ORTHOGONALITY TESTS IN LINEAR MODELS SEUNG CHAN AHN ARIZONA STATE UNIVERSITY ABSTRACT

Similar documents
Efficient Estimation of Dynamic Panel Data Models: Alternative Assumptions and Simplified Estimation

1. GENERAL DESCRIPTION

Increasing the Power of Specification Tests. November 18, 2018

Generalized Method of Moments (GMM) Estimation

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

An estimate of the long-run covariance matrix, Ω, is necessary to calculate asymptotic

Multiple Equation GMM with Common Coefficients: Panel Data

Chapter 2. GMM: Estimating Rational Expectations Models

Birkbeck Working Papers in Economics & Finance

Instrumental Variables and GMM: Estimation and Testing. Steven Stillman, New Zealand Department of Labour

Markov-Switching Models with Endogenous Explanatory Variables. Chang-Jin Kim 1

A Course on Advanced Econometrics

Economic modelling and forecasting

Econ 582 Fixed Effects Estimation of Panel Data

Generalized Method of Moment

Asymptotic Distributions of Instrumental Variables Statistics with Many Instruments

1 Outline. 1. Motivation. 2. SUR model. 3. Simultaneous equations. 4. Estimation

Lecture 4: Heteroskedasticity

Finite Sample Performance of A Minimum Distance Estimator Under Weak Instruments

Instrumental Variables Estimation in Stata

ASSET PRICING MODELS

Economics 582 Random Effects Estimation

Notes on Generalized Method of Moments Estimation

Improving GMM efficiency in dynamic models for panel data with mean stationarity

Chapter 1. GMM: Basic Concepts

This chapter reviews properties of regression estimators and test statistics based on

IV and IV-GMM. Christopher F Baum. EC 823: Applied Econometrics. Boston College, Spring 2014

IDENTIFICATION OF THE BINARY CHOICE MODEL WITH MISCLASSIFICATION

The Power of the KPSS Test for Cointegration when Residuals are Fractionally Integrated 1

Department of Econometrics and Business Statistics

Testing an Autoregressive Structure in Binary Time Series Models

Time Invariant Variables and Panel Data Models : A Generalised Frisch- Waugh Theorem and its Implications

Financial Econometrics Lecture 6: Testing the CAPM model

GMM Estimation and Testing

Econometrics of Panel Data

Testing Overidentifying Restrictions with Many Instruments and Heteroskedasticity

A Non-Parametric Approach of Heteroskedasticity Robust Estimation of Vector-Autoregressive (VAR) Models

Generalized Method of Moments Estimation

Missing dependent variables in panel data models

generalized method of moments estimation

Darmstadt Discussion Papers in Economics

EFFICIENT ESTIMATION USING PANEL DATA 1. INTRODUCTION

Maximum Likelihood (ML) Estimation

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

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

Advanced Econometrics

DSGE Methods. Estimation of DSGE models: GMM and Indirect Inference. Willi Mutschler, M.Sc.

Regression with time series

On asymptotic properties of Quasi-ML, GMM and. EL estimators of covariance structure models

DSGE-Models. Limited Information Estimation General Method of Moments and Indirect Inference

Estimating and Identifying Vector Autoregressions Under Diagonality and Block Exogeneity Restrictions

GMM Based Tests for Locally Misspeci ed Models

On Variance Estimation for 2SLS When Instruments Identify Different LATEs

Modified Generalized Instrumental Variables Estimation of Panel Data Models with Strictly Exogenous Instrumental Variables

GMM and SMM. 1. Hansen, L Large Sample Properties of Generalized Method of Moments Estimators, Econometrica, 50, p

A better way to bootstrap pairs

Slide Set 14 Inference Basded on the GMM Estimator. Econometrics Master in Economics and Finance (MEF) Università degli Studi di Napoli Federico II

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

the error term could vary over the observations, in ways that are related

GENERALIZEDMETHODOFMOMENTS I

Økonomisk Kandidateksamen 2004 (I) Econometrics 2. Rettevejledning

A Test of Cointegration Rank Based Title Component Analysis.

Simple Estimators for Semiparametric Multinomial Choice Models

An Introduction to Generalized Method of Moments. Chen,Rong aronge.net

Identification and Estimation Using Heteroscedasticity Without Instruments: The Binary Endogenous Regressor Case

xtseqreg: Sequential (two-stage) estimation of linear panel data models

The Estimation of Simultaneous Equation Models under Conditional Heteroscedasticity

Are Forecast Updates Progressive?

