Econometrics Note

ECONOMETRICS
Bruce E. Hansen
c 2000, 20141
University of Wisconsin
Department of Economics
This Revision: January 3, 2014 Comments Welcome
This manuscript may be printed and reproduced for individual or instructional use, but may not be printed for commercial purposes.
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viii 1 Introduction 1.1 What is Econometrics? . . . . . . . . . . . . 1.2 The Probability Approach to Econometrics 1.3 Econometric Terms and Notation . . . . . . 1.4 Observational Data . . . . . . . . . . . . . . 1.5 Standard Data Structures . . . . . . . . . . 1.6 Sources for Economic Data . . . . . . . . . 1.7 Econometric Software . . . . . . . . . . . . 1.8 Reading the Manuscript . . . . . . . . . . . 1.9 Common Symbols . . . . . . . . . . . . . . 2 Conditional Expectation and Projection 2.1 Introduction . . . . . . . . . . . . . . . . 2.2 The Distribution of Wages . . . . . . . . 2.3 Conditional Expectation . . . . . . . . . 2.4 Log Dierences* . . . . . . . . . . . . . 2.5 Conditional Expectation Function . . . 2.6 Continuous Variables . . . . . . . . . . . 2.7 Law of Iterated Expectations . . . . . . 2.8 CEF Error . . . . . . . . . . . . . . . . . 2.9 Intercept-Only Model . . . . . . . . . . 2.10 Regression Variance . . . . . . . . . . . 2.11 Best Predictor . . . . . . . . . . . . . . 2.12 Conditional Variance . . . . . . . . . . . 2.13 Homoskedasticity and Heteroskedasticity 2.14 Regression Derivative . . . . . . . . . . 2.15 Linear CEF . . . . . . . . . . . . . . . . 2.16 Linear CEF with Nonlinear Eects . . . 2.17 Linear CEF with Dummy Variables . . . 2.18 Best Linear Predictor . . . . . . . . . . 2.19 Linear Predictor Error Variance . . . . . 2.20 Regression Coecients . . . . . . . . . . 2.21 Regression Sub-Vectors . . . . . . . . . 2.22 Coecient Decomposition . . . . . . . . 2.23 Omitted Variable Bias . . . . . . . . . . 2.24 Best Linear Approximation . . . . . . . 2.25 Normal Regression . . . . . . . . . . . . 2.26 Regression to the Mean . . . . . . . . . 2.27 Reverse Regression . . . . . . . . . . . . 2.28 Limitations of the Best Linear Predictor i 1 1 1 2 3 4 5 6 7 8 9 9 9 11 13 14 15 16 18 19 20 20 21 23 23 24 25 26 28 34 35 35 36 37 38 38 39 40 41
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . .
CONTENTS 2.29 Random Coecient Model . . . . . . . . . . . . . . . . . . . 2.30 Causal Eects . . . . . . . . . . . . . . . . . . . . . . . . . . 2.31 Expectation: Mathematical Details* . . . . . . . . . . . . . 2.32 Existence and Uniqueness of the Conditional Expectation* 2.33 Identication* . . . . . . . . . . . . . . . . . . . . . . . . . . 2.34 Technical Proofs* . . . . . . . . . . . . . . . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 The Algebra of Least Squares 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . 3.2 Random Samples . . . . . . . . . . . . . . . . . . . 3.3 Sample Means . . . . . . . . . . . . . . . . . . . . . 3.4 Least Squares Estimator . . . . . . . . . . . . . . . 3.5 Solving for Least Squares with One Regressor . . . 3.6 Solving for Least Squares with Multiple Regressors 3.7 Illustration . . . . . . . . . . . . . . . . . . . . . . 3.8 Least Squares Residuals . . . . . . . . . . . . . . . 3.9 Model in Matrix Notation . . . . . . . . . . . . . . 3.10 Projection Matrix . . . . . . . . . . . . . . . . . . 3.11 Orthogonal Projection . . . . . . . . . . . . . . . . 3.12 Estimation of Error Variance . . . . . . . . . . . . 3.13 Analysis of Variance . . . . . . . . . . . . . . . . . 3.14 Regression Components . . . . . . . . . . . . . . . 3.15 Residual Regression . . . . . . . . . . . . . . . . . 3.16 Prediction Errors . . . . . . . . . . . . . . . . . . . 3.17 Inuential Observations . . . . . . . . . . . . . . . 3.18 Normal Regression Model . . . . . . . . . . . . . . 3.19 CPS Data Set . . . . . . . . . . . . . . . . . . . . . 3.20 Programming . . . . . . . . . . . . . . . . . . . . . 3.21 Technical Proofs* . . . . . . . . . . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Least Squares Regression 4.1 Introduction . . . . . . . . . . . . . . 4.2 Sample Mean . . . . . . . . . . . . . 4.3 Linear Regression Model . . . . . . . 4.4 Mean of Least-Squares Estimator . . 4.5 Variance of Least Squares Estimator 4.6 Gauss-Markov Theorem . . . . . . . 4.7 Residuals . . . . . . . . . . . . . . . 4.8 Estimation of Error Variance . . . . 4.9 Mean-Square Forecast Error . . . . . 4.10 Covariance Matrix Estimation Under 4.11 Covariance Matrix Estimation Under 4.12 Standard Errors . . . . . . . . . . . . 4.13 Computation . . . . . . . . . . . . . 4.14 Measures of Fit . . . . . . . . . . . . 4.15 Empirical Example . . . . . . . . . . 4.16 Multicollinearity . . . . . . . . . . . 4.17 Normal Regression Model . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
ii 41 43 47 49 50 51 55 57 57 57 58 58 59 60 62 62 63 65 66 67 68 68 70 71 72 74 76 78 82 83 86 86 86 87 88 89 91 92 93 95 96 97 100 101 102 103 105 108 110
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Homoskedasticity Heteroskedasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
CONTENTS 5 An Introduction to Large Sample Asymptotics 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . 5.2 Asymptotic Limits . . . . . . . . . . . . . . . . . 5.3 Convergence in Probability . . . . . . . . . . . . 5.4 Weak Law of Large Numbers . . . . . . . . . . . 5.5 Almost Sure Convergence and the Strong Law* . 5.6 Vector-Valued Moments . . . . . . . . . . . . . . 5.7 Convergence in Distribution . . . . . . . . . . . . 5.8 Higher Moments . . . . . . . . . . . . . . . . . . 5.9 Functions of Moments . . . . . . . . . . . . . . . 5.10 Delta Method . . . . . . . . . . . . . . . . . . . . 5.11 Stochastic Order Symbols . . . . . . . . . . . . . 5.12 Uniform Stochastic Bounds* . . . . . . . . . . . . 5.13 Semiparametric Eciency . . . . . . . . . . . . . 5.14 Technical Proofs* . . . . . . . . . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . .
iii 112 112 112 114 115 116 117 118 120 121 123 124 126 127 130 134 135 135 136 137 141 144 144 145 147 147 148 151 153 154 155 157 158 158 159 160 162 163 164 166 169 169 170 171 172 173 174 175 177 177
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
. . . . . . . . . . . . . . .
6 Asymptotic Theory for Least Squares 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Consistency of Least-Squares Estimation . . . . . . . . . . . . . . . . 6.3 Asymptotic Normality . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 Joint Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5 Consistency of Error Variance Estimators . . . . . . . . . . . . . . . 6.6 Homoskedastic Covariance Matrix Estimation . . . . . . . . . . . . . 6.7 Heteroskedastic Covariance Matrix Estimation . . . . . . . . . . . . 6.8 Summary of Covariance Matrix Notation . . . . . . . . . . . . . . . . 6.9 Alternative Covariance Matrix Estimators* . . . . . . . . . . . . . . 6.10 Functions of Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 6.11 Asymptotic Standard Errors . . . . . . . . . . . . . . . . . . . . . . . 6.12 t statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.13 Condence Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.14 Regression Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.15 Forecast Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.16 Wald Statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.17 Homoskedastic Wald Statistic . . . . . . . . . . . . . . . . . . . . . . 6.18 Condence Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.19 Semiparametric Eciency in the Projection Model . . . . . . . . . . 6.20 Semiparametric Eciency in the Homoskedastic Regression Model* . 6.21 Uniformly Consistent Residuals* . . . . . . . . . . . . . . . . . . . . 6.22 Asymptotic Leverage* . . . . . . . . . . . . . . . . . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7 Restricted Estimation 7.1 Introduction . . . . . . . . . . . . . . . . 7.2 Constrained Least Squares . . . . . . . . 7.3 Exclusion Restriction . . . . . . . . . . . 7.4 Minimum Distance . . . . . . . . . . . . 7.5 Asymptotic Distribution . . . . . . . . . 7.6 Ecient Minimum Distance Estimator . 7.7 Exclusion Restriction Revisited . . . . . 7.8 Variance and Standard Error Estimation 7.9 Misspecication . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
CONTENTS 7.10 Nonlinear Constraints 7.11 Inequality Restrictions 7.12 Constrained MLE . . . 7.13 Technical Proofs* . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
iv 179 180 181 181 183 185 185 186 187 187 188 189 190 192 193 194 194 195 196 197 198 199 202 204 205 207 210 211 213 215 215 218 221 221 222 224 227 229 229 229 231 231 232 234 234 235 237 238 239
8 Hypothesis Testing 8.1 Hypotheses . . . . . . . . . . . . . . . . . . . . . . 8.2 Acceptance and Rejection . . . . . . . . . . . . . . 8.3 Type I Error . . . . . . . . . . . . . . . . . . . . . 8.4 t tests . . . . . . . . . . . . . . . . . . . . . . . . . 8.5 Type II Error and Power . . . . . . . . . . . . . . . 8.6 Statistical Signicance . . . . . . . . . . . . . . . . 8.7 P-Values . . . . . . . . . . . . . . . . . . . . . . . . 8.8 t-ratios and the Abuse of Testing . . . . . . . . . . 8.9 Wald Tests . . . . . . . . . . . . . . . . . . . . . . 8.10 Homoskedastic Wald Tests . . . . . . . . . . . . . . 8.11 Criterion-Based Tests . . . . . . . . . . . . . . . . 8.12 Minimum Distance Tests . . . . . . . . . . . . . . . 8.13 Minimum Distance Tests Under Homoskedasticity 8.14 F Tests . . . . . . . . . . . . . . . . . . . . . . . . 8.15 Likelihood Ratio Test . . . . . . . . . . . . . . . . 8.16 Problems with Tests of NonLinear Hypotheses . . 8.17 Monte Carlo Simulation . . . . . . . . . . . . . . . 8.18 Condence Intervals by Test Inversion . . . . . . . 8.19 Power and Test Consistency . . . . . . . . . . . . . 8.20 Asymptotic Local Power . . . . . . . . . . . . . . . 8.21 Asymptotic Local Power, Vector Case . . . . . . . 8.22 Technical Proofs* . . . . . . . . . . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . 9 Regression Extensions 9.1 NonLinear Least Squares . . . . . 9.2 Generalized Least Squares . . . . . 9.3 Testing for Heteroskedasticity . . . 9.4 Testing for Omitted NonLinearity . 9.5 Least Absolute Deviations . . . . . 9.6 Quantile Regression . . . . . . . . Exercises . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
. . . . . . .
