thus h(T ) = T (T − 1)/2 is the unique UMVU estimator of θ2 when n = 2. For n > 2, we set S = I(X1 + X2 = 2) and note that this is an unbiased estimator of θ2 . By the LehmanScheﬀ´e Theorem, S ∗ = E(ST ) is the unique UMVU estimator of θ2 . We have E(ST = t) = P (X1 + X2 = 2T = t) Pθ (X1 + X2 = 2, X3 + · · · + Xn = t − 2) = Pθ (T = t) 6 0 if t ≤ 1 n−2 7n = if t ≥ 2 t−2 t =
t(t − 1) . n(n − 1)
Thus S ∗ = T (T − 1)/[n(n − 1)] is the UMVU estimator of θ2 . ✸ EXAMPLE 6.11: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with unknown mean and variance, µ and σ 2 . It was shown earlier that the statistic (T1 , T2 ) =
n i=1
Xi ,
n
¯ 2 (Xi − X)
i=1
is suﬃcient and complete for (µ, σ). Consider unbiased estimators of µ/σ. It is not obvious that an unbiased estimator exists; however, T1 is independent of T2 with T2 /σ 2 ∼ χ2n−1 , which suggests that µ −1/2 = kn E T1 / T2 = E(T1 )E T2 σ where the constant kn depends only on n and not on µ or σ. It follows that E(T1 ) = nµ while
−1/2
E T2
=√
Γ(n/2 − 1) σ −1 2Γ ((n − 1)/2)
c 2000 by Chapman & Hall/CRC
(for n ≥ 3)
and so kn = √
nΓ(n/2 − 1) . 2Γ ((n − 1)/2)
√ Thus kn−1 T1 / T2 is an unbiased estimator of µ/σ when n ≥ 3 and has ﬁnite variance when n ≥ 4; since the estimator is a function of the suﬃcient and complete statistic, it is the unique UMVU estimator. ✸ 6.4 The Cram´ erRao lower bound In the previous section, we saw that UMVU estimators of g(θ) could be found if a complete and suﬃcient statistic existed. However, in many problems, the minimal suﬃcient statistic is not complete and so we cannot appeal to the LehmannScheﬀ´e Theorem to ﬁnd UMVU estimators. In this section, we will derive a lower bound for the variance of an unbiased estimator of g(θ); if the variance of some unbiased estimator achieves this lower bound, then the estimator will be UMVU. Suppose that X1 , · · · , Xn are random variables with joint density or frequency function depending on a realvalued parameter θ and consider unbiased estimators of g(θ) (if they exist). Suppose that S = S(X1 , · · · , Xn ) is an unbiased estimator of g(θ). Under fairly weak regularity conditions, we would like to ﬁnd a function φ(θ) such that Varθ (S) ≥ φ(θ) (for all θ) for any statistic S with Eθ (S) = g(θ). Moreover, we would like to be able to ﬁnd unbiased estimators such that the lower bound is achieved or comes close to being achieved. (For example, 0 is always a lower bound for Varθ (S) but is typically unattainable.) The CauchySchwarz inequality states that if U and V are random variables with E(U 2 ) < ∞ and E(V 2 ) < ∞ then [Cov(U, V )]2 ≤ Var(U )Var(V ). Using this result, we obtain the following lower bound for Varθ (S), Varθ (S) ≥
[Covθ (S, U )]2 , Varθ (U )
which is valid for any random variable U . However, as it stands, this lower bound is not particularly useful since Covθ (S, U ) will c 2000 by Chapman & Hall/CRC
generally depend on S and not merely on θ. Fortunately, it is possible to ﬁnd a random variable U such that Covθ (S, U ) depends only on g(θ) = Eθ (S). We will assume that θ is realvalued and make the following assumptions about the joint density or frequency function of X = (X1 , · · · , Xn ): (1) The set A = {x : f (x; θ) > 0} does not depend on θ. (2) For all x ∈ A, f (x; θ) is diﬀerentiable with respect to θ. (3) Eθ [Uθ (X1 , · · · , Xn )] = 0 where Uθ (x) =
∂ ln f (x; θ). ∂θ
(4) For a statistic T = T (X) with Eθ (T ) < ∞ (for all θ) and g(θ) = Eθ (T ) diﬀerentiable, g (θ) = Eθ [T Uθ (X)] for all θ. At ﬁrst glance, assumptions (3) and (4) may not appear to make sense. However, suppose that f (x; θ) is a density function. Then for some statistic S, d Eθ (S) = dθ
d dθ
=
=
··· ···
···
S(x)f (x; θ) dx
S(x)
∂ f (x; θ) dx ∂θ
S(x)Uθ (x) dx
= Eθ [SUθ ] provided that the derivative may be taken inside the integral sign. Setting S = 1 and S = T , it is easy to see that assumptions (3) and (4) hold. THEOREM 6.5 (Cram´ erRao lower bound) Suppose that X = (X1 , · · · , Xn ) has a joint density (frequency) function f (x; θ) satisfying assumptions (1), (2), and (3). If the statistic T satisﬁes assumption (4) then Varθ (T ) ≥ where I(θ) = Eθ (Uθ2 ). c 2000 by Chapman & Hall/CRC
[g (θ)]2 I(θ)
Proof. By the CauchySchwarz inequality, Varθ (T ) ≥
Cov2θ (T, Uθ ) . Varθ (Uθ )
Since Eθ (Uθ ) = 0, it follows that Varθ (Uθ ) = I(θ). Also Covθ (T, Uθ ) = Eθ (T Uθ ) − Eθ (T )Eθ (Uθ ) = Eθ (T Uθ ) d = Eθ (T ) = g (θ), dθ which completes the proof. When is the Cram´erRao lower bound attained? If Varθ (T ) = then Varθ (T ) =
[g (θ)]2 I(θ)
Cov2θ (T, Uθ ) , Varθ (Uθ )
which occurs if, and only if, Uθ is a linear function of T ; that is, with probability 1, Uθ = A(θ)T + B(θ) or
∂ ln f (x; θ) = A(θ)T (x) + B(θ) ∂θ for all x ∈ A. Hence ln f (x; θ) = A∗ (θ)T (x) + B ∗ (θ) + S(x)
and so Varθ (T ) attains the Cram´erRao lower bound if, and only if, the density or frequency function of (X1 , · · · , Xn ) has the oneparameter exponential family form given above. In particular, the estimator T must be a suﬃcient statistic. If X1 , · · · , Xn are i.i.d. random variables with common density f (x; θ) then n ∂ Uθ (x) = ln f (xi ; θ) ∂θ i=1 and so
I(θ) =
Varθ (Uθ2 )
c 2000 by Chapman & Hall/CRC
= nVarθ
∂ ln f (X1 ; θ) . ∂θ
An alternative method for computing I(θ) can frequently be used. If Eθ (T ) can be diﬀerentiated twice with respect to θ under the integral or sum signs then I(θ) = Eθ [Hθ (X)] where
∂2 . ln f (x; θ) ∂θ2 A similar result was proved in Chapter 5 in the case of ﬁnding the asymptotic variance of the MLE. This alternative method works in many models including oneparameter exponential families. Hθ (x) = −
EXAMPLE 6.12: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ. A simple calculation yields Uλ (x) =
n ∂ i=1
∂λ
ln f (xi ; λ) = −n +
and so I(λ) = Varλ (Uλ ) =
n 1 xi λ i=1
n 1 n Varλ (Xi ) = . 2 λ i=1 λ
Also note that Hλ (x) = −
n ∂2 i=1
∂λ2
ln f (xi ; λ) =
n 1 xi λ2 i=1
and so Eλ (Hλ ) = n/λ = I(λ). The Cram´erRao lower bound for unbiased estimators of λ is simply λ/n and this lower bound is attained by the estimator ¯ In fact, the lower bound can only be attained by estimators T = X. of the form aT + b for constants a and b. Now consider unbiased estimators of λ2 ; the Cram´erRao lower bound in this case is (2λ)2 4λ3 = . I(λ) n From above, we know that no unbiased estimator attains this lower bound. However, we would like to see how close diﬀerent estimators come to attaining the lower bound. Using the fact that Eλ [Xi (Xi − 1)] = λ2 , it follows that T1 = c 2000 by Chapman & Hall/CRC
n 1 Xi (Xi − 1) n i=1
is an unbiased estimator of λ2 ; a tedious calculation yields 4λ3 2λ2 + . n n ¯ is suﬃcient and complete for λ, the UMVU estimator of Since X 2 ¯ with λ is T2 = E(T1 X) Varλ (T1 ) =
4λ3 2λ2 + 2. n n ¯ 2 , which is the MLE Finally consider the biased estimator T3 = (X) ¯ has a Poisson distribution with mean nλ, another of λ2 . Since nX tedious calculation yields Varλ (T2 ) =
Eλ (T3 ) = λ2 + and
Varλ (T3 ) =
λ n
4λ3 5λ2 λ + 2 + 3. n n n
The diﬀerence between the UMVE estimator and the MLE of λ2 becomes negligible for large n. ✸ EXAMPLE 6.13: Suppose that X1 , · · · , Xn are i.i.d. Logistic random variables with density function f (x; θ) =
exp(x − θ) . [1 + exp(x − θ)]2
It is easy to verify that I(θ) = n/3 and so if θ/ is an unbiased estimator of θ based on X1 , · · · , Xn , we have / ≥ 3. Varθ (θ) n Since the model is not a oneparameter exponential family, no ¯ is an unbiased estimator attains the lower bound. For example, X 2 ¯ unbiased estimator of θ with Varθ (X) = π /(3n) ≈ 3.29/n. The sample median is also unbiased; asymptotic theory predicts that its variance is approximately 4/n. It follows from standard asymptotic √ theory for MLEs that n(θ/n − θ) →d N (0, 3) for the MLE, which suggests that for large n, Varθ (θ/n ) ≈ 3/n. ✸ 6.5 Asymptotic eﬃciency Suppose that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) where θ is a realvalued parameter. In c 2000 by Chapman & Hall/CRC
Chapter 4, we showed that, subject to some regularity conditions, if θ/n is the MLE of θ then √ / n(θn − θ) →d N (0, 1/I(θ)) where
)
*
∂ I(θ) = Varθ ln f (X1 ; θ) . ∂θ This result suggests that for suﬃciently large n, Eθ (θ/n ) ≈ θ
and
Varθ (θ/n ) ≈
1 . nI(θ)
On the other hand if T is any statistic (based on X1 , · · · , Xn ) with Eθ (T ) = θ) then 1 Varθ (T ) ≥ . nI(θ) Juxtaposing the Cram´erRao lower bound with the asymptotic theory for MLEs (as developed in Chapter 5) raises the following question: If θ0n is a sequence of estimators with √ 0 n(θn − θ) →d N (0, σ 2 (θ)) then is σ 2 (θ) ≥ 1/I(θ) for all θ? The answer to the question is a qualiﬁed “yes” although we will show it is possible to ﬁnd estimators for which σ 2 (θ) < 1/I(θ) for some θ. What qualiﬁcations must be made in order to conclude that the asymptotic variance σ 2 (θ) ≥ 1/I(θ)? It will follow from results given below that if σ 2 (θ) is a continuous function of θ then σ 2 (θ) ≥ 1/I(θ). The following example shows that σ 2 (θ) < 1/I(θ) for some θ if σ 2 (θ) is not continuous. EXAMPLE 6.14: Suppose X1 , X2 , · · · , Xn are i.i.d. Normal random variables with mean θ and variance 1; for this model, I(θ) = 1 for all θ. Consider the estimator 6
θ0n =
¯ n  ≥ n−1/4 ¯ n if X X ¯ n if X ¯ n  < n−1/4 aX
where a is a constant with a < 1. We will show that N (0, σ 2 (θ)) where
σ 2 (θ) = c 2000 by Chapman & Hall/CRC
1 if θ = 0 ; a2 if θ = 0
√
n(θ0n −θ) →d
√ ¯ thus, σ 2 (0) < 1/I(0). To prove this, note that n(X n − θ) =d Z where Z ∼ N (0, 1). It follows then that √ 0 √ ¯ n − θ)I(X ¯ n  ≥ n−1/4 ) n(θn − θ) = n(X √ ¯ n − θ)I(X ¯ n  < n−1/4 ) + n(aX √ √ ¯ n − θ)I( nX ¯ n − θ + θ ≥ n1/4 ) n(X = √ √ ¯ n − θ)I( nX ¯ n − θ + θ < n1/4 ) + n(aX √ =d ZI(Z + nθ ≥ n1/4 ) √ √ +[aZ + nθ(a − 1)]I(Z + nθ < n1/4 ). √ √ Now note that Z + nθ ∼ N ( nθ, 1) and so I(Z +
√
nθ ≥ n
1/4
) →p
Thus we have ZI(Z +
√
nθ ≥ n1/4 ) →p
0 if θ = 0 1 if θ = 0. Z if θ = 0 0 if θ = 0
and also [aZ + Thus
√
nθ(a − 1)]I(Z + √
√
nθ < n1/4 ) →p
0 if θ = 0 aZ if θ = 0.
