Lecture 28: Asymptotic confidence sets

Similar documents
Lecture 32: Asymptotic confidence sets and likelihoods

Stat 710: Mathematical Statistics Lecture 31

Chapter 9: Interval Estimation and Confidence Sets Lecture 16: Confidence sets and credible sets

Chapter 7. Confidence Sets Lecture 30: Pivotal quantities and confidence sets

parameter space Θ, depending only on X, such that Note: it is not θ that is random, but the set C(X).

Chapter 2: Fundamentals of Statistics Lecture 15: Models and statistics

Statistics Ph.D. Qualifying Exam

Statistics Ph.D. Qualifying Exam: Part I October 18, 2003

Chapter 3: Unbiased Estimation Lecture 22: UMVUE and the method of using a sufficient and complete statistic

Lecture 17: Likelihood ratio and asymptotic tests

SOLUTION FOR HOMEWORK 4, STAT 4352

Stat 5101 Lecture Notes

McGill University. Faculty of Science. Department of Mathematics and Statistics. Part A Examination. Statistics: Theory Paper

ST495: Survival Analysis: Hypothesis testing and confidence intervals

Stat 710: Mathematical Statistics Lecture 40

Lecture 21: Convergence of transformations and generating a random variable

Final Examination Statistics 200C. T. Ferguson June 11, 2009

Multivariate Analysis and Likelihood Inference

Lecture 20: Linear model, the LSE, and UMVUE

Chapter 8 - Statistical intervals for a single sample

Lecture 26: Likelihood ratio tests

Maximum Likelihood Large Sample Theory

Asymptotic Statistics-VI. Changliang Zou

Lecture 1: August 28

simple if it completely specifies the density of x

A Very Brief Summary of Statistical Inference, and Examples

Hypothesis Testing. 1 Definitions of test statistics. CB: chapter 8; section 10.3

Chapter 4: Asymptotic Properties of the MLE (Part 2)

f(x θ)dx with respect to θ. Assuming certain smoothness conditions concern differentiating under the integral the integral sign, we first obtain

Chapter 6: Large Random Samples Sections

HYPOTHESIS TESTING: FREQUENTIST APPROACH.

Interval Estimation. Chapter 9

Test Problems for Probability Theory ,

Testing Hypothesis. Maura Mezzetti. Department of Economics and Finance Università Tor Vergata

Stat 5102 Final Exam May 14, 2015

Lecture 11. Multivariate Normal theory

Institute of Actuaries of India

STAT 512 sp 2018 Summary Sheet

Limiting Distributions

Asymptotic Statistics-III. Changliang Zou

Lecture 3. Inference about multivariate normal distribution

Statistics - Lecture One. Outline. Charlotte Wickham 1. Basic ideas about estimation

Stat 5101 Notes: Brand Name Distributions

MVE055/MSG Lecture 8

Testing Algebraic Hypotheses

Limiting Distributions

Stat 704 Data Analysis I Probability Review

Preliminaries. Copyright c 2018 Dan Nettleton (Iowa State University) Statistics / 38

Chapters 9. Properties of Point Estimators

Qualifying Exam CS 661: System Simulation Summer 2013 Prof. Marvin K. Nakayama

Continuous Distributions

Mathematical statistics

Delta Method. Example : Method of Moments for Exponential Distribution. f(x; λ) = λe λx I(x > 0)

Inference on reliability in two-parameter exponential stress strength model

Regression #4: Properties of OLS Estimator (Part 2)

Lecture 16: Sample quantiles and their asymptotic properties

Stat 710: Mathematical Statistics Lecture 12

STAT 830 Non-parametric Inference Basics

ST5215: Advanced Statistical Theory

Severity Models - Special Families of Distributions

Some Assorted Formulae. Some confidence intervals: σ n. x ± z α/2. x ± t n 1;α/2 n. ˆp(1 ˆp) ˆp ± z α/2 n. χ 2 n 1;1 α/2. n 1;α/2

