The Exact Distribution of the t-ratio with Robust and Clustered Standard Errors

Transcription

1 The Exact Distribution of the t-ratio with Robust and Clustered Standard Errors by Bruce E. Hansen Department of Economics University of Wisconsin October 2018 Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

2 Peter Phillips Contributions to Exact Distribution Theory Before unit roots, Peter was the master of exact distribution theory Econometrica 1977 (asymptotic expansions) Journal of Econometrics 1977 (SUR) Econometrica 1980 (IV) Biometrika 1982 (F) Handbook of Econometrics 1983 (simultaneous equations) International Economic Review 1984 (LIML) Journal of Econometrics 1984 (Stein Rule) Journal of Econometrics 1984 (exogenous case) International Economic Review 1985 (LIML) Econometrica 1985 (SUR) International Economic Review 1986 (FIML) Econometrica 1986 (Wald statistic) Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

3 t-ratio Gosset (1908) The t-ratio of the sample mean has the exact t n 1 distribution A fundamental intellectual achievement Linear regression Gosset s result extends to classical t-ratios (classical standard errors) Classical t-ratios have t n k distribution Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

4 But... Classical standard errors are no longer used in economic research Papers use either Heteroskedasticity-consistent (HC) Cluster-robust (CR) Heteroskedasticity-and-autocorrelation-consistent (HAC) Justification is asymptotic Most assess significance (testing and confidence intervals) using finite sample distribution: t n k distribution (HC) t G 1 distribution (CR) THIS IS WRONG!!! Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

5 Most applied regressions use Stata with Regression: reg y x, r Uses HC1 variance estimator White estimator scaled by n/(n k) Uses t n k distribution for p-values and confidence intervals UNJUSTIFIED! Clustered: reg y x, cluster(id ) Uses CR1 variance estimator Described later, ad hoc Uses t G 1 distribution for p-values and confidence intervals No finite sample justification Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

6 This paper Provides an exact theory of inference Linear regression with robust standard errors Linear regression with clustered standard errors Exact distribution of HC and CR t-ratios under i.i.d. normality Computable Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

7 Linear Regression with Heteroskedasticity y i = x i β + e i E (e i x i ) = 0 E ( e 2 i x i ) = σ 2 i n observations k regressors Core model in applied econometrics Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

8 Heteroskedastic (HC) Variance Estimation: Some History Eicker (1963): HC0 Horn, Horn and Duncan (1975): HC2 Hinkley (1977): HC1 White (1980): HC0 for econometrics MacKinnon and White (1985): HC3 Chesher and Jewitt (1987): Bias can be large Bera, Suprayitno and Premaratne (2002): Unbiased estimator Bell-McCaffrey (2002): Distributional approximation Cribari-Neto (2004): HC4 Cribari-Neto, Souza and Vasconcellos (2007): HC5 Cattaneo, Jansson and Newey (2017): Many regressors Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

9 HC Variance Estimation OLS: Residuals: HC0 HC1 V 1 = V 0 = ( X X ) 1 β = ( X X ) 1 X Y ê i = y i x i β ( n x i x i êi 2 i=1 ) (X X ) 1 n ( X X ) ( 1 n ) (X x i x i êi 2 X ) 1 n k i=1 robust covariance matrix in Stata Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

10 HC2: V 2 = ( X X ) 1 ( n ) (X x i x i êi 2 (1 h i ) 1 X ) 1 i=1 h i = x i (X X ) 1 x i Unbiased under homoskedasticity HC3: V 3 = ( X X ) 1 ( n ) (X x i x i êi 2 (1 h i ) 2 X ) 1 i=1 jackknife Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

11 HC3 (jackknife) is a conservative estimator Theorem. In the linear regression model, ( ) E V 3 X ( ) ) V = E ( β β) ( β β X (However, inference using HC3 is not necessarily conservative.) Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

12 HC t-ratios t-ratio for R β: ( β T = R β) R V R Distribution theory Asymptotic: T d N(0, 1) This is what we (typically) teach Distribution used in practical applications Finite Sample: T t n k This is what most applied papers use Incorrect Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