The properties of L p -GMM estimators

Heteroskedasticity and Autocorrelation

Heteroskedasticity-Robust Inference in Finite Samples

Advanced Econometrics

Ma 3/103: Lecture 25 Linear Regression II: Hypothesis Testing and ANOVA

INTERNALLY STUDENTIZED RESIDUALS TEST FOR MULTIPLE NON-NESTED REGRESSION MODELS

Ch 2: Simple Linear Regression

AN EFFICIENT GLS ESTIMATOR OF TRIANGULAR MODELS WITH COVARIANCE RESTRICTIONS*

Lecture 3 Stationary Processes and the Ergodic LLN (Reference Section 2.2, Hayashi)

Measurement Errors and the Kalman Filter: A Unified Exposition

Reliability of inference (1 of 2 lectures)

GMM-based inference in the AR(1) panel data model for parameter values where local identi cation fails

MIT Spring 2015

Empirical Economic Research, Part II

A forward demeaning transformation for a dynamic count panel data model *

Instrumental variables estimation using heteroskedasticity-based instruments

The regression model with one stochastic regressor.

[y i α βx i ] 2 (2) Q = i=1

APEC 8212: Econometric Analysis II

Department of Economics. Working Papers

A Robust Approach to Estimating Production Functions: Replication of the ACF procedure

Chapter 6. Panel Data. Joan Llull. Quantitative Statistical Methods II Barcelona GSE

Specification testing in panel data models estimated by fixed effects with instrumental variables

Econometric Analysis of Cross Section and Panel Data

Exogeneity tests and weak-identification

PANEL DATA RANDOM AND FIXED EFFECTS MODEL. Professor Menelaos Karanasos. December Panel Data (Institute) PANEL DATA December / 1

EFFICIENT ESTIMATION OF PANEL DATA MODELS WITH STRICTLY EXOGENOUS EXPLANATORY VARIABLES

Cointegration Lecture I: Introduction

EFFICIENT ESTIMATION OF MODELS FOR DYNAMIC PANEL DATA. Seung C. Ahn. Arizona State University. Peter Schmidt. Michigan State University

Motivation Non-linear Rational Expectations The Permanent Income Hypothesis The Log of Gravity Non-linear IV Estimation Summary.

GMM - Generalized method of moments

Volume 30, Issue 1. Measuring the Intertemporal Elasticity of Substitution for Consumption: Some Evidence from Japan

Transcription:

ORTHOGONALITY TESTS IN LINEAR MODELS SEUNG CHAN AHN ARIZONA STATE UNIVERSITY ABSTRACT This paper considers several tests of orthogonality conditions in linear models where stochastic errors may be heteroskedastic or autocorrelated. It is shown that these tests can be performed with Wald statistics obtained from simple auxiliary regressions. JEL Classification Number: C20 Word Count: 1460 Address for correspondence: Seung Chan Ahn, Department of Economics, Arizona State University, Tempe, AZ 85287-3806, USA, Phone: 602-965-6574, Fax: 602-965-0748, E-mail: AASCA.ASUACAD.BITNET August, 1994 Revised March, 1995 Revised August, 1995 The author gratefully acknowledges the financial support of the College of Business, Arizona State University.

I. INTRODUCTION This paper considers tests of orthogonality conditions for linear models. The tests of interest are the GMM (generalized method of moments) tests of Hansen (1982), Newey (1985), and Eichenbaum, Hansen and Singleton (1988, hereafter, EHS). These tests are convenient for models in which stochastic error terms are independently and identically distributed (i.i.d.), because each of the test statistics can be computed as the number of observations times the uncentered R 2 from an auxiliary regression. [See Newey (1985) and Pesaran and Smith (1990).] This paper considers cases in which the errors are not i.i.d., and shows that the GMM tests can be easily performed with Wald statistics obtained from auxiliary regressions. This paper is organized as follows. Section II considers the Hansen test, and derives its Wald-type representation. Section III considers the tests of Newey (1985) and EHS (1988), establishes the numerical equivalence of the two statistics, and shows how they can be obtained by a simple auxiliary regression. The relation of these two tests to the Hausman test (1978, 1984) is also examined. Some concluding remarks are contained in Section IV. II. THE HANSEN TEST AND ITS WALD-TYPE ALTERNATIVE Consider the linear regression model: (1) where y is the T 1vector of a dependent variable, X is the T pmatrix of regressors, and ε is the vector of errors. Here we allow the errors to be heteroskedastic and/or autocorrelated. Some variables in X are assumed to be correlated with ε, so that consistent estimation of β requires instrumental-variables methods. Let Z be at q(q p) matrix of instrumental