10 The Bootstrap 10.1 Denition of the Bootstrap . . . . . . . . . 10.2 The Empirical Distribution Function . . . . 10.3 Nonparametric Bootstrap . . . . . . . . . . 10.4 Bootstrap Estimation of Bias and Variance 10.5 Percentile Intervals . . . . . . . . . . . . . . 10.6 Percentile-t Equal-Tailed Interval . . . . . . 10.7 Symmetric Percentile-t Intervals . . . . . . 10.8 Asymptotic Expansions . . . . . . . . . . . 10.9 One-Sided Tests . . . . . . . . . . . . . . . 10.10Symmetric Two-Sided Tests . . . . . . . . . 10.11Percentile Condence Intervals . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
CONTENTS
10.12Bootstrap Methods for Regression Models . . . . . . . . . . . . . . . . . . . . . . . . 240 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242 11 NonParametric Regression 11.1 Introduction . . . . . . . . . . . . . . . . 11.2 Binned Estimator . . . . . . . . . . . . . 11.3 Kernel Regression . . . . . . . . . . . . . 11.4 Local Linear Estimator . . . . . . . . . . 11.5 Nonparametric Residuals and Regression 11.6 Cross-Validation Bandwidth Selection . 11.7 Asymptotic Distribution . . . . . . . . . 11.8 Conditional Variance Estimation . . . . 11.9 Standard Errors . . . . . . . . . . . . . . 11.10Multiple Regressors . . . . . . . . . . . . 243 243 243 245 246 247 249 252 255 255 256 259 259 259 261 261 261 263 263 266 266 267 268 269 270 271 272 272 278 278 279 280 281 282 282 283 284 285 287 289 289 291 292 293 294
. . . . . . . . . . . . Fit . . . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
12 Series Estimation 12.1 Approximation by Series . . . . . . . . . . . . 12.2 Splines . . . . . . . . . . . . . . . . . . . . . . 12.3 Partially Linear Model . . . . . . . . . . . . . 12.4 Additively Separable Models . . . . . . . . . 12.5 Uniform Approximations . . . . . . . . . . . . 12.6 Runges Phenomenon . . . . . . . . . . . . . . 12.7 Approximating Regression . . . . . . . . . . . 12.8 Residuals and Regression Fit . . . . . . . . . 12.9 Cross-Validation Model Selection . . . . . . . 12.10Convergence in Mean-Square . . . . . . . . . 12.11Uniform Convergence . . . . . . . . . . . . . . 12.12Asymptotic Normality . . . . . . . . . . . . . 12.13Asymptotic Normality with Undersmoothing 12.14Regression Estimation . . . . . . . . . . . . . 12.15Kernel Versus Series Regression . . . . . . . . 12.16Technical Proofs . . . . . . . . . . . . . . . . 13 Generalized Method of Moments 13.1 Overidentied Linear Model . . . . . . . . . 13.2 GMM Estimator . . . . . . . . . . . . . . . 13.3 Distribution of GMM Estimator . . . . . . 13.4 Estimation of the Ecient Weight Matrix . 13.5 GMM: The General Case . . . . . . . . . . 13.6 Over-Identication Test . . . . . . . . . . . 13.7 Hypothesis Testing: The Distance Statistic 13.8 Conditional Moment Restrictions . . . . . . 13.9 Bootstrap GMM Inference . . . . . . . . . . Exercises . . . . . . . . . . . . . . . . . . . . . . 14 Empirical Likelihood 14.1 Non-Parametric Likelihood . . . . . . . . 14.2 Asymptotic Distribution of EL Estimator 14.3 Overidentifying Restrictions . . . . . . . . 14.4 Testing . . . . . . . . . . . . . . . . . . . . 14.5 Numerical Computation . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . . . . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
. . . . .
CONTENTS 15 Endogeneity 15.1 Instrumental Variables . . . 15.2 Reduced Form . . . . . . . 15.3 Identication . . . . . . . . 15.4 Estimation . . . . . . . . . 15.5 Special Cases: IV and 2SLS 15.6 Bekker Asymptotics . . . . 15.7 Identication Failure . . . . Exercises . . . . . . . . . . . . .
vi 296 297 298 299 299 299 301 302 304
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
. . . . . . . .
16 Univariate Time Series 16.1 Stationarity and Ergodicity . . . . . . 16.2 Autoregressions . . . . . . . . . . . . . 16.3 Stationarity of AR(1) Process . . . . . 16.4 Lag Operator . . . . . . . . . . . . . . 16.5 Stationarity of AR(k) . . . . . . . . . 16.6 Estimation . . . . . . . . . . . . . . . 16.7 Asymptotic Distribution . . . . . . . . 16.8 Bootstrap for Autoregressions . . . . . 16.9 Trend Stationarity . . . . . . . . . . . 16.10Testing for Omitted Serial Correlation 16.11Model Selection . . . . . . . . . . . . . 16.12Autoregressive Unit Roots . . . . . . . 17 Multivariate Time Series 17.1 Vector Autoregressions (VARs) . . . . 17.2 Estimation . . . . . . . . . . . . . . . 17.3 Restricted VARs . . . . . . . . . . . . 17.4 Single Equation from a VAR . . . . . 17.5 Testing for Omitted Serial Correlation 17.6 Selection of Lag Length in an VAR . . 17.7 Granger Causality . . . . . . . . . . . 17.8 Cointegration . . . . . . . . . . . . . . 17.9 Cointegrated VARs . . . . . . . . . . . 18 Limited Dependent Variables 18.1 Binary Choice . . . . . . . . 18.2 Count Data . . . . . . . . . 18.3 Censored Data . . . . . . . 18.4 Sample Selection . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
306 . 306 . 308 . 309 . 309 . 310 . 310 . 311 . 312 . 312 . 313 . 314 . 314 316 316 317 317 317 318 318 319 319 320 322 322 323 324 325
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . . . . . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
. . . .
19 Panel Data 327 19.1 Individual-Eects Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 19.2 Fixed Eects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327 19.3 Dynamic Panel Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 329 20 Nonparametric Density Estimation 330 20.1 Kernel Density Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330 20.2 Asymptotic MSE for Kernel Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 332
CONTENTS A Matrix Algebra A.1 Notation . . . . . . . . . . . A.2 Matrix Addition . . . . . . A.3 Matrix Multiplication . . . A.4 Trace . . . . . . . . . . . . . A.5 Rank and Inverse . . . . . . A.6 Determinant . . . . . . . . . A.7 Eigenvalues . . . . . . . . . A.8 Positive Deniteness . . . . A.9 Matrix Calculus . . . . . . . A.10 Kronecker Products and the A.11 Vector and Matrix Norms . A.12 Matrix Inequalities . . . . .
vii 335 335 336 336 337 338 339 340 341 342 342 343 343 348 348 350 350 351 352 354 356 358 359 361 364
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vec Operator . . . . . . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
. . . . . . . . . . . .
B Probability B.1 Foundations . . . . . . . . . . . . . . . . . . B.2 Random Variables . . . . . . . . . . . . . . B.3 Expectation . . . . . . . . . . . . . . . . . . B.4 Gamma Function . . . . . . . . . . . . . . . B.5 Common Distributions . . . . . . . . . . . . B.6 Multivariate Random Variables . . . . . . . B.7 Conditional Distributions and Expectation . B.8 Transformations . . . . . . . . . . . . . . . B.9 Normal and Related Distributions . . . . . B.10 Inequalities . . . . . . . . . . . . . . . . . . B.11 Maximum Likelihood . . . . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
. . . . . . . . . . .
C Numerical Optimization C.1 Grid Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.2 Gradient Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C.3 Derivative-Free Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
369 . 369 . 369 . 371
Preface
This book is intended to serve as the textbook for a rst-year graduate course in econometrics. It can be used as a stand-alone text, or be used as a supplement to another text. Students are assumed to have an understanding of multivariate calculus, probability theory, linear algebra, and mathematical statistics. A prior course in undergraduate econometrics would be helpful, but not required. Two excellent undergraduate textbooks are Wooldridge (2009) and Stock and Watson (2010). For reference, some of the basic tools of matrix algebra, probability, and statistics are reviewed in the Appendix. For students wishing to deepen their knowledge of matrix algebra in relation to their study of econometrics, I recommend Matrix Algebra by Abadir and Magnus (2005). An excellent introduction to probability and statistics is Statistical Inference by Casella and Berger (2002). For those wanting a deeper foundation in probability, I recommend Ash (1972) or Billingsley (1995). For more advanced statistical theory, I recommend Lehmann and Casella (1998), van der Vaart (1998), Shao (2003), and Lehmann and Romano (2005). For further study in econometrics beyond this text, I recommend Davidson (1994) for asymptotic theory, Hamilton (1994) for time-series methods, Wooldridge (2002) for panel data and discrete response models, and Li and Racine (2007) for nonparametrics and semiparametric econometrics. Beyond these texts, the Handbook of Econometrics series provides advanced summaries of contemporary econometric methods and theory. The end-of-chapter exercises are important parts of the text and are meant to help teach students of econometrics. Answers are not provided, and this is intentional. I would like to thank Ying-Ying Lee for providing research assistance in preparing some of the empirical examples presented in the text. As this is a manuscript in progress, some parts are quite incomplete, and there are many topics which I plan to add. In general, the earlier chapters are the most complete while the later chapters need signicant work and revision.
viii
Chapter 1
Introduction
1.1 What is Econometrics?
The term econometrics is believed to have been crafted by Ragnar Frisch (1895-1973) of Norway, one of the three principle founders of the Econometric Society, rst editor of the journal Econometrica, and co-winner of the rst Nobel Memorial Prize in Economic Sciences in 1969. It is therefore tting that we turn to Frischs own words in the introduction to the rst issue of Econometrica to describe the discipline. A word of explanation regarding the term econometrics may be in order. Its denition is implied in the statement of the scope of the [Econometric] Society, in Section I of the Constitution, which reads: The Econometric Society is an international society for the advancement of economic theory in its relation to statistics and mathematics.... Its main object shall be to promote studies that aim at a unication of the theoreticalquantitative and the empirical-quantitative approach to economic problems.... But there are several aspects of the quantitative approach to economics, and no single one of these aspects, taken by itself, should be confounded with econometrics. Thus, econometrics is by no means the same as economic statistics. Nor is it identical with what we call general economic theory, although a considerable portion of this theory has a deninitely quantitative character. Nor should econometrics be taken as synonomous with the application of mathematics to economics. Experience has shown that each of these three view-points, that of statistics, economic theory, and mathematics, is a necessary, but not by itself a sucient, condition for a real understanding of the quantitative relations in modern economic life. It is the unication of all three that is powerful. And it is this unication that constitutes econometrics. Ragnar Frisch, Econometrica, (1933), 1, pp. 1-2. This denition remains valid today, although some terms have evolved somewhat in their usage. Today, we would say that econometrics is the unied study of economic models, mathematical statistics, and economic data. Within the eld of econometrics there are sub-divisions and specializations. Econometric theory concerns the development of tools and methods, and the study of the properties of econometric methods. Applied econometrics is a term describing the development of quantitative economic models and the application of econometric methods to these models using economic data.