18.175: Lecture 13 Infinite divisibility and Lévy processes

Comparing two independent samples

Stat 135, Fall 2006 A. Adhikari HOMEWORK 6 SOLUTIONS

Chapter 6. Hypothesis Tests Lecture 20: UMP tests and Neyman-Pearson lemma

STATS 200: Introduction to Statistical Inference. Lecture 29: Course review

Chapter 8: Hypothesis Testing Lecture 9: Likelihood ratio tests

Statistics. Statistics

Mathematical statistics

EXAMINATIONS OF THE ROYAL STATISTICAL SOCIETY

Asymptotic Theory. L. Magee revised January 21, 2013

MAS223 Statistical Inference and Modelling Exercises

Statistical Inference

STA 2101/442 Assignment 3 1

Economics 583: Econometric Theory I A Primer on Asymptotics

COMPSCI 240: Reasoning Under Uncertainty

Lecture 6: Linear models and Gauss-Markov theorem

MISCELLANEOUS TOPICS RELATED TO LIKELIHOOD. Copyright c 2012 (Iowa State University) Statistics / 30

DA Freedman Notes on the MLE Fall 2003

Central Limit Theorem ( 5.3)

Lecture 15: Multivariate normal distributions

Qualifying Exam in Probability and Statistics.

Lecture 23: UMPU tests in exponential families

Economics 583: Econometric Theory I A Primer on Asymptotics: Hypothesis Testing

Lecture 34: Properties of the LSE

Inference in Constrained Linear Regression

Interval estimation. October 3, Basic ideas CLT and CI CI for a population mean CI for a population proportion CI for a Normal mean

A Very Brief Summary of Statistical Inference, and Examples

Lecture 3 September 1

Master s Written Examination

KRUSKAL-WALLIS ONE-WAY ANALYSIS OF VARIANCE BASED ON LINEAR PLACEMENTS

Course information: Instructor: Tim Hanson, Leconte 219C, phone Office hours: Tuesday/Thursday 11-12, Wednesday 10-12, and by appointment.

Lecture 13: Subsampling vs Bootstrap. Dimitris N. Politis, Joseph P. Romano, Michael Wolf

1 Exercises for lecture 1

Notes on the Multivariate Normal and Related Topics


Chapter 4 Multiple Random Variables

Part IA Probability. Definitions. Based on lectures by R. Weber Notes taken by Dexter Chua. Lent 2015

Statistics & Data Sciences: First Year Prelim Exam May 2018

Information in a Two-Stage Adaptive Optimal Design

Transcription:

Lecture 28: Asymptotic confidence sets 1 α asymptotic confidence sets Similar to testing hypotheses, in many situations it is difficult to find a confidence set with a given confidence coefficient or level 1 α. A common approach is to find a confidence set whose confidence coefficient or level is approximately 1 α when the sample size n is large. Based on a sample X, a confidence set C(X) for ϑ has asymptotic level 1 α if lim n P θ (ϑ C(X)) 1 α Often, we can replace by = in the previous expression. In any case, we call C(X) a 1 α asymptotic confidence set. Asymptotic pivotal quantities A known function q n (X,ϑ) is said to be asymptotically pivotal iff the limiting distribution of q n (X,ϑ) does not depend on any unknown quantity. UW-Madison (Statistics) Stat 610 Lecture 28 2014 1 / 10

Confidence sets based on asymptotic pivots Like a pivotal quantity in constructing confidence sets with a given confidence coefficient or level, an asymptotically pivotal quantity can be used in constructing 1 α asymptotic confidence sets. Because of the wide applications of the CLT, Slutsky s theorem, and the delta-method, many asymptotic pivots can be constructed with the 1/2 form V n ( ϑ n ϑ), where ϑ n is an estimator of ϑ that is asymptotically normal, i.e., V 1/2 n ( ϑ n ϑ) converges in distribution to N(0,I k ) and V n is a consistent estimator of V n (called the asymptotic covariance matrix of ϑ n ) in the sense that if v ij and v ij are the (i,j)th elements of V n and V n, respectively, then n( v ij v ij ) converges in probability to 0. The resulting 1 α asymptotically correct confidence sets are of the form C(X) = {ϑ : V 1/2 n ( ϑ n ϑ) 2 χk,α 2 }, UW-Madison (Statistics) Stat 610 Lecture 28 2014 2 / 10