Z if θ = 0 aZ if θ = 0. This example was ﬁrst given by J.L. Hodges in 1952. n(θ0n − θ) →d
✸
Asymptotically Normal estimators whose limiting variance σ 2 (θ) satisﬁes 1 σ 2 (θ0 ) < I(θ0 ) for some θ0 are often called supereﬃcient estimators. Bahadur (1964) showed that (subject to some weak regularity conditions) the set of θ for which σ 2 (θ) < 1/I(θ) is at most countable. We can also deﬁne the notion of “regularity” of a sequence of estimators, an idea dating back to H´ ajek (1970). We say that a sequence of √ / estimators {θn } is regular at θ if, for θn = θ + c/ n, √ lim Pθn n(θ/n − θn ) ≤ x = Gθ (x) n→∞
where the limiting distribution function Gθ (x) can depend on θ but not c. It is possible to verify that the sequence of estimators c 2000 by Chapman & Hall/CRC
{θ/n } given in Example 6.14 is not regular at θ = 0. Virtually all “standard” estimators such as maximum likelihood and substitution principle estimators are regular. EXAMPLE 6.15: Suppose that X1 , X2 , · · · , Xn are i.i.d. Normal ¯ n . If random variables with mean θ and variance 1; deﬁne θ/n = X √ / θn = θ + c/ n is the true parameter value then θn has a Normal √ distribution with mean θn and variance 1/n; hence n(θ/n − θn ) has a standard Normal distribution for all n and so {θ/n } is regular (at any θ). ✸ EXAMPLE 6.16: Suppose that X1 , X2 , · · · , Xn are i.i.d. Expo/ n = 1/X ¯n. nential random√variables with parameter λ and deﬁne λ If λn = λ + c/ n, it is possible to show (using, for example, the Lyapunov Central Limit Theorem) that √ ¯ n − λ−1 ) ≤ x → Φ(λx) Pλn n(X n where Φ(·) is the standard Normal distribution function. Then using a “Delta Method”type argument, we get √ / n − λn ) ≤ x → Φ(x/λ) Pλn n(λ / n } is a regular sequence of estimators. and so {λ
✸
The following theorem gives a representation for regular estimators based on i.i.d. random variables in the case where the loglikelihood function can be approximated by a quadratic function in a neighbourhood of the true parameter θ. THEOREM 6.6 Let X1 , X2 , · · · , Xn be i.i.d. random variables with density or frequency function f (x; θ) and suppose that {θ/n } is a sequence of estimators that is regular at θ. If n ' i=1
√ ( 1 ln f (Xi ; θ + c/ n) − ln f (Xi ; θ) = cSn (θ) − c2 I(θ) + Rn (c, θ) 2
where Sn (θ) →d N (0, I(θ)) and Rn (c, θ) →p 0 for all c then √ / n(θn − θ) →d Z1 + Z2 where Z1 ∼ N (0, 1/I(θ)) and Z2 is independent of Z1 . √ In most cases, n(θ/n − θ) →d N (0, σ 2 (θ)) so that the random variable Z2 in Theorem 6.6 has a Normal distribution with mean c 2000 by Chapman & Hall/CRC
0 and variance σ 2 (θ) − 1/I(θ). If σ 2 (θ) = 1/I(θ) then the sequence {θ/n } is said to be asymptotically eﬃcient. Theorem 6.6 can also be extended to handle an estimator of a vector parameter θ in which case Z1 becomes a multivariate Normal random vector with variancecovariance matrix [I(θ)]−1 . It is important to note that, under the conditions Theorem √ of / / 6.6, the sequence of MLEs {θn } typically satisﬁes n(θn − θ) →d N (0, 1/I(θ)), which establishes the MLE as the most eﬃcient of all regular estimators. However, it is also important to recognize that there may be other regular estimators that, while not MLEs, have the same asymptotic variance as the MLE and, at the same time, have superior ﬁnite sample properties. The proof of Theorem 6.6 is quite technical and will not be given here. We will however sketch a proof of the result under somewhat stronger assumptions. In particular, we will assume that √ Eθ exp t1 n(θ/n − θ) + t2 Sn (θ) → m(t1 , t2 ) and that as n → ∞
√ Eθn exp( t1 n(θ/n − θn ) → m(t1 , 0) √ (where θn = θ + c/ n) for t1  ≤ b, t2  ≤ b where b > 0. We need to show that m(t1 , 0) is the product of two moment generating functions, one of which is the moment generating function √ of a N (0, 1/I(θ)) random variable. First of all, for θn = θ + c/ n, we have that √ Eθn exp t1 n(θ/n − θ) √ = exp(−t1 c)Eθn exp t1 n(θ/n − θn )
→ exp(−t1 c)m(t1 , 0). On the other hand, if we set Wn (θ, c) =
n '
√ ( ln f (Xi ; θ + c/ n) − ln f (Xi ; θ) ,
i=1
we also have
√ Eθn exp t1 n(θ/n − θ) √ = Eθ exp t1 n(θ/n − θ) + Wn (θ, c)
1 → m(t1 , c) exp − c2 I(θ) , 2 c 2000 by Chapman & Hall/CRC
which follows by substituting the approximately quadratic function (in c) for Wn (θ, c). Equating the two limits above, it follows that
1 m(t1 , 0) = m(t1 , c) exp −t1 c − c2 I(θ) 2 and setting c = −t1 /I(θ), we get
t21 m(t1 , 0) = m(t1 , −t1 /I(θ)) exp . 2I(θ) It is easy to verify that m(t 1 , −t1 /I(θ)) is a moment generating function and exp t21 /(2I(θ)) is the moment generating function of a N (0, 1/I(θ)) random variable. It should be noted that the rigorous proof of Theorem 6.6 is very similar to that given above except that characteristic functions are used and a few other technical diﬃculties must be addressed. Theorem 6.6 also holds in the multiparameter case where Z1 and Z2 are random vectors and I(θ) is a matrix. Verifying the regularity of a sequence of estimators is generally a tedious process. The assumption of regularity in Theorem 6.6 can be replaced by a more natural condition (Tierney, 1987), namely that √ lim Pθ n(θ/n − θ) ≤ x = Gθ (x) n→∞
where the limiting distribution function Gθ (x) is “continuous” in θ in the sense that ∞ h(x) dGθ (x) −∞
is a continuous function of θ for all bounded, continuous functions h(x). For example, if Gθ (x) is the N (0, σ 2 (θ)) distribution function and σ 2 (θ) is a continuous function of θ then this condition is satisﬁed. Sequences of estimators satisfying this “continuous limit” condition are typically regular and vice versa although exceptions in both directions can be found. It is important to view Theorem 6.6 in its proper context, that is, as an asymptotic optimality result for MLEs within a particular class of estimators, namely regular estimators. In particular, it is somewhat tempting to dismiss nonregular estimators as contrivances that would never be used in practice. While this is probably true of the estimator in Example 6.14, there are estimators used in practice (particularly in multiparameter problems) that c 2000 by Chapman & Hall/CRC
are similar in spirit to the Hodges estimator of Example 6.14; for example, in regression models, it is quite common to set certain parameter estimators to 0 if some preliminary estimator of that parameter falls below a speciﬁed threshold. 6.6 Problems and complements 6.1: Suppose that X = (X1 , · · · , Xn ) have joint density or frequency function f (x; θ) where θ is a realvalued parameter with a proper prior density function π(θ). For squared error loss, deﬁne the Bayes risk of an estimator θ/ = S(X): / θ) = RB (θ,
Θ
Eθ [(θ/ − θ)2 ]π(θ) dθ.
The Bayes estimator of minimizes the Bayes risk. (a) Show that the Bayes estimator is the mean of the posterior distribution of θ. (b) Suppose that the Bayes estimator in (a) is also an unbiased estimator. Show that the Bayes risk of this estimator must be 0. (This result implies that Bayes estimators and unbiased estimators agree only in pathological examples.) 6.2: Suppose that X ∼ Bin(n, θ) where θ has a Beta prior: π(θ)
Γ(α + β) α−1 (1 − θ)β−1 θ Γ(α)Γ(β)
for 0 < θ < 1. (a) Show that the Bayes estimator of θ under squared error loss is (X + α)/(α + β + n). (b) Find the mean square error of the Bayes estimator in (a). Compare the mean square error to that of the unbiased estimator X/n. 6.3: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean θ where θ has a Gamma (α , β ) prior distribution. (a) Show that
α + ni=1 Xi / θ= β+n is the Bayes estimator of θ under squared error loss. (b) Use the result of (a) to show that any estimator of the form c 2000 by Chapman & Hall/CRC
¯ + b for 0 < a < 1 and b > 0 is an admissible estimator of θ aX under squared error loss. 6.4: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean θ and variance 1 where θ has the improper prior density π(θ) = exp(θ) for −∞ < θ < ∞. (a) Find the posterior density of θ given X1 , · · · , Xn . (b) Find the posterior mean of θ. Is this estimator admissible under squared error loss? 6.5: Given a loss function L, we want to ﬁnd a minimax estimator of a parameter θ. (a) Suppose that θ/ is a Bayes estimator of θ for some prior distribution π(θ) with / = sup R (θ). / RB (θ) θ θ∈Θ
Show that θ/ is a minimax estimator. (The prior distribution π is called a least favourable distribution.) (b) Let {πn (θ)} be a sequence of prior density functions on Θ and suppose that {θ/n } are the corresponding Bayes estimators. If θ/0 is an estimator with sup Rθ (θ/0 ) = lim θ∈Θ
n→∞ Θ
Rθ (θ/n )π(θ) dθ,
show that θ/0 is a minimax estimator. (c) Suppose that X ∼ Bin(n, θ). Assuming squared error loss, ﬁnd a minimax estimator of θ. (Hint: Use a Beta prior as in Problem 6.2.) 6.6: The RaoBlackwell Theorem can be extended to convex loss functions. Let L(a, b) be a loss function that is convex in a for each ﬁxed b. Let S = S(X) be some estimator and suppose that T = T (X) is suﬃcient for θ. Show that Eθ [L(E(ST ), θ)] ≤ Eθ [L(S, θ)]. (Hint: Write Eθ [L(S, θ)] = Eθ [E[L(S, θ)T ]]. Then apply Jensen’s inequality to the conditional expected value.) 6.7: Suppose that X = (X1 , · · · , Xn ) are random variables with joint density or frequency function f (x; θ) and suppose that T = T (X) is suﬃcient for θ. c 2000 by Chapman & Hall/CRC
(a) Suppose that there exists no function φ(t) such that φ(T ) is an unbiased estimator of g(θ). Show that no unbiased estimator of g(θ) (based on X) exists. 6.8: Suppose that X has a Binomial distribution with parameters n and θ where θ is unknown. Consider unbiased estimators of g(θ) = θ2 (1 − θ)2 . (a) Show that no unbiased estimator of g(θ) exists if n ≤ 3. (b) Find the UMVU estimator of g(θ) when n ≥ 4. (Hint: Consider the case n = 4 ﬁrst.) 6.9: Suppose that X = (X1 , · · · , Xn ) have a joint distribution depending on a parameter θ where T = T (X) is suﬃcient for θ. (a) Prove Basu’s Theorem: If S = S(X) is an ancillary statistic and the suﬃcient statistic T is complete then T and S are independent. (Hint: It suﬃces to show that P (S ∈ AT ) = P (S ∈ A) for any set A; note that neither P (S ∈ AT ) nor P (S ∈ A) depends on θ. Use the completeness of T to argue that P (S ∈ AT ) = P (S ∈ A).) (b) Suppose that X and Y are independent Exponential random variables with parameter λ. Use Basu’s Theorem to show that X + Y and X/(X + Y ) are independent. (c) Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance σ 2 . Let T = T (X1 , · · · , Xn ) be a statistic such that T (X1 + a, X2 + a, · · · , Xn + a) = T (X1 , · · · , Xn ) + a and E(T ) = µ. Show that ¯ + E[(T − X) ¯ 2 ]. Var(T ) = Var(X) ¯ is ancillary for µ with σ 2 known.) (Hint: Show that T − X 6.10: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean µ and variance σ 2 , both unknown. We want to ﬁnd the UMVU estimator of
c−µ gc (µ, σ) = Φ σ
= Pµ,σ (Xi ≤ c)
for some speciﬁed c. ¯ 2 ) is suﬃcient and complete ¯ n (Xi − X) (a) State why (X, i=1 for (µ, σ). c 2000 by Chapman & Hall/CRC
¯ is independent of X ¯ for any i. (Hint: (b) Show that Xi − X It suﬃces to show (why?) that the covariance between the two r.v.’s is 0; alternatively, we could use Basu’s Theorem.) (c) Using the preliminary estimator S = I(X1 ≤ c), ﬁnd the UMVU estimator of gc (µ, σ) by “RaoBlackwellizing” S. 6.11: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ. (a) Use the fact that ∞
ck xk = 0
for all a < x < b
k=0
if, and only if, c0 = c1 = c2 = · · · = 0 to show that T = ni=1 Xi is complete for λ. (b) Find the unique UMVU estimator of λ2 . (Hint: Find g(0), g(1), g(2), · · · to solve ∞
g(k)
k=0
exp(−nλ)(nλ)k = λ2 k!
by multiplying both sides by exp(nλ) and matching the coeﬃcients of λk .) (c) Find the unique UMVU estimator of λr for any integer r > 2. 6.12: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. Deﬁne g(λ) = Pλ (Xi > t) for some speciﬁed t > 0. (a) Show that T = X1 + · · · + Xn is independent of X1 /T . (b) Find the UMVU estimator of g(λ). (Hint: “RaoBlackwellize” the unbiased estimator S = I(X1 > t) using the result of part (a).) 6.13: Suppose that X = (X1 , · · · , Xn ) has a joint distribution that depends on an unknown parameter θ and deﬁne U = {U : Eθ (U ) = 0, Eθ (U 2 ) < ∞} to be the space of all statistics U = U (X) that are unbiased estimators of 0 with ﬁnite variance. (a) Suppose that T = T (X) is an unbiased estimator of g(θ) with Varθ (T ) < ∞. Show that any unbiased estimator S of g(θ) c 2000 by Chapman & Hall/CRC
with Varθ (S) < ∞ can be written as S =T +U for some U ∈ U. (b) Let T be an unbiased estimator of g(θ) with Varθ (T ) < ∞. Suppose that Covθ (T, U ) = 0 for all U ∈ U (and all θ). Show that T is a UMVU estimator of g(θ). (Hint: Use the result of part (a).) (c) Suppose that T is a UMVU estimator of g(θ). Show that Covθ (T, U ) = 0 for all U ∈ U. (Hint: Let Sλ = T + λU for some U ∈ U and ﬁnd the minimum value of Varθ (Sλ ) for −∞ < λ < ∞.) 6.14: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean θ > 0 and variance θ2 . A (minimal) suﬃcient statistic for θ is T =
n i=1
Xi ,
n
Xi2
.
i=1
Show that T is not complete. 6.15: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with mean θ and variance θ2 where θ > 0. Deﬁne
n ¯ n )2 − nX ¯2 (X − X i n ¯ n 1 + i=1 θ/n = X ¯ n )2 3 n (Xi − X i=1
¯ n is the sample mean of X1 , · · · , Xn . where X (a) Show that θ/n →p θ as n → ∞. √ (b) Find the asymptotic distribution of n(θ/n − θ). Is θ/n asymptotically eﬃcient? (c) Find the Cram´erRao lower bound for unbiased estimators of θ. (Assume all regularity conditions are satisﬁed.) (d) Does there exist an unbiased estimator of θ that achieves the lower bound in (a)? Why or why not? 6.16: Suppose that X1 , · · · , Xn are independent random variables where the density function of Xi is 1 fi (x; β) = exp(−x/(βti )) for x ≥ 0 βti where t1 , · · · , tn are known constants. (Note that each Xi has an Exponential distribution.) c 2000 by Chapman & Hall/CRC
(a) Show that β/ =
n 1 Xi /ti n i=1
is an unbiased estimator of β. (b) Compute the Cram´erRao lower bound for the variance of unbiased estimators of β. Does the estimator in (a) achieve this lower bound? (Hint: Write the joint density as a oneparameter exponential family.) 6.17: Suppose that X1 , · · · , Xn are i.i.d. random variables with frequency function
f (x; θ) =
θ for x = −1 2 x (1 − θ) θ for x = 0, 1, 2, · · ·
where 0 < θ < 1. (a) Find the Cram´erRao lower bound for unbiased estimators of θ based on X1 , · · · , Xn . (b) Show that the maximum likelihood estimator of θ based on X1 , · · · , Xn is θ/n =
2
n
n
= −1) +
2n + ni=1 Xi
i=1 I(Xi
i=1 Xi
and show that {θ/n } is consistent for θ. √ (c) Show that n(θ/n − θ) →d N (0, σ 2 (θ)) and ﬁnd the value of σ 2 (θ). Compare σ 2 (θ) to the Cram´erRao lower bound found in part (a). 6.18: Suppose that X = (X1 , · · · , Xn ) are random variables with joint density or frequency function f (x; θ) where θ is a onedimensional parameter. Let T = T (X) be some statistic with Varθ (T ) < ∞ for all θ and suppose that (i) A = {x : f (x; θ) > 0} does not depend on θ (ii) Eθ (T ) = g(θ) Show that Varθ (T ) ≥
[g(θ + ∆) − g(θ)]2 Varθ (ψ(X; θ)
(provided that θ + ∆ lies in the parameter space) where ψ(x; θ) = c 2000 by Chapman & Hall/CRC
f (x; θ + ∆) − f (x; θ) . f (x; θ)
(This lower bound for Varθ (T ) is called the ChapmanRobbins lower bound.) 6.19: Suppose that X1 , · · · , Xn be i.i.d. Bernoulli random variables with parameter θ. (a) Indicate why S = X1 + · · · + Xn is a suﬃcient and complete statistic for θ. (b) Find the UMVU estimator of θ(1 − θ). (Hint: I(X1 = 0, X2 = 1) is an unbiased estimator of θ(1 − θ).) 6.20: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables /n to with mean µ and variance 1. Given λn > 0, deﬁne µ minimize n gn (t) =
(Xi − t)2 + λn t
i=1
(a) Show that /n = µ
¯ n  ≤ λn /(2n) 0 if X ¯ ¯ ¯ n  > λn /(2n). Xn − λn sgn(Xn )/(2n) if X
√ (b) Suppose that √ λn / n → λ0 > 0. Find the limiting /n − µ) for µ = 0 and µ = 0. Is {µ /n } a distribution of n(µ sequence of regular estimators in either the H´ajek or Tierney sense? 6.21: Suppose that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) satisfying the conditions of Theorem 6.6. Let θ/n be the MLE of θ and θ0n be another (regular) estimator of θ such that √
n
θ/n − θ θ0n − θ
→d N2 (0, C(θ)).