1.2
The Probability Approach to Econometrics
The unifying methodology of modern econometrics was articulated by Trygve Haavelmo (19111999) of Norway, winner of the 1989 Nobel Memorial Prize in Economic Sciences, in his seminal 1
CHAPTER 1. INTRODUCTION
paper The probability approach in econometrics, Econometrica (1944). Haavelmo argued that quantitative economic models must necessarily be probability models (by which today we would mean stochastic ). Deterministic models are blatently inconsistent with observed economic quantities, and it is incoherent to apply deterministic models to non-deterministic data. Economic models should be explicitly designed to incorporate randomness; stochastic errors should not be simply added to deterministic models to make them random. Once we acknowledge that an economic model is a probability model, it follows naturally that an appropriate tool way to quantify, estimate, and conduct inferences about the economy is through the powerful theory of mathematical statistics. The appropriate method for a quantitative economic analysis follows from the probabilistic construction of the economic model. Haavelmos probability approach was quickly embraced by the economics profession. Today no quantitative work in economics shuns its fundamental vision. While all economists embrace the probability approach, there has been some evolution in its implementation. The structural approach is the closest to Haavelmos original idea. A probabilistic economic model is specied, and the quantitative analysis performed under the assumption that the economic model is correctly specied. Researchers often describe this as taking their model seriously. The structural approach typically leads to likelihood-based analysis, including maximum likelihood and Bayesian estimation. A criticism of the structural approach is that it is misleading to treat an economic model as correctly specied. Rather, it is more accurate to view a model as a useful abstraction or approximation. In this case, how should we interpret structural econometric analysis? The quasistructural approach to inference views a structural economic model as an approximation rather than the truth. This theory has led to the concepts of the pseudo-true value (the parameter value dened by the estimation problem), the quasi-likelihood function, quasi-MLE, and quasi-likelihood inference. Closely related is the semiparametric approach. A probabilistic economic model is partially specied but some features are left unspecied. This approach typically leads to estimation methods such as least-squares and the Generalized Method of Moments. The semiparametric approach dominates contemporary econometrics, and is the main focus of this textbook. Another branch of quantitative structural economics is the calibration approach. Similar to the quasi-structural approach, the calibration approach interprets structural models as approximations and hence inherently false. The dierence is that the calibrationist literature rejects mathematical statistics (deeming classical theory as inappropriate for approximate models) and instead selects parameters by matching model and data moments using non-statistical ad hoc 1 methods.
1.3
Econometric Terms and Notation
In a typical application, an econometrician has a set of repeated measurements on a set of variables. For example, in a labor application the variables could include weekly earnings, educational attainment, age, and other descriptive characteristics. We call this information the data, dataset, or sample. We use the term observations to refer to the distinct repeated measurements on the variables. An individual observation often corresponds to a specic economic unit, such as a person, household, corporation, rm, organization, country, state, city or other geographical region. An individual observation could also be a measurement at a point in time, such as quarterly GDP or a daily interest rate.
Ad hoc means for this purpose a method designed for a specic problem and not based on a generalizable principle.
1
Economists typically denote variables by the italicized roman characters y , x, and/or z. The convention in econometrics is to use the character y to denote the variable to be explained, while the characters x and z are used to denote the conditioning (explaining) variables. Following mathematical convention, real numbers (elements of the real line R) are written using lower case italics such as y , and vectors (elements of Rk ) by lower case bold italics such as x, e.g. x1 x2 x = . . . . xk
Upper case bold italics such as X are used for matrices. We typically denote the number of observations by the natural number n, and subscript the variables by the index i to denote the individual observation, e.g. yi , xi and z i . In some contexts we use indices other than i, such as in time-series applications where the index t is common, and in panel studies we typically use the double index it to refer to individual i at a time period t.
The ith observation is the set (yi , xi , z i ). The sample is the set {(yi , xi , z i ) : i = 1, ..., n}. It is proper mathematical practice to use upper case X for random variables and lower case x for realizations or specic values. Since we use upper case to denote matrices, the distinction between random variables and their realizations is not rigorously followed in econometric notation. Thus the notation yi will in some places refer to a random variable, and in other places a specic realization. This is an undesirable but there is little to be done about it without terrically complicating the notation. Hopefully there will be no confusion as the use should be evident from the context. We typically use Greek letters such as , and 2 to denote unknown parameters of an econometric model, and will use boldface, e.g. or , when these are vector-valued. Estimates are typically denoted by putting a hat ^, tilde ~ or bar - over the corresponding letter, e.g. are estimates of . and The covariance matrix of an econometric estimator will typically be written using the capital b boldface V , often with a subscript to denote the estimator, e.g. V = var as the covariance b . Hopefully without causing confusion, we will use the notation V = avar( b ) to denote matrix for b the asymptotic covariance matrix of n (the variance of the asymptotic distribution). b is an estimate of V . Estimates will be denoted by appending hats or tildes, e.g. V
1.4
Observational Data
A common econometric question is to quantify the impact of one set of variables on another variable. For example, a concern in labor economics is the returns to schooling the change in earnings induced by increasing a workers education, holding other variables constant. Another issue of interest is the earnings gap between men and women. Ideally, we would use experimental data to answer these questions. To measure the returns to schooling, an experiment might randomly divide children into groups, mandate dierent levels of education to the dierent groups, and then follow the childrens wage path after they mature and enter the labor force. The dierences between the groups would be direct measurements of the effects of dierent levels of education. However, experiments such as this would be widely condemned as immoral! Consequently, we see few non-laboratory experimental data sets in economics.
Instead, most economic data is observational. To continue the above example, through data collection we can record the level of a persons education and their wage. With such data we can measure the joint distribution of these variables, and assess the joint dependence. But from observational data it is dicult to infer causality, as we are not able to manipulate one variable to see the direct eect on the other. For example, a persons level of education is (at least partially) determined by that persons choices. These factors are likely to be aected by their personal abilities and attitudes towards work. The fact that a person is highly educated suggests a high level of ability, which suggests a high relative wage. This is an alternative explanation for an observed positive correlation between educational levels and wages. High ability individuals do better in school, and therefore choose to attain higher levels of education, and their high ability is the fundamental reason for their high wages. The point is that multiple explanations are consistent with a positive correlation between schooling levels and education. Knowledge of the joint distibution alone may not be able to distinguish between these explanations. Most economic data sets are observational, not experimental. This means that all variables must be treated as random and possibly jointly determined. This discussion means that it is dicult to infer causality from observational data alone. Causal inference requires identication, and this is based on strong assumptions. We will discuss these issues on occasion throughout the text.
1.5
Standard Data Structures
There are three major types of economic data sets: cross-sectional, time-series, and panel. They are distinguished by the dependence structure across observations. Cross-sectional data sets have one observation per individual. Surveys are a typical source for cross-sectional data. In typical applications, the individuals surveyed are persons, households, rms or other economic agents. In many contemporary econometric cross-section studies the sample size n is quite large. It is conventional to assume that cross-sectional observations are mutually independent. Most of this text is devoted to the study of cross-section data. Time-series data are indexed by time. Typical examples include macroeconomic aggregates, prices and interest rates. This type of data is characterized by serial dependence so the random sampling assumption is inappropriate. Most aggregate economic data is only available at a low frequency (annual, quarterly or perhaps monthly) so the sample size is typically much smaller than in cross-section studies. The exception is nancial data where data are available at a high frequency (weekly, daily, hourly, or by transaction) so sample sizes can be quite large. Panel data combines elements of cross-section and time-series. These data sets consist of a set of individuals (typically persons, households, or corporations) surveyed repeatedly over time. The common modeling assumption is that the individuals are mutually independent of one another, but a given individuals observations are mutually dependent. This is a modied random sampling environment. Data Structures Cross-section Time-series Panel
Some contemporary econometric applications combine elements of cross-section, time-series, and panel data modeling. These include models of spatial correlation and clustering. As we mentioned above, most of this text will be devoted to cross-sectional data under the assumption of mutually independent observations. By mutual independence we mean that the ith observation (yi , xi , z i ) is independent of the j th observation (yj , xj , z j ) for i 6= j . (Sometimes the label independent is misconstrued. It is a statement about the relationship between observations i and j , not a statement about the relationship between yi and xi and/or z i .) Furthermore, if the data is randomly gathered, it is reasonable to model each observation as a random draw from the same probability distribution. In this case we say that the data are independent and identically distributed or iid. We call this a random sample. For most of this text we will assume that our observations come from a random sample.
Denition 1.5.1 The observations (yi , xi , z i ) are a random sample if they are mutually independent and identically distributed (iid) across i = 1, ..., n.
In the random sampling framework, we think of an individual observation (yi , xi , z i ) as a realization from a joint probability distribution F (y, x, z ) which we can call the population. This population is innitely large. This abstraction can be a source of confusion as it does not correspond to a physical population in the real world. Its an abstraction since the distribution F is unknown, and the goal of statistical inference is to learn about features of F from the sample. The assumption of random sampling provides the mathematical foundation for treating economic statistics with the tools of mathematical statistics. The random sampling framework was a major intellectural breakthrough of the late 19th century, allowing the application of mathematical statistics to the social sciences. Before this conceptual development, methods from mathematical statistics had not been applied to economic data as they were viewed as inappropriate. The random sampling framework enabled economic samples to be viewed as homogenous and random, a necessary precondition for the application of statistical methods.