2 variables which satisfy the orthogonality hypothesis: (2) Under H o, the optimal GMM estimator is β =(X ZV -1 Z X) -1 X ZV -1 Z y, where V = cov(z ε). Following White (1982), we will call any estimator of this form a two-stage instrumentalvariables (2SIV) estimator. In practice, a consistent estimator of V should be used for the computation of β. For the appropriate estimates of V, see Newey and West (1987b). Once β is computed, H o can be tested by the Hansen (1982) statistic, which is defined by J T =(y- Xβ ) ZV -1 Z (y-xβ ). Under H o, this statistic has a χ 2 distribution with (q - p) degrees of freedom. For an alternative way to compute J T, consider an auxiliary regression model: (3) where Z A isanyt (q-p)submatrix of Z whose columns are not in X. Since model (2) is exactly identified under H o, the 2SLS estimator of ξ A equals ξˆ A =[βˆa,δˆa ] =(Z X A ) -1 Z y. Define W A T = δˆ A [cov(δˆ A)] -1 δˆ A, which is the Wald statistic for testing the restriction δ A =0. Then, we obtain the following: Proposition 1. W A T=J T, if the same estimator of V is used for both statistics. It is important to note that the numerical equivalence of W A T and J T may not hold if different estimators of V are used. However, the difference between these statistics is asymptotically negligible. In practice, W A T is easier to compute than J T. Most available software packages can compute 2SLS estimates and their covariance matrix by the method of Newey and West (1987b). Using such software, researchers can perform the Hansen test with

3 the Wald statistic for the significance of Z A in model (3). Testing for the orthogonality of the whole set X is further simplified. For this case, Z coincides with X A, so that the Hansen statistic can be computed as a Wald statistic based on the OLS estimator of model (3). III. TESTING A SUBSET OF ORTHOGONALITY CONDITIONS Researchers may wish to test the orthogonality of a subset of instrumental variables, when prior information about the orthogonality of other instruments is available. For such cases, assume that the alternative hypothesis of H o is given by: (4) where Z=[Z 1,Z 2 ], and Z 1 and Z 2 include q 1 and q 2 (=q-q 1 ) instruments, respectively. According to Z 1 and Z 2, partition V into [V ij ], where i, j =1, 2. We assume q 1 p, so that β can be estimated by 2SIV using Z 1 only. We denote this 2SIV estimator by βˆ =[X Z 1 (V 11 ) - 1 Z 1 X] -1 X Z 1 (V 11 ) -1 Z 1 y. Newey (1985) considers an optimal GMM statistic for testing H o against HA, s which is a Wald statistic based on r=z 2 (y-xβˆ)-v 21 (V 11 ) -1 Z 1 (y-xβˆ); that is, M T =r R -1 r, where F = Z 2 X-V 21 (V 11 ) -1 Z 1 X and R (= cov(r)) = V 22 -V 21 (V 11 ) -1 V 12 + F[X Z 1 (V 11 ) -1 Z 1 X] -1 F. Under H o, this statistic has a χ 2 distribution with q 2 degrees of freedom. The advantage of M T over J T is that its power is focused on the orthogonality of Z 2. Following EHS (1988), we can also consider a statistic of the likelihood-ratio type, which is given by D T =J T -J 1 T, where J 1 T =(y-xβˆ) Z 1 (V 11 ) -1 Z 1 (y-xβˆ). Proposition 1 suggests an alternative way to compute D T. Similarly to J T,J 1 T can be obtained by a Wald statistic from

4 an artificial 2SLS regression. Thus, D T can be computed as the difference between two Wald statistics. Another interesting and novel property of D T, which we formally state below, is that it is in fact numerically equivalent to the optimal GMM statistic M T. Alternatively, both of M T and D T can be obtained from a 2SIV estimation of the model: (5) where Q 1 =I T -Z 1 (Z 1 Z 1 ) -1 Z 1. Pesaran and Smith (1990) consider 2SLS of the same model with i.i.d. errors, and find that the Wald statistic of the restriction δ B = 0 is equivalent to M T. Their result can be extended to the cases with non-i.i.d. errors. Denoting the 2SIV estimator of ξ B by ˆξ B =[ˆβ B,δˆB ] =[X B ZV -1 Z X B ] -1 X B ZV -1 Z y, define the Wald statistic based on δˆ B by W B T = δˆ B [cov(δˆ B)] -1 δˆ B. Then, we obtain the following result: Proposition 2. M T =D T =W B T, if the same estimator of V is used for the statistics. Finally, we consider a Hausman (1978, 1984) statistic for testing H o against HA, s which is given by H T =(ˆβ-β ) [cov(ˆβ - β )] - (ˆβ - β ), where ( ) - is a g-inverse. Under H o,h T has a χ 2 distribution with degrees of freedom equal to Rank[cov( ˆβ - β )]. Newey (1985) has shown that H T =M T if Rank[cov(ˆβ - β )]=q 2. The following proposition reveals a general link between H T and M T (= D T =W B T): Proposition 3. Let r H =F R -1 r and R H (= cov(r H ))=F R -1 F. Define a Wald statistic based on r H by M H T =r H (R H ) - r H. Then, H T =M H T, when the same estimator of V is used. Comparing the forms of M H T and M T, we can see that M H T (and H T ) is the statistic for testing the hypothesis, H H o: E[F R -1 E(Z 2 ε Z)] = 0. Indeed, it can be shown that β is consistent under H H o. This implies that H T is for testing the consistency of β, not H o (against HA) s itself. Ruud (1984) finds a similar result in a maximum-likelihood framework.

