Maximum Likelihood Tests and Quasi-Maximum-Likelihood

Similar documents
Introduction Large Sample Testing Composite Hypotheses. Hypothesis Testing. Daniel Schmierer Econ 312. March 30, 2007

Greene, Econometric Analysis (6th ed, 2008)

Chapter 4: Constrained estimators and tests in the multiple linear regression model (Part III)

Maximum Likelihood (ML) Estimation

Introduction to Estimation Methods for Time Series models Lecture 2

Max. Likelihood Estimation. Outline. Econometrics II. Ricardo Mora. Notes. Notes

Statistics and econometrics

The outline for Unit 3

Econ 583 Homework 7 Suggested Solutions: Wald, LM and LR based on GMM and MLE

Final Exam. 1. (6 points) True/False. Please read the statements carefully, as no partial credit will be given.

Binary choice 3.3 Maximum likelihood estimation

LECTURE 10: NEYMAN-PEARSON LEMMA AND ASYMPTOTIC TESTING. The last equality is provided so this can look like a more familiar parametric test.

Advanced Quantitative Methods: maximum likelihood

ECON 5350 Class Notes Nonlinear Regression Models

Discrete Dependent Variable Models

(θ θ ), θ θ = 2 L(θ ) θ θ θ θ θ (θ )= H θθ (θ ) 1 d θ (θ )

Ch. 5 Hypothesis Testing

8. Hypothesis Testing

Advanced Quantitative Methods: maximum likelihood

Theory of Maximum Likelihood Estimation. Konstantin Kashin

Chapter 4. Theory of Tests. 4.1 Introduction

Recall that in order to prove Theorem 8.8, we argued that under certain regularity conditions, the following facts are true under H 0 : 1 n

System Identification, Lecture 4

Likelihood-based inference with missing data under missing-at-random

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

Advanced Econometrics

A Very Brief Summary of Statistical Inference, and Examples

DA Freedman Notes on the MLE Fall 2003

Spatial Regression. 9. Specification Tests (1) Luc Anselin. Copyright 2017 by Luc Anselin, All Rights Reserved

1.5 Testing and Model Selection

Economics 582 Random Effects Estimation

Statistical Methods for Handling Incomplete Data Chapter 2: Likelihood-based approach

Econometrics I, Estimation

System Identification, Lecture 4

To appear in The American Statistician vol. 61 (2007) pp

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

Lecture 6: Hypothesis Testing

Practical Econometrics. for. Finance and Economics. (Econometrics 2)

Exercises Chapter 4 Statistical Hypothesis Testing

Graduate Econometrics I: Maximum Likelihood I

Econometric Analysis of Cross Section and Panel Data

Asymptotics for Nonlinear GMM

Mathematical statistics

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

Let us first identify some classes of hypotheses. simple versus simple. H 0 : θ = θ 0 versus H 1 : θ = θ 1. (1) one-sided

Inference using structural equations with latent variables

Chapter 3: Maximum Likelihood Theory

Bootstrap Testing in Nonlinear Models

Some General Types of Tests

1 One-way analysis of variance

Inference Based on the Wild Bootstrap

ECONOMETRICS I. Assignment 5 Estimation

What s New in Econometrics. Lecture 15

A Course on Advanced Econometrics

Generalized Method of Moment

Lecture 1: Introduction

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

Testing Algebraic Hypotheses

Chapter 1: A Brief Review of Maximum Likelihood, GMM, and Numerical Tools. Joan Llull. Microeconometrics IDEA PhD Program

Economic modelling and forecasting

Econometrics II - EXAM Answer each question in separate sheets in three hours

Cointegration Lecture I: Introduction

Testing and Model Selection

Econometrics of Panel Data

Link lecture - Lagrange Multipliers

Eksamen på Økonomistudiet 2006-II Econometrics 2 June 9, 2006

Likelihood based Statistical Inference. Dottorato in Economia e Finanza Dipartimento di Scienze Economiche Univ. di Verona

Optimization. The value x is called a maximizer of f and is written argmax X f. g(λx + (1 λ)y) < λg(x) + (1 λ)g(y) 0 < λ < 1; x, y X.

Lecture 14 More on structural estimation

6. MAXIMUM LIKELIHOOD ESTIMATION

Mathematical statistics

Parameter Estimation

Graduate Econometrics I: Maximum Likelihood II

Estimating AR/MA models

AGEC 661 Note Eleven Ximing Wu. Exponential regression model: m (x, θ) = exp (xθ) for y 0

LESLIE GODFREY LIST OF PUBLICATIONS

Spring 2017 Econ 574 Roger Koenker. Lecture 14 GEE-GMM

Graduate Econometrics I: Unbiased Estimation

Composite Hypotheses and Generalized Likelihood Ratio Tests

GARCH Models Estimation and Inference

Hypothesis Testing (May 30, 2016)

Testing Restrictions and Comparing Models

Lecture 10 Maximum Likelihood Asymptotics under Non-standard Conditions: A Heuristic Introduction to Sandwiches

ML Testing (Likelihood Ratio Testing) for non-gaussian models

GARCH Models Estimation and Inference. Eduardo Rossi University of Pavia

Parametric Modelling of Over-dispersed Count Data. Part III / MMath (Applied Statistics) 1

Statistics Ph.D. Qualifying Exam

Lecture 17: Likelihood ratio and asymptotic tests

Topic 12 Overview of Estimation

ECON 4160, Autumn term Lecture 1

Mathematics Ph.D. Qualifying Examination Stat Probability, January 2018

y it = α i + β 0 ix it + ε it (0.1) The panel data estimators for the linear model are all standard, either the application of OLS or GLS.