Show that C(θ) must have the form
C(θ) =
I −1 (θ) I −1 (θ) I −1 (θ) σ 2 (θ)
.
(Hint: Consider estimators of the form tθ/n + (1 − t)θ0n ; by Theorem 6.6, the minimum asymptotic variance must occur at t = 1.) c 2000 by Chapman & Hall/CRC
CHAPTER 7
Interval Estimation and Hypothesis Testing 7.1 Conﬁdence intervals and regions Suppose X1 , · · · , Xn are random variables with some joint distribution depending on a parameter θ that may be real or vectorvalued. To this point, we have dealt exclusively with the problem of ﬁnding point estimators of θ. The obvious problem with point estimation is the fact that typically Pθ (θ/ = θ) is small (if not 0) for a given point estimator / Of course, in practice, we usually attach to any point estimator θ. an estimator of its variability (for example, its standard error); however, this raises the question of exactly how to interpret such an estimate of variability. An alternative approach to estimation is interval estimation. Rather than estimating θ by a single statistic, we instead give a range of values for θ that we feel are consistent with observed values of X1 , · · · , Xn , in the sense, that these parameter values could have produced (with some degree of plausibility) the observed data. We will start by considering interval estimation for a single (that is, realvalued) parameter. DEFINITION. Let X = (X1 , · · · , Xn ) be random variables with joint distribution depending on a realvalued parameter θ and let L(X) < U (X) be two statistics. Then the (random) interval [L(X), U (X)] is called a 100p% conﬁdence interval for θ if Pθ [L(X) ≤ θ ≤ U (X)] ≥ p for all θ with equality for at least one value of θ. The number p is called the coverage probability (or simply coverage) or conﬁdence level of the conﬁdence interval. In many cases, we will be able to ﬁnd an interval [L(X), U (X)] with Pθ [L(X) ≤ θ ≤ U (X)] = p c 2000 by Chapman & Hall/CRC
for all θ. We can also deﬁne upper and lower conﬁdence bounds for θ. For example, suppose that Pθ [θ ≥ L(X)] = p for some statistic L(X) and for all θ; then L(X) is called a 100p% lower conﬁdence bound for θ. Likewise, if Pθ [θ ≤ U (X)] = p for some statistic U (X) and all θ then U (X) is called a 100p% upper conﬁdence bound for θ. It is easy to see that if L(X) is a 100p1 % lower conﬁdence bound and U (X) a 100p2 % upper conﬁdence bound for θ then the interval [L(X), U (X)] is a 100p% conﬁdence interval for θ where p = p1 + p2 − 1 (provided that L(X) < U (X)). The interpretation of conﬁdence intervals is frequently misunderstood. Much of the confusion stems from the fact that conﬁdence intervals are deﬁned in terms of the distribution of X = (X1 , · · · , Xn ) but, in practice, are stated in terms of the observed values of these random variables leaving the impression that a probability statement is being made about θ rather than about the random interval. However, given data X = x, the interval [L(x), U (x)] will either contain the true value of θ or not contain the true value of θ; under repeated sampling, 100p% of these intervals will contain the true value of θ. This distinction is important but poorly understood by many nonstatisticians. In many problems, it is diﬃcult or impossible to ﬁnd an exact conﬁdence interval; this is particularly true if a model is not completely speciﬁed. However, it may be possible to ﬁnd an interval [L(X), U (X)] for which Pθ [L(X) ≤ θ ≤ U (X)] ≈ p, in which case the resulting interval is called an approximate 100p% conﬁdence interval for θ. i.i.d. Normal EXAMPLE 7.1: Suppose that X1 , · · · , Xn are √ ¯ − µ) ∼ random variables with mean µ and variance 1. Then n(X N (0, 1) and so √ ' ( ¯ − µ) ≤ 1.96 = 0.95. Pµ −1.96 ≤ n(X c 2000 by Chapman & Hall/CRC
√ ¯ The event [−1.96√≤ n(X − µ) ≤ 1.96] is clearly the same as the ¯ − 1.96/ n ≤ µ ≤ X ¯ + 1.96/√n] and so we have event [X )
*
¯ − 1.96 ¯ + 1.96 √ ≤µ≤X √ = 0.95. Pµ X n n ¯ ± 1.96/√n is a 95% Thus the interval whose endpoints are X conﬁdence interval for µ. Note in this example, if we assume only that X1 , · · · , Xn are i.i.d. with mean µ and variance 1 (not necessarily normally distributed), we have (by the CLT), √ ' ( ¯ − µ) ≤ 1.96 ≈ 0.95 Pµ −1.96 ≤ n(X if n is suﬃciently large. Using the same argument used √ above, it ¯ follows that the interval whose endpoints are X ± 1.96/ n is an approximate 95% conﬁdence interval for µ. ✸ Pivotal method Example 7.1 illustrates a simple but useful approach to ﬁnding conﬁdence intervals; this approach is called the pivotal method. Suppose X = (X1 , · · · , Xn ) have a joint distribution depending on a realvalued parameter θ and let g(X; θ) be a random variable whose distribution does not depend on θ; that is, the distribution function Pθ [g(X; θ) ≤ x] = G(x) is independent of θ. Thus we can ﬁnd constants a and b such that p = Pθ [a ≤ g(X; θ) ≤ b] for all θ. The event [a ≤ g(X; θ) ≤ b] can (hopefully) be manipulated to yield p = Pθ [L(X) ≤ θ ≤ U (X)] and so the interval [L(X), U (X)] is a 100p% conﬁdence interval for θ. The random variable g(X; θ) is called a pivot for the parameter θ. EXAMPLE 7.2: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on [0, θ]. The MLE of θ is X(n) , the sample maximum, and the distribution function of X(n) /θ is G(x) = xn
for 0 ≤ x ≤ 1.
Thus X(n) /θ is a pivot for θ. To ﬁnd a 100p% conﬁdence interval c 2000 by Chapman & Hall/CRC
for θ, we need to ﬁnd a and b such that )
Pθ
*
X(n) a≤ ≤ b = p. θ
There are obviously inﬁnitely many choices for a and b; however, it can be shown that setting b = 1 and a = (1 − p)1/n results in the shortest possible conﬁdence interval using the pivot X(n) /θ, namely [X(n) , X(n) /(1 − p)1/n ]. ✸ EXAMPLE 7.3: Suppose that X1 , · · · , X10 are i.i.d. Exponential random variables with parameter λ. Then the random variable
λ 10 X i=1 i is a pivot for λ having a Gamma distribution with shape parameter 10 and scale parameter 1. (Alternatively, we can use
2 2λ 10 i=1 Xi , which has a χ distribution with 20 degrees of freedom, as our pivot.) To ﬁnd a 90% conﬁdence interval for λ, we need to ﬁnd a and b such that ,
Pλ a ≤ λ
10

Xi ≤ b = 0.90.
i=1
Again there are inﬁnitely many choices for a and b; one approach is to choose a and b so that ,
Pλ λ ,
Pλ λ
10 i=1 10

Xi < a
= 0.05
and

Xi > b
= 0.05;
i=1
this yields a = 5.425 and b = 15.705. We thus get ,
5.425 15.705 [L(X), U (X)] = 10 , 10 i=1 Xi i=1 Xi

as a 90% conﬁdence interval for λ. As one might expect, this conﬁdence interval is not the shortest possible based on the pivot used here; in fact, by using a = 4.893 and b = 14.938 we obtain the shortest possible 90% conﬁdence interval based on the pivot
10 λ i=1 Xi . ✸ It is easy to extend the pivotal method to allow us to ﬁnd conﬁdence intervals for a single realvalued parameter when there c 2000 by Chapman & Hall/CRC
are other unknown parameters. Suppose that θ = (θ1 , · · · , θp ) and consider ﬁnding a conﬁdence interval for θ1 . Let g(X; θ1 ) be a random variable that depends on θ1 but not on θ2 , · · · , θp and suppose that Pθ [g(X; θ1 ) ≤ x] = G(x) where G(x) is independent of θ; the random variable g(X; θ1 ) is then a pivot for θ1 and can be used to obtain a conﬁdence interval for θ1 in exactly the same way as before. EXAMPLE 7.4: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with unknown mean and variance µ and σ 2 , respectively. To ﬁnd a conﬁdence interval for µ, deﬁne S2 =
n 1 ¯ 2 (Xi − X) n − 1 i=1
and note that the random variable √ ¯ n(X − µ) S has Student’s t distribution with n − 1 degrees of freedom; this distribution is independent of both µ and σ 2 and hence √ ¯ − µ)/S n(X is a pivot for µ. To ﬁnd conﬁdence intervals for σ 2 , note that n (n − 1)S 2 1 ¯ 2 ∼ χ2 (n − 1) = (Xi − X) σ2 σ 2 i=1
and is therefore a pivot for σ 2 .
✸
In many problems, it is diﬃcult to ﬁnd exact pivots or to determine the distribution of an exact pivot if it does exist. However, in these cases, it is often possible to ﬁnd an approximate pivot, that is, a random variable g(X; θ) for which Pθ [g(X; θ) ≤ x] ≈ G(x) where G(x) is independent of θ; almost inevitably approximate pivots are justiﬁed via asymptotic arguments and so we assume that n is large enough to justify the approximation. In such cases, the approximate pivot can be used to ﬁnd approximate conﬁdence intervals for θ. The classic example of this occurs when we have a c 2000 by Chapman & Hall/CRC
point estimator θ/ whose distribution is approximately Normal with mean θ and variance σ 2 (θ); in this case, we have θ/ − θ ∼ N (0, 1) σ(θ) and so (θ/ − θ)/σ(θ) is an approximate pivot for θ. (Note that σ(θ) / If σ(θ) depends on θ, it is essentially the standard error of θ.) / / as may be desirable to substitute θ for θ and use (θ/ − θ)/σ(θ) the approximate pivot. This approximate pivot is particularly easy to use; if zp satisﬁes Φ(zp ) − Φ(−zp ) = p then ,
Pθ −zp ≤
θ/ − θ / σ(θ)

≤ zp ≈ p,
which yields an approximate 100p% conﬁdence interval whose / endpoints are θ/ ± zp σ(θ). EXAMPLE 7.5: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with unknown parameter λ. The MLE of λ is / = 1/X / − λ) is approximately ¯ and, if n is suﬃciently large, √n(λ λ Normal with mean 0 and variance λ2 (from asymptotic theory for MLEs). Thus the random variable √ / n(λ − λ) λ is an approximate pivot for λ and has a standard Normal distribution. To ﬁnd an approximate 95% conﬁdence interval for λ, we note that , √ / n(λ − λ) ≤ 1.96 ≈ 0.95 Pλ −1.96 ≤ λ and so the interval
√ √ / − 1.96/ n)−1 / + 1.96/ n)−1 , λ(1 λ(1
is 95% conﬁdence interval for λ. We can also use √ an/ approximate / as an approximate pivot; using the same argument n(λ − λ)/λ as before, we obtain an approximate 95% conﬁdence interval whose √ / / endpoints are λ ± 1.96λ/ n. The two conﬁdence intervals are quite similar when n is large since √ √ (1 − 1.96/ n)−1 = 1 + 1.96/ n + (1.96)2 /n + · · · √ √ and (1 + 1.96/ n)−1 = 1 − 1.96/ n + (1.96)2 /n + · · · ; c 2000 by Chapman & Hall/CRC
a more careful analysis is needed in order to determine which interval gives a coverage closer to 95%. ✸ EXAMPLE 7.6: Suppose that X has a Binomial distribution with parameters n and θ where θ is unknown. The MLE of θ is √ θ/ = X/n and if n is suﬃciently large, n(θ/ − θ)/[θ(1 − θ)]1/2 has approximately a standard Normal distribution and is an approximate pivot for θ. To ﬁnd an approximate 95% conﬁdence interval for θ, note that , , √ / n(θ − θ) n(θ/ − θ)2 2 Pθ −1.96 ≤ ≤ 1.96 = Pθ ≤ 1.96 θ(1 − θ) [θ(1 − θ)]1/2 ≈ 0.95. Thus an approximate 95% conﬁdence interval for θ will consist of all values of t for which g(t) = n(θ/ − t)2 − 1.962 t(1 − t) ≤ 0. Note that g(t) is a quadratic function and will have zeros at
t=
1/2
/ − θ) / + 1.962 /(4n2 ) θ/ + 1.962 /(2n) ± 1.96 θ(1
1 + 1.962 /n
and so the region between these two zeros becomes the approximate 95% conﬁdence interval for θ. Figure 7.1 shows the function g(t) and the conﬁdence interval for θ when n = 20 and X = 13. Alternatively, we can use √ / n(θ − θ) / − θ)] / 1/2 [θ(1
as an approximate pivot, which leads to an approximate 95% 1/2 . / − θ)/n] / conﬁdence interval whose endpoints are θ/ ± 1.96[θ(1 It is easy to see that the diﬀerence between these two conﬁdence intervals will be small when n is large. (We can also reﬁne the conﬁdence intervals in this example by using the continuity correction for the Normal approximation discussed in Chapter 3; / the result of this is that we replace θ/ by θ+1/(2n) in the expressions / / for the upper limits and replace θ by θ − 1/(2n) in the expressions for the lower limits.) ✸ c 2000 by Chapman & Hall/CRC
0.2 0.0 0.8
0.4
g(t)
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx
0.4
0.5
0.6
0.7
0.8
t
Figure 7.1 The function g(t) in Example 7.6; the approximate 95% conﬁdence interval for θ consists of the values of t for which g(t) ≤ 0.