where χk,α 2 is the 100(1 α)th percentile of the chi-square distribution with k degrees of freedom and b 2 = b b for any vector b. If ϑ is real-valued, then this C(X) is a 1 α asymptotic confidence interval; ϑ is multivariate, C(X) is an ellipsoid. Eample. Let X 1,...,X n be iid random variables with E(X 1 ) = µ and Var(X 1 ) = σ 2, and let ϑ = µ. Then n( X µ)/s is an asymptotic pivot, since by the CLT and Slutsky s theorem, it converges in distribution to N(0, 1). Based on this asymptotic pivot, we obtain a 1 α asymptotic confidence interval [ X z α/2 S/ n, X + z α/2 S/ n] If we also know that X i has the Poisson(λ) distribution with unknown λ > 0, then µ = λ and σ 2 = λ. Hence, we have two other asymptotic pivots: UW-Madison (Statistics) Stat 610 Lecture 28 2014 3 / 10

where χk,α 2 is the 100(1 α)th percentile of the chi-square distribution with k degrees of freedom and b 2 = b b for any vector b. If ϑ is real-valued, then this C(X) is a 1 α asymptotic confidence interval; ϑ is multivariate, C(X) is an ellipsoid. Eample. Let X 1,...,X n be iid random variables with E(X 1 ) = µ and Var(X 1 ) = σ 2, and let ϑ = µ. Then n( X µ)/s is an asymptotic pivot, since by the CLT and Slutsky s theorem, it converges in distribution to N(0, 1). Based on this asymptotic pivot, we obtain a 1 α asymptotic confidence interval [ X z α/2 S/ n, X + z α/2 S/ n] If we also know that X i has the Poisson(λ) distribution with unknown λ > 0, then µ = λ and σ 2 = λ. Hence, we have two other asymptotic pivots: UW-Madison (Statistics) Stat 610 Lecture 28 2014 3 / 10

n( X λ) X and n( X λ) λ. Based on the first one we obtain a 1 α asymptotic confidence interval [ X zα/2 X/ n, X + zα/2 X/ n ] For the second asymptotic pivot, we can construct a 1 α asymptotic confidence set C(X) = {λ : ( X λ) 2 } { zα/2 2 = λ : λ 2 (2 λ/n X + n 1 zα/2 2 )λ + X } 2 0 Note that Q(λ) = λ 2 (2 X + n 1 z 2 α/2 )λ + X 2 is a convex quadratic function. Thus, C(X) is also a confidence interval [L(X),U(X)] whose two limits are the two solutions to Q(λ) = 0: L(X) = X + 2 1 n 1 z 2 α/2 Xz 2 α/2 /n + z 4 α/2 /n2 U(X) = X + 2 1 n 1 z 2 α/2 + Xz 2 α/2 /n + z 4 α/2 /n2 UW-Madison (Statistics) Stat 610 Lecture 28 2014 4 / 10

Taking the difference of the lower limits of the two confidence intervals we get zα/2 2 2n z 2 X α/2 + z4 α/2 n n 2 zα/2 2 X n / = z2 α/2 2n z4 α/2 z 2 X α/2 n 2 + z4 α/2 n n 2 + zα/2 2 X n which converges to 0 at the rate n 1 and, hence, the two intervals are asymptotically equivalent. From the previous example, we find that, if ϑ = θ is the only unknown parameter in the population and thus V n = V n (θ), then we may not need to estimate V n in order to obtain a confidence interval. We can directly solve ( ϑ n ϑ) 2 /V n (ϑ) χ1,α 2 and obtain a confidence set { C(X) = ϑ : ( ϑ } n ϑ) 2 χ1,α 2 V n(ϑ) Often, C(X) is a confidence interval. UW-Madison (Statistics) Stat 610 Lecture 28 2014 5 / 10