1.6
Sources for Economic Data
Fortunately for economists, the internet provides a convenient forum for dissemination of economic data. Many large-scale economic datasets are available without charge from governmental agencies. An excellent starting point is the Resources for Economists Data Links, available at rfe.org. From this site you can nd almost every publically available economic data set. Some specic data sources of interest include Bureau of Labor Statistics US Census Current Population Survey Survey of Income and Program Participation Panel Study of Income Dynamics Federal Reserve System (Board of Governors and regional banks) National Bureau of Economic Research
CHAPTER 1. INTRODUCTION U.S. Bureau of Economic Analysis CompuStat International Financial Statistics
Another good source of data is from authors of published empirical studies. Most journals in economics require authors of published papers to make their datasets generally available. For example, in its instructions for submission, Econometrica states: Econometrica has the policy that all empirical, experimental and simulation results must be replicable. Therefore, authors of accepted papers must submit data sets, programs, and information on empirical analysis, experiments and simulations that are needed for replication and some limited sensitivity analysis. The American Economic Review states: All data used in analysis must be made available to any researcher for purposes of replication. The Journal of Political Economy states: It is the policy of the Journal of Political Economy to publish papers only if the data used in the analysis are clearly and precisely documented and are readily available to any researcher for purposes of replication. If you are interested in using the data from a published paper, rst check the journals website, as many journals archive data and replication programs online. Second, check the website(s) of the papers author(s). Most academic economists maintain webpages, and some make available replication les complete with data and programs. If these investigations fail, email the author(s), politely requesting the data. You may need to be persistent. As a matter of professional etiquette, all authors absolutely have the obligation to make their data and programs available. Unfortunately, many fail to do so, and typically for poor reasons. The irony of the situation is that it is typically in the best interests of a scholar to make as much of their work (including all data and programs) freely available, as this only increases the likelihood of their work being cited and having an impact. Keep this in mind as you start your own empirical project. Remember that as part of your end product, you will need (and want) to provide all data and programs to the community of scholars. The greatest form of attery is to learn that another scholar has read your paper, wants to extend your work, or wants to use your empirical methods. In addition, public openness provides a healthy incentive for transparency and integrity in empirical analysis.
1.7
Econometric Software
Economists use a variety of econometric, statistical, and programming software. STATA (www.stata.com) is a powerful statistical program with a broad set of pre-programmed econometric and statistical tools. It is quite popular among economists, and is continuously being updated with new methods. It is an excellent package for most econometric analysis, but is limited when you want to use new or less-common econometric methods which have not yet been programed. R (www.r-project.org), GAUSS (www.aptech.com), MATLAB (www.mathworks.com), and Ox (www.oxmetrics.net) are high-level matrix programming languages with a wide variety of built-in statistical functions. Many econometric methods have been programed in these languages and are available on the web. The advantage of these packages is that you are in complete control of your
analysis, and it is easier to program new methods than in STATA. Some disadvantages are that you have to do much of the programming yourself, programming complicated procedures takes signicant time, and programming errors are hard to prevent and dicult to detect and eliminate. Of these languages, Gauss used to be quite popular among econometricians, but now Matlab is more popular. A smaller but growing group of econometricians are enthusiastic fans of R, which of these languages is uniquely open-source, user-contributed, and best of all, completely free! For highly-intensive computational tasks, some economists write their programs in a standard programming language such as Fortran or C. This can lead to major gains in computational speed, at the cost of increased time in programming and debugging. As these dierent packages have distinct advantages, many empirical economists end up using more than one package. As a student of econometrics, you will learn at least one of these packages, and probably more than one.
1.8
Reading the Manuscript
I have endeavored to use a unied notation and nomenclature. The development of the material is cumulative, with later chapters building on the earlier ones. Never-the-less, every attempt has been made to make each chapter self-contained, so readers can pick and choose topics according to their interests. To fully understand econometric methods, it is necessary to have a mathematical understanding of its mechanics, and this includes the mathematical proofs of the main results. Consequently, this text is self-contained, with nearly all results proved with full mathematical rigor. The mathematical development and proofs aim at brevity and conciseness (sometimes described as mathematical elegance), but also at pedagogy. To understand a mathematical proof, it is not sucient to simply read the proof, you need to follow it, and re-create it for yourself. Never-the-less, many readers will not be interested in each mathematical detail, explanation, or proof. This is okay. To use a method it may not be necessary to understand the mathematical details. Accordingly I have placed the more technical mathematical proofs and details in chapter appendices. These appendices and other technical sections are marked with an asterisk (*). These sections can be skipped without any loss in exposition.
1.9
Common Symbols
y x X R Rk E (y ) var (y ) cov (x, y ) var (x) corr(x, y) Pr p
d
scalar vector matrix real line Euclidean k space mathematical expectation variance covariance covariance matrix correlation probability limit convergence in probability convergence in distribution probability limit normal distribution standard normal distribution chi-square distribution with k degrees of freedom identity matrix trace matrix transpose matrix inverse positive denite, positive semi-denite Euclidean norm matrix (Frobinius) norm approximate equality denitional equality is distributed as natural logarithm
plimn N(, 2 ) N(0, 1) 2 k In tr A A0 A1 A > 0, A 0 kak kAk = log

def
Chapter 2
Conditional Expectation and Projection

2.1 Introduction
The most commonly applied econometric tool is least-squares estimation, also known as regression. As we will see, least-squares is a tool to estimate an approximate conditional mean of one variable (the dependent variable) given another set of variables (the regressors, conditioning variables, or covariates). In this chapter we abstract from estimation, and focus on the probabilistic foundation of the conditional expectation model and its projection approximation.
2.2
The Distribution of Wages
Suppose that we are interested in wage rates in the United States. Since wage rates vary across workers, we cannot describe wage rates by a single number. Instead, we can describe wages using a probability distribution. Formally, we view the wage of an individual worker as a random variable wage with the probability distribution F (u) = Pr(wage u). When we say that a persons wage is random we mean that we do not know their wage before it is measured, and we treat observed wage rates as realizations from the distribution F. Treating unobserved wages as random variables and observed wages as realizations is a powerful mathematical abstraction which allows us to use the tools of mathematical probability. A useful thought experiment is to imagine dialing a telephone number selected at random, and then asking the person who responds to tell us their wage rate. (Assume for simplicity that all workers have equal access to telephones, and that the person who answers your call will respond honestly.) In this thought experiment, the wage of the person you have called is a single draw from the distribution F of wages in the population. By making many such phone calls we can learn the distribution F of the entire population. When a distribution function F is dierentiable we dene the probability density function f (u) = d F (u). du
The density contains the same information as the distribution function, but the density is typically easier to visually interpret.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION
10
Wage Distribution
0.6
0.7
0.8
0.9
1.0
0.0
0.1
0.2
0.3
0.4
10
20
30
40
50
60
70
Wage Density
0.5
10
20
30
40
50
60
70
80
90
100
Dollars per Hour
Dollars per Hour
Figure 2.1: Wage Distribution and Density. All full-time U.S. workers In Figure 2.1 we display estimates1 of the probability distribution function (on the left) and density function (on the right) of U.S. wage rates in 2009. We see that the density is peaked around $15, and most of the probability mass appears to lie between $10 and $40. These are ranges for typical wage rates in the U.S. population. Important measures of central tendency are the median and the mean. The median m of a continuous2 distribution F is the unique solution to 1 F (m) = . 2 The median U.S. wage ($19.23) is indicated in the left panel of Figure 2.1 by the arrow. The median is a robust3 measure of central tendency, but it is tricky to use for many calculations as it is not a linear operator. The expectation or mean of a random variable y with density f is Z uf (u)du. = E (y) =
A general denition of the mean is presented in Section 2.31. The mean U.S. wage ($23.90) is indicated in the right panel of Figure 2.1 by the arrow. Here we have used the common and convenient convention of using the single character y to denote a random variable, rather than the more cumbersome label wage. We sometimes use the notation the notation Ey instead of E (y) when the variable whose expectation is being taken is clear from the context. There is no distinction in meaning. The mean is a convenient measure of central tendency because it is a linear operator and arises naturally in many economic models. A disadvantage of the mean is that it is not robust4 especially in the presence of substantial skewness or thick tails, which are both features of the wage
The distribution and density are estimated nonparametrically from the sample of 50,742 full-time non-military wage-earners reported in the March 2009 Current Population Survey. The wage rate is constructed as annual individual wage and salary earnings divided by hours worked. 1 2 If F is not continuous the denition is m = inf {u : F (u) } 2 3 The median is not sensitive to pertubations in the tails of the distribution. 4 The mean is sensitive to pertubations in the tails of the distribution.
1
11
distribution as can be seen easily in the right panel of Figure 2.1. Another way of viewing this is that 64% of workers earn less that the mean wage of $23.90, suggesting that it is incorrect to describe the mean as a typical wage rate.
Log Wage Density
Log Dollars per Hour
Figure 2.2: Log Wage Density In this context it is useful to transform the data by taking the natural logarithm5 . Figure 2.2 shows the density of log hourly wages log(wage) for the same population, with its mean 2.95 drawn in with the arrow. The density of log wages is much less skewed and fat-tailed than the density of the level of wages, so its mean E (log(wage)) = 2.95 is a much better (more robust) measure6 of central tendency of the distribution. For this reason, wage regressions typically use log wages as a dependent variable rather than the level of wages. Another useful way to summarize the probability distribution F (u) is in terms of its quantiles. For any (0, 1), the th quantile of the continuous7 distribution F is the real number q which satises F (q ) = . The quantile function q , viewed as a function of , is the inverse of the distribution function F. The most commonly used quantile is the median, that is, q0.5 = m. We sometimes refer to quantiles by the percentile representation of , and in this case they are often called percentiles, e.g. the median is the 50th percentile.
2.3
Conditional Expectation
We saw in Figure 2.2 the density of log wages. Is this distribution the same for all workers, or does the wage distribution vary across subpopulations? To answer this question, we can compare wage distributions for dierent groups for example, men and women. The plot on the left in Figure 2.3 displays the densities of log wages for U.S. men and women with their means (3.05 and 2.81) indicated by the arrows. We can see that the two wage densities take similar shapes but the density for men is somewhat shifted to the right with a higher mean.
