Chapter 5: The Generalized Linear Regression Model and Heteroscedasticity Advanced Econometrics - HEC Lausanne Christophe Hurlin University of Orléans December 15, 2013 Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 1 / 153 Section 1 Introduction Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 2 / 153 1. Introduction The outline of this chapter is the following: Section 2. The generalized linear regression model Section 3. Ine¢ ciency of the Ordinary Least Squares Section 4. Generalized Least Squares (GLS) Section 5. Heteroscedasticity Section 6. Testing for heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 3 / 153 1. Introduction References Amemiya T. (1985), Advanced Econometrics. Harvard University Press. Greene W. (2007), Econometric Analysis, sixth edition, Pearson - Prentice Hil (recommended) Pelgrin, F. (2010), Lecture notes Advanced Econometrics, HEC Lausanne (a special thank) Ruud P., (2000) An introduction to Classical Econometric Theory, Oxford University Press. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 4 / 153 1. Introduction Notations: In this chapter, I will (try to...) follow some conventions of notation. fY ( y ) probability density or mass function FY ( y ) cumulative distribution function Pr () probability y vector Y matrix Be careful: in this chapter, I don’t distinguish between a random vector (matrix) and a vector (matrix) of deterministic elements (except in section 2). For more appropriate notations, see: Abadir and Magnus (2002), Notation in econometrics: a proposal for a standard, Econometrics Journal. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 5 / 153 Section 2 The generalized linear regression model Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 6 / 153 2. The generalized linear regression model Objectives The objective of this section are the following: 1 De…ne the generalized linear regression model 2 De…ne the concept of heteroscedasticity 3 De…ne the concept of autocorrelation (or correlation) of disturbances Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 7 / 153 2. The generalized linear regression model Consider the (population) multiple linear regression model: y = Xβ + ε where (cf. chapter 3): y is a N 1 vector of observations yi for i = 1, .., N X is a N K matrix of K explicative variables xik for k = 1, ., K and i = 1, .., N ε is a N 1 vector of error terms εi . β = ( β1 ..βK )> is a K Christophe Hurlin (University of Orléans) 1 vector of parameters Advanced Econometrics - HEC Lausanne December 15, 2013 8 / 153 2. The generalized linear regression model In chapter 3 (linear regression model), we assume spherical disturbances (assumption A4): V ( ε j X ) = σ 2 IN In this chapter, we will relax the assumption that the errors are independent and/or identically distributed and we will study: 1 Heteroscedasticity 2 Autocorrelation or correlation. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 9 / 153 2. The generalized linear regression model De…nition (Generalized linear regression model) The generalized linear regression model is de…ned as to be: y = Xβ + ε where X is a matrix of …xed or random regressors, β 2 RK , and the error term ε satis…es: E ( εj X) = 0N 1 V ( ε j X) = Σ = σ 2 Ω where Ω and Σ are symmetric positive de…nite matrices. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 10 / 153 2. The generalized linear regression model Reminder V ( εj X) = E εε> X | {z } | {z } N N 0 N N V ε21 X Cov ( ε1 ε2 j X) B E ( ε2 ε1 j X) V ε22 X =B @ .. .. Cov ( εN ε1 j X) .. Christophe Hurlin (University of Orléans) 1 .. Cov ( ε1 εN j X) .. Cov ( ε2 εN j X) C C A .. .. 2 .. V εN X Advanced Econometrics - HEC Lausanne December 15, 2013 11 / 153 2. The generalized linear regression model Remark In the generalized linear regression model, we have V ( ε j X) = Σ = σ 2 Ω with 0 σ21 σ12 B σ21 σ2 2 Σ=B @ .. .. σN 1 .. and ω ij = σij /σ2 . Christophe Hurlin (University of Orléans) 1 0 .. σ1N ω 11 ω 12 C B .. σ2N C ω 21 ω 22 = σ2 B @ .. .. .. A .. 2 .. σN ωN 1 .. Advanced Econometrics - HEC Lausanne 1 .. ω 1N .. ω 2N C C .. .. A .. ω NN December 15, 2013 12 / 153 2. The generalized linear regression model De…nition (Heteroscedasticity) Disturbances are heteroscedastic when they have di¤erent (conditional) variances: V ( εi j X) 6= V ( εj j X) for i 6= j Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 13 / 153 2. The generalized linear regression model Remarks 1 Heteroscedasticity often arises in volatile high-frequency time-series data such as daily observations in …nancial markets. 2 Heteroscedasticity often arises in cross-section data where the scale of the dependent variable and the explanatory power of the model tend to vary across observations. Microeconomic data such as expenditure surveys are typical Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 14 / 153 2. The generalized linear regression model Example (Heteroscedasticity) If the disturbances are heteroscedastic but they are still assumed uncorrelated across observations, so Ω and Σ would be: 0 2 1 0 σ1 0 .. 0 ω 1 0 .. 0 B 0 σ2 .. 0 C B ω 2 .. 0 2 2B 0 2 C Σ=B @ .. .. .. .. A = σ Ω = σ @ .. .. .. .. 0 .. .. σ2N 0 .. .. ω N to be with ω i = σ2i /σ2 for i = 1, .., N. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 1 C C A 15 / 153 2. The generalized linear regression model De…nition (Autocorrelation) Disturbances are autocorrelated (or correlated) when: Cov ( εi , εj j X) 6= 0 Christophe Hurlin (University of Orléans) for i 6= j Advanced Econometrics - HEC Lausanne December 15, 2013 16 / 153 2. The generalized linear regression model Example (Autocorrelation) For instance, time-series data are usually homoscedastic, but autocorrelated, so Ω and Σ would be: 0 2 1 0 σ σ12 .. σ1N 1 ω 12 .. ω 1N B σ21 σ2 .. σ2N C B ω 1 .. ω 2N C = σ2 Ω = σ2 B 21 Σ=B @ .. A @ .. .. .. .. .. .. .. σN 1 .. .. σ2 ωN 1 .. .. 