13 Clustered Samples Observations are (y ig, x ig ) g = 1,..., G indexes cluster (group) i = 1,..., n n indexes observation within g th cluster Clusters are mutually independent Observations within a cluster have unknown dependence In panels, (y ig, x ig ) could be demeaned observations Assumptions fully allow for this Number of observations n g per cluster may vary across cluster Total number of observations n = G g =1 n g Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

14 Cluster Regression y g = (y 1g,..., y ng g ) is n g 1 vector of dependent variables X g = (x 1g,..., x ng g ) is n g K regressor matrix for g th cluster. Linear regression model y g = X g β + e g E (e g X g ) = 0 E ( e g e g ) X g = Sg Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

15 Cluster-Robust (CR) Variance Estimation OLS: β = Residual: Variance estimator V 0 = ( G ) 1 ( G X g X g g =1 g =1 ê g = y g X g β X g y g ) ( G ) 1 ( G ) ( G X g X g X g ê g ê g X g g =1 g =1 g =1 X g X g ) 1 Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

16 Adjustments Chris Hansen (2007) adjustment ( ) G V = V 0 G 1 Justified in Large homogenous clusters framework Stata adjusment No justification V 1 = ( ) ( ) n 1 G V 0 n k G 1 Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

17 Cluster-Robust (CR) Variance Estimation: Some History Methods: Moulton (1986, 1990), Arellano (1987) Popularization: Rogers (1993), Bertrand, Duflo and Mullainathan (2004) Large G asymptotics: White (1984), C. Hansen (2007), Carter, Schnepel and Steigerwald (2017) Fixed G asymptotics: C. Hansen (2007), Bester, Conley and C. Hansen (2011), Conley and Taber (2011), Ibragimov and Mueller (2010, 2016) Small Sample: Donald and Lang (2007), Imbens and Kolesar (2016), Young (2017), Canay, Romano, and Shaikh (2017) Bootstrap: Cameron, Gelbach and Miller (2008), MacKinnon and Webb (2017abc, 2018), MacKinnon, Nielsen, and Webb (2017), Djogbenou, MacKinnon, and Nielsen (2018), Canay, Santos, and Shaikh (2018) Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

18 Illustration: Heteroskedastic Dummy Variable Regression Dummy variable model Angrist and Pinchke (2009) Imbens and Kolesar (2016) y i = β 0 + β 1 x i + e i n i=1 x i = 3 e i N(0, 1) Coeffi cient of interest: β 1 Simulation with 100,000 replications Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

19 Large Size Distortion with HC Standard Errors Rejection Probability of Nominal 5% Tests Using t n k Critical Values n = 30 HC HC HC HC Notice that even conservative HC3 t-ratio over-rejects. That is because the t n k distribution is incorrect. Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

20 Distortion increases with Sample size! Rejection Probability of Nominal 5% Tests Using t n k Critical Values n = 30 n = 100 n = 500 HC HC HC HC Reason: Highly Leveraged Design Matrix Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

21 Simulation Results All procedures over-reject HC1 correction doesn t help Unbiased estimator HC2 over-rejects Conservative estimator HC3 over-rejects t n k vs N(0, 1) ineffective Conclusion: Distributional approximation needs improvement Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

22 Exact Distribution of White t-ratio Assumption: Observations are i.i.d., e i x i N ( 0, σ 2) Goal: Calculate exact density and distribution of White t-ratio T Parameter of interest: θ = R β The distribution is unknown Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

23 Notation M = I X (X X ) 1 X D = diag { d1 2,..., d n 2 } d i = R (X X ) 1 x i λ 1,..., λ K are the non-zero eigenvalues of D 1/2 MD 1/2 w i = λ i /R (X X ) 1 R Distribution depends on normalized eigenvalues w i Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

24 Theorem: Distribution of White Variance Estimator where G r (u) is the χ 2 r distribution, R V R ( u ) R VR b m G K +2m m=0 δ δ = min w m m N ( δ b 0 = i=1 b m = 1 m a m = w i m l=1 N 1 i=1 2 ) ki /2 b m l a l, m 1 (1 δwi ) m Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

25 Comments The Theorem shows that the distribution of the variance estimator can be written as an infinite mixture of chi-square distributions The weights b m are non-negative, sum to one Weights are determined by a simple recursion in known parameters Weights determined by eigenvalues of matrix D 1/2 MD 1/2 Specializes to conventional chi-square when eigenvalues are all equal This theorem is a refinement of Castano and Lopez (2005) Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