5 IV. CONCLUSION This paper has considered several popular GMM tests of orthogonality conditions in linear models. A remarkable feature of the GMM tests is that they are equivalent to Wald tests of exclusive restrictions imposed in artificial models. The test strategies considered in this paper are convenient for models with heteroskedastic and/or autocorrelated errors. APPENDIX: Proofs of Propositions We briefly sketch the proofs of Propositions 1-3. More detailed proofs can be found in the earlier version of this paper. Proposition 1. Define J A T(ξ A )=(y-x A ξ A ) ZV -1 Z (y-x A ξ A ). ξˆ A is the unrestricted GMM estimator minimizing J A T(ξ A ) [because Z (y-x A ξˆ A) = 0], while ξ A =( β,0 ) is the restricted GMM estimator that minimizes J A T(ξ A ) subject to the restriction δ A = 0. Then, Proposition 4 of Newey and West (1987a) implies that W A T =J A T(ξ A)-J A T(ξˆA)=J T. Proposition 2. Applying the separability result of Ahn and Schmidt (1995), we can show that δˆ B =L -1 r, where L=Z 2 Q 1 Z 2. Some tedious but straightforward algebra also shows that cov(δˆ B) =C B (X B ZV -1 Z X B ) -1 C B =L -1 RL -1, where C B =[0 q p,i q q2 ]. Substituting these results into W B T yields M T =W B T. The equality of D T and W B T can be shown similarly to the proof of Proposition 1. Proposition 3. Letting A=X Z 1 (V 11 ) -1 Z 1 X, we can show that βˆ - β = -A -1 F R -1 r and cov(βˆ - β) = cov(βˆ) - cov( β) =[X Z 1 (V 11 ) -1 Z 1 X] -1 -[X ZV -1 Z X] -1 =A -1 F R -1 FA -1. Substituting these results into H T gives us the desired equality.

REFERENCES Ahn, S.C. and Schmidt, P. (1995). A Separability Result for GMM Estimation, with Applications to GLS Prediction and Conditional Moment Tests, Econometric Reviews, Vol. 14, pp. 19-34. Eichenbaum, M.S., Hansen, L.P. and Singleton, K.J. (1988). A Time Series Analysis of Representative Agent Models of Consumption and Leisure Choice under Uncertainty, Quarterly Journal of Economics, Vol. 103, pp. 51-78. Hansen, L. (1982). Large Sample Properties of Generalized Method of Moments Estimators, Econometrica, Vol. 50, pp. 1029-1054. Hausman, J.A. (1978). Specification Tests in Econometrics, Econometrica, Vol. 46, pp. 1251-1271. Hausman, J.A. (1984). Specification and Estimation of Simultaneous Equations Model, Chapter 13 in: Griliches, Z. and Intriligator, M.D., eds., Handbook of Econometrics (Amsterdam: North-Holland). Newey, W. (1985). Generalized Method of Moments Specification Testing, Journal of Econometrics, Vol. 29, pp. 229-256. Newey, W. and West, K.D. (1987a). Hypothesis Testing with Efficient Method of Moments Estimation, International Economic Review, Vol. 28, pp. 777-787. Newey, W. and West, K.D. (1987b). A Simple, Positive Definite, Heteroskedasticity and Autocorrelation Consistent Covariance Matrix, Econometrica, Vol. 55, pp. 703-708. Pesaran, M.H. and Smith, R.J. (1990). A Unified Approach to Estimation and Orthogonality Tests in Linear Single-Equation Econometric Models, Journal of Econometrics, Vol. 44, pp. 41-66. Ruud, P.A. (1984). Tests of Specification in Econometrics, Econometric Reviews, Vol. 3, pp. 211-242. White, H. (1982). Instrumental Variables Regression with Independent Observations, Econometrica, Vol. 50, pp. 483-499.