1 1 C C A with ω ij = σij /σ2 for i = 1, .., N denotes the correlation (autocorrelation) ω ij = Christophe Hurlin (University of Orléans) σij = cor (εi , εj ) σ2 Advanced Econometrics - HEC Lausanne December 15, 2013 17 / 153 2. The generalized linear regression model Key Concepts 1 The generalized linear regression model 2 Heteroscedasticity 3 Autocorrelation (or correlation) of disturbances Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 18 / 153 Section 3 Ine¢ ciency of the Ordinary Least Squares Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 19 / 153 3. Ine¢ ciency of the Ordinary Least Squares Objectives The objective of this section are the following: 1 Study the properties of the OLS estimator in the generalized linear regression model 2 Study the …nite sample properties of the OLS 3 Study the asymptotic properties of the OLS 4 Introduce the concept of robust / non-robust inference Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 20 / 153 3. Ine¢ ciency of the Ordinary Least Squares Introduction Assume that the data are generated by the generalized linear regression model: y = Xβ + ε E ( εj X) = 0N 1 V ( ε j X) = σ 2 Ω = Σ b Now consider the OLS estimator, denoted β OLS , of the parameters β: > b β OLS = X X 1 X> y We will study its …nite sample and asymptotic properties. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 21 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Assumption 3: Strict exogeneity of the regressors) The regressors are exogenous in the sense that: E ( εj X) = 0N Christophe Hurlin (University of Orléans) 1 Advanced Econometrics - HEC Lausanne December 15, 2013 22 / 153 3. Ine¢ ciency of the Ordinary Least Squares Finite sample properties of the OLS estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 23 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Bias) In the generalized linear regression model, under the assumption A3 (exogeneity), the OLS estimator is unbiased: b E β OLS = β0 where β0 denotes the true value of the parameters. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 24 / 153 3. Ine¢ ciency of the Ordinary Least Squares Remark Heteroscedasticity and/or autocorrelation don’t induce a bias for the OLS estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 25 / 153 3. Ine¢ ciency of the Ordinary Least Squares Proof > b β OLS = X X 1 X> y = β 0 + X> X 1 X> ε So we have: > b E β OLS X = β0 + X X 1 X> E ( ε j X ) Under assumption A3 (exogeneity), E ( εj X) = 0. Then, we get: b E β OLS X = β0 Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 26 / 153 3. Ine¢ ciency of the Ordinary Least Squares Proof (cont’d) b E β OLS X = β0 So, we have: b E β OLS b = EX E β OLS X = EX ( β 0 ) = β 0 where EX denotes the expectation with respect to the distribution of X. The OLS estimator is unbiased: b E β OLS Christophe Hurlin (University of Orléans) = β0 Advanced Econometrics - HEC Lausanne December 15, 2013 27 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Bias) In the generalized linear regression model, under the assumption A3 (exogeneity), the OLS estimator has a conditional variance covariance matrix given by 2 > b V β OLS X = σ0 X X 1 X> ΩX X> X 1 and a variance covariance matrix given by: b V β OLS Christophe Hurlin (University of Orléans) b = EX V β OLS X Advanced Econometrics - HEC Lausanne December 15, 2013 28 / 153 3. Ine¢ ciency of the Ordinary Least Squares Proof > b β OLS = X X 1 X> y = β 0 + X> X 1 X> ε So we have: b V β OLS X = E = X> X = σ20 X> X Christophe Hurlin (University of Orléans) 1 X> X 1 X> εε> X X> X 1 X> E εε> X X X> X 1 X> ΩX X> X Advanced Econometrics - HEC Lausanne X 1 1 December 15, 2013 29 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Variance estimator) An estimator of the variance covariance matrix of the OLS estimator b β OLS is given by b b β V OLS b 2 X> X =σ 1 b X> ΩX X> X 1 b is a consistent estimator of Σ = σ 2 Ω. This estimator holds b2 Ω where σ whether X is stochastic or non-stochastic. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 30 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Normality assumption) Under assumptions A3 (exogeneity) and A6 (normality), the OLS estimator obtained in the generalized linear regression model has an (exact) normal conditional distribution: b β OLS X N Christophe Hurlin (University of Orléans) β 0 , σ 2 X> X 1 X> ΩX X> X Advanced Econometrics - HEC Lausanne 1 December 15, 2013 31 / 153 3. Ine¢ ciency of the Ordinary Least Squares Asymptotic properties of the OLS estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 32 / 153 3. Ine¢ ciency of the Ordinary Least Squares Assumptions plim plim 1 > X X=Q N 1 > X ΩX = Q N where: 1 Q is a K 2 Q is a K K …nite (non null) de…nite positive matrix K …nite (non null) de…nite positive matrix with rank (Q) = K Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 33 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Consistency of the OLS estimator) If plim N 1 X> ΩX and plim N 1 X> X are both …nite positive de…nite b matrices, then β OLS is a consistent estimator of β: p b β OLS ! β0 Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 34 / 153 3. Ine¢ ciency of the Ordinary Least Squares Proof > b β OLS = β0 + X X 1 X> ε We know that under assumption A3 (exogeneity): 1 > X ε = 0K 1 N 1 plim X> X = Q N plim So, we have b plim β OLS = β0 b is consistent. So, the estimator β Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 35 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Asymptotic distribution of the OLS) If the regressors are su¢ ciently well behaved and the o¤-diagonal terms in diminish su¢ ciently rapidly, then the least squares estimator is asymptotically normally distributed with p where b N β OLS Q = plim Christophe Hurlin (University of Orléans) d β0 ! N 0, σ2 Q 1 > X X N Q = plim Advanced Econometrics - HEC Lausanne 1 Q Q 1 1 > X ΩX N December 15, 2013 36 / 153 3. Ine¢ ciency of the Ordinary Least Squares Remark 1 Regularity conditions include the exogeneity conditions, but also (i) the regressors are su¢ ciently well-behaved and (ii) the o¤-diagonal terms of the variance-covariance matrix diminish su¢ ciently rapidly (relative to the diagonal elements). 