26 Theorem: Distribution of White t-ratio T ) b m F K +2m (u (K + 2m) δ m=0 where F r (u) is the student t distribution, and weights b m are from the previous theorem. Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

27 Comments The Theorem shows that the distribution of T can be written as an infinite mixture of student t distributions The weights b m are non-negative, sum to one Weights are determined by a simple recursion in known parameters Weights determined by eigenvalues of matrix D 1/2 MD 1/2 Specializes to conventional student t when eigenvalues are all equal Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

28 Finite Sample Distribution This is the exact finite sample distribution of the White HC t-ratio under normality. The distribution is determined by the design matrix X X This result is entirely new The exact distribution is not student t. It is a mixture of student t distributions. The difference can be large when the design matrix is highly leveraged. Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

29 Exact Distribution Advantages Computatable exact distribution under normality Improved accuracy when regressor matrix is highly leveraged Disadvantages Increased computation cost relative to classical methods Reliable algorithm in development Limitations Assumes homoskedasticity Assumes normality Linear parameters Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

30 Alternative Bell-McCaffrey (2002) Satterthwaite (1946) approximation for Q is αχ 2 K match first two moments of Q Approximate distribution of t by t K Endorsed by Imbens-Kolesar (2016) An approximation but no formal theory where α and K Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

31 Simulation Experiement Dummy variable model Angrist and Pinchke (2009) Imbens and Kolesar (2016) y i = β 0 + β 1 x i + e i n i=1 x i = 3 Coeffi cient of interest: β 1 n = 50, 100, 500 Compare: HC1, HC2, HC3 t n k, Bell-McCaffrey, and T distributions Size and median length of confidence regions e i N(0, 1), Heteroskedastic, and student-t errors 100,000 replications Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

32 Rejection Probability of Nominal 5% Tests Median Length of 95% Confidence Intervals Normal Homoskedastic Errors t n k Bell-McCaffrey Exact T size Length size Length n = 50 HC HC HC n = 100 HC HC HC n = 500 HC HC HC Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

33 Rejection Probability of Nominal 5% Tests Median Length of 95% Confidence Intervals Normal Heteroskedastic Errors σ 2 (x) = 1(x = 1) + 0.5(x = 0) t n k Bell-McCaffrey T size Length size Length n = 50 HC HC HC n = 100 HC HC HC n = 500 HC HC HC Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

34 Rejection Probability of Nominal 5% Tests Median Length of 95% Confidence Intervals Normal Heteroskedastic Errors σ 2 (x) = 1(x = 1) + 2(x = 0) t n k Bell-McCaffrey T size Length size Length n = 50 HC HC HC n = 100 HC HC HC n = 500 HC HC HC Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

35 Rejection Probability of Nominal 5% Tests Median Length of 95% Confidence Intervals t 5 Errors t n k Bell-McCaffrey T size Length size Length n = 50 HC HC HC n = 100 HC HC HC n = 500 HC HC HC Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

36 Simulation Summary t n k criticals inappropriate Bell-McCaffrey can be quite conservative T is precise under homoskedastic normality (as expected) Both Bell-McCaffrey and T sensitive to heteroskedasticity and non-normality HC3 appears least sensitive HC3 with T distribution reasonably reliable Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

37 Clustered Samples Same analysis applies to clustered regression and CR standard errors Under i.i.d. normality, clustered t-ratios have exact T distributions Weights are determined by regressor matrix Distortions from normality when design matrix is highly leveraged When clusters are heterogeneous When only a few clusters are treated Accuracy of conventional distribution theory depends on the number of clusters G and degree of leverage Conventional asymptotics requires a large G, not large n Many applied papers don t even report G G should be reported, along with sample size! Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38

38 Conclusion In 1908, Gosset revolutionized statistical inference by providing the exact distribution of the classical t-ratio Peter Phillips extended Gosset s theory to a broad set of econometric estimators and statistics Applied econometrics relies on heteroskedasticity-robust and cluster-robust standard errors There is no finite sample theory for HC and CR t-ratios This paper provides the first exact distribution theory Bruce Hansen (University of Wisconsin) Exact Inference for Robust t ratio October / 38