5 6
Throughout the text, we will use log(y ) to denote the natural logarithm of y. More precisely, the geometric mean exp (E (log w)) = $19.11 is a robust measure of central tendency. 7 If F is not continuous the denition is q = inf {u : F (u) }
12
white men white women black men black women
Log Wage Density
Women
Men
3 Log Dollars per Hour
Log Wage Density
3 Log Dollars per Hour
(a) Women and Men
(b) By Gender and Race
Figure 2.3: Log Wage Density by Gender and Race The values 3.05 and 2.81 are the mean log wages in the subpopulations of men and women workers. They are called the conditional means (or conditional expectations) of log wages given gender. We can write their specic values as E (log(wage) | gender = man) = 3.05 E (log(wage) | gender = woman) = 2.81. (2.1)
(2.2)
We call these means conditional as they are conditioning on a xed value of the variable gender. While you might not think of a persons gender as a random variable, it is random from the viewpoint of econometric analysis. If you randomly select an individual, the gender of the individual is unknown and thus random. (In the population of U.S. workers, the probability that a worker is a woman happens to be 43%.) In observational data, it is most appropriate to view all measurements as random variables, and the means of subpopulations are then conditional means. As the two densities in Figure 2.3 appear similar, a hasty inference might be that there is not a meaningful dierence between the wage distributions of men and women. Before jumping to this conclusion let us examine the dierences in the distributions of Figure 2.3 more carefully. As we mentioned above, the primary dierence between the two densities appears to be their means. This dierence equals E (log(wage) | gender = man) E (log(wage) | gender = woman) = 3.05 2.81 = 0.24
(2.3)
A dierence in expected log wages of 0.24 implies an average 24% dierence between the wages of men and women, which is quite substantial. (For an explanation of logarithmic and percentage dierences see Section 2.4.) Consider further splitting the men and women subpopulations by race, dividing the population into whites, blacks, and other races. We display the log wage density functions of four of these groups on the right in Figure 2.3. Again we see that the primary dierence between the four density functions is their central tendency.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION men 3.07 2.86 3.03 women 2.82 2.73 2.86
13
white black other
Table 2.1: Mean Log Wages by Sex and Race Focusing on the means of these distributions, Table 2.1 reports the mean log wage for each of the six sub-populations. The entries in Table 2.1 are the conditional means of log(wage) given gender and race. For example E (log(wage) | gender = man, race = white) = 3.07 and E (log(wage) | gender = woman, race = black) = 2.73 One benet of focusing on conditional means is that they reduce complicated distributions to a single summary measure, and thereby facilitate comparisons across groups. Because of this simplifying property, conditional means are the primary interest of regression analysis and are a major focus in econometrics. Table 2.1 allows us to easily calculate average wage dierences between groups. For example, we can see that the wage gap between men and women continues after disaggregation by race, as the average gap between white men and white women is 25%, and that between black men and black women is 13%. We also can see that there is a race gap, as the average wages of blacks are substantially less than the other race categories. In particular, the average wage gap between white men and black men is 21%, and that between white women and black women is 9%.
2.4
Log Dierences*
log (1 + x) x. (2.4)
A useful approximation for the natural logarithm for small x is
This can be derived from the innite series expansion of log (1 + x) : log (1 + x) = x x2 x3 x4 + + 2 3 4 = x + O(x2 ).
The symbol O(x2 ) means that the remainder is bounded by Ax2 as x 0 for some A < . A plot of log (1 + x) and the linear approximation x is shown in Figure 2.4. We can see that log (1 + x) and the linear approximation x are very close for |x| 0.1, and reasonably close for |x| 0.2, but the dierence increases with |x|. Now, if y is c% greater than y, then y = (1 + c/100)y. Taking natural logarithms, log y = log y + log(1 + c/100) c 100 where the approximation is (2.4). This shows that 100 multiplied by the dierence in logarithms is approximately the percentage dierence between y and y , and this approximation is quite good for |c| 10. log y log y = log(1 + c/100) or
14
Figure 2.4: log(1 + x)
2.5
Conditional Expectation Function
An important determinant of wage levels is education. In many empirical studies economists measure educational attainment by the number of years of schooling, and we will write this variable as education 8 . The conditional mean of log wages given gender, race, and education is a single number for each category. For example E (log(wage) | gender = man, race = white, education = 12) = 2.84 We display in Figure 2.5 the conditional means of log(wage) for white men and white women as a function of education. The plot is quite revealing. We see that the conditional mean is increasing in years of education, but at a dierent rate for schooling levels above and below nine years. Another striking feature of Figure 2.5 is that the gap between men and women is roughly constant for all education levels. As the variables are measured in logs this implies a constant average percentage gap between men and women regardless of educational attainment. In many cases it is convenient to simplify the notation by writing variables using single characters, typically y, x and/or z . It is conventional in econometrics to denote the dependent variable (e.g. log(wage)) by the letter y, a conditioning variable (such as gender ) by the letter x, and multiple conditioning variables (such as race, education and gender ) by the subscripted letters x1 , x2 , ..., xk . Conditional expectations can be written with the generic notation E (y | x1 , x2 , ..., xk ) = m(x1 , x2 , ..., xk ). We call this the conditional expectation function (CEF). The CEF is a function of (x1 , x2 , ..., xk ) as it varies with the variables. For example, the conditional expectation of y = log(wage) given (x1 , x2 ) = (gender , race ) is given by the six entries of Table 2.1. The CEF is a function of (gender , race ) as it varies across the entries.
8 Here, education is dened as years of schooling beyond kindergarten. A high school graduate has education =12, a college graduate has education =16, a Masters degree has education =18, and a professional degree (medical, law or PhD) has education =20.
15
2.0
2.5
3.0
3.5
4.0
white men white women

4 6 8 10 12 14 16 18 20
Years of Education
Figure 2.5: Mean Log Wage as a Function of Years of Education For greater compactness, we will typically write the conditioning variables as a vector in Rk : x1 x2 (2.5) x = . . . . xk
Here we follow the convention of using lower case bold italics x to denote a vector. Given this notation, the CEF can be compactly written as E (y | x) = m (x) .
The CEF E (y | x) is a random variable as it is a function of the random variable x. It is also sometimes useful to view the CEF as a function of x. In this case we can write m (u) = E (y | x = u), which is a function of the argument u. The expression E (y | x = u) is the conditional expectation of y, given that we know that the random variable x equals the specic value u. However, sometimes in econometrics we take a notational shortcut and use E (y | x) to refer to this function. Hopefully, the use of E (y | x) should be apparent from the context.
2.6
Continuous Variables
In the previous sections, we implicitly assumed that the conditioning variables are discrete. However, many conditioning variables are continuous. In this section, we take up this case and assume that the variables (y, x) are continuously distributed with a joint density function f (y, x). As an example, take y = log(wage) and x = experience, the number of years of potential labor market experience9 . The contours of their joint density are plotted on the left side of Figure 2.6 for the population of white men with 12 years of education. Given the joint density f (y, x) the variable x has the marginal density Z f (y, x)dy. fx (x) =
R
9
Here, experience is dened as potential labor market experience, equal to age education 6
16
4.0
3.5
Exp=5 Exp=10 Exp=25 Exp=40
Conditional Mean Linear Projection Quadratic Projection 0 10 20 30 40 50

1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
2.0
Labor Market Experience (Years)

(a) Joint density of log(wage) and experience and conditional mean
Log Wage Conditional Density
2.5
3.0
(b) Conditional density
Figure 2.6: White men with education =12 For any x such that fx (x) > 0 the conditional density of y given x is dened as fy|x (y | x) = f (y, x) . fx (x) (2.6)
The conditional density is a slice of the joint density f (y, x) holding x xed. We can visualize this by slicing the joint density function at a specic value of x parallel with the y -axis. For example, take the density contours on the left side of Figure 2.6 and slice through the contour plot at a specic value of experience. This gives us the conditional density of log(wage) for white men with 12 years of education and this level of experience. We do this for four levels of experience (5, 10, 25, and 40 years), and plot these densities on the right side of Figure 2.6. We can see that the distribution of wages shifts to the right and becomes more diuse as experience increases from 5 to 10 years, and from 10 to 25 years, but there is little change from 25 to 40 years experience. The CEF of y given x is the mean of the conditional density (2.6) Z yfy|x (y | x) dy. (2.7) m (x) = E (y | x) =
R
Intuitively, m (x) is the mean of y for the idealized subpopulation where the conditioning variables are xed at x. This is idealized since x is continuously distributed so this subpopulation is innitely small. In Figure 2.6 the CEF of log(wage) given experience is plotted as the solid line. We can see that the CEF is a smooth but nonlinear function. The CEF is initially increasing in experience, attens out around experience = 30, and then decreases for high levels of experience.
2.7
Law of Iterated Expectations
An extremely useful tool from probability theory is the law of iterated expectations. An important special case is the known as the Simple Law.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION Theorem 2.7.1 Simple Law of Iterated Expectations If E |y | < then for any random vector x, E (E (y | x)) = E (y)
17
The simple law states that the expectation of the conditional expectation is the unconditional expectation. In other words, the average of the conditional averages is the unconditional average. When x is discrete X E (y | xj ) Pr (x = xj ) E (E (y | x)) =
j =1
and when x is continuous E (E (y | x)) =
Going back to our investigation of average log wages for men and women, the simple law states that E (log(wage) | gender = man) Pr (gender = man) = E (log(wage)) .
Rk
E (y | x) fx (x)dx.
+ E (log(wage) | gender = woman) Pr (gender = woman)
Or numerically, 3.05 0.57 + 2.79 0.43 = 2.92. The general law of iterated expectations allows two sets of conditioning variables.
Theorem 2.7.2 Law of Iterated Expectations If E |y | < then for any random vectors x1 and x2 , E (E (y | x1 , x2 ) | x1 ) = E (y | x1 )
Notice the way the law is applied. The inner expectation conditions on x1 and x2 , while the outer expectation conditions only on x1 . The iterated expectation yields the simple answer E (y | x1 ) , the expectation conditional on x1 alone. Sometimes we phrase this as: The smaller information set wins. As an example E (log(wage) | gender = man, race = white) Pr (race = white|gender = man)
+ E (log(wage) | gender = man, race = other) Pr (race = other|gender = man) = E (log(wage) | gender = man) or numerically 3.07 0.84 + 2.86 0.08 + 3.05 0.08 = 3.05. A property of conditional expectations is that when you condition on a random vector x you can eectively treat it as if it is constant. For example, E (x | x) = x and E (g (x) | x) = g (x) for any function g (). The general property is known as the conditioning theorem.