2 For a formal proof in a general case, see Amemiya (1985, p. 187). Amemiya T. (1985), Advanced Econometrics. Harvard University Press. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 37 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Asymptotic variance) Under suitable regularity conditions, the asymptotic variance covariance b is given by: matrix of the OLS estimator β with b Vasy β OLS Q = plim Christophe Hurlin (University of Orléans) 1 > X X N = σ2 Q N 1 Q Q Q = plim Advanced Econometrics - HEC Lausanne 1 1 > X ΩX N December 15, 2013 38 / 153 3. Ine¢ ciency of the Ordinary Least Squares Fact (Non-robust inference) Because the variance of the least squares estimator is not σ 2 X> X b 2 X> X σ 1 1 statistical inference ( non-robust inference) based on may be misleading. For instance the t-test-statistic: t βk = b β pk b mkk σ where mkk is kth diagonal element of X> X do not have a Student distribution. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 39 / 153 3. Ine¢ ciency of the Ordinary Least Squares Robust / Non-robust inference As a consequence, the familiar inference procedures based on the F and t distributions will no longer be appropriate. b The question is to know how to estimate V β OLS in the context of the linear generalized regression model in order to make robust inference. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 40 / 153 3. Ine¢ ciency of the Ordinary Least Squares De…nition (Estimator of the asymptotic variance covariance matrix) If Σ = σ2 Ω were known, the consistent estimator of the (asymptotic) b variance covariance of β OLS would be: b b asy β V OLS = Christophe Hurlin (University of Orléans) σ2 N 1 > X X N 1 1 > X ΩX N Advanced Econometrics - HEC Lausanne 1 > X X N 1 December 15, 2013 41 / 153 3. Ine¢ ciency of the Ordinary Least Squares Proof By de…nition: 1 > X X N 1 Q = plim X> ΩX N Q = plim So, b b asy β plim V OLS = plim = σ2 N σ2 Q N 1 1 > X X N Q Q 1 1 > X ΩX N 1 > X X N 1 1 Or equivalently b b asy β V OLS Christophe Hurlin (University of Orléans) p b ! Vasy β OLS Advanced Econometrics - HEC Lausanne December 15, 2013 42 / 153 3. Ine¢ ciency of the Ordinary Least Squares Reminder X> X = N ∑ xi xi> i =1 X> ΩX = N N ∑ ∑ ωij xi xi> i =1 j =1 X> ΣX = N ∑ N ∑ σij xi xi> = σ2 i =1 j =1 Christophe Hurlin (University of Orléans) N N ∑ ∑ ωij xi xi> i =1 j =1 Advanced Econometrics - HEC Lausanne December 15, 2013 43 / 153 3. Ine¢ ciency of the Ordinary Least Squares Remark The estimator b b asy β V OLS σ2 = N 1 1 > X X N 1 > X ΩX N 1 1 > X X N can also be written as b b asy β V OLS σ2 = N 1 N Christophe Hurlin (University of Orléans) N ∑ i =1 xi xi> ! 1 1 N N N ∑∑ i =1 j =1 Advanced Econometrics - HEC Lausanne ω ij xi xi> ! 1 N N ∑ i =1 December 15, 2013 xi xi> ! 1 44 / 153 3. Ine¢ ciency of the Ordinary Least Squares Remark b b asy β In the next section, we will give a feasible estimator V OLS speci…c case of an heteroscedastic model. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne in the December 15, 2013 45 / 153 3. Ine¢ ciency of the Ordinary Least Squares Summary In the GLR model, under some regularity conditions: 1 The OLS estimator is unbiased 2 The OLS estimator is (weakly) consistent 3 The OLS estimator is asymptotically normally distributed Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 46 / 153 3. Ine¢ ciency of the Ordinary Least Squares Summary But... 1 The inference based on the estimator σ2 X> X 2 The OLS is ine¢ cient. b V β OLS Christophe Hurlin (University of Orléans) 1 is misleading. I N 1 ( β0 ) is a positive de…nite matrix Advanced Econometrics - HEC Lausanne December 15, 2013 47 / 153 3. Ine¢ ciency of the Ordinary Least Squares Key Concepts 1 OLS estimator in the generalized regression model 2 Finite sample properties 3 Asymptotic variance covariance matrix of the OLS estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 48 / 153 Section 4 Generalized Least Squares (GLS) Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 49 / 153 4. Generalized Least Squares (GLS) Objectives The objective of this section are the following: 1 De…ne the Generalized Least Squares (GLS) 2 De…ne the Feasible Generalized Least Squares (FGLS) 3 Study the statistical properties of the GLS and FGLS estimators Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 50 / 153 4. Generalized Least Squares (GLS) Consider the generalized linear regression model with V ( ε j X) = Σ = σ 2 Ω We will distinguish two cases: Case 1: the variance covariance matrix Σ is known (unrealistic case) Case 2: the variance covariance matrix Σ is unknown Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 51 / 153 4. Generalized Least Squares (GLS) Case 1: Σ is known The Generalized Least Squares (GLS) estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 52 / 153 4. Generalized Least Squares (GLS) De…nition (Factorisation) Since Ω is a positive de…nite matrix, it can factored as follows: Ω = CΛC> where the columns of C are the characteristics vectors of Ω, the characteristic roots of Ω are arrayed in the diagonal matrix Λ, and C> C = CC> = IN where I denotes the identity matrix. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 53 / 153 4. Generalized Least Squares (GLS) De…nition We de…ne the matrix P such that P> = CΛ 1/2 so that Ω Christophe Hurlin (University of Orléans) 1 = P> P Advanced Econometrics - HEC Lausanne December 15, 2013 54 / 153 4. Generalized Least Squares (GLS) Proof P> = CΛ Since Λ is diagonal, Λ 1/2 Λ P> P = CΛ 1/2 1/2 1/2 =Λ 1 Λ C> = CΛ 1/2 , and we have: 1 C> Consider the quantity P> PΩ: P> PΩ = CΛ 1 C> CΛC> = CΛ 1 ΛC> = CC> = IN Since C satis…es CC> = IN . Then, P> P = Ω Christophe Hurlin (University of Orléans) 1 Advanced Econometrics - HEC Lausanne December 15, 2013 55 / 153 4. Generalized Least Squares (GLS) GLS estimator Premultiply the generalized linear regression model by P to obtain Py = PXβ + Pε or equivalently y = X β+ε The conditional variance of ε is V ( ε j X) = E ε ε > X = PE εε> X P> = σ2 PΩP> = σ2 Λ 1/2 C> CΛC> CΛ = σ 2 IN Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne 1/2 December 15, 2013 56 / 153 4. Generalized Least Squares (GLS) GLS estimator (cont’d) y = X β+ε V ( ε j X ) = σ 2 IN The classical regression model applies to this transformed model. If Ω is assumed to be known, y = Py and X = PX are observed data. So, we can apply the ordinary least squares to this transformed model: b= X β Christophe Hurlin (University of Orléans) > X 1 X > y Advanced Econometrics - HEC Lausanne December 15, 2013 57 / 153 4. Generalized Least Squares (GLS) GLS estimator (cont’d) b = β X > 1 X = X> P> PX = X> Ω 1 X X 1 1 > y X> P> Py X> Ω 1 y This estimator is the generalized least squares (GLS) estimator of β. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 58 / 153 4. Generalized Least Squares (GLS) De…nition (GLS estimator) The Generalized Least Squares (GLS) estimator of β is de…ned as to be: > b β GLS = X Ω Christophe Hurlin (University of Orléans) 1 X 1 X> Ω Advanced Econometrics - HEC Lausanne 1 y December 15, 2013 59 / 153 4. Generalized Least Squares (GLS) De…nition (Bias) b Under the exogeneity assumption (A3), the estimator β GLS is unbiased: b E β GLS = β0 where β0 denotes the true value of the parameters. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 60 / 153 4. Generalized Least Squares (GLS) Proof We have: > b β GLS = X Ω 1 1 X X> Ω 1 y = β 0 + X> Ω So, > b E β GLS X = β0 + X Ω 1 X 1 X> Ω 1 X 1 1 X> Ω 1 ε E ( ε j X) Under the exogeneity assumption A3, E ( εj X) = 0, so we have and b E β GLS b E β GLS X = β0 b = EX E β GLS X Christophe Hurlin (University of Orléans) = EX ( β 0 ) = β 0 Advanced Econometrics - HEC Lausanne December 15, 2013 61 / 153 4. Generalized Least Squares (GLS) De…nition (Variance covariance matrix) b The conditional variance covariance matrix of the estimator β GLS is de…ned as to be: 2 b V β X> Ω GLS X = σ 1 X 1 X 1 The variance covariance matrix is given by b V β GLS Christophe Hurlin (University of Orléans) = σ2 EX X> Ω Advanced Econometrics - HEC Lausanne 1 December 15, 2013 62 / 153 4. Generalized Least Squares (GLS) Proof b Consider the de…nition of β GLS in the transformed model: b β GLS = β0 + X b V β GLS X = X Since E ε ε > > X 1 > X 1 X > E ε ε X > > X X ε X > X 1 X = σ2 IN , we have b V β GLS X = σ2 X > X = σ2 X > X 1 Christophe Hurlin (University of Orléans) > X X > X 1 1 = σ2 X> P> PX = σ 2 X> Ω X 1 X 1 1 Advanced Econometrics - HEC Lausanne December 15, 2013 63 / 153 4. Generalized Least Squares (GLS) De…nition (Consistency) b Under the exogeneity assumption A3, the GLS estimator β GLS is (weakly) consistent: p b β GLS ! β0 as soon as 1 > X X =Q N where Q is a …nite positive de…nite matrix. plim Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 64 / 153 4. Generalized Least Squares (GLS) Proof > b β GLS = β0 + X Ω 1 X 1 X> Ω 1 ε Under the assumption A3 (exogeneity): plim 1 > X Ω N plim 1 > X Ω N 1 ε = 0K 1 1 X=Q So, we have b plim β GLS = β0 b The estimator β GLS is weakly consistent. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 65 / 153 4. Generalized Least Squares (GLS) De…nition (Asymptotic distribution) b Under some regularity conditions, the GLS estimator β GLS is asymptotically normally distributed: p where b N β GLS Q = plim Christophe Hurlin (University of Orléans) 1 X N d β0 ! N 0, σ2 Q > X = plim 1 > X Ω N Advanced Econometrics - HEC Lausanne 1 1 X December 15, 2013 66 / 153 4. Generalized Least Squares (GLS) De…nition (Asymptotic variance covariance matrix) b The asymptotic variance covariance matrix of the estimator β GLS is: b Vasy β GLS = σ2 Q N 1 If Σ = σ2 Ω is known, a consistent estimator is given by: b b asy β V GLS = σ2 X> Ω N 1 X 1 This estimator holds whether X is stochastic or non-stochastic. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 67 / 153 4. Generalized Least Squares (GLS) Theorem (BLUE estimator) b The GLS estimator β GLS is the minimum variance linear unbiased estimator ( BLUE estimator) in the semi-parametric generalized linear regression model. In particular, the matrix de…ned by: b Vasy β OLS is a positive semi de…nite matrix. Christophe Hurlin (University of Orléans) b Vasy β GLS Advanced Econometrics - HEC Lausanne December 15, 2013 68 / 153 4. Generalized Least Squares (GLS) Theorem (E¢ ciency) Under suitable regularity conditions, in a parametric generalized linear b regression model, the GLS estimator β GLS is e¢ cient b V β GLS = I N 1 ( β0 ) where I N 1 ( β0 ) denotes the FDCR or Cramer-Rao bound. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 69 / 153 4. Generalized Least Squares (GLS) Remark In a Gaussian generalized linear regression model (under assumption A6), the likelihood of the sample is given by: LN (θ; y j x ) = 2πσ2 exp N /2 jΩj 1 (y 2σ2 N /2 Xβ)> Ω 1 (y Xβ) The log-likelihood is de…ned as to be: `N (θ; y j x ) = Christophe Hurlin (University of Orléans) N N ln 2πσ2 log (jΩj) 2 2 1 (y Xβ)> Ω 1 (y Xβ) 2σ2 Advanced Econometrics - HEC Lausanne December 15, 2013 70 / 153 4. Generalized Least Squares (GLS) Remark For testing hypotheses, we can apply the full set of results in Chapter 4 to the transformed model. For instance, for testing the p linear constraints H0 : Rβ = q, the appropriate test-statistic is: F= 1 b Rβ GLS p q Christophe Hurlin (University of Orléans) > σ 2 R X> Ω 1 X 1 1 R> Advanced Econometrics - HEC Lausanne b Rβ GLS q December 15, 2013 71 / 153 4. Generalized Least Squares (GLS) Fact To summarize, all the results for the classical model, including the usual inference procedures, apply to the transformed model. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 72 / 153 4. Generalized Least Squares (GLS) Case 2: Σ is unknown The Feasible Generalized Least Squares (FGLS) estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 73 / 153 4. Generalized Least Squares (GLS) Introduction 1 2 If Σ contains unknown parameters that must be estimated, then generalized least squares is not feasible. With an unrestricted matrix Σ = σ2 Ω, there are N (N + 1) /2 additional parameters (since Σ is symmetric) to estimate 3 This number is far too many to estimate with N observations. 4 Obviously, some structure must be imposed on the model if we are to proceed. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 74 / 153 4. Generalized Least Squares (GLS) De…nition (Structure of variance covariance matrix) We assume that the conditional variance covariance matrix of the disturbances can be expressed as a function of a small set of parameters α: V ( ε j X) = σ 2 Ω ( α ) Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 75 / 153 4. Generalized Least Squares (GLS) Example (Time series) For instance, a commonly used formula in time-series 0 1 ρ ρ2 ρ3 .. B ρ 1 ρ ρ2 .. B B ρ2 ρ 1 ρ .. Ω (ρ) = B 2 B ρ3 ρ ρ 1 .. B @ .. .. .. .. .. ρN 1 ρN 2 ρN 3 .. .. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne settings is 1 ρN 1 ρN 2 C C ρN 3 C C .. C C .. A 1 December 15, 2013 76 / 153 4. Generalized Least Squares (GLS) Example (Heteroscedascticity) If we consider a heteroscedastic model, where the variance of εi depends on a variable zi , with V ( εi j X) = σ2 ziθ we have 0 B B Ω (θ ) = B B @ Christophe Hurlin (University of Orléans) z1θ 0 0 0 z2θ 0 0 0 z3θ .. .. .. 0 0 0 .. 0 .. 0 .. 0 .. .. .. zNθ Advanced Econometrics - HEC Lausanne 1 C C C C A December 15, 2013 77 / 153 4. Generalized Least Squares (GLS) De…nition (Feasible Generalized Least Squares (FGLS)) Consider a consistent estimator b α of α, then the Feasible Least Generalized Squares (FGLS) estimator of β is de…ned as to be: >b b β FGLS = X Ω 1 X 1 b X> Ω 1 y b = Ω (b where Ω α) is a consistent estimator of Ω (α) . Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 78 / 153 4. Generalized Least Squares (GLS) Remark If plim plim 1 >b X Ω N 1 1 >b X Ω N X 1 y 1 > X Ω N 1 X =0 1 > X Ω N 1 y =0 Then the GLS and FGLS estimators are asymptotically equivalent b β FGLS Christophe Hurlin (University of Orléans) p b β GLS ! 0K 1 Advanced Econometrics - HEC Lausanne December 15, 2013 79 / 153 4. Generalized Least Squares (GLS) Theorem (E¢ ciency) An asymptotically e¢ cient FGLS estimator does not require that we have an e¢ cient estimator of α; only a consistent one is required to achieve full e¢ ciency for the FGLS estimator. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 80 / 153 4. Generalized Least Squares (GLS) Remark If the estimator b α is consistent p b α!α then the FGLS estimator has the same asymptotic properties (consistency, e¢ ciency, asymptotic distribution etc.) than the GLS estimator. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 81 / 153 4. Generalized Least Squares (GLS) Key Concepts 1 Factorisation of the variance covariance matrix 2 Generalized Least Squares (GLS) estimator 3 Feasible Generalized Least Squares (FGLS) estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 82 / 153 Section 5 Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 83 / 153 5. Heteroscedasticity Objectives The objective of this section are the following: 1 To determine the properties of the OLS in presence of heteroscedasticity 2 To estimate the asymptotic variance covariance matrix of the OLS estimator in presence of heteroscedasticity 3 To introduce the concept of robust inference (to heteroscedasticity) Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 84 / 153 5. Heteroscedasticity Introduction In the rest of this chapter, we will focus on the case of heteroscedastic disturbances. V ( εi j X) = σ2i for i = 1, .., N Heteroscedasticity arises in numerous applications, in both cross-section and time-series data. For example, even after accounting for …rm sizes, we expect to observe greater variation in the pro…ts of large …rms than in those of small ones. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 85 / 153 5. Heteroscedasticity Assumption: We assume that the disturbances are pairwise uncorrelated and heteroscedastic: V ( ε j X) = Σ = σ 2 Ω with 0 σ21 0 B 0 σ2 2 Σ=B @ .. .. 0 .. 1 0 .. 0 ω1 0 B .. 0 C C = σ2 Ω = σ2 B 0 ω 2 @ .. .. .. A .. 2 .. σN 0 .. with ω i = σ2i /σ2 for i = 1, .., N. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne 1 .. 0 .. 0 C C .. .. A .. ω N December 15, 2013 86 / 153 5. Heteroscedasticity De…nition (Scaling) The fact to scale the variances as σ2i = σ2 ω i for i = 1, .., N allows us to use a normalisation on Ω trace (Ω) = N ∑ ωi = N i =1 Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 87 / 153 5. Heteroscedasticity Introduction (cont’d) We will consider three cases: Case 1: the heteroscedasticity form (structure) is unknown: OLS estimator and robust inference Case 2: the variance covariance matrix Σ is known: GLS or Weighted Least Square (WLS) Case 3: the variance covariance matrix Σ is unknown but its form (structure) is known: two-steps or iterated FGLS estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 88 / 153 5. Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 89 / 153 5. Heteroscedasticity Case 1: Heteroscedasticity of unknown form OLS and robust inference Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 90 / 153 5. Heteroscedasticity Assumption: We assume that the variances σ2i are unknown for i = 1, ..N and no particular form (structure) is imposed on Ω (or Σ). Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 91 / 153 5. Heteroscedasticity Introduction 1 The GLS cannot be implemented since Σ is unknown. 2 The FGLS estimator requires to estimate (in a …rst step) N parameters σ21 , .., σ2N . With N observations, the FGLS is not feasible. 3 The only solution to estimate β consists in using the OLS. 4 Under suitable regularity conditions, the OLS estimator is unbiased, consistent, asymptotically normally distributed but... ine¢ cient. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 92 / 153 5. Heteroscedasticity Introduction (cont’d) Consider the OLS estimator: > b β OLS = X X We know that asy b β OLS with N b Vasy β OLS Q = plim Christophe Hurlin (University of Orléans) 1 > X X N β0 , σ2 Q N = σ2 Q N 1 X> y 1 Q Q 1 Q Q Q = plim Advanced Econometrics - HEC Lausanne 1 1 1 > X ΩX N December 15, 2013 93 / 153 5. Heteroscedasticity Problem (Robust inference with OLS) The conventionally estimated covariance matrix for the least squares estimator σ2 X> X 1 1 is inappropriate; the appropriate matrix is 1 1 σ 2 X> X X> ΩX X> X . It is unlikely that these two would coincide, so the usual estimators of the standard errors are likely to be erroneous. The inference (test-statistics) based σ2 X> X misleading. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne 1 is December 15, 2013 94 / 153 5. Heteroscedasticity Question b How to estimate Vasy β OLS and to make robust inference? b Vasy β OLS Q = plim Christophe Hurlin (University of Orléans) 1 > X X N = σ2 Q N 1 Q Q Q = plim Advanced Econometrics - HEC Lausanne 1 1 > X ΩX N December 15, 2013 95 / 153 5. Heteroscedasticity We seek an estimator for Q = plim 1 > 1 X ΩX = plim N N N ∑ ωi xi xi> = EX ω i xi xi> i =1 or equivalently of Q 1 1 = plim X> ΣX = plim N N N ∑ σ2i xi xi> = EX σi xi xi> i =1 with Q Christophe Hurlin (University of Orléans) = σ2 Q Advanced Econometrics - HEC Lausanne December 15, 2013 96 / 153 5. Heteroscedasticity Q 1 1 = plim X> ΣX = plim N N N ∑ σ2i xi xi> i =1 White (1980) shows that under very general condition, the estimator S0 = where bεi = yi 1 N N ∑ bε2i xi xi> i =1 b xi> β OLS , converges to Q p S0 ! Q = σ2 Q = σ2 Q White, H. “A Heteroscedasticity-Consistent Covariance Matrix Estimator and a Direct Test for Heteroscedasticity.” Econometrica, 48, 1980, pp. 817–838. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 97 / 153 5. Heteroscedasticity We know that: b Vasy β OLS S0 = 1 1 > X X N So, 1 N 1 > X X N Christophe Hurlin (University of Orléans) 1 N S0 σ2 Q N N 1 Q Q 1 p ∑ bε2i xi xi> ! σ2 Q i =1 = 1 = 1 N N ∑ xi xi> i =1 1 > X X N 1 ! 1 p !Q 1 p b ! Vasy β OLS Advanced Econometrics - HEC Lausanne December 15, 2013 98 / 153 5. Heteroscedasticity De…nition (White heteroscedasticity consistent estimator) The White consistent estimator of the asymptotic variance-covariance b matrix of the ordinary least squares estimator β OLS in the generalized linear regression model is de…ned to be: b b asy β V OLS with = N X> X b b asy β V OLS S0 = Christophe Hurlin (University of Orléans) 1 N 1 S0 X> X 1 p b ! Vasy β OLS N ∑ bε2i xi xi> i =1 Advanced Econometrics - HEC Lausanne December 15, 2013 99 / 153 5. Heteroscedasticity Corollary (White heteroscedasticity consistent estimator) The White consistent estimator can written as: ! 1 ! N N 1 1 1 2 > > b b asy β bεi xi xi V xi xi OLS = N N i∑ N i∑ =1 =1 Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne 1 N N ∑ xi xi> i =1 December 15, 2013 ! 1 100 / 153 5. Heteroscedasticity Remarks 1 This result is extremely important and useful. It implies that without actually specifying the type of heteroscedasticity, we can still make appropriate inferences based on the results of least squares. 2 This implication is especially useful if we are unsure of the precise nature of the heteroscedasticity (which is probably most of the time). Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 101 / 153 5. Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 102 / 153 5. Heteroscedasticity Remark Given the normalisation trace(Ω) = N, we have: σ2 = Christophe Hurlin (University of Orléans) 1 N N ∑ σ2i i =1 Advanced Econometrics - HEC Lausanne December 15, 2013 103 / 153 5. Heteroscedasticity De…nition (SSR) b2 de…ned by: The least squares estimator σ b2 = σ b ε>b ε 1 = N K N K N ∑ bε2i i =1 converges to the probability limit of the average variance of the disturbances 1 N 2 p b2 ! lim σ2 = lim σ ∑ σi N !∞ N !∞ N i =1 Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 104 / 153 5. Heteroscedasticity Example (White robust estimator. Source: Greene (2012)) Consider the generalized linear regression model: AVGEXPi = β1 + β2 AGEi + β3 Ownrenti + β4 Incomei + β5 Income2i + εi where AVGEXP denotes the Avg. monthly credit card expenditure, Ownrent denotes a binary variable (individual owns (1) or rents (0) home), Age denotes the age in years, Income denotes the income divided by 10,000. The data are available in …le Chapter5_data.xls. Question: write a Matlab code to (1) estimate the parameters by OLS, (2) compute the standard errors and the robust standard errors and (3) compare your results with Eviews. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 105 / 153 5. Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 106 / 153 5. Heteroscedasticity 2000 OLS residuals 1500 1000 500 0 -500 1 2 3 4 5 6 Income 7 8 9 10 This graph is the sign of heteroscedasticity.. the variance of the residuals seems to depend on the income. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 107 / 153 5. Heteroscedasticity The values are the same.. perfect Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 108 / 153 5. Heteroscedasticity The values are di¤erent... Why? Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 109 / 153 5. Heteroscedasticity Remark This di¤erence is due to the fact that Eviews uses a …nite sample correction for S0 (Davidson and MacKinnon, 1993) S0 = N 1 N K ∑ bε2i xi xi> i =1 Davidson, R. and J. MacKinnon. Estimation and Inference in Econometrics. New York: Oxford University Press, 1993. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 110 / 153 5. Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 111 / 153 5. Heteroscedasticity The values are now identical. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 112 / 153 5. Heteroscedasticity Case 2: Heteroscedasticity with known Σ GLS and Weighted Least Squares Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 113 / 153 5. Heteroscedasticity Assumption: We assume that the disturbances are heteroscedastic with V ( ε j X) = Σ = σ 2 Ω with 0 σ21 0 B 0 σ2 2 Σ=B @ .. .. 0 .. 1 0 .. 0 ω1 0 B .. 0 C C = σ2 Ω = σ2 B 0 ω 2 @ .. .. .. A .. 2 .. σN 0 .. where the parameters σ2i and ω i are known for i = 1, ..N. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne 1 .. 0 .. 0 C C .. .. A .. ω N December 15, 2013 114 / 153 5. Heteroscedasticity De…nition (GLS estimator) In presence of heteroscedasticity, the Generalized Least Squares (GLS) estimator of β is de…ned as to: or equivalently by b β GLS = b β GLS = Christophe Hurlin (University of Orléans) N xi xi> ∑ i =1 ω i N xi xi> ∑ 2 i =1 σ i ! ! 1 1 N xi yi ∑ ωi i =1 N xi yi ∑ σ2 i i =1 Advanced Econometrics - HEC Lausanne ! ! December 15, 2013 115 / 153 5. Heteroscedasticity Proof In general, whatever the form of Σ = σ2 Ω, we have: > b β GLS = X Ω 1 1 X X> Ω 1 y Since Ω is diagonal: X> Ω X> Ω 1 N X= 1 xi xi> ∑ i =1 ω i N y= xi yi i =1 ω i ∑ As a consequence: b β GLS = Christophe Hurlin (University of Orléans) N xi x> ∑ ωii i =1 ! 1 N xi yi ∑ ωi i =1 Advanced Econometrics - HEC Lausanne ! December 15, 2013 116 / 153 5. Heteroscedasticity Remark b β GLS = N xi x> ∑ ωii i =1 ! 1 N xi yi ∑ ωi i =1 ! This formula is similar to that obtained for a Weighted Least Squares (WLS). ! 1 ! N N > b β ∑ δi xi xi ∑ δi xi yi WLS = i =1 Christophe Hurlin (University of Orléans) i =1 Advanced Econometrics - HEC Lausanne December 15, 2013 117 / 153 5. Heteroscedasticity Fact (GLS and WLS) In presence of heteroscedasticity, the GLS estimator is a particular case of the Weighted Least Squares (WLS) estimator. b β WLS = N ∑ δi xi xi> i =1 ! 1 N ∑ δi xi yi i =1 ! b b where δi is an arbitrary weight. For δi = 1/ω i , we have β WLS = βGLS . Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 118 / 153 5. Heteroscedasticity Remark 1 The WLS estimator is consistent regardless of the weights used, as long as the weights are uncorrelated with the disturbances. 2 In general, we consider a weight which is proportional to one explicative variable (the income in the last example): σ2i = σ2 xik2 () δi = Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne 1 xik2 December 15, 2013 119 / 153 5. Heteroscedasticity Case 3: Heteroscedasticity for a given structure FGLS and two-step or iterated estimators Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 120 / 153 5. Heteroscedasticity Assumption: We assume that the disturbances are heteroscedastic with V ( ε j X) = Σ ( α ) = σ 2 Ω ( α ) where α denotes a set of parameters. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 121 / 153 5. Heteroscedasticity Example (Restriction) We assume that V ( εi j X) = σ2i (α) = σ2 zi> α 2 where α = (α1 : .. : αH )> is a H 1 vector of parameters and zi is H of explicative variables (not necessarily the same as in xi ). Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 1 122 / 153 5. Heteroscedasticity Example (Harvey’s (1976) restriction) Harvey (1976) considers a restriction of the form: V ( εi j X) = σ2i (α) = exp xi> α where α = (α1 : .. : αH )> is a H 1 vector of parameters and zi is H of explicative variables (not necessarily the same as in xi ). Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 1 123 / 153 5. Heteroscedasticity We know that the GLS estimator is de…ned by: N ∑ b β GLS = i =1 xi xi> σ2i (α) ! 1 N ! ∑ xi yi σ2i (α) N xi yi σ2i (b α) i =1 S, the feasible GLS (FGLS) estimator is: b β FGLS = Christophe Hurlin (University of Orléans) N ∑ i =1 xi xi> σ2i (b α) ! 1 ∑ i =1 Advanced Econometrics - HEC Lausanne ! December 15, 2013 124 / 153 5. Heteroscedasticity If we assume for instance that V ( εi j X) = σ2i (α) = exp zi> α where zi is a vector of H variables, a way to estimate α consists in considering the model: ln bε2i = zi> α + vi and to estimate α by OLS. The OLS is consistent even it is ine¢ cient (due to the heteroscedasticity). Given b α, we have a consistent estimator for σ2i : p b2i = exp zi> b σ α ! σ2i (α) Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 125 / 153 5. Heteroscedasticity Problem In order to estimate β by the GLS, we need b α, and to estimate α, we need b ... the residuals bεi = yi xi> β GLS Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 126 / 153 5. Heteroscedasticity Two solutions 1 A two steps FGLS estimator 2 An iterative FGLS estimator Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 127 / 153 5. Heteroscedasticity De…nition (Two-steps FGLS estimator) First step: estimate the parameters β by OLS. Compute the residuals b bεi = yi xi> β OLS and estimate the parameters α according to the α) appropriate model. Second step: compute the estimated variances σ2i (b and compute the FGLS estimator: b β FGLS = Christophe Hurlin (University of Orléans) N ∑ i =1 xi xi> σ2i (b α) ! 1 N ∑ i =1 xi yi σ2i (b α) Advanced Econometrics - HEC Lausanne ! December 15, 2013 128 / 153 5. Heteroscedasticity De…nition (Iterated FGLS estimator) Estimate the parameters β by OLS. Compute the residuals b bεi = yi xi> β OLS and estimate the parameters α according to the appropriate model. Compute the estimated variances σ2i (b α) and compute the FGLS estimator: ! 1 ! N N > x x (1 ) x y i i i b β ∑ 2 αi ) ∑ 2 α) FGLS = i =1 σ i ( b i =1 σ i ( b (1 ) b xi> β FGLS and estimate the parameters α b (2 ) and so according to the appropriate model. Compute the FGLS β FGLS on...The procedure stop when Compute the residuals bεi = yi sup j =1,..,K b (i ) β j ,FGLS Christophe Hurlin (University of Orléans) b (i 1 ) < threshold (ex: 0.001) β j ,FGLS Advanced Econometrics - HEC Lausanne December 15, 2013 129 / 153 5. Heteroscedasticity Example (Harvey’s (1976) multiplicative model of heteroscedasticity) Consider the generalized linear regression model: AVGEXPi = β1 + β2 AGEi + β3 Ownrenti + β4 Incomei + β5 Income2i + εi where the heteroscedasticity satis…es the Harvey’s (1976) speci…cation V ( εi j X) = σ2i = exp (α1 + α2 Incomei ) The data are available in …le Chapter5_data.xls. Question: write a Matlab code to estimate the parameters by FGLS by using a two-step and an iterative estimator. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 130 / 153 5. Heteroscedasticity Remark A way to get the estimates of the parameters α1 and α2 is to consider the regression: ln bε2i = α1 + α2 Incomei + vi and to estimate the parameters by OLS. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 131 / 153 5. Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 132 / 153 5. Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 133 / 153 5. Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 134 / 153 5. Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 135 / 153 5. Heteroscedasticity Key Concepts 1 OLS and robust inference 2 White heteroscedasticity consistent estimator 3 GLS and Weighted Least Squares (WLS) 4 FGLS: two-steps and iterated estimators Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 136 / 153 Section 6 Testing for Heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 137 / 153 6. Testing for heteroscedasticity Objectives The objective of this section are to introduce the following tests for heteroscedasticity: 1 White general test 2 The Breusch-Pagan / Godfrey LM test Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 138 / 153 6. Testing for heteroscedasticity De…nition (White test for heteroscedasticity) The White test for heteroscedasticity is based on: H0 : σ2i = σ2 H1 : σ2i 6= σ2j Christophe Hurlin (University of Orléans) for i = 1, .., N for at least one pair (i, j ) Advanced Econometrics - HEC Lausanne December 15, 2013 139 / 153 6. Testing for heteroscedasticity The intuition of the test is based on the following idea: 1 If there is no heteroscedasticity (under the null H0 ): b Vasy β OLS 2 = σ2 Q 1 b b asy β V OLS = σ 2 X> X b Vasy β OLS = σ2 Q 1 Under the alternative (heteroscedasticity): b b asy β V OLS Christophe Hurlin (University of Orléans) = σ 2 X> X 1 1 Q Q 1 X> ΩX X> X Advanced Econometrics - HEC Lausanne 1 December 15, 2013 140 / 153 6. Testing for heteroscedasticity White (1980) proposes the following procedure and test-statistic: Step 1: Estimation of the model using the OLS estimator of β. Step 2: Determine the residuals bεi = yi b xi> β OLS . Step 3: Regress bε2i on a constant and all unique columns vectors contained in X and all the squares and cross-products of the column vectors in X. Step 4: Determine the coe¢ cient of determination, R2 , of the previous regression. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 141 / 153 6. Testing for heteroscedasticity De…nition (White test for heteroscedasticity) Under the null, the White test-statistic N R2 converges: N d R2 ! χ2 (m H0 1) where m is the number of explanatory variables in the regression of bε2i . The critical region of size α is W= y :N where χ21 α R2 > χ21 α denotes the 1-α critical value of the χ2 (m Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne 1) distribution. December 15, 2013 142 / 153 6. Testing for heteroscedasticity Example (White’s (1980) test for heteroscedasticity) Consider the generalized linear regression model: AVGEXPi = β1 + β2 AGEi + β3 Ownrenti + β4 Incomei + β5 Income2i + εi The data are available in …le Chapter5_data.xls. Question: write a Matlab code to compute the White test-statistic for heteroscedasticity and its p-value. What is you conclusion for a signi…cance level of 5%? Compare your results with Eviews. Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 143 / 153 6. Testing for heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 144 / 153 6. Testing for heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 145 / 153 6. Testing for heteroscedasticity De…nition (Breusch and Pagan test) Breusch and Pagan (1979) have devised a Lagrange multiplier test of the hypothesis that σ2i = σ2 f α0 + zi> α where zi = (zi 1 ..zip )> is a p 1 vector of independent variables. The test is: H0 : α = 0p 1 (homoscedasticity) H1 : α 6= 0p Christophe Hurlin (University of Orléans) 1 (heteroscedasticity) Advanced Econometrics - HEC Lausanne December 15, 2013 146 / 153 6. Testing for heteroscedasticity The test can be carried out with a simple regression of gi = N bε2i b ε>b ε 1=N bε2i 2 εi ∑N i =1 b 1 on the variables zik for k = 1, ., N and a constant term. gi = α0 + α1 zi 1 + ... + αp zip + vi Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 147 / 153 6. Testing for heteroscedasticity De…nition (Breusch and Pagan test-statistic) De…ne Z the N (p + 1) matrix of observations on (1, zi ) and let g be the N 1 vector of observations gi = N bε2i b ε>b ε 1 Then, the Breusch and Pagan’s test-statistic is de…ned by: LM = 1 > g Z Z> Z 2 1 Z> g Under the null, we have: d LM ! χ2 (p ) H0 Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 148 / 153 6. Testing for heteroscedasticity Example (Breusch and Pagan’s (1979) test for heteroscedasticity) Consider the generalized linear regression model: AVGEXPi = β1 + β2 AGEi + β3 Ownrenti + β4 Incomei + β5 Income2i + εi The data are available in …le Chapter5_data.xls. Question: write a Matlab code to compute the Breusch and Pagan test-statistic for heteroscedasticity with zi = xi and its p-value. What is you conclusion for a signi…cance level of 5%? Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 149 / 153 6. Testing for heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 150 / 153 6. Testing for heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 151 / 153 6. Testing for heteroscedasticity Key Concepts 1 White test for heteroscedasticity 2 Breusch and Pagan test for heteroscedasticity Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 152 / 153 End of Chapter 5 Christophe Hurlin (University of Orléans) Christophe Hurlin (University of Orléans) Advanced Econometrics - HEC Lausanne December 15, 2013 153 / 153