Taking the difference of the lower limits of the two confidence intervals we get zα/2 2 2n z 2 X α/2 + z4 α/2 n n 2 zα/2 2 X n / = z2 α/2 2n z4 α/2 z 2 X α/2 n 2 + z4 α/2 n n 2 + zα/2 2 X n which converges to 0 at the rate n 1 and, hence, the two intervals are asymptotically equivalent. From the previous example, we find that, if ϑ = θ is the only unknown parameter in the population and thus V n = V n (θ), then we may not need to estimate V n in order to obtain a confidence interval. We can directly solve ( ϑ n ϑ) 2 /V n (ϑ) χ1,α 2 and obtain a confidence set { C(X) = ϑ : ( ϑ } n ϑ) 2 χ1,α 2 V n(ϑ) Often, C(X) is a confidence interval. UW-Madison (Statistics) Stat 610 Lecture 28 2014 5 / 10

A sufficient condition for C(X) to be an interval is that, as a function of ϑ, χ 2 1,α V n(ϑ) ( ϑ n ϑ) 2 is unimodal. This is actually the case for the second confidence interval in the previous example for the Poisson case. We may replace V n (ϑ) by V n ( ϑ n ), which is consistent if V n (ϑ) is a continuous function of ϑ. This ensures that we obtain a confidence interval, and the first interval in the previous example for the Poisson case is this kind. Suppose now that ϑ is not the only parameter, i.e., θ = (ϑ,ϕ) and we want a confidence set or interval for ϑ. If V n (θ) does not depend on ϕ, then the procedures in the previous discussion can still be applied. If V n (θ) = V n (ϑ,ϕ) depends on ϕ, then we first find an estimator ϕ n of ϕ, and then use V n (ϑ, ϕ n ) or V n ( ϑ n, ϕ n ) and apply the previous discussed procedure. We may also find an estimator of V n (θ) in a different way. UW-Madison (Statistics) Stat 610 Lecture 28 2014 6 / 10

Example. Let X 1,...,X n be iid from Double-exponential(µ,σ) with unknown µ R and σ 2 > 0. Consider setting a confidence interval for ϑ = µ based on the asymptotic distribution of X (not the MLE of µ). By the CLT and Var(X 1 ) = 2σ 2, n( X µ) converges in distribution to N(0,1) 2σ If we estimator 2σ 2 by the sample variance S 2, then we obtain the following 1 α asymptotic confidence interval (noting that zα/2 2 = χ2 1,α ) C(X) = {µ : n( X µ) 2 χ1,α 2 S2} = [ X zα/2 S/ n, X + z α/2 S/ n ] Suppose now that we know σ 2 = µ 2. If we estimate 2σ 2 by 2 X 2 and inverting n( X µ) 2 χ1,α 2 (2 X 2 ), the resulting confidence interval is the same as C(X) with S replaced by 2 X. UW-Madison (Statistics) Stat 610 Lecture 28 2014 7 / 10

If we invert n( X µ) 2 χ 2 1,α (2µ2 ), the resulting confidence interval is [L(X),U(X)] with L(X) = X 2z α/2 X / n 1 2z 2 α/2 /n, U(X) = X + 2z α/2 X / n 1 2z 2 α/2 /n If we estimate 2σ 2 = 2µ 2 by S 2, then we can still use interval C(X). Typically, in a given problem there exist many different asymptotic pivots that lead to different 1 α asymptotic confidence sets for θ. The following result indicates that we can compare volumes of confidence sets constructed using two asymptotic pivots. Theorem. Let C j (X), j = 1,2, be the confidence sets for ϑ using asymptotic V 1/2 pivots jn ( ϑ jn ϑ), j = 1,2, respectively. If V 1n < V 2n for sufficiently large n, where A is the determinant of the matrix A, then lim n P( vol(c 1 (X)) < vol(c 2 (X)) ) = 1. UW-Madison (Statistics) Stat 610 Lecture 28 2014 8 / 10