It is possible to use variance stabilizing transformations (see Chapter 3) to obtain approximate pivots. For example, let θ/ be √ an estimator of θ and suppose that the distribution of n(θ/ − θ) is approximately Normal with mean 0 and variance σ 2 (θ). By the Delta Method, if g is a monotone function then the distribution of √ / − g(θ)) is approximately normal with mean 0 and variance n(g(θ) 2 [g (θ)] σ 2 (θ); we choose the variance stabilizing transformation g √ / − g(θ)) as the so that [g (θ)]2 σ 2 (θ) = 1. Now using n(g(θ) approximate pivot, we obtain an approximate 100p% conﬁdence √ / interval for g(θ) with endpoints g(θ) ± zp / n. Since we can take g to be strictly increasing, our approximate 100p% conﬁdence interval for θ is the interval
√ √ / − zp / n , g −1 g(θ) / + zp / n . g −1 g(θ)
It should be noted that it is not necessary to take g to be a variance c 2000 by Chapman & Hall/CRC
stabilizing transformation. In general, it is possible to use √ √ / − g(θ)) / − g(θ)) n(g(θ) n(g(θ) or / θ) / g (θ)σ(θ) g (θ)σ( as approximate pivots for θ. If the parameter space Θ is not the entire realline then it is sometimes useful to take g to be a function mapping Θ onto the realline. EXAMPLE 7.7: Suppose that X1 , · · · , Xn are i.i.d. Exponential / = 1/X, ¯ we random variables with parameter λ. For the MLE λ √ / have n(λ − λ) approximately Normal with mean 0 and variance λ2 . It is easy to verify that a variance stabilizing transformation is g(t) = ln(t); we get an approximate 95% conﬁdence interval for / ± 1.96/√n and so an approximate 95% ln(λ) with endpoints ln(λ) conﬁdence interval for λ is √ / √ / exp(−1.96/ n), λ λ exp(1.96/ n) . Note that the function g(t) = ln(t) maps the parameter space (the positive realnumbers) onto the realline; thus both endpoints of this conﬁdence interval always lie in the parameter space (unlike the conﬁdence intervals given in Example 7.5). ✸ Conﬁdence regions Up to this point, we have considered interval estimation only for realvalued parameters. However, it is often necessary to look at two or more parameters and hence extend the notion of conﬁdence intervals (for a single parameter) to conﬁdence regions for multiple parameters. DEFINITION. Let X = (X1 , · · · , Xn ) be random variables with joint distribution depending on a (possibly) vectorvalued parameter θ ∈ Θ and let R(X) be a subset of Θ depending on X. Then R(X) is called a 100p% conﬁdence region for θ if Pθ [θ ∈ R(X)] ≥ p for all θ with equality at least one value of θ ∈ Θ. Again it is important to keep in mind that it is R(X) that is random and not θ in the probability statement above. c 2000 by Chapman & Hall/CRC
Note that there is nothing in the deﬁnition above that dictates that R(X) be a contiguous set; thus it is possible to have a conﬁdence region that consists of two or more disjoint regions. Thus for a single parameter, the deﬁnition of a conﬁdence region is somewhat more general than that for a conﬁdence interval. However, in practice, conﬁdence regions are typically (but not always) contiguous sets. Many of the procedures for conﬁdence intervals extend mutatis mutandis to conﬁdence regions. For example, the pivotal method can be easily extended to derive conﬁdence regions. A random variable g(X; θ) is called a pivot (as before) if its distribution is independent of θ; if so, we have p = Pθ [a ≤ g(X; θ) ≤ b] = Pθ [θ ∈ R(X)] where R(x) = {θ : a ≤ g(x; θ) ≤ b}. EXAMPLE 7.8: Suppose that X 1 , · · · , X n are i.i.d. kvariate Normal random vectors with mean µ and variancecovariance matrix C (where we assume that C −1 exists). Unbiased estimators of µ and C are given by / = µ
n 1 Xi n i=1
C/ =
n 1 / / T, (X i − µ)(X i − µ) n − 1 i=1
which are simply the natural analogs of the unbiased estimators in the univariate case. (We have assumed here that the X i ’s and µ are column vectors.) To obtain a conﬁdence region for µ, we will use the pivot / µ) = / C; g(µ,
n(n − k) / −1 (µ / − µ)T C / − µ), (µ k(n − 1)
which (Johnson and Wichern, 1992) has an F distribution with k and n − k degrees of freedom. Let fp be the p quantile of the F(k, n − k) distribution. Then a 100p% conﬁdence region for µ is given by the set
/ = µ: / C) R(µ,
3
n(n − k) / −1 (µ / − µ)T C / − µ) ≤ fp . (µ k(n − 1)
c 2000 by Chapman & Hall/CRC
This conﬁdence region contains all the points lying within a certain ellipsoid in kdimensional Euclidean space. If n is large then we can obtain an approximate 100p% conﬁdence region for µ that will be valid for i.i.d. (not necessarily kvariate Normal) X i ’s having (ﬁnite) variancecovariance matrix C. In this case, / µ) = n(µ / −1 (µ / C; / − µ)T C / − µ) g ∗ (µ,
is approximately χ2 distributed with k degrees of freedom and so is an approximate pivot. If cp is the p quantile of the χ2 (k) distribution then an approximate 100p% conﬁdence region for µ is given by the set 8 9 / = µ : n(µ / −1 (µ / C) / − µ)T C / − µ) ≤ cp . R∗ (µ, The validity of √ this conﬁdence region depends on the fact that the / is / − µ) is approximately Nk (0, C) and that C distribution of n(µ a consistent estimator of C. ✸ While it is as conceptually simple to construct conﬁdence regions as it is to construct conﬁdence intervals, conﬁdence regions lose the ease of interpretation that conﬁdence intervals have as the dimension of the parameter increases; it is straightforward to graphically represent a conﬁdence region in two dimensions and feasible in three dimensions but for four or more dimensions, it is practical impossible to give a useful graphical representation of a conﬁdence region. One exception to this is when a conﬁdence region is a rectangle in the kdimensional space; unfortunately, such regions do not seem to arise naturally! However, it is possible to construct such regions by combining conﬁdence intervals for each of the parameters. Suppose that θ = (θ1 , · · · , θk ) and suppose that [Li (X), Ui (X)] is a 100pi % conﬁdence interval for θi . Now deﬁne R(X) = [L1 (X), U1 (X)] × · · · × [Lk (X), Uk (X)] = {θ : Li (X) ≤ θi ≤ Ui (X) for i = 1, · · · , k} . R(X) is a conﬁdence region for θ but the coverage of this region is unclear. However, it is possible to give a lower bound for the coverage of R(X). Using Bonferroni’s inequality (see Example 1.2), it follows that Pθ [θ ∈ R(X)] ≥ 1 −
k i=1
c 2000 by Chapman & Hall/CRC
P [θi ∈ / [Li (X), Ui (X)]]
= 1−
k
(1 − pi ).
i=1
This suggests a simple procedure for obtaining conservative conﬁdence regions. Given a desired coverage probability p, we construct conﬁdence intervals for each of the k parameters with coverage pi = 1 − (1 − p)/k. Then the resulting region R(X) (which is called a Bonferroni conﬁdence region) will have coverage of at least p. 7.2 Highest posterior density regions Conﬁdence intervals represent the classic frequentist approach to interval estimation. Within the Bayesian framework, posterior distributions provide a natural analog to frequentist conﬁdence intervals. Suppose that X = (X1 , · · · , Xn ) are random variables with joint density or frequency function f (x; θ) where θ is a realvalued parameter. We will take π(θ) to be a prior density for the parameter θ; recall that this density reﬂects the statistician’s beliefs about the parameter prior to observing the data. The statistician’s beliefs after observing X = x are reﬂected by the posterior density π(θ)f (x; θ) ; −∞ π(t)f (x; t) dt
π(θx) = . ∞
the posterior density of θ is proportional to the product of the prior density and the likelihood function. DEFINITION. Let π(θx) be a posterior density for θ on Θ ⊂ R. A region C = C(x) is called a highest posterior density (HPD) region . of content p if (a) C π(θx) dθ = p; (b) for any θ ∈ C and θ∗ ∈ C, we have π(θx) ≥ π(θ∗ x). HPD regions are not necessarily contiguous intervals; however, if the posterior density is unimodal (as is typically the case) then the HPD region will be an interval. It is important to note that HPD regions and conﬁdence intervals (or conﬁdence regions) are very diﬀerent notions derived from diﬀerent philosophies of statistical inference. More precisely, the c 2000 by Chapman & Hall/CRC
conﬁdence interval is constructed using the distribution of X: if C(X) = [L(X), U (X)] is the conﬁdence interval, we have Pθ (θ ∈ C(X)) = p
for all θ
where Pθ is the joint probability distribution of X for a given parameter value θ. On the other hand, a HPD region is constructed using the posterior distribution of θ given X = x: if C = C(x) is a HPD region, we have
C
π(θx) dθ = P (θ ∈ CX = x) = p.
Interestingly enough, in many cases, conﬁdence intervals and HPD regions show a remarkable agreement even for modest sample sizes. EXAMPLE 7.9: Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean θ and assume a Gamma prior density for θ: λα θα−1 exp(−λθ) π(θ) = for θ > 0 Γ(α) (λ and α are hyperparameters). The posterior density of θ is also Gamma: (n + λ)t+α θt+α+1 exp(−(n + λ)θ) π(θx) = for θ > 0 Γ(t + α) where t = x1 + · · · + xn . It is easy to verify that the posterior mode is θ/ = (t + α − 1)/(n + λ). What happens to the posterior density when n is large? First of all, note that the variance of the posterior distribution is (t+α)/(n+λ)2 , which tends to 0 as n tends to ∞; thus the posterior density becomes more and more concentrated around the posterior mode as n increases. We will deﬁne ψ(θ) = ln π(θx); making a / we Taylor series expansion of ψ(θ) around the posterior mode θ, have (for large n) / + (θ − θ)ψ / (θ) / + 1 (θ − θ) / 2 ψ (θ) / ψ(θ) ≈ ψ(θ) 2
/ − 1 (θ − θ) / 2 n+λ = ψ(θ) 2 θ/ / = 0 and ψ (θ) / = −(n + λ)/θ. / We also have since ψ (θ) / = (t + α) ln(n + λ) + (t + α − 1) ln(θ) / − (n − λ)θ/ − ln Γ(t + α) ψ(θ) c 2000 by Chapman & Hall/CRC
and (using Stirling’s approximation) ln Γ(t + α) = ln(t + α − 1) + ln Γ(t + α − 1) / = ln(t + α − 1) + ln Γ((n − λ)θ) 1 ≈ ln(t + α − 1) + ln(2π) 2 * ) 1 / / − (n − λ)θ. / ln((n − λ)θ) + (n − λ)θ − 2
Putting the pieces together, we get
n+λ 1 1 ψ(θ) ≈ − ln(2π) + ln 2 2 θ/
1 / 2 n+λ − (θ − θ) 2 θ/
and since π(θx) = exp(ψ(θ)),
π(θx) ≈
n+λ
1/2
2π θ/
)
exp −
n+λ 2θ/
* 2 / (θ − θ) .
Thus, for large n, the posterior distribution is approximately / + λ). This suggests, for Normal with mean θ/ and variance θ/(n example, that an approximate 95% HPD interval for θ is
θ/ θ/ − 1.96 n+λ
1/2
θ/ , θ/ + 1.96 n+λ
1/2 ;
note that this interval is virtually identical to the approximate 95% ¯/n since conﬁdence interval for θ whose endpoints are x ¯ ± 1.96 x θ/ ≈ x ¯ and n/(n + λ) ≈ 1 for large n. ✸ Example 7.9 illustrates that there may exist a connection (at least for large sample sizes) between conﬁdence intervals and HPD intervals. We will now try to formalize this connection. Suppose that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) where f (x; θ) satisﬁes the regularity conditions for asymptotic normality of MLEs given in Chapter 5. If θ has prior density π(θ) then the posterior density of θ is given by π(θ)Ln (θ) −∞ π(t)Ln (t) dt
π(θx) = . ∞ 1
where Ln (θ) = ni=1 f (xi ; θ). We want to try to show that the posterior density can be approximated by a Normal density when n is large; the key result needed to do this is the following. c 2000 by Chapman & Hall/CRC
PROPOSITION 7.1 (Laplace’s approximation) Deﬁne In =
∞
g(θ) exp(nh(θ)) dθ
−∞
/ where g and h are “smooth” functions and h(θ) is maximized at θ. Then In = I/n (1 + rn ) where / − I/n = g(θ)
1/2
2π / nh (θ)
/ exp(nh(θ))
and nrn  ≤ M < ∞. We have not precisely speciﬁed the conditions on the functions g and h necessary to give a rigorous proof of Laplace’s approximation. However, it is quite easy to give a heuristic proof by expanding h(θ) / Then in Taylor series around θ. / + 1 (θ − θ) / 2 h (θ) / h(θ) ≈ h(θ) 2 / < 0 since θ/ maximizes h. Then where h (θ) / In ≈ exp(nh(θ))
∞
−∞
g(θ) exp
n / /2 h (θ)(θ − θ) 2
dθ.
Laplace’s approximation follows if we make the change of variable / 1/2 (θ − θ) / and assume suﬃcient smoothness for g. s = [−nh (θ)] Laplace’s approximation is quite crude but will be suﬃcient for our purposes; we will use it to approximate the integral ∞
−∞
π(t)Ln (t) dt.
Note that we can write
,

n 1 ln f (xi ; θ) π(θ)Ln (θ) = π(θ) exp n n i=1
and so if θ/n maximizes Ln (θ), we have by Laplace’s approximation ∞ −∞
π(t)Ln (t) dt ≈ π(θ/n )
where Hn (θ) = − c 2000 by Chapman & Hall/CRC
2π
1/2
Hn (θ/n ) d2 ln Ln (θ). dθ2
exp(Ln (θ/n ))
Thus
π(θx) ≈
π(θ) π(θ/n )
Hn (θ/n ) 2π
1/2
exp ln Ln (θ) − ln Ln (θ/n ) .
Since the posterior is concentrated around θ/n for large n, we need only worry about values of θ close to θ/n ; if the prior density is continuous everywhere then π(θ)/π(θ/n ) ≈ 1 for θ close to θ/n . We also have (under the regularity conditions of Chapter 5) that 1 exp ln Ln (θ) − ln Ln (θ/n ) ≈ − (θ − θ/n )Hn (θ/n ), 2
which leads to the ﬁnal approximation
π(θx) ≈
Hn (θ/n ) 2π
1/2
,

Hn (θ/n ) exp − (θ − θ/n )2 . 2
This heuristic development suggests that the posterior density can be approximated by a Normal density with mean θ/n and variance 1/Hn (θ/n ) when n is suﬃciently large; note that Hn (θ/n ) is simply the observed information for θ (deﬁned in Chapter 5) and that 1/Hn (θ/n ) is an estimate of the variance of the maximum likelihood estimate θ/n . Thus the interval who endpoints are θ/n ±zp Hn (θ/n )−1/2 is an approximate 100p% HPD interval for θ (where zp satisﬁes Φ(zp ) − Φ(−zp ) = p); this is exactly the same as the approximate 100p% conﬁdence interval for θ based on the MLE. 7.3 Hypothesis testing Suppose that X = (X1 , · · · , Xn ) are random variables with joint distribution density or frequency function f (x; θ) for some θ ∈ Θ. Let Θ = Θ0 ∪Θ1 for two disjoint sets Θ0 and Θ1 ; given the outcome of X, we would like to decide if θ lies in Θ0 or Θ1 . In practice, Θ0 is typically taken to be a lower dimensional subspace of the parameter space Θ. Thus θ ∈ Θ0 represents a simpliﬁcation of the model in the sense that the model contains fewer parameters. EXAMPLE 7.10: Suppose that X1 , · · · , Xm and Y1 , · · · , Yn are independent random variables where Xi ∼ N (µ1 , σ 2 ) and Yi ∼ N (µ2 , σ 2 ). The parameter space is then Θ = {(µ1 , µ2 , σ) : −∞ < µ1 , µ2 < ∞, σ > 0}. c 2000 by Chapman & Hall/CRC
In many applications, it is of interest to determine whether the Xi ’s and Yi ’s have the same distribution (that is, µ1 = µ2 ), for example, when the Xi ’s and Yi ’s represent measurements from two diﬀerent groups. In this case, we can represent Θ0 as Θ0 = {(µ1 , µ2 , σ) : −∞ < µ1 = µ2 < ∞, σ > 0}. Note that the parameter space Θ is threedimensional while Θ0 is only twodimensional. ✸ Given X = (X1 , · · · , Xn ), we need to ﬁnd a rule (based on X) that determines if we decide that θ lies in Θ0 or in Θ1 . This rule essentially is a twovalued function φ and (without loss of generality) we will assume that φ can take the values 0 and 1; if φ(X) = 0 then we will decide that θ ∈ Θ0 while if φ(X) = 1, we will decide that θ ∈ Θ1 . The function φ will be called a test function. In many cases, φ will depend on X only through some realvalued statistic T = T (X), which we will call the test statistic for the test. It is unlikely that any given test function will be perfect. Thus for a given test function φ, we must examine the probability of making an erroneous decision as θ varies over Θ. If θ ∈ Θ0 then an error will occur if φ(X) = 1 and the probability of this error (called a type I error) is Pθ [φ(X) = 1] = Eθ [φ(X)]
(θ ∈ Θ0 ).