1. Fisher Information

Testing Linear Restrictions: cont.

Outline of GLMs. Definitions

Statistics. Lecture 2 August 7, 2000 Frank Porter Caltech. The Fundamentals; Point Estimation. Maximum Likelihood, Least Squares and All That

Maximum Likelihood Estimation

Information in a Two-Stage Adaptive Optimal Design

Economics 536 Lecture 7. Introduction to Specification Testing in Dynamic Econometric Models

Sampling distribution of GLM regression coefficients

Transcription:

Maximum Likelihood Tests and Quasi-Maximum-Likelihood Wendelin Schnedler Department of Economics University of Heidelberg 10. Dezember 2007 Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 1 / 17

Objectives After this lecture, you should be able to test restrictions using the maximum likelihood method by examining the restriction (WALD test) comparing the value of the likelihood (likelihood ratio test) checking the first order condition (Lagrange multiplier test) decide when to apply which test explain twhat happens if the wrong density is used for the likelihood Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 2 / 17

Motivation often linear model not reasonable last time: non-linear relationships next time: limited dependent variables How can we estimate θ in such models? GMM, non-linear regression, or maximum likelihood today: maximum likelihood H 0 : Rθ = r, how can we test this? in linear model: F-test and now? Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 3 / 17

Maximum-Likelihood revisited (Y, X) f (y, x, θ) + data (y i, X i ) for i = 1,..., n, i.i.d. joint density evaluated at data: i f i(y i, X i, θ) this is called: likelihood taking the logarithm: l(θ) = i l i(θ) = i log(f i(y i, X i, θ)) gives log-likelihood maximum-likelihood θ ML estimator maximises (log-)likelihood necessary condition: l(θml ) = 0 or score is equal to zero properties: θ ML consistent n(θ ML θ) is asymptotically normal distributed with variation V = J 1, J : information matrix equal to Cramer-Rao lower bound = asymptotically efficient. Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 4 / 17

Consistent estimates of V Hessian notation of V : V (θ) = J 1 = ( 1 [ ] [ n n i=1 E 2 log(f i (y i,x i,θ)) ) 1 i.i.d = ( E 2 log(f (y,x,θ)) ]) 1. estimate: V (θ) H = ( 1 n n i=1 consistent estimator of V! 2 log(f i (y i,x i,θ ML )) ) 1 gradient notation of V : V (θ) = ( 1 ] n n i=1 [( E log(f i (y i,x i,θ ML )) )( log(f i (y i,x i,θ ML )) ) n i=1 ( log(f i (y i,x i,θ ML )) estimate: V (θ) G = ( 1 n consistent estimator of V ˆV G : Berndt Hall, Hall and Hausman (BHHH) estimator ) 1 )( log(f i (y i,x i,θ ML )) ) ) 1 Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 5 / 17

The Wald test H 0 : Rθ r = 0, number of restrictions: q test idea: how well does this equation hold? estimate without restriction θ ML How is n(rθ ML r) asymptotically distributed under H 0? under H 0 : E [ n(rθ ML r) ] = 0 asymptotically Why? because θ ML close to θ asy. variation of n(rθ ML r) is RVR n(rθ ML r) asy. N(0, RVR ) replace unknown V by consistent ˆV W := n (Rθ ML r) (R ˆV R ) 1 (Rθ ML r) asy. χ 2 (q) q number of restrictions (elements in r) Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 6 / 17

More on Wald test alternatively (but equivalent) test take q ˆV instead of V asy. q ˆV χ 2 (n k), k number of estimated parameters so qw = n(rθ ML r) (Rq ˆV R ) 1 (Rθ ML r) asy. F(q, n k) qw and W are asymptotically equivalent in linear model with normally distributed errors: ML estimator = OLS estimator variance of ML estimator: (X X) 1 σ 2 can be estimated by (X X) 1 s 2 then qw F(q, n k) normal linear model: wald test also valid for finite samples Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 7 / 17

Do restrictions have to be linear? until now H 0 : Rθ r = 0 mighty class of restrictions which includes θ1 = 4 with R = (1, 0,..., 0) and r = 4 θ1 = θ 2 with R = (1, 1,..., 0) and r = 0 θ1 = 2θ 2 with R = (1, 2,..., 0) and r = 0 θ 1 = θ2 2 with R = no! not possible extend tests such that restriction is h(θ) = 0 requirements: h(θ) full rank θ ML R in interior of parameter space replace Rθ r by h(θ) in all formulas and variance R ˆV R by h(θ) ˆV h(θ) Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 8 / 17