+ E (log(wage) | gender = man, race = black) Pr (race = black|gender = man)
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION Theorem 2.7.3 Conditioning Theorem If E |g (x) y | < then E (g (x) y | x) = g (x) E (y | x) and E (g (x) y ) = E (g (x) E (y | x)) . (2.10) (2.9)
18
(2.8)
The proofs of Theorems 2.7.1, 2.7.2 and 2.7.3 are given in Section 2.34.
2.8
CEF Error
The CEF error e is dened as the dierence between y and the CEF evaluated at the random vector x: e = y m(x). By construction, this yields the formula y = m(x) + e. (2.11)
In (2.11) it is useful to understand that the error e is derived from the joint distribution of (y, x), and so its properties are derived from this construction. A key property of the CEF error is that it has a conditional mean of zero. To see this, by the linearity of expectations, the denition m(x) = E (y | x) and the Conditioning Theorem E (e | x) = E ((y m(x)) | x) = m(x) m(x) = E (y | x) E (m(x) | x)
= 0.
This fact can be combined with the law of iterated expectations to show that the unconditional mean is also zero. E (e) = E (E (e | x)) = E (0) = 0. We state this and some other results formally. Theorem 2.8.1 Properties of the CEF error If E |y | < then 1. E (e | x) = 0. 2. E (e) = 0. 3. If E |y |r < for r 1 then E |e|r < . 4. For any function h (x) such that E |h (x) e| < then E (h (x) e) = 0. The proof of the third result is deferred to Section 2.34. The fourth result, whose proof is left to Exercise 2.3, says that e is uncorrelated with any function of the regressors.
19
2.0
0
2.5
3.0
3.5
4.0
10
20
30
40
50
Figure 2.7: Joint density of CEF error e and experience for white men with education =12. The equations y = m(x) + e E (e | x) = 0 together imply that m(x) is the CEF of y given x. It is important to understand that this is not a restriction. These equations hold true by denition. The condition E (e | x) = 0 is implied by the denition of e as the dierence between y and the CEF m (x) . The equation E (e | x) = 0 is sometimes called a conditional mean restriction, since the conditional mean of the error e is restricted to equal zero. The property is also sometimes called mean independence, for the conditional mean of e is 0 and thus independent of x. However, it does not imply that the distribution of e is independent of x. Sometimes the assumption e is independent of x is added as a convenient simplication, but it is not generic feature of the conditional mean. Typically and generally, e and x are jointly dependent, even though the conditional mean of e is zero. As an example, the contours of the joint density of e and experience are plotted in Figure 2.7 for the same population as Figure 2.6. The error e has a conditional mean of zero for all values of experience, but the shape of the conditional distribution varies with the level of experience. As a simple example of a case where x and e are mean independent yet dependent, let e = x where x and are independent N(0, 1). Then conditional on x, the error e has the distribution N(0, x2 ). Thus E (e | x) = 0 and e is mean independent of x, yet e is not fully independent of x. Mean independence does not imply full independence.
2.9
Intercept-Only Model
A special case of the regression model is when there are no regressors x . In this case m(x) = E (y ) = , the unconditional mean of y. We can still write an equation for y in the regression format: y =+e E (e) = 0 This is useful for it unies the notation.
20
2.10
Regression Variance
An important measure of the dispersion about the CEF function is the unconditional variance of the CEF error e. We write this as 2 = var (e) = E (e Ee)2 = E e2 . Theorem 2.8.1.3 implies the following simple but useful result.
Theorem 2.10.1 If Ey 2 < then 2 < . We can call 2 the regression variance or the variance of the regression error. The magnitude of 2 measures the amount of variation in y which is not explained or accounted for in the conditional mean E (y | x) . The regression variance depends on the regressors x. Consider two regressions y = E (y | x1 ) + e1
y = E (y | x1 , x2 ) + e2 .
We write the two errors distinctly as e1 and e2 as they are dierent changing the conditioning information changes the conditional mean and therefore the regression error as well. In our discussion of iterated expectations, we have seen that by increasing the conditioning set, the conditional expectation reveals greater detail about the distribution of y. What is the implication for the regression error? It turns out that there is a simple relationship. We can think of the conditional mean E (y | x) as the explained portion of y. The remainder e = y E (y | x) is the unexplained portion. The simple relationship we now derive shows that the variance of this unexplained portion decreases when we condition on more variables. This relationship is monotonic in the sense that increasing the amont of information always decreases the variance of the unexplained portion.
Theorem 2.10.2 If Ey 2 < then var (y ) var (y E (y | x1 )) var (y E (y | x1 , x2 )) .
Theorem 2.10.2 says that the variance of the dierence between y and its conditional mean (weakly) decreases whenever an additional variable is added to the conditioning information. The proof of Theorem 2.10.2 is given in Section 2.34.
2.11
Best Predictor
Suppose that given a realized value of x, we want to create a prediction or forecast of y. We can write any predictor as a function g (x) of x. The prediction error is the realized dierence y g (x). A non-stochastic measure of the magnitude of the prediction error is the expectation of its square E (y g (x))2 . (2.12)
We can dene the best predictor as the function g (x) which minimizes (2.12). What function is the best predictor? It turns out that the answer is the CEF m(x). This holds regardless of the joint distribution of (y, x).
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION To see this, note that the mean squared error of a predictor g (x) is E (y g (x))2 = E (e + m (x) g (x))2 = Ee2 + 2E (e (m (x) g (x))) + E (m (x) g (x))2 = Ee2 + E (m (x) g (x))2 Ee2 = E (y m (x))2
21
where the rst equality makes the substitution y = m(x) + e and the third equality uses Theorem 2.8.1.4. The right-hand-side after the third equality is minimized by setting g (x) = m (x), yielding the inequality in the fourth line. The minimum is nite under the assumption Ey 2 < as shown by Theorem 2.10.1. We state this formally in the following result. Theorem 2.11.1 Conditional Mean as Best Predictor If Ey2 < , then for any predictor g (x), E (y g (x))2 E (y m (x))2 where m (x) = E (y | x). It may be helpful to consider this result in the context of the intercept-only model y =+e E(e) = 0. Theorem 2.11.1 shows that the best predictor for y (in the class of constant parameters) is the unconditional mean = E(y ), in the sense that the mean minimizes the mean squared prediction error.
2.12
Conditional Variance
While the conditional mean is a good measure of the location of a conditional distribution, it does not provide information about the spread of the distribution. A common measure of the dispersion is the conditional variance. Denition 2.12.1 If Ey 2 < , the conditional variance of y given x is 2 (x) = var (y | x) = E (y E (y | x))2 | x = E e2 | x . Generally, 2 (x) is a non-trivial function of x and can take any form subject to the restriction that it is non-negative. One way to think about 2 (x) is that it is the conditional mean of e2 given x.
22
The variance is in a dierent unit of measurement than the original variable. To convert the variance back to the same unit of measure we dene the conditional standard deviation as its p square root (x) = 2 (x). As an example of how the conditional variance depends on observables, compare the conditional log wage densities for men and women displayed in Figure 2.3. The dierence between the densities is not purely a location shift, but is also a dierence in spread. Specically, we can see that the density for mens log wages is somewhat more spread out than that for women, while the density for womens wages is somewhat more peaked. Indeed, the conditional standard deviation for mens wages is 3.05 and that for women is 2.81. So while men have higher average wages, they are also somewhat more dispersed. The unconditional error variance and the conditional variance are related by the law of iterated expectations 2 = E e2 = E E e2 | x = E 2 (x) . That is, the unconditional error variance is the average conditional variance. Given the conditional variance, we can dene a rescaled error = e . (x)
(2.13)
We can calculate that since (x) is a function of x e 1 |x = E (e | x) = 0 E ( | x) = E (x) (x) and var ( | x) = E 2 | x = E 2 2 (x) 1 e2 = | x E e = 1. | x = 2 2 (x) 2 (x) (x)
Thus has a conditional mean of zero, and a conditional variance of 1. Notice that (2.13) can be rewritten as e = (x). and substituting this for e in the CEF equation (2.11), we nd that y = m(x) + (x).
(2.14)
This is an alternative (mean-variance) representation of the CEF equation. Many econometric studies focus on the conditional mean m(x) and either ignore the conditional variance 2 (x), treat it as a constant 2 (x) = 2 , or treat it as a nuisance parameter (a parameter not of primary interest). This is appropriate when the primary variation in the conditional distribution is in the mean, but can be short-sighted in other cases. Dispersion is relevant to many economic topics, including income and wealth distribution, economic inequality, and price dispersion. Conditional dispersion (variance) can be a fruitful subject for investigation. The perverse consequences of a narrow-minded focus on the mean has been parodied in a classic joke:
An economist was standing with one foot in a bucket of boiling water and the other foot in a bucket of ice. When asked how he felt, he replied, On average I feel just ne.
Clearly, the economist in question ignored variance!
23
2.13
Homoskedasticity and Heteroskedasticity
An important special case obtains when the conditional variance 2 (x) is a constant and independent of x. This is called homoskedasticity. Denition 2.13.1 The error is homoskedastic if E e2 | x = 2 does not depend on x. In the general case where 2 (x) depends on x we say that the error e is heteroskedastic. Denition 2.13.2 The error is heteroskedastic if E e2 | x = 2 (x) depends on x. It is helpful to understand that the concepts homoskedasticity and heteroskedasticity concern the conditional variance, not the unconditional variance. By denition, the unconditional variance 2 is a constant and independent of the regressors x. So when we talk about the variance as a function of the regressors, we are talking about the conditional variance 2 (x). Some older or introductory textbooks describe heteroskedasticity as the case where the variance of e varies across observations. This is a poor and confusing denition. It is more constructive to understand that heteroskedasticity means that the conditional variance 2 (x) depends on observables. Older textbooks also tend to describe homoskedasticity as a component of a correct regression specication, and describe heteroskedasticity as an exception or deviance. This description has inuenced many generations of economists, but it is unfortunately backwards. The correct view is that heteroskedasticity is generic and standard, while homoskedasticity is unusual and exceptional. The default in empirical work should be to assume that the errors are heteroskedastic, not the converse. In apparent contradiction to the above statement, we will still frequently impose the homoskedasticity assumption when making theoretical investigations into the properties of estimation and inference methods. The reason is that in many cases homoskedasticity greatly simplies the theoretical calculations, and it is therefore quite advantageous for teaching and learning. It should always be remembered, however, that homoskedasticity is never imposed because it is believed to be a correct feature of an empirical model, but rather because of its simplicity.