Likewise if θ ∈ Θ1 then an error will occur if φ(X) = 0 and the probability of this error (called a type II error) is Pθ [φ(X) = 0] = 1 − Eθ [φ(X)]
(θ ∈ Θ1 ).
It is tempting to try to ﬁnd a test function φ whose error probabilities are uniformly small over the parameter space. While in certain problems this is possible to do, it should be realized that there is necessarily a tradeoﬀ between the probabilities of error for θ ∈ Θ0 and θ ∈ Θ1 . For example, let φ1 and φ2 be two test functions and deﬁne R1 = {x : φ1 (x) = 1} and
R2 = {(x) : φ2 (x) = 1} where R1 ⊂ R2 ; note that this implies that Eθ [φ1 (X)] ≤ Eθ [φ2 (X)] c 2000 by Chapman & Hall/CRC
for all θ ∈ Θ and so 1 − Eθ [φ1 (X)] ≥ 1 − Eθ [φ2 (X)] . Hence by attempting to decrease the probability of error when θ ∈ Θ0 , we run the risk of increasing the probability of error for θ ∈ Θ1 . The classic approach to testing is to specify a test function φ(X) such that for some speciﬁed α > 0, Eθ [φ(X)] ≤ α
for all θ ∈ Θ0 .
The hypothesis that θ belongs to Θ0 is called the null hypothesis and will be denoted by H0 ; likewise, the hypothesis that θ lies in Θ1 is called the alternative hypothesis and will be denoted by H1 . The constant α given above is called the level (or size) of the test; if φ(X) = 1, we say that H0 is rejected at level α. The level of the test is thus the maximum probability of “rejection” of the null hypothesis H0 when H0 is true. This particular formulation gives us a reasonably welldeﬁned mathematical problem for ﬁnding a test function φ: for a given level α, we would like to ﬁnd a test function φ so that Eθ [φ(X)] is maximized for θ ∈ Θ1 . For a given test function φ, we deﬁne π(θ) = Eθ [φ(X)] to be the power of the test at θ; for a speciﬁed level α, we require π(θ) ≤ α for all θ ∈ Θ0 and so we are most interested in π(θ) for θ ∈ Θ1 . The rationale for the general procedure given above is as follows. Suppose that we test H0 versus H1 at level α where α is small and suppose that, given data X = x, φ(x) = 1. If H0 is true then this event is quite rare (it occurs with probability at most α) and so this gives us some evidence to believe that H0 is false (and hence that H1 is true). Of course, this “logic” assumes that the test is chosen so that Pθ (φ(X) = 1) is larger when H1 is true. Conversely, if φ(x) = 0 then the test is very much inconclusive; this may tell us that H0 is true or, alternatively, that H1 is true but that the test used does not have suﬃcient power to detect this. Since the dimension of Θ0 is typically lower than that of Θ (and so the model under H0 is simpler), this approach to testing protects us against choosing unnecessarily complicated models (since the probability of doing so is at most α when the simpler model holds) but, depending c 2000 by Chapman & Hall/CRC
on the power of the test, may prevent us from identifying more complicated models when such models are appropriate. To get some idea of how to ﬁnd “good” test functions, we will consider the simple case where the joint density or frequency function of X1 , · · · , Xn is either f0 or f1 where both f0 and f1 are known and depend on no unknown parameters. We will then test H0 : f = f0
versus
H1 : f = f1
at some speciﬁed level α; the null and alternative hypothesis are called simple in this case as they both consist of a single density or frequency function. The problem now becomes a straightforward optimization problem: ﬁnd a test function φ with level α to maximize the power under H1 . The following result, the NeymanPearson Lemma, is important because it suggests an important principle for ﬁnding “good” test functions. THEOREM 7.2 (NeymanPearson Lemma) Suppose that X = (X1 , · · · , Xn ) have joint density or frequency function f (x) where f is one of f0 or f1 and suppose we test H0 : f = f0
versus
H1 : f = f1 .
Then the test whose test function is
φ(X) =
1 if f1 (X) ≥ kf0 (X) 0 otherwise
(for some 0 < k < ∞) is a most powerful (MP) test of H0 versus H1 at level α = E0 [φ(X)]. Proof. In this proof, we will let P0 , P1 and E0 , E1 denote probability and expectation under H0 and H1 . It suﬃces to show that if ψ is any function with 0 ≤ ψ(x) ≤ 1 and E0 [ψ(X)] ≤ E0 [φ(X)] then
E1 [ψ(X)] ≤ E1 [φ(X)].
We will assume that f0 and f1 are density functions; the same proof carries over to frequency functions with obvious modiﬁcations. First of all, note that
f1 (x) − kf0 (x) c 2000 by Chapman & Hall/CRC
≥ 0 if φ(x) = 1 < 0 if φ(x) = 0.
Thus
ψ(x)(f1 (x) − kf0 (x)) ≤ φ(x)(f1 (x) − kf0 (x))
and so
··· ≤
ψ(x)(f1 (x) − kf0 (x)) dx ···
φ(x)(f1 (x) − kf0 (x)) dx
or rearranging terms,
···
(ψ(x) − φ(x))f1 (x) dx ≤ k
···
(ψ(x) − φ(x))f0 (x) dx.
The lefthand side above is simply E1 [ψ(X)] − E1 [φ(X)] and the righthand side is k(E0 [ψ(X)] − E0 [φ(X)]). Since E0 [ψ(X)] − E0 [φ(X)] ≤ 0, it follows that E1 [ψ(X)] − E1 [φ(X)] ≤ 0 and so φ(X) is the test function of an MP test of H0 versus H1 . The NeymanPearson Lemma essentially states that an optimal test statistic for testing H0 : f = f0 versus H1 : f = f1 is T (X) =
f1 (X) f0 (X)
and that for a given level α, we should reject the null hypothesis H0 is T (X) ≥ k where k is chosen so that the test has level α. However, note that the NeymanPearson Lemma as stated here does not guarantee the existence of an MP α level test but merely states that the test that rejects H0 for T (X) ≥ k will be an MP test for some level α. Moreover, the NeymanPearson Lemma does guarantee uniqueness of an MP test when one exists; indeed, there may be inﬁnitely many test functions having the same power as the MP test function prescribed by the NeymanPearson Lemma. A more general form of the NeymanPearson Lemma gives a solution to the following optimization problem: Suppose we want to maximize E1 [φ(X)] subject to the constraints E0 [φ(X)] = α
and
0 ≤ φ(X) ≤ 1.
The optimal φ is given by
1 if f1 (X) > kf0 (X)
φ(X) =
c 2000 by Chapman & Hall/CRC
c if f1 (X) = kf0 (X) 0 if f1 (X) < kf0 (X)
where k and 0 ≤ c ≤ 1 are chosen so that the constraints are satisﬁed. The function φ described above need not be a test function since it can possibly take values other than 0 and 1; however, in the case where the statistic T (X) =
f1 (X) f0 (X)
is a continuous random variable (which implies that X1 , · · · , Xn are continuous random variables), we can take the optimal φ to be either 0 or 1 for all possible values of X1 , · · · , Xn and so an MP test of H0 : f = f0 versus H1 : f = f1 exists for all levels α > 0. Moreover, even if φ(X) given above is not a test function, it may be possible to ﬁnd a valid test function φ∗ (X) such that Ek [φ∗ (X)] = Ek [φ(X)]
for k = 0, 1
(see Example 7.12 below). However, for a given level α, an MP test of H0 versus H1 need not exist if T (X) is discrete unless we are willing to consider socalled randomized tests. (Randomized tests are tests where, for some values of a test statistic, rejection or acceptance of H0 is decided by some external random mechanism that is independent of the data.) EXAMPLE 7.11: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ that is either λ0 or λ1 (where λ1 > λ0 ). We want to test H0 : λ = λ 0
versus
H1 : λ = λ1
at level α. For a given λ, the joint density of X = (X1 , · · · , Xn ) is
f (x; λ) = λ exp −λ n
n
xi
i=1
and we will base our test on the statistic ,

n f (X; λ1 ) n T = Xi , = (λ1 /λ0 ) exp (λ0 − λ1 ) f (X; λ0 ) i=1
rejecting H0 if T ≥ k where k is chosen so that the test has level
α. Note, however, that T is a decreasing function of S = ni=1 Xi (since λ0 − λ1 < 0) and so T ≥ k if, and only if, S ≤ k for some c 2000 by Chapman & Hall/CRC
constant k (which will depend on k). We can choose k so that Pλ0
, n

Xi ≤ k
= α.
i=1 in this For given values of λ0 and α, it is quite feasible to ﬁnd k problem. Under the null hypothesis, the test statistic S = ni=1 Xi has a Gamma distribution with parameters n and λ0 and 2λ0 S has a χ2 distribution with 2n degrees of freedom. For example, if n = 5, α = 0.01 and λ0 = 5 then the 0.01 quantile of a χ2 distribution with 10 degrees of freedom is 2.56 and so
k =
2.56 = 0.256. 2×5
Thus we would reject H0 at the 1% level if S ≤ 0.256.
✸
EXAMPLE 7.12: Suppose that X1 , · · · , Xn are i.i.d. Uniform random variables on the interval [0, θ] where θ is either θ0 or θ1 (where θ0 > θ1 ). We want to test H0 : θ = θ 0
versus
H1 : θ = θ 1
at level α. The joint density of X = (X1 , · · · , Xn ) is 1 I(max(x1 , · · · , xn ) ≤ θ) θn and an MP test of H0 versus H1 will be based on the test statistic f (x; θ) =
T =
f (X; θ1 ) = (θ0 /θ1 )n I(X(n) ≤ θ1 ). f (X; θ0 )
Note that T can take only two possible values, 0 or (θ0 /θ1 )n depending on whether X(n) is greater than θ1 or not. It follows then that the test that rejects H0 when X(n) ≤ θ1 will be an MP test of H0 versus H1 with level
α = Pθ0 X(n) ≤ θ1 =
θ1 θ0
n
and the power of this test under H1 is
Pθ1 X(n) ≤ θ1 = 1; note that this test will also be the MP test of for any level α > (θ1 /θ0 )n since its power is 1. If we want to ﬁnd the MP test for α < (θ1 /θ0 )n , the situation is more complicated since the c 2000 by Chapman & Hall/CRC
NeymanPearson Lemma does not tell us what to do. Nonetheless, intuition suggests that the appropriate test statistic is X(n) (which is suﬃcient for θ in this model) and that we should reject H0 for X(n) ≤ k where
Pθ0 X(n) ≤ k =
k θ0
n
= α.
Solving the equation above, we get k = θ0 α1/n and so the power of this test is
Pθ1 X(n) ≤ θ0 α
1/n
=
θ0 α1/n θ1
n
=α
θ0 θ1
n
.
To show that this is the MP test for α < (θ1 /θ0 )n , we use the more general form of the NeymanPearson Lemma; a function φ that maximizes Eθ1 [φ(X)] subject to the constraints
Eθ0 [φ(X)] = α <
is φ(X) = It is easy to verify that
θ1 θ0
n
and 0 ≤ φ(X) ≤ 1
α(θ0 /θ1 )n if X(n) ≤ θ1 0 otherwise.
Eθ1 [φ(X)] = α
θ0 θ1
n
= Pθ1 X(n) ≤ θ0 α1/n
and so the test that rejects H0 if X(n) ≤ θ0 α1/n is an MP test for level α < (θ1 /θ0 )n . ✸ Uniformly most powerful tests The NeymanPearson Lemma gives us a simple criterion for determining the test function of an MP test of a simple null hypothesis when the alternative hypothesis is also simple. However, as we noted before, this particular testing scenario is rarely applicable in practice as the alternative hypothesis (and often the null hypothesis) are usually composite. Let X = (X1 , · · · , Xn ) be random variables with joint density or frequency function f (x; θ) and suppose we want to test H0 : θ ∈ Θ0
versus
H1 : θ ∈ Θ1
at level α. Suppose that φ is a test function such that Eθ [φ(X)] ≤ α c 2000 by Chapman & Hall/CRC
for all θ ∈ Θ0
and that φ is the MP α level test of H0 : θ = θ0
versus
H1 : θ = θ1
for some θ0 ∈ Θ0 and all θ1 ∈ Θ1 ; then the test function φ describes the uniformly most powerful (UMP) α level test of H0 versus H1 . Unfortunately, UMP tests only exist for certain testing problems. For example, if we want to test H0 = θ = θ 0
versus
H1 : θ = θ0
then UMP tests typically do not exist. The reason for this is quite simple: in order for a given test to be UMP, it must be an MP test of H0 : θ = θ0 versus H1 : θ = θ1 for any θ1 = θ0 . However, the form of the MP test (given by the NeymanPearson Lemma) typically diﬀers for θ1 > θ0 and θ1 < θ0 as the following examples indicate. EXAMPLE 7.13: Let X be a Binomial random variable with parameters n and θ and suppose we want to test H0 : θ = θ 0
versus
H1 : θ = θ0
at some level α. First consider testing H0 : θ = θ0
versus
H1 : θ = θ1
where θ1 = θ0 . The NeymanPearson Lemma suggests that the MP test of H0 versus H1 will be based on the statistic T =
f (X; θ1 ) = f (X; θ0 )
1 − θ0 1 − θ1
n
θ1 (1 − θ0 ) θ0 (1 − θ1 )
X
.