Likelihood ratio test H 0 : h(θ) = 0 test idea: how much larger is likelihood without H 0? estimate without restriction θ ML estimate with restriction θ ML R how? maximise likelihood with side-constraint h(θ) = 0 note that l(θ ML ) l(θ ML R ) 0 what is the distribution? expand using Taylor series: 2(l(θ ML ) l(θr ML = (θ ML θ ML (θ ML θ ML R )+rest χ2 (q) q number of restrictions (elements in r) )) Taylor useful for any two nested models R ) 2 l Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 9 / 17

Score test (Preparation) H 0 : h(θ) = 0 recall: s(θ ML ) := l(θml ) = 0 if H 0 valid: l(θml R ) 0 test idea: examine size of score how is the score distributed? ) = l(θml R ) = i ( log(f i (y i,x i,θr ML)) ) = i s i(θr ML). ) is asymptotically normal with s(θr ML thus 1 n sθml R E [ 1 n s(θml R )] = 0 and [ VAR 1 n s(θml R )] = ( 1 n estimate for J 1? as before [ ] n i=1 E 2 log(f i (y i,x i,θ ML R )) ) 1 = J 1 Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 10 / 17

Score test H 0 : h(θ) = 0 consider 1 n s(θml R ) V (θr ML) 1 n s(θml R number of restrictions in H 0. ) asy. score test or lagrange multiplier test why the name lagrange multiplier (LM-) test? maximise l(θ) under restriction H 0 : h(θ) = 0 χ 2 (q), where q is the lagrange approach:l(θ) λh(θ) first-order condition: l(θml R ) = λ h(θ), same as ) = λ h(θ) s(θ ML R tests based on s and λ are equivalent (as h(θ) has full rank) Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 11 / 17

Similarities of Wald, Score and Likelihood ratio test log likelihood Likelihood ratio Score unrestricted estimate θ ML Wald restricted estimate θ ML R parameter Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 12 / 17

Overview on tests how do you test hypothesis H 0 : h(θ) = 0 all three tests ( holy trinity ): asymptotically equivalent problem: which test should be used? criteria: ease of computation (number of maximisation problems): Wald: one (unrestricted estimate) Score test: one (restricted estimate) Likelihood ratio test: two (both) finite sample properties: few analytical results, Wald test is exact for linear model with normal errors robustness, sensitive to...... algebraically equivalent parameterisations? Wald: yes; Score: sometimes; LR: always... wrong specification of density? next Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 13 / 17

Maximum likelihood as a moment estimator assumed density: g(y i x i, θ), true density: f (y i x i, θ) what if ˆθ is moment estimator? if moment condition is valid, estimator is still correct can we write likelihood estimator as moment estimator? definition of density f (y i x i ; θ)dy i = 1 differentiate f (y i x i ; θ)dy i = 0 observe that log(f (y i x i ; θ)) = f (y i x i ;θ) f (y i x i ;θ) so that f (y i x i ; θ)dy i = log(f (y i x i ; θ))f (y i x i ; θ)dy i which means E [ log(f (y i x i ; θ)) ] = 0 valid moment condition for any density identical to first-order condition to identify ML estimator Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 14 / 17

Saving maximum likelihood estimation assumed density: g(y i x i, θ), true density: f (y i x i, θ) but E [ log(g(y i x i ; θ)) ] = 0 use method of moment techniques: replace theoretical moment by sample moment resulting moment estimator: quasi-maximum likelihood estimator result by Gourieroux/Monfort/Trognon (1984): n(ˆθ θ) asy N(0, V) with V = I 1 JI 1 and I = ( 1 [ n n i=1 E 2 log(g i (y i,x i,θ)) ]) 2 log(g i (y i,x i,ˆθ)) ) estimate: Î = ( 1 n n i=1 estimate J = ( 1 n n i=1 [( E log(f i (y i,x i,ˆθ)) Ĵ = 1 n n i=1 ( log(f i (y i,x i,ˆθ)) )( log(f i (y i,x i,ˆθ)) ) )( log(f i (y i,x i,ˆθ)) ) ] ) 1 by Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 15 / 17

More results on quasi-maximum likelihood if g = f it follows that J = I back to ordinary ML V is called sandwich formula; true information matrix is sandwiched by correction matrices estimating V by Î 1 ĴÎ 1 : sandwich estimator sandwich estimator leads to robust standard errors also called White or Huber estimates here: robustness with respect to density misspecification can be used in Wald test how? replace ˆV by ˆV by the way: linear model sandwich estimator is old friend: heteroscedasticity-consistent covariance estimator Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 16 / 17

Where are we? when estimating using maximum likelihood how can we test hypotheses? what happens if density is wrong? next: models where dependent variable is limited clue: can only be estimated with ML Wendelin Schnedler (AWI) Maximum Likelihood Tests and Quasi-Maximum-Likelihood10. Dezember 2007 17 / 17