2.14
Regression Derivative
One way to interpret the CEF m(x) = E (y | x) is in terms of how marginal changes in the regressors x imply changes in the conditional mean of the response variable y. It is typical to consider marginal changes in a single regressor, say x1 , holding the remainder xed. When a regressor x1 is continuously distributed, we dene the marginal eect of a change in x1 , holding the variables x2 , ..., xk xed, as the partial derivative of the CEF m(x1 , ..., xk ). x1 When x1 is discrete we dene the marginal eect as a discrete dierence. For example, if x1 is binary, then the marginal eect of x1 on the CEF is m(1, x2 , ..., xk ) m(0, x2 , ..., xk ).
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION We can unify the continuous and discrete cases with the notation m(x1 , ..., xk ), if x1 is continuous x1 1 m(x) = m(1, x , ..., x ) m(0, x , ..., x ), if x is binary.
2 k 2 k 1
24
m(x), the When all elements of x are continuous, then we have the simplication m(x) = x vector of partial derivatives. There are two important points to remember concerning our denition of the regression derivative. First, the eect of each variable is calculated holding the other variables constant. This is the ceteris paribus concept commonly used in economics. But in the case of a regression derivative, the conditional mean does not literally hold all else constant. It only holds constant the variables included in the conditional mean. This means that the regression derivative depends on which regressors are included. For example, in a regression of wages on education, experience, race and gender, the regression derivative with respect to education shows the marginal eect of education on mean wages, holding constant experience, race and gender. But it does not hold constant an individuals unobservable characteristics (such as ability), or variables not included in the regression (such as the quality of education). Second, the regression derivative is the change in the conditional expectation of y , not the change in the actual value of y for an individual. It is tempting to think of the regression derivative as the change in the actual value of y , but this is not a correct interpretation. The regression derivative m(x) is the change in the actual value of y only if the error e is unaected by the change in the regressor x. We return to a discussion of causal eects in Section 2.30.
Collecting the k eects into one k 1 vector, we dene the regression derivative with respect to x: 1 m(x) 2 m(x) m(x) = . . . k m(x)
2.15
Linear CEF
An important special case is when the CEF m (x) = E (y | x) is linear in x. In this case we can write the mean equation as m(x) = x1 1 + x2 2 + + xk k + k+1 . Notationally it is convenient to write this as a simple function of the vector x. An easy way to do so is to augment the regressor vector x by listing the number 1 as an element. We call this the constant and the corresponding coecient is called the intercept. Equivalently, specify that the nal element10 of the vector x is xk = 1. Thus (2.5) has been redened as the k 1 vector x1 x2 . (2.15) x= . . . xk1 1
10
The order doesnt matter. It could be any element.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION With this redenition, the CEF is m(x) = x1 1 + x2 2 + + xk k = x0 where 1 . = . . k
25
(2.16)
(2.17)
is a k 1 coecient vector. This is the linear CEF model. It is also often called the linear regression model, or the regression of y on x. In the linear CEF model, the regression derivative is simply the coecient vector. That is m(x) = . This is one of the appealing features of the linear CEF model. The coecients have simple and natural interpretations as the marginal eects of changing one variable, holding the others constant.
Linear CEF Model y = x0 + e E (e | x) = 0
If in addition the error is homoskedastic, we call this the homoskedastic linear CEF model.
Homoskedastic Linear CEF Model y = x0 + e E (e | x) = 0 E e2 | x = 2
2.16
Linear CEF with Nonlinear Eects
The linear CEF model of the previous section is less restrictive than it might appear, as we can include as regressors nonlinear transformations of the original variables. In this sense, the linear CEF framework is exible and can capture many nonlinear eects. For example, suppose we have two scalar variables x1 and x2 . The CEF could take the quadratic form 2 (2.18) m(x1 , x2 ) = x1 1 + x2 2 + x2 1 3 + x2 4 + x1 x2 5 + 6 . This equation is quadratic in the regressors (x1 , x2 ) yet linear in the coecients = (1 , ..., 6 )0 . We will descriptively call (2.18) a quadratic CEF, and yet (2.18) is also a linear CEF in the sense of being linear in the coecients. The key is to understand that (2.18) is quadratic in the variables (x1 , x2 ) yet linear in the coecients .
26
2 To simplify the expression, we dene the transformations x3 = x2 1 , x4 = x2 , x5 = x1 x2 , and 0 x6 = 1, and redene the regressor vector as x = (x1 , ..., x6 ) . With this redenition,
m(x1 , x2 ) = x0 which is linear in . For most econometric purposes (estimation and inference on ) the linearity in is all that is important. An exception is in the analysis of regression derivatives. In nonlinear equations such as (2.18), the regression derivative should be dened with respect to the original variables, not with respect to the transformed variables. Thus m(x1 , x2 ) = 1 + 2x1 3 + x2 5 x1 m(x1 , x2 ) = 2 + 2x2 4 + x1 5 x2 We see that in the model (2.18), the regression derivatives are not a simple coecient, but are functions of several coecients plus the levels of (x1, x2 ). Consequently it is dicult to interpret the coecients individually. It is more useful to interpret them as a group. We typically call 5 the interaction eect. Notice that it appears in both regression derivative equations, and has a symmetric interpretation in each. If 5 > 0 then the regression derivative with respect to x1 is increasing in the level of x2 (and the regression derivative with respect to x2 is increasing in the level of x1 ), while if 5 < 0 the reverse is true. It is worth noting that this symmetry is an articial implication of the quadratic equation (2.18), and is not a general feature of nonlinear conditional means m(x1 , x2 ).
2.17
Linear CEF with Dummy Variables
When all regressors takes a nite set of values, it turns out the CEF can be written as a linear function of regressors. This simplest example is a binary variable, which takes only two distinct values. For example, the variable gender takes only the values man and woman. Binary variables are extremely common in econometric applications, and are alternatively called dummy variables or indicator variables. Consider the simple case of a single binary regressor. In this case, the conditional mean can only take two distinct values. For example, 0 if gender=man E (y | gender) = 1 if gender=woman To facilitate a mathematical treatment, we typically record dummy variables with the values {0, 1}. For example 0 if gender=man (2.19) x1 = 1 if gender=woman
Given this notation we can write the conditional mean as a linear function of the dummy variable x1 , that is E (y | x1 ) = 1 x1 + 2 where 1 = 1 0 and 2 = 0 . In this simple regression equation the intercept 2 is equal to the conditional mean of y for the x1 = 0 subpopulation (men) and the slope 1 is equal to the dierence in the conditional means between the two subpopulations.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION Equivalently, we could have dened x1 as 1 if gender=man x1 = 0 if gender=woman
27
(2.20)
In this case, the regression intercept is the mean for women (rather than for men) and the regression slope has switched signs. The two regressions are equivalent but the interpretation of the coecients has changed. Therefore it is always important to understand the precise denitions of the variables, and illuminating labels are helpful. For example, labelling x1 as gender does not help distinguish between denitions (2.19) and (2.20). Instead, it is better to label x1 as women or female if denition (2.19) is used, or as men or male if (2.20) is used. Now suppose we have two dummy variables x1 and x2 . For example, x2 = 1 if the person is married, else x2 = 0. The conditional mean given x1 and x2 takes at most four possible values: 00 if x1 = 0 and x2 = 0 (unmarried men) (married men) 01 if x1 = 0 and x2 = 1 E (y | x1 , x2 ) = if x1 = 1 and x2 = 0 (unmarried women) 10 11 if x1 = 1 and x2 = 1 (married women) In this case we can write the conditional mean as a linear function of x1 , x2 and their product x1 x2 : E (y | x1 , x2 ) = 1 x1 + 2 x2 + 3 x1 x2 + 4
where 1 = 10 00 , 2 = 01 00 , 3 = 11 10 01 + 00 , and 4 = 00 . We can view the coecient 1 as the eect of gender on expected log wages for unmarried wages earners, the coecient 2 as the eect of marriage on expected log wages for men wage earners, and the coecient 3 as the dierence between the eects of marriage on expected log wages among women and among men. Alternatively, it can also be interpreted as the dierence between the eects of gender on expected log wages among married and non-married wage earners. Both interpretations are equally valid. We often describe 3 as measuring the interaction between the two dummy variables, or the interaction eect, and describe 3 = 0 as the case when the interaction eect is zero. In this setting we can see that the CEF is linear in the three variables (x1 , x2 , x1 x2 ). Thus to put the model in the framework of Section 2.15, we would dene the regressor x3 = x1 x2 and the regressor vector as x1 x2 x= x3 . 1 So even though we started with only 2 dummy variables, the number of regressors (including the intercept) is 4. If there are 3 dummy variables x1 , x2 , x3 , then E (y | x1 , x2 , x3 ) takes at most 23 = 8 distinct values and can be written as the linear function E (y | x1 , x2 , x3 ) = 1 x1 + 2 x2 + 3 x3 + 4 x1 x2 + 5 x1 x3 + 6 x2 x3 + 7 x1 x2 x3 + 8 which has eight regressors including the intercept. In general, if there are p dummy variables x1 , ..., xp then the CEF E (y | x1 , x2 , ..., xp ) takes at most 2p distinct values, and can be written as a linear function of the 2p regressors including x1 , x2 , ..., xp and all cross-products. This might be excessive in practice if p is modestly large. In the next section we will discuss projection approximations which yield more parsimonious parameterizations. We started this section by saying that the conditional mean is linear whenever all regressors take only a nite number of possible values. How can we see this? Take a categorical variable,
28
This is not a linear function of x3 itself, but it can be made a linear function by constructing dummy variables for two of the three categories. For example 1 if black x4 = 0 if not black x5 = 1 if other 0 if not other
When doing so, the values of x3 have no meaning in terms of magnitude, they simply indicate the relevant category. When the regressor is categorical the conditional mean of y given x3 takes a distinct value for each possibility: 1 if x3 = 1 if x3 = 2 E (y | x3 ) = 2 3 if x3 = 3
such as race. For example, we earlier divided race into three categories. We can record categorical variables using numbers to indicate each category, for example 1 if white 2 if black x3 = 3 if other
Given these transformations, we can write the conditional mean of y as a linear function of x4 and x5 E (y | x3 ) = E (y | x4 , x5 ) = 1 x4 + 2 x5 + 3 We can write the CEF as either E (y | x3 ) or E (y | x4 , x5 ) (they are equivalent), but it is only linear as a function of x4 and x5 . This setting is similar to the case of two dummy variables, with the dierence that we have not included the interaction term x4 x5 . This is because the event {x4 = 1 and x5 = 1} is empty by construction, so x4 x5 = 0 by denition.