If θ1 > θ0 , it is easy to verify that T is an increasing function of X; hence, an MP test of H0 versus H1 will reject H0 for large values of X. On the other hand, if θ1 < θ0 then T is a decreasing function of X and so an MP test will reject H0 for small values of X. From this, we can see that no UMP test of H0 versus H1 will exist. ✸ EXAMPLE 7.14: Let X1 , · · · , Xn be i.i.d. Exponential random variables with parameter λ and suppose that we want to test H0 : λ ≤ λ 0
versus
H1 : λ > λ 0
at level α. In Example 7.11, we saw that the MP α level test of H0 : λ = λ0 c 2000 by Chapman & Hall/CRC
versus
H1 : λ = λ1
rejects H0 for
n
i=1 Xi
≤ k when λ1 > λ0 where k is determined by Pλ0
, n

Xi ≤ k = α.
i=1
It is also easy to verify that if λ < λ0 then Pλ
, n

Xi ≤ k < α.
i=1
Thus the test that rejects H0 when ni=1 Xi ≤ k is a level α test and since it is an MP α level test of H0 versus H1 for every λ1 > λ0 , it is a UMP test of H0 versus H1 . ✸ Examples 7.13 and 7.14 give us some insight as to when UMP tests can exist; typically, we need the following conditions: • θ is a realvalued parameter (with no other unknown parameters), and • the testing problem is “onesided”; that is, we are testing H0 : θ ≤ θ 0
(θ ≥ θ0 )
versus
H1 : θ > θ 0
(θ < θ0 )
for some speciﬁed value θ0 . (In fact, UMP tests may also exist if H1 : θL < θ < θU for some speciﬁed θL and θU .) Suppose that X = (X1 , · · · , Xn ) are random variables with joint density or frequency function depending on some realvalued θ and suppose that we want to test H0 : θ ≤ θ 0
versus
H1 : θ > θ 0
at level α. We noted above that this testing setup is essentially necessary for the existence of a UMP test; however, a UMP test need not exist for a particular model. We would thus like to ﬁnd a suﬃcient condition for the existence of a UMP test of H0 versus H1 ; this is guaranteed if the family {f (x; θ) : θ ∈ Θ} has a property known as monotone likelihood ratio. DEFINITION. A family of joint density (frequency) functions {f (x; θ)} (where θ ∈ Θ ⊂ R) is said to have monotone likelihood ratio if there exists a realvalued function T (x) such that for any θ1 < θ2 , f (x; θ2 ) f (x; θ1 ) c 2000 by Chapman & Hall/CRC
is a nondecreasing function of T (x). (By the Factorization Criterion, the statistic T (X) is necessarily suﬃcient for θ.) If the family {f (x; θ)} has monotone likelihood ratio then the test statistic for the UMP test is T (X); for example, if we test H0 : θ ≤ θ0
versus
H1 : θ > θ 0
then the test that rejects H0 (that is, φ(X) = 1) if T (X) ≥ k will be a UMP α level test where α = Pθ0 [T (X) ≥ k] . Similarly, for H0 : θ ≥ θ0 , the test rejecting H0 if T (X) ≤ k is a UMP test. EXAMPLE 7.15: Suppose the joint density or frequency function of X = (X1 , · · · , Xn ) is a oneparameter exponential family f (x; θ) = exp [c(θ)T (x) − b(θ) + S(x)]
for x ∈ A
and assume (with loss of generality) that c(θ) is strictly increasing in θ. Then for θ1 < θ2 , we have f (x; θ2 ) = exp [(c(θ2 ) − c(θ1 ))T (x) + b(θ1 ) − b(θ2 )] , f (x; θ1 ) which is an increasing function of T (x) since c(θ2 )−c(θ1 ) > 0. Thus this oneparameter exponential family has monotone likelihood ratio and so if we test H0 : θ ≤ θ0 versus H1 : θ > θ0 , we would reject H0 for T (X) ≥ k where k is chosen so that Pθ0 [T (X) ≥ k] = α. ✸ Other most powerful tests How do we ﬁnd “good” tests if a UMP test does not exist? One approach is to ﬁnd the most powerful test among some restricted class of tests. One possible restriction that can be applied is unbiasedness: a test with test function φ is said to be an unbiased level α test if Eθ [φ(X)] ≤ α for all θ ∈ Θ0 and Eθ [φ(X)] ≥ α for all θ ∈ Θ1 . Essentially, unbiasedness of a test requires that the power of the test for θ ∈ Θ1 is greater than the level of the test. This would c 2000 by Chapman & Hall/CRC
seem to be a very reasonable criterion to expect of a test; however, in many situations, unbiased tests do not exist. When they do, it is sometimes possible to ﬁnd uniformly most powerful unbiased (UMPU) tests. Suppose we want to test H0 : θ = θ0 versus H1 : θ = θ0 at level α, it may be possible to construct a UMPU test by combining UMP tests of H0 : θ ≤ θ0 and H0 : θ ≥ θ0 . More precisely, suppose that φ1 (X) is a UMP level α1 test function of H0 and φ2 (X) is a UMP level α2 test function of H0 such that α1 +α2 = α. Then φ = φ1 +φ2 will be a level α test function provided that φ1 (X) + φ2 (X) ≤ 1 and by judiciously choosing α1 and α2 , it may be possible to make φ an unbiased α level test function; the resulting test will typically be a UMPU test. The natural choice for α1 and α2 is α1 = α2 = α/2; in general, however, this will not lead to an unbiased test as the following example indicates. EXAMPLE 7.16: Let X be a continuous random variable with density function f (x; θ) = θxθ−1
for 0 ≤ x ≤ 1
and suppose we want to test H0 : θ = 1
and H1 : θ = 1
at the 5% level. We will reject H0 if either X ≤ 0.025 or X ≥ 0.975; clearly this is a 5% level test since P1 (X ≤ 0.025) = P1 (X ≥ 0.975) = 0.025. The power function is then 0.025
π(θ) = 0
θxθ−1 dx +
1
θxθ−1 dx
0.975
= 1 + 0.025θ − 0.975θ . Evaluating π(θ) for θ close to 1 reveals that this test is not unbiased; in fact, π(θ) < 0.05 for 1 < θ < 2. However, it is possible to ﬁnd an unbiased 5% level test of H0 versus H1 . This test rejects H0 if either X ≤ 0.008521 or if X ≥ 0.958521. The power functions for both tests are shown in Figure 7.2. ✸ c 2000 by Chapman & Hall/CRC
0.15 0.10 0.0
0.05
power
0.5
1.0
1.5
2.0
2.5
theta
Figure 7.2 The power functions of the test in Example 7.16; the power function for the unbiased test is indicated by the dashed line.
In Example 7.16, the unbiased test has higher power for θ > 1 but lower power for θ < 1. This illustrates an important point, namely that in choosing an unbiased test, we are typically sacriﬁcing power in some region of the parameter space relative to other (biased) tests. The following example illustrates how to determine a UMPU test.
EXAMPLE 7.17: Let X1 , · · · , Xn be i.i.d. Exponential random variables with parameter λ and suppose we want to test H0 : λ = λ 0
versus
H1 : λ = λ0
at some level α. We know, of course, that we can ﬁnd UMP tests for the null hypotheses H0 : λ ≤ λ0 using the test statistic T = c 2000 by Chapman & Hall/CRC
and H0 : λ ≥ λ0
n
i=1 Xi ;
we will reject H0 for T ≤ k1
and H0 for T ≥ k2 . If k1 < k2 , we can deﬁne the test function
φ(T ) =
1 if T ≤ k1 or T ≥ k2 , 0 otherwise
which will give a level α test if Pλ0 [k1 < T < k2 ] = 1 − α. We would now like to choose k1 and k2 to make this test unbiased. The power of the test is given by π(λ) = 1 − Pλ [k1 < T < k2 ] k2 n n−1 λ x exp(−λx) = 1− dx (n − 1)! k1 since T has a Gamma distribution with parameters n and λ. For given k1 and k2 , the power function π(λ) is diﬀerentiable with derivative k2 n (λ x − nλn−1 )xn−1 exp(−λx) π (λ) = dx (n − 1)! k1 and it is easy to see that the requirement of unbiasedness is equivalent to requiring that π (λ0 ) = 0. Thus k1 and k2 must satisfy the equations 1−α = and 0 =
k2 1 λn xn−1 exp(−λ0 x) dx (n − 1)! k1 0 k2 1 (λn x − nλn−1 )xn−1 exp(−λ0 x) dx. 0 (n − 1)! k1 0
(The equations given here for k1 and k2 are similar to the equations needed to obtain the shortest conﬁdence interval in Example 7.3.) For example, if n = 5, λ0 = 3, and α = 0.05, we obtain k1 = 0.405 and k2 = 3.146. This test turns out to be the UMPU 5% level test of H0 versus H1 . ✸ UMPU tests also exist for certain tests in kparameter exponential families. Suppose X = (X1 , · · · , Xn ) are random variables with joint density or frequency function ,
f (x; θ, η) = exp θT0 (x) +
k−1

ηi Ti (x) − d(θ, η) + S(x)
i=1
where θ is realvalued and η is a vector of “nuisance” parameters; c 2000 by Chapman & Hall/CRC
we want to test H0 : θ ≤ θ 0
versus
H1 : θ > θ 0
for some speciﬁed value θ0 where the value of η is arbitrary. It turns out that a UMPU α level test of H0 versus H1 will reject H0 if T1 (X) ≥ c where α = Pθ0 [T0 (X) ≥ cT1 = t1 , · · · , Tk−1 = tk−1 ] where t1 , · · · , tk−1 are the observed values of T1 , · · · , Tk−1 ; that is, ti = Ti (x) where x is the observed value of X. Thus the test function is based not on the marginal distribution of the statistic T0 but rather the conditional distribution of T0 given the observed values of the other suﬃcient statistics T1 , · · · , Tk−1 . See Lehmann (1991) for more details. In the case where θ is realvalued, it is often possible to ﬁnd locally most powerful (LMP) tests. For example, suppose that we want to test H0 : θ ≤ θ0
versus
H1 : θ > θ 0 .
Our intuition tells us that if the true value of θ is suﬃciently far from θ0 , any reasonable test of H0 versus H1 will have power close to 1. Thus we should be concerned mainly about ﬁnding a test whose power is as large as possible for values of θ close to θ0 . Let φ(X) be a test function and deﬁne the power function π(θ) = Eθ [φ(X)] ; we will require that π(θ0 ) = α. We would like to ﬁnd φ so that π(θ) is maximized for θ close to θ0 . As a mathematical problem, this is somewhat ambiguous; however, if θ is very close to θ0 then we have π(θ) ≈ π(θ0 ) + π (θ0 )(θ − θ0 ) = α + π (θ0 )(θ − θ0 ) (provided, of course, that π(θ) is diﬀerentiable). This suggests that we try to ﬁnd a test function φ to maximize π (θ0 ) subject to the constraint π(θ0 ) = α. (In the case where H0 : θ ≥ θ0 and H1 : θ < θ0 , we would want to minimize π (θ0 ).) Suppose that f (x; θ) is the joint density function of continuous random variables X = (X1 , · · · , Xn ). Then for a given test function c 2000 by Chapman & Hall/CRC
φ, we have
···
π(θ) =
φ(x)f (x; θ) dx
and if we can diﬀerentiate under the integral sign, we have
∂ f (x; θ) dx ∂θ
∂ · · · φ(x) = ln f (x; θ) f (x; θ) dx ∂θ )
* ∂ = Eθ φ(X) ln f (X; θ) . ∂θ
π(θ) =
···
φ(x)
(The same result holds if X1 , · · · , Xn are discrete provided we can diﬀerentiate inside the summation sign.) The development given above suggests that to ﬁnd the form of the LMP test we must solve the following optimization problem: maximize Eθ0 [φ(X)S(X; θ0 )] subject to the constraint Eθ0 [φ(X)] = α where
∂ ln f (x; θ). ∂θ The following result provides a solution to this problem. S(x; θ) =
PROPOSITION 7.3 Suppose that X = (X1 , · · · , Xn ) has joint density or frequency function f (x; θ) and deﬁne the test function
φ(X) =
1 if S(X; θ0 ) ≥ k 0 otherwise
where k is such that Eθ0 [φ(X)] = α. Then φ(X) maximizes Eθ0 [ψ(X)S(X; θ0 )] over all test functions ψ with Eθ0 [ψ(X)] = α. Proof. This proof parallels the proof of the NeymanPearson Lemma. Suppose ψ(x) is a function with 0 ≤ ψ(x) ≤ 1 for all x such that Eθ0 [ψ(X)] = α. Note that
φ(X) − ψ(X) c 2000 by Chapman & Hall/CRC
≥ 0 if S(X; θ0 ) ≥ k ≤ 0 if S(X; θ0 ) ≤ k
and so
Eθ0 [(φ(X) − ψ(X))(S(X; θ0 ) − k)] ≥ 0. Since Eθ0 [φ(X) − ψ(X)] = 0, it follows that Eθ0 [φ(X)S(X; θ0 )] ≥ Eθ0 [ψ(X)S(X; θ0 )] , which completes the proof. The test described in Proposition 7.3 is often called the score test since it is based on the score function (see Chapter 5); we will discuss a more general form of the score test in section 7.5. In deriving this LMP test, we have been concerned about the behaviour of the power function only for values of θ near θ0 ; it is, in fact, conceivable that this LMP test function does not result in a level α test since Eθ [φ(X)] may be greater than α for some θ < θ0 . In the case where X1 , · · · , Xn are i.i.d. with common density or frequency function f (x; θ) then S(X; θ) =
n
: (Xi ; θ)
i=1
where : (x; θ) is the partial derivative with respect to θ of ln f (x; θ). Subject to the regularity conditions given in Chapter 5, we have '
(
Eθ0 : (Xi ; θ0 ) = 0
'
(
Varθ0 : (Xi ; θ0 ) = I(θ)
and
and so, if n is suﬃciently large, the distribution (under θ = θ0 ) of the test statistic S(X; θ0 ) will be approximately Normal with mean 0 and variance nI(θ). Thus the critical values for the LMP test of H0 : θ ≤ θ0 versus H1 : θ > θ0 can be determined approximately from this result. EXAMPLE 7.18: Let X1 , · · · , Xn be i.i.d. random variables with a Cauchy distribution with density 1 f (x; θ) = π(1 + (x − θ)2 and suppose that we want to test H0 : θ ≥ 0
versus
H1 : θ < 0
at the 5% level. We then have S(X; 0) =
n i=1
c 2000 by Chapman & Hall/CRC
2Xi 1 + Xi2
and the LMP test of H0 versus H1 rejects H0 if S(X; 0) ≤ k where k is chosen so that P0 [S(X; 0) ≤ k] = 0.05. The exact distribution is diﬃcult to obtain; however, since each of the summands of S(X; 0) has mean 0 and variance 1/2, the distribution of S(X; 0) (for θ = 0) will be approximately Normal with mean0 and variance n/2 if n is suﬃciently large. Thus we will have k 2/n ≈ −1.645 and so we can take k = −1.645 n/2 to obtain a test whose level is approximately 5%. ✸ 7.4 Likelihood ratio tests Our discussion of UMP and LMP tests has involved only one parameter models since these type of optimal tests do not generally exist for models with more than one parameter. It is therefore desirable to develop a general purpose method for developing reasonable test procedures for more general situations. We return to our general hypothesis testing problem where X = (X1 , · · · , Xn ) has a joint density or frequency function f (x; θ) where θ ∈ Θ with Θ = Θ0 ∪ Θ1 ; we wish to test H0 : θ ∈ Θ0
versus
H1 : θ ∈ Θ1 .
Earlier in this chapter, we saw that the NeymanPearson Lemma was useful in oneparameter problems for ﬁnding various optimal tests based on the ratio of joint density (frequency) functions. In the more general testing problem, we will use the NeymanPearson paradigm along with maximum likelihood estimation to give us a general purpose testing procedure. DEFINITION. The likelihood ratio (LR) statistic Λ is deﬁned to be supθ∈Θ f (X; θ) supθ∈Θ L(θ) Λ= = supθ∈Θ0 f (X; θ) supθ∈Θ0 L(θ) where L(θ) is the likelihood function. A likelihood ratio test of H0 : θ ∈ Θ0 versus H1 : θ ∈ Θ1 will reject H0 for large values of Λ. In our deﬁnition of Λ, we take the supremum of L(θ) = f (X; θ) over θ ∈ Θ rather than θ ∈ Θ1 in the numerator of Λ. We do this mainly for convenience; as mentioned earlier, Θ0 is often a c 2000 by Chapman & Hall/CRC
lower dimensional subspace of Θ, which makes the calculation of the MLE under H0 typically no more diﬃcult than the calculation of the unrestricted MLE. In order to use LR tests, we must know (either exactly or approximately) the distribution of the statistic Λ when H0 is true. In some cases, Λ is a function of some other statistic T whose distribution is known; in such cases, we can use T (rather than Λ) as our test statistic. In other cases, we can approximate the distribution of Λ (or some function of Λ) by a standard distribution such as the χ2 distribution. EXAMPLE 7.19: Let X1 , · · · , Xn be i.i.d. Normal random variables with mean µ and variance σ 2 (both unknown) and suppose that we want to test H0 : µ = µ0
versus
H1 : µ = µ0 .