If we invert n( X µ) 2 χ 2 1,α (2µ2 ), the resulting confidence interval is [L(X),U(X)] with L(X) = X 2z α/2 X / n 1 2z 2 α/2 /n, U(X) = X + 2z α/2 X / n 1 2z 2 α/2 /n If we estimate 2σ 2 = 2µ 2 by S 2, then we can still use interval C(X). Typically, in a given problem there exist many different asymptotic pivots that lead to different 1 α asymptotic confidence sets for θ. The following result indicates that we can compare volumes of confidence sets constructed using two asymptotic pivots. Theorem. Let C j (X), j = 1,2, be the confidence sets for ϑ using asymptotic V 1/2 pivots jn ( ϑ jn ϑ), j = 1,2, respectively. If V 1n < V 2n for sufficiently large n, where A is the determinant of the matrix A, then lim n P( vol(c 1 (X)) < vol(c 2 (X)) ) = 1. UW-Madison (Statistics) Stat 610 Lecture 28 2014 8 / 10

Proof. The result follows from the consistency of V jn and the fact that the volume of the ellipsoid C(X) is equal to vol(c(x)) = π k/2 (χ 2 k,α )k/2 V n 1/2/ Γ(1 + k/2). If ϑ 1n is asymptotically more efficient than ϑ 2n, then V 1n V 2n and ϑ 1n results in a better confidence set in terms of volume. If ϑ n is asymptotically efficient (such as the MLE), then C(X) obtained by using the pivot based on ϑ n is asymptotically optimal in terms of volume. In the example for the Poisson distribution, X is the MLE and hence the first confidence interval is asymptotically optimal in terms of length. Although the theorem is not directly applicable to the second confidence interval in the Poisson example, actually we have shown in the end of the example that the second confidence interval is also asymptotically optimal in terms of length, because the difference between the lengths of the two intervals is of the order n 1, whereas the length is of the order n 1/2. UW-Madison (Statistics) Stat 610 Lecture 28 2014 9 / 10

Proof. The result follows from the consistency of V jn and the fact that the volume of the ellipsoid C(X) is equal to vol(c(x)) = π k/2 (χ 2 k,α )k/2 V n 1/2/ Γ(1 + k/2). If ϑ 1n is asymptotically more efficient than ϑ 2n, then V 1n V 2n and ϑ 1n results in a better confidence set in terms of volume. If ϑ n is asymptotically efficient (such as the MLE), then C(X) obtained by using the pivot based on ϑ n is asymptotically optimal in terms of volume. In the example for the Poisson distribution, X is the MLE and hence the first confidence interval is asymptotically optimal in terms of length. Although the theorem is not directly applicable to the second confidence interval in the Poisson example, actually we have shown in the end of the example that the second confidence interval is also asymptotically optimal in terms of length, because the difference between the lengths of the two intervals is of the order n 1, whereas the length is of the order n 1/2. UW-Madison (Statistics) Stat 610 Lecture 28 2014 9 / 10

In the double exponential example, the MLE of µ is the sample median Q 0.5, which has an asymptotic distribution derived previously: ( ) 1 n( Q0.5 µ) converges in distribution to N 0, 4(2σ) 2 = σ 2 By inverting n( Q 0.5 µ) 2 χ1,α 2 S2 /2 (σ 2 should be estimated by S 2 /2, not S 2 ), we obtain the 1 α asymptotic confidence interval [ C 1 (X) = X zα/2 S/ 2n, X + z α/2 S/ ] 2n which is shorter than the C(X) previously derived, by a factor of 2. We may replace the estimator S 2 /2 of σ 2 by the MLE of σ 2, which is equal to (verify) σ 2 = 1 n X i n Q 0.5 i=1 However, replacing S 2 /2 by σ 2 does not reduce the length. Finally, there are situations where an MLE does not exist or is very hard to calculate and, thus, we have to use other asymptotic pivots. UW-Madison (Statistics) Stat 610 Lecture 28 2014 10 / 10