In general, the MLEs of µ and σ 2 are ¯ /=X µ
/2 = and σ
n 1 ¯ 2 (Xi − X) n i=1
while under H0 the MLE of σ 2 is /02 σ
n 1 = (Xi − µ0 )2 . n i=1
Substituting the respective MLEs yields
Λ=
/02 σ /2 σ
n/2
and so the LR test will reject H0 when Λ ≥ k where k is chosen so the level of the test is some speciﬁed α. The distribution of Λ is /02 /σ /2 not obvious; however, note that Λ is a monotone function of σ and
n 2 /02 σ i=1 (Xi − µ0 )
= n ¯ 2 /2 σ i=1 (Xi − X) ¯ − µ0 )2 n(X = 1 + n ¯ 2 i=1 (Xi − X) ¯ − µ0 )2 n(X 1 = 1+ n−1 S2 = 1+ c 2000 by Chapman & Hall/CRC
T2 n−1
where S2 =
n 1 ¯ 2 (Xi − X) n − 1 i=1
and T =
√
¯ − µ0 )/S. n(X
From this, we can conclude that Λ is a monotone function of T 2 and we know (from Chapter 2) that, when H0 is true, T has Student’s t distribution with (n − 1) degrees of freedom and so T 2 has an F distribution with 1 and (n − 1) degrees of freedom. ✸ EXAMPLE 7.20: Let X1 , · · · , Xm be i.i.d. Exponential random variables with parameter λ and Y1 , · · · , Yn be i.i.d. random variables with parameter θ; we also assume that Xi ’s are independent of the Yi ’s. Suppose we want to test H0 : λ = θ
versus
H1 : λ = θ
The (unrestricted) MLEs of λ and θ are / = 1/X ¯ λ
and θ/ = 1/Y¯
while the MLEs under H0 are , / 0 = θ/0 = λ
¯ + nY¯ mX n+m
−1
.
Substituting these MLEs, we obtain the LR statistic
Λ=
m n Y¯ + ¯ m+n m+nX
m
¯ n m X + ¯ m+n m+nY
n
.
¯ Y¯ and we can make Clearly, Λ depends only on the statistic T = X/ Λ large by making T large or T small. Moreover, it is quite easy to see that, when H0 is true, T has an F distribution with 2m and 2n degrees of freedom and so, for a given value of α, it is quite simple to base a test of H0 versus H1 on the statistic T . ✸ Asymptotic distribution of the LR statistic As we mentioned previously, we can often approximate the distribution of the LR test statistic. In this section, we will assume that Θ is an open subset of Rp and the H0 parameter space Θ0 is eic 2000 by Chapman & Hall/CRC
ther a single point or an open subset of Rs where s < p. We will concentrate on the case where X1 , · · · , Xn are i.i.d. random variables although many of the results will hold under more general conditions. First of all, suppose that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) where the parameter space Θ is an open subset of the real line. We will consider the testing H0 : θ = θ0 and H1 : θ = θ0 . In this case, the LR statistic is simply Λn =
n f (Xi ; θ/n ) i=1
f (Xi ; θ0 )
where θ/n is the MLE of θ. Assuming the regularity conditions in Chapter 5, we obtain the following result. THEOREM 7.4 Suppose that X1 , · · · , Xn are i.i.d. random variables with a density or frequency function satisfying conditions A1 to A6 in Chapter 5 with I(θ) = J(θ). If the MLE θ/n satisﬁes √ / n(θn − θ) →d N (0, 1/I(θ) then the LR statistic Λn for testing H0 : θ = θ0 satisﬁes 2 ln(Λn ) →d V ∼ χ2 (1) when H0 is true. Proof. Let :(x; θ) = ln f (x; θ) and : (x; θ), : (x; θ) be its derivatives with respect to θ. Under the conditions of the theorem, √ / n(θn − θ0 ) →d N (0, 1/I(θ0 )) when H0 is true. Taking logarithms and doing a Taylor series expansion, we get ln(Λn ) =
n
[:(Xi ; θ/n ) − :(Xi ; θ0 )]
i=1
= (θ0 − θ/n )
n
n 1 : (Xi ; θ/n ) − (θ/n − θ0 )2 : (Xi ; θn∗ ) 2 i=1 i=1
n 1 1 : (Xi ; θn∗ ) = − n(θ/n − θ0 )2 2 n i=1 c 2000 by Chapman & Hall/CRC
where θn∗ lies between θ0 and θ/n . Now under assumptions A5 and A6 of Chapter 5, it follows that when H0 is true n 1 : (Xi ; θn∗ ) →p −Eθ0 [: (Xi ; θ0 )] = I(θ0 ). n i=1
Since
V I(θ0 ) the conclusion follows by applying Slutsky’s Theorem. n(θ/n − θ0 )2 →d
Henceforth, we will refer to both Λn and 2 ln(Λn ) as LR statistics depending on the situation; there is no real ambiguity in doing so since they are equivalent from a hypothesis testing viewpoint. Theorem 7.4 can be extended fairly easily to the multiparameter case. Let θ = (θ1 , · · · , θp ) and consider testing H0 : θ1 = θ10 , · · · , θr = θr0 where r ≤ p for some speciﬁed θ10 , · · · , θr0 . The LR statistic is deﬁned by n /n) f (Xi ; θ Λn = / i=1 f (Xi ; θ n0 ) / n0 is the MLE of θ under H0 (and thus whose ﬁrst r where θ components are θ10 , · · · , θr0 ). Again we will assume the regularity conditions B1 to B6 in Chapter 5 to obtain the following theorem.
THEOREM 7.5 Suppose that X1 , · · · , Xn are i.i.d. random variables with a density or frequency function satisfying conditions B1 to B6 in Chapter 5 with I(θ) = J(θ) where θ = (θ1 , · · · , θp ). If the / n satisﬁes √n(θ / n −θ) → N (0, I −1 (θ)) then the LR statistic MLE θ d Λn for testing H0 : θ1 = θ10 , · · · , θr = θr0 satisﬁes 2 ln(Λn ) →d V ∼ χ2 (r) when H0 is true. The proof of Theorem 7.5 will be left as an exercise. However, note that when r = p, the proof of Theorem 7.4 can be easily adapted to give a proof of Theorem 7.5. In the general case, the result can be deduced from the fact that the loglikelihood function can be approximated by a quadratic function in a neighbourhood of the true parameter value; this quadratic approximation is discussed brieﬂy below. Theorem 7.5 can be applied to testing null hypotheses of the c 2000 by Chapman & Hall/CRC
form H0 : g1 (θ) = a1 , · · · , gr (θ) = ar for some realvalued functions g1 , · · · , gr . To see this, deﬁne parameters φ1 , · · · , φp such that φk = gk (θ) where g1 , · · · , gr are as given in H0 and gr+1 , · · · , gp are deﬁned so that the vectorvalued function g(θ) = (g1 (θ), · · · , gp (θ)) is a onetoone function. Then provided that this function is diﬀerentiable, Theorem 7.5 can be applied using the parameters φ1 , · · · , φp . EXAMPLE 7.21: Suppose that (X1 , Y1 ), · · · , (Xn , Yn ) are i.i.d. pairs of continuous random variables with the joint density function of (Xi , Yi ) given by f (x, y; θ, λ, α) =
2θλα (θx + λy + α)3
for x, y > 0
where θ, λ, α > 0. The marginal densities of Xi and Yi are fX (x; θ, α) = and fY (y; λ, α) =
θα (θx + α)2 λα (λy + α)2
for x > 0 for y > 0,
which are equal if θ = λ. Thus, we may be interested in testing H0 : θ = λ. We can reparametrize in a number of ways. For example, we could deﬁne η1 = θ−λ, η2 = θ, and η3 = α, or alternatively, η1 = θ/λ with η2 , η3 deﬁned as before. With either reparametrization, we could express H0 in terms of η1 and so we would expect our likelihood test to have an asymptotic χ2 distribution with one degree of freedom. ✸ Other likelihood based tests While LR tests are motivated by the NeymanPearson paradigm, there are in fact other commonly used tests based on the likelihood function. Let X1 , · · · , Xn be i.i.d. random variables with density or frequency function f (x; θ) where θ = (θ1 , · · · , θp ) and suppose that we want to test the null hypothesis H0 : θ1 = θ10 , · · · , θr = θr0 . c 2000 by Chapman & Hall/CRC
To make the notation more compact, we will set φ = (θ1 , · · · , θr ) and τ = (θr+1 , · · · , θp ) so that θ = (φ, τ ); thus, H0 becomes H0 : φ = φ0 where φ0 = (θ10 , · · · , θr0 ). The Wald test of H0 compares the (unrestricted) MLE of φ to its value under the null hypothesis; if the distance between the two is large, this might indicate that H0 is false and so our test should / is the MLE (based on X1 , · · · , Xn ) then if H0 is reﬂect this. If φ n true, we have √ / n(φn − φ0 ) →d Nr (0, C(φ0 , τ )) where the variancecovariance matrix C(φ0 , τ ) can be obtained from the Fisher information matrix
I11 (φ0 , τ ) I12 (φ0 , τ ) I(φ0 , τ ) = I21 (φ0 , τ ) I22 (φ0 , τ ) by
−1
−1 (φ0 , τ )I12 (φ0 , τ ) C(φ0 , τ ) = I11 (φ0 , τ ) − I21 (φ0 , τ )I22
.
The Wald test statistic is / − φ )T C / −φ ) /n (φ Wn = n(φ n 0 n 0
where C/n is some estimator of C(φ0 , τ ) that is consistent under H0 . There are several possibilities for C/n ; for example, we could / , τ/ n ) (or C(φ , τ/ n )) or we could set C /n equal to set C/n = C(φ n 0 2 the observed information matrix. Under H0 , Wn →d χ (r) and we reject H0 for large values of this statistic. The score test (or Lagrange multiplier test as it is called by econometricians) uses the fact that if the null hypothesis is false then the gradient of the loglikelihood function should not be close to the 0 vector. To be more precise, let S i (φ, τ ) be the gradient of ln f (Xi ; φ, τ ) with respect to φ. Then under H0 , we have (subject to the regularity conditions of Chapter 5), n 1 √ S i (φ0 , τ/ n0 ) →d Nr (0, I11 (φ0 , τ )) n i=1
where τ/ n0 is the MLE of τ under H0 . The score statistic is
T
n 1 S i (φ0 , τ/ n0 ) Sn = n i=1 c 2000 by Chapman & Hall/CRC
−1 I11 (φ0 , τ/ n0 )
n i=1
S i (φ0 , τ/ n0 )
0 10 20 30 40
Scaled loglikelihood
3.5
4.0
4.5
5.0
5.5
6.0
lambda
Figure 7.3 The loglikelihood (multiplied by two and rescaled to have a maximum of 0) in Example 7.22; for testing H0 : λ = 4, the LR statistic is the length of the vertical line, the Wald statistic is proportional to the square of the length of the horizontal line while the score statistic is proportional to the square of the slope of the tangent line at λ = 4.
As for the Wald statistic, we reject H0 for large values of Sn and, under H0 , we have Sn →d χ2 (r). EXAMPLE 7.22: We will give an illustration of the three tests in a very simple setting. Suppose that X1 , · · · , Xn are i.i.d. Poisson random variables with mean λ and we want to test the null hypothesis H0 : λ = λ0 versus the alternative hypothesis H1 : λ = λ0 . For the Poisson distribution, we have I(λ) = 1/λ and the MLE /n = X ¯ n . Therefore, the LR, Wald, and score statistics have the is λ following formulas:
¯ n /λ0 ) − (X ¯ n − λ0 ) ¯ n ln(X 2 ln(Λn ) = 2n X Wn = Sn =
¯ n − λ0 )2 n(X λ0 ¯ n − λ0 )2 n(X
c 2000 by Chapman & Hall/CRC
λ0
.
Note that Sn and Wn are the same in this case. Figure 7.3 shows a graph of the (scaled) likelihood function for a sample of 100 observations. For the purpose of illustration, we take λ0 = 4 in which case the value of the LR statistic (for these data) is 13.60 while the value for the score and Wald statistic is 14.44; using the asymptotic χ2 (1) null distribution, we would reject H0 at the 0.05 level for values greater than 3.84 and at the 0.01 level for values greater than 6.63. ✸ Although both the Wald and the score statistics have the same limiting distribution (under H0 ) as the LR statistic, there are some practical diﬀerences in the use of these statistics. First, suppose that we reparametrize the model by setting θ ∗ = g(θ) where g is a onetoone transformation on Θ. In this case, the LR statistic remains unchanged since the maxima (both restricted and unrestricted) of the likelihood function remain unchanged. On the other hand, the Wald and score statistics will not remain invariant to reparametrization although the dependence on the parametrization of the model becomes smaller as the sample size increases. Second, both the LR and Wald statistics require the computation of restricted and unrestricted MLEs; on the other hand, the score statistic requires only computation of the restricted MLE (that is, the MLE under H0 ). Thus the score statistic is potentially simpler from a computational point of view and for this reason, the score test is often used in deciding whether or not parameters should be added to a model. As mentioned above, the score test is often called the Lagrange multiplier test, particularly in econometrics. The reason for this is the following. Suppose that we want to maximize a loglikelihood function ln L(θ) subject to the constraint g(θ) = 0 for some function g. To solve this problem, we can introduce a vector of Lagrange multipliers λ and maximize the function h(θ, λ) = ln L(θ) + λT g(θ). / λ) / maximizes h(θ, λ) then λ / ≈ 0 implies that the constrained If (θ, maximum and unconstrained maximum of ln L(θ) are close; on the / is not close to the 0 vector then the two maxima other hand, if λ can be very diﬀerent. This suggests that a test statistic can be based / in the problem considered above, this statistic is simply the on λ; score statistic. c 2000 by Chapman & Hall/CRC
EXAMPLE 7.23: Suppose that X1 , · · · , Xn are i.i.d. Gamma random variables with shape parameter α and scale parameter λ. We want to test H0 : α = 1
versus
H1 : α = 1.
Under H0 , note that the Xi ’s have an Exponential distribution. For this model, estimating both α and λ via maximum likelihood is nontrivial (although not diﬃcult) as there is no closedform expression for the MLEs (see Example 5.15); however, for ﬁxed α, there is a simple closedform expression for the MLE of λ. Thus the score test seems an attractive approach to testing H0 . For α = 1, the MLE / n = 1/X ¯ n and so the score statistic is of λ is λ
n 1 ¯ n ) − ψ(1) Sn = ln(Xi /X nψ (1) i=1
2
where ψ(α) and ψ (α) are the ﬁrst and second derivatives of ln Γ(α) with ψ(1) = −0.57722 and ψ (1) = 1.64493; see Example 5.15 for more details. The limiting distribution of Sn is χ2 with 1 degree of freedom. ✸ In addition to having the same limiting distribution, the LR statistic as well as the Wald and score statistics are asymptotically equivalent (under the null hypothesis) in the sense that the diﬀerence between any two of them tends in probability to 0 as n → ∞. This asymptotic equivalence is a consequence of the fact that the loglikelihood function is a quadratic function in a neighbourhood of the true parameter value (assuming the regularity conditions of Theorem 7.5). More precisely, if ln Ln (θ) is the loglikelihood function and θ 0 is the true value of the parameter then we have √ Zn (u) = ln(Ln (θ 0 + u/ n)/Ln (θ 0 )) 1 = uT V n − uT I(θ 0 )u + Rn (u) 2 where Rn (u) →p 0 for each u and V n →d Np (0, I(θ 0 )); Zn (u) is √ / maximized at u = n(θ n − θ 0 ) and the quadratic approximation to Zn is maximized at u = I −1 (θ 0 )V n . Thus if we are interested in testing H0 : θ = θ 0 , the quadratic approximation to Zn suggests the following approximations to our test statistics: √ / 2 ln(Λn ) = 2Zn ( n(θ n − θ 0 )) c 2000 by Chapman & Hall/CRC
≈ V Tn I −1 (θ 0 )V n Wn = = Sn = ≈
/ n − θ 0 )T I(θ 0 )(θ / n − θ0 ) n(θ V Tn I −1 (θ 0 )V n [∇Zn (0)]T I −1 (θ 0 )[∇Zn (0)] V n I −1 (θ 0 )V n
where ∇Zn (0) is the gradient √ of Zn (u) at u = 0, or equivalently, the score vector divided by n. Thus V n I −1 (θ 0 )V n serves as an approximation for each of 2 ln(Λn ) (the LR statistic), Wn (the Wald statistic), and Sn (the score statistic); note that V n I −1 (θ 0 )V n →d χ2 (p). Similarly approximations hold for these test statistics applied to other null hypotheses considered in this section. 7.5 Other issues Pvalues Up to this point, we have assume a ﬁxed level α when discussing hypothesis tests. That is, given α, we deﬁne a test function φ; then given data X = x, we reject the null hypothesis at level α if φ(x). An alternative approach (which is more in line with current practice) is to consider a family of test functions φα for 0 < α < 1 where the test function φα has level α. We will assume the test functions {φα } satisfy the condition φα1 (x) = 1
implies φα2 (x) = 1
for any α1 < α2 . We then deﬁne the pvalue (or observed signiﬁcance level) to be p(x) = inf{α : φα (x) = 1}. The pvalue p(x) is the smallest value of α for which the null hypothesis would be rejected at level α given X = x. In the case where the hypothesis test is framed in terms of a single test statistic T = T (X) such that φα (X) = 1 for T > kα then it is straightforward to evaluate pvalues. If G(x) is the null distribution function of T , then given T (x) = t, the pvalue is p(x) = 1 − G(t). EXAMPLE 7.24: Suppose that X1 , · · · , Xm and Y1 , · · · , Yn are two samples of i.i.d. random variables with Xi ∼ N (µ1 , σ 2 ) and Yi ∼ N (µ2 , σ 2 ). The LR test of the null hypothesis H0 : µ1 = µ2 c 2000 by Chapman & Hall/CRC
(versus the alternative H1 = µ1 = µ2 ) rejects H0 for large values of the test statistic T  where T =
¯ − Y¯ X S m−1 + n−1 √
and S 2 is the socalled “pooled” estimator of variance: ,

m n 1 ¯ 2+ S = (Xi − X) (Yi − Y¯ )2 . m + n − 2 i=1 i=1 2
It is easy to verify that T ∼ T (m+n−2) under the null hypothesis. Thus given T = t, the pvalue is p(t) = 1 + G(−t) − G(t) where G(x) is the distribution function of the t distribution. However, if the alternative hypothesis is H1 : µ1 > µ2 , we would typically reject H0 for large values of T . In this case, given T = t, the pvalue is p(t) = 1 − G(t). ✸ The pvalue is often used as a measure of evidence against the null hypothesis: the smaller p(x), the more evidence against the null hypothesis. While this use of pvalues is quite common in statistical practice, its use as a measure of evidence is quite controversial. In particular, it is diﬃcult to calibrate pvalues as measures of evidence (Goodman, 1999a). Pvalues are often (erroneously) interpreted as the probability (given the observed data) that the null hypothesis is true. However, if we put a prior distribution on the parameter space then it may be possible to compute such a probability (from a Bayesian perspective) using the posterior distribution. Some care is required in interpreting these probabilities though; for example, the posterior probability that the null hypothesis is true necessarily depends on the prior probability of the null hypothesis. As a Bayesian alternative to pvalues, some authors have proposed using Bayes factors, which essentially measure the change in the odds of the null hypothesis from the prior to the posterior; see, for example, Kass and Raftery (1995), and DiCiccio et al (1997). It is often argued that Bayes factors are more easily intepretable than pvalues (Goodman, 1999b) although this view is not universally shared. We can also view pvalues, in their own right, as test statistics. Suppose that we want to test H0 : θ ∈ Θ0 c 2000 by Chapman & Hall/CRC
versus
H1 : θ ∈ Θ1
using a family of tests whose test functions are {φα }. Suppose that X = (X1 , · · · , Xn ) are continuous random variables and that, for some θ0 ∈ Θ0 , Eθ0 [φα (X)] = α for all 0 < α < 1. Then if the pvalue p(X) has a continuous distribution, it follows that Pθ0 [p(X) ≤ x] = x for 0 < x < 1; that is, p(X) has a Uniform distribution on [0, 1] when θ0 is the true value of the parameter. (Note that, for given θ ∈ Θ0 , we have Pθ [p(X) ≤ x] ≥ x for 0 < x < 1.) The fact that p(X) is uniformly distributed under H0 can be useful in practice. For example, suppose that we have pvalues P1 , · · · , Pk from k independent tests of the same null hypothesis. Assuming that the Pi ’s are uniformly distributed when the null hypothesis is true, we can combine the pvalues using the test statistic T = −2
k
ln(Pi );
i=1
under the null hypothesis, T has a χ2 with 2k degrees of freedom. This simple approach to metaanalysis (that is, combining results of diﬀerent studies) is due to R.A. Fisher. Obtaining conﬁdence regions from hypothesis tests Our discussion of conﬁdence intervals and regions gave essentially no theoretical guidance on how to choose a “good” conﬁdence procedure; in contrast, for hypothesis testing, the NeymanPearson Lemma provides a useful paradigm for deriving “good” hypothesis tests in various situations. In fact, there turns out to be a very close relationship between conﬁdence intervals (or regions) and hypothesis tests; we can exploit this relationship to turn “good” hypothesis tests into “good” conﬁdence procedures. Suppose that R(X) is an exact 100p% conﬁdence region for a parameter θ and we want to test the null hypothesis H0 : θ = θ 0 c 2000 by Chapman & Hall/CRC
versus
H1 : θ = θ 0 .
Deﬁne the test function
φ(X) =
1 if θ 0 ∈ / R(X) 0 if θ 0 ∈ R(X).
It is easy to verify that the size of this test is Eθ0 [φ(X)] = 1 − Pθ0 [θ 0 ∈ R(X)] ≤ 1 − p. Thus we can use a 100p% conﬁdence region to construct a test of H0 whose level is at most 1 − p. On the other hand, suppose that we have αlevel tests of H0 for each θ 0 ∈ Θ; deﬁne φ(X; θ 0 ) to be the test function for a given θ0 . Now deﬁne R∗ (X) = {θ 0 : φ(X; θ 0 ) = 0} ; the coverage of R∗ (X) is Pθ [θ ∈ R∗ (X)] = Pθ [φ(X; θ) = 0] ≥ 1 − α. Thus we can construct a (possibly conservative) 100p% conﬁdence region for θ by considering a family of α = 1 − p level tests and deﬁning the conﬁdence region to be the set of θ’s for which we cannot reject the null hypothesis at level α. This “duality” between hypothesis tests and conﬁdence intervals or regions can be very useful in practice. For example, suppose that X1 , · · · , Xn are i.i.d. random variables with density or frequency function f (x; θ) and we want to ﬁnd a conﬁdence interval for a single parameter (call it φ) in θ. Writing θ = (φ, τ ), the LR statistic for H0 : φ = φ0 is 2 ln(Λn ) = 2
n
ln[f (Xi ; φ/n , τ/ n )/f (Xi ; φ0 , τ/ n (φ0 ))]
i=1
where τ/ n (φ0 ) is the MLE of τ under H0 (that is, assuming that φ = φ0 is known). According to Theorem 7.5, under H0 , 2 ln(Λn ) →d χ2 (1) and H0 is rejected for large values of the statistic 2 ln(Λn ). Thus if kp is the p quantile of a χ2 (1) distribution, an approximate 100p% conﬁdence interval for φ is R(X) = {φ : g(X; φ) ≤ kp } c 2000 by Chapman & Hall/CRC
where g(X; φ) is the “likelihood ratio” pivot: g(X; φ) = 2
n
ln[f (Xi ; φ/n , τ/ n )/f (Xi ; φ, τ/ n (φ))].
i=1
Note that g(X; φ) ≥ 0 with g(X; φ/n ) = 0. EXAMPLE 7.25: Suppose that X1 , · · · , Xn are i.i.d. Gamma random variables with shape parameter α and scale parameter λ. We will derive an approximate 95% conﬁdence interval for α based on the LR test procedure. First of all, we start with the loglikelihood function ln L(α, λ) = nα ln(λ) + (α − 1)
n
ln(Xi ) − λ
i=1
n
Xi − n ln Γ(α).
i=1
The LR test of H0 : α = α0 compares the maximized likelihood with α = α0 to the unrestricted maximized likelihood; if α is assumed / / ¯ Substituting λ(α) known then the MLE of λ is λ(α) = α/X. for λ in the loglikelihood, we obtain the proﬁle loglikelihood / ln Lp (α) = ln L(α, λ(α)) '
(
¯ − 1 + (α − 1) = nα ln(α/X)
n
ln(Xi ) − n ln Γ(α).
i=1
/ which is the MLE of The proﬁle loglikelihood is maximized at α, α; to obtain a conﬁdence interval for α we look at the approximate pivot / g(X; α) = 2 ln[Lp (α)/L p (α)].
We know that, for a given α, g(X; α) is approximately χ2 distributed with 1 degree of freedom; thus an approximate 95% conﬁdence interval for α is R(X) = {α : g(X; α) ≤ 3.841} . For a sample of 50 i.i.d. Gamma random variables, a graph of g(X; α) and the approximate 95% conﬁdence interval for α are shown in Figure 7.4. ✸ A similar approach can be followed to obtain conﬁdence intervals (or, indeed, conﬁdence regions) from the Wald and score tests. c 2000 by Chapman & Hall/CRC
5 4 3 1
2
pivot
0
xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx 4
5
6
7
8
alpha
Figure 7.4 The proﬁle likelihood pivot in Example 7.25 and an approximate 95% conﬁdence for α.
Conﬁdence intervals and tests based on nonparametric likelihood
Suppose that X1 , · · · , Xn are i.i.d. random variables with unknown distribution function F . In section 5.6, we deﬁned a notion of nonparametric maximum likelihood estimation and noted that the empirical distribution function was the nonparametric MLE of F under this formulation. We will indicate here how the nonparametric likelihood function can be used to obtain a conﬁdence interval for a functional parameter θ(F ) or to test a null hypothesis of the form H0 : θ(F ) = θ0 . The idea is to extend the idea of LR tests to the nonparametric setting. We will consider the hypothesis testing problem ﬁrst. According to our discussion in section 5.6, we consider only distributions putting all their probability mass at the points X1 , · · · , Xn . For a given vector of probabilities p = (p1 , · · · , pn ), we deﬁne Fp to be the (discrete) distribution with probability pi at Xi (i = 1, · · · , n). c 2000 by Chapman & Hall/CRC
The nonparametric (or empirical) loglikelihood function is ln L(p) =
n
ln(pi )
i=1
and so to implement the LR test procedure, we must consider the unrestricted MLEs of p1 , · · · , pn (which are p/i = 1/n for i = 1, · · · , n) as well as the restricted MLEs assuming that the null hypothesis is true; the restricted MLEs maximize ln L(p) subject / 0 is this to the constraints p1 + · · · + pn = 1 and θ(Fp ) = θ0 . If p restricted MLE of p then the LR test statistic is given by ln(Λn ) = −
n
ln(np/i0 );
i=1
as before, we would reject H0 for large values of ln(Λn ) (or, equivalently, Λn ). .
EXAMPLE 7.26: Suppose that θ(F ) = x dF (x) = E(Xi ). Under H0 : θ(F ) = θ0 , the nonparametric MLEs of p1 , · · · , pn can be obtained by maximizing ln L(p) subject to the constraints p 1 + · · · + pn = 1 p 1 X 1 + · · · + pn X n = θ 0 . The MLEs (under H0 ) can be determined by introducing two Lagrange multipliers, λ1 and λ2 , and maximizing n
ln(pi ) + λ1 (p1 + · · · + pn − 1) + λ2 (p1 X1 + · · · + pn Xn − θ0 )
i=1
with respect to p1 , · · · , pn , λ1 and λ2 . After diﬀerentiating and setting the partial derivatives to 0, we obtain p/i =
1 n − λ2 (Xi − θ0 )
where λ2 is deﬁned so that the constraints on the pi ’s are satisﬁed. ✸ The limiting null distribution of the LR test statistic is not clear; we do not have the standard conditions for this statistic to have a limiting χ2 distribution. Nonetheless, if θ(F ) is a suﬃciently “smooth” functional parameter then the null distribution 2 ln(Λn ) will be approximately χ2 with 1 degree of freedom for large n. For c 2000 by Chapman & Hall/CRC
example, the following . result can be found in Owen (1988) for the case where θ(F ) = g(x) dF (x). .
THEOREM 7.6 Suppose. that θ(F ) = g(x) dF (x) and we are testing H0 : θ(F ) = θ0 . If g(x)3 dF (x) < ∞ then 2 ln(Λn ) →d χ2 (1) if H0 is true. Similar results can be given for other functional parameters θ(F ); see Owen (1988) for details. We can use the result of Theorem 7.6 to derive approximate conﬁdence intervals for functional parameters. θ(F ) satisfying its conditions. For example, suppose that θ(F ) = x dF (x) and deﬁne g(X; θ) = −2
n
ln(np/i (θ))
i=1
/ (θ) maximizes the nonparametric likelihood subject to the where p constraint n pi Xi = θ. i=1
Then g(X; θ(F )) is an approximate pivot for θ(F ) and hence we can deﬁne an approximate 100p% conﬁdence interval for θ(F ) to be R(X) = {θ : g(X; θ) ≤ kp } where kp is the p quantile of a χ2 distribution with 1 degree of freedom. 7.6 Problems and complements 7.1: Suppose that X1 , · · · , Xn are i.i.d. Normal random variables with unknown mean µ and variance σ 2 . (a) Using the pivot (n − 1)S 2 /σ 2 where S2 =
n 1 ¯ 2, (Xi − X) n − 1 i=1
we can obtain a 95% conﬁdence interval [k1 S 2 , k2 S 2 ] for some constants k1 and k2 . Find expressions for k1 and k2 if this conﬁdence interval has minimum length. Evaluate k1 and k2 when n = 10. c 2000 by Chapman & Hall/CRC
(b) When n is suﬃciently large, we can approximate the distribution of the pivot by a Normal distribution (why?). Find approximations for k1 and k2 that are valid for large n. 7.2: Suppose that X1 , · · · , Xn are i.i.d. Exponential random variables with parameter λ. (a) Show that 2λXi has a χ2 distribution with 2 degrees of freedom and hence that 2λ ni=1 Xi has a χ2 distribution with 2n degrees of freedom. (b) Suppose that n = 5. Give a 90% conﬁdence interval for θ using the result of part (a). (c) Let θ/ be the MLE of θ. Find a function g such that √ / − g(θ)) is approximately standard the distribution of n(g(θ) Normal and use this to give an approximate 90% conﬁdence interval for θ valid for large n. 7.3: Suppose that X1 , · · · , Xn are i.i.d. continuous random variables with median θ.
(a) What is the distribution of ni=1 I(Xi ≤ θ)? (b) Let X(1) < · · · < X(n) be the order statistics of X1 , · · · , Xn . Show that the interval [X() , X(u) ] is a 100p% conﬁdence interval for θ and ﬁnd an expression for p in terms of : and u. (Hint: use the random variable in part (a) as a pivot for θ.) (c) Suppose that for large n, we set :
:=
n 0.98 − √ 2 n
;