In this case, the categorical variable x3 is equivalent to the pair of dummy variables (x4 , x5 ). The explicit relationship is 1 if x4 = 0 and x5 = 0 x3 = 2 if x4 = 1 and x5 = 0 3 if x4 = 0 and x5 = 1
2.18
Best Linear Predictor
While the conditional mean m(x) = E (y | x) is the best predictor of y among all functions of x, its functional form is typically unknown. In particular, the linear CEF model is empirically unlikely to be accurate unless x is discrete and low-dimensional so all interactions are included. Consequently in most cases it is more realistic to view the linear specication (2.16) as an approximation. In this section we derive a specic approximation with a simple interpretation. Theorem 2.11.1 showed that the conditional mean m (x) is the best predictor in the sense that it has the lowest mean squared error among all predictors. By extension, we can dene an approximation to the CEF by the linear function with the lowest mean squared error among all linear predictors. For this derivation we require the following regularity condition.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION Assumption 2.18.1 1. Ey 2 < . 2. E kxk2 < . 3. Qxx = E (xx0 ) is positive denite.
29
The best linear predictor of y given x, written P (y | x), is found by selecting the vector to minimize S ().
In Assumption 2.18.1.2 we use the notation kxk = (x0 x)1/2 to denote the Euclidean length of the vector x. The rst two parts of Assumption 2.18.1 imply that the variables y and x have nite means, variances, and covariances. The third part of the assumption is more technical, and its role will become apparent shortly. It is equivalent to imposing that the columns of Qxx = E (xx0 ) are linearly independent, or equivalently that the matrix Qxx is invertible. A linear predictor for y is a function of the form x0 for some Rk . The mean squared prediction error is 2 S () = E y x0 .
Denition 2.18.1 The Best Linear Predictor of y given x is P (y | x) = x0 where minimizes the mean squared prediction error 2 S () = E y x0 . = argmin S ( )
Rk
The minimizer
(2.21)
is called the Linear Projection Coecient.
The quadratic structure of S () means that we can solve explicitly for the minimizer. The rstorder condition for minimization (from Appendix A.9) is 0= Rewriting (2.22) as S () = 2E (xy) + 2E xx0 . 2E (xy ) = 2E xx0 Qxy = Qxx (2.23) (2.22)
We now calculate an explicit expression for its value. The mean squared prediction error can be written out as a quadratic function of : S () = Ey 2 20 E (xy) + 0 E xx0 .
and dividing by 2, this equation takes the form
30
where Qxy = E (xy ) is k 1 and Qxx = E (xx0 ) is k k. The solution is found by inverting the matrix Qxx , and is written 1 = Q xx Qxy or 1 E (xy) . = E xx0 (2.24)
It is worth taking the time to understand the notation involved in the expression (2.24). Qxx is a E(xy ) k k matrix and Qxy is a k 1 column vector. Therefore, alternative expressions such as E (xx0 )
This expression is also referred to as the linear projection of y on x. The projection error is e = y x0 .
or E (xy ) (E (xx0 ))1 are incoherent and incorrect. We also can now see the role of Assumption 2.18.1.3. It is necessary in order for the solution (2.24) to exist. Otherwise, there would be multiple solutions to the equation (2.23). We now have an explicit expression for the best linear predictor: 1 E (xy ) . P (y | x) = x0 E xx0
(2.25)
This equals the error from the regression equation when (and only when) the conditional mean is linear in x, otherwise they are distinct. Rewriting, we obtain a decomposition of y into linear predictor and error y = x0 + e. (2.26)
In general we call equation (2.26) or x0 the best linear predictor of y given x, or the linear projection of y on x. Equation (2.26) is also often called the regression of y on x but this can sometimes be confusing as economists use the term regression in many contexts. (Recall that we said in Section 2.15 that the linear CEF model is also called the linear regression model.) An important property of the projection error e is E (xe) = 0. (2.27)
To see this, using the denitions (2.25) and (2.24) and the matrix properties AA1 = I and Ia = a, E (xe) = E x y x0 1 = E (xy ) E xx0 E xx0 E (xy ) =0 (2.28) as claimed. Equation (2.27) is a set of k equations, one for each regressor. In other words, (2.27) is equivalent to (2.29) E (xj e) = 0 for j = 1, ..., k. As in (2.15), the regressor vector x typically contains a constant, e.g. xk = 1. In this case (2.29) for j = k is the same as E (e) = 0. (2.30) Thus the projection error has a mean of zero when the regressor vector contains a constant. (When x does not have a constant, (2.30) is not guaranteed. As it is desirable for e to have a zero mean, this is a good reason to always include a constant in any regression model.)
31
It is also useful to observe that since cov(xj , e) = E (xj e) E (xj ) E (e) , then (2.29)-(2.30) together imply that the variables xj and e are uncorrelated. This completes the derivation of the model. We summarize some of the most important properties. Theorem 2.18.1 Properties of Linear Projection Model Under Assumption 2.18.1, 1. The moments E (xx0 ) and E (xy) exist with nite elements. 2. The Linear Projection Coecient (2.21) exists, is unique, and equals 1 = E xx0 E (xy ) .
3. The best linear predictor of y given x is 1 P (y | x) = x0 E xx0 E (xy ) . 4. The projection error e = y x0 exists and satises E e2 < and E (xe) = 0. 5. If x contains an constant, then E (e) = 0. 6. If E |y |r < and E kxkr < for r 2 then E |e|r < .
A complete proof of Theorem 2.18.1 is given in Section 2.34. It is useful to reect on the generality of Theorem 2.18.1. The only restriction is Assumption 2.18.1. Thus for any random variables (y, x) with nite variances we can dene a linear equation (2.26) with the properties listed in Theorem 2.18.1. Stronger assumptions (such as the linear CEF model) are not necessary. In this sense the linear model (2.26) exists quite generally. However, it is important not to misinterpret the generality of this statement. The linear equation (2.26) is dened as the best linear predictor. It is not necessarily a conditional mean, nor a parameter of a structural or causal economic model.
Linear Projection Model y = x0 + e. E (xe) = 0 1 E (xy) = E xx0 We illustrate projection using three log wage equations introduced in earlier sections.
32
For our rst example, we consider a model with the two dummy variables for gender and race similar to Table 2.1. As we learned in Section 2.17, the entries in this table can be equivalently expressed by a linear CEF. For simplicity, lets consider the CEF of log(wage) as a function of Black and Female. E(log(wage) | Black, F emale) = 0.20Black 0.24F emale + 0.10Black F emale + 3.06. (2.31) This is a CEF as the variables are dummys and all interactions are included. Now consider a simpler model omitting the interaction eect. This is the linear projection on the variables Black and F emale P (log(wage) | Black, F emale) = 0.15Black 0.23F emale + 3.06. (2.32)
What is the dierence? The full CEF (2.31) shows that the race gap is dierentiated by gender: it is 20% for black men (relative to non-black men) and 10% for black women (relative to non-black women). The projection model (2.32) simplies this analysis, calculating an average 15% wage gap for blacks, ignoring the role of gender. Notice that this is despite the fact that the gender variable is included in (2.32).
2.0
2.5
3.0
3.5
4.0
10
12
14
16
18
20
Years of Education
Figure 2.8: Projections of log(wage) onto Education For our second example we consider the CEF of log wages as a function of years of education for white men which was illustrated in Figure 2.5 and is repeated in Figure 2.8. Superimposed on the gure are two projections. The rst (given by the dashed line) is the linear projection of log wages on years of education P (log(wage) | Education) = 0.11Education + 1.5 This simple equation indicates an average 11% increase in wages for every year of education. An inspection of the Figure shows that this approximation works well for education 9, but underpredicts for individuals with lower levels of education. To correct this imbalance we use a linear spline equation which allows dierent rates of return above and below 9 years of education: P (log(wage) | Education, (Education 9) 1 (Education > 9))
= 0.02Education + 0.10 (Education 9) 1 (Education > 9) + 2.3
This equation is displayed in Figure 2.8 using the solid line, and appears to t much better. It indicates a 2% increase in mean wages for every year of education below 9, and a 12% increase in
33
4.0
Conditional Mean Linear Projection Quadratic Projection
2.0
0
2.5
3.0
3.5
10
20
30
40
50
Figure 2.9: Linear and Quadratic Projections of log(wage) onto Experience mean wages for every year of education above 9. It is still an approximation to the conditional mean but it appears to be fairly reasonable. For our third example we take the CEF of log wages as a function of years of experience for white men with 12 years of education, which was illustrated in Figure 2.6 and is repeated as the solid line in Figure 2.9. Superimposed on the gure are two projections. The rst (given by the dot-dashed line) is the linear projection on experience P (log(wage) | Experience) = 0.011Experience + 2.5 and the second (given by the dashed line) is the linear projection on experience and its square P (log(wage) | Experience) = 0.046Experience 0.0007Experience2 + 2.3. It is fairly clear from an examination of Figure 2.9 that the rst linear projection is a poor approximation. It over-predicts wages for young and old workers, and under-predicts for the rest. Most importantly, it misses the strong downturn in expected wages for older wage-earners. The second projection ts much better. We can call this equation a quadratic projection since the function is quadratic in experience.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION Invertibility and Identication The linear projection coecient = (E (xx0 ))1 E (xy) exists and is unique as long as the k k matrix Qxx = E (xx0 ) is invertible. The matrix Qxx is sometimes called the design matrix, as in experimental settings the researcher is able to control Qxx by manipulating the distribution of the regressors x. Observe that for any non-zero Rk , 2 0 Qxx = E 0 xx0 = E 0 x 0
34
so Qxx by construction is positive semi-denite. The assumption that it is positive denite means that this is a strict inequality, E (0 x)2 > 0. Equivalently, there cannot exist a non-zero vector such that 0 x = 0 identically. This occurs when redundant variables are included in x. Positive semi-denite matrices are invertible if and only if they are positive denite. When Qxx is invertible then = (E (xx0 ))1 E (xy ) exists and is uniquely dened. In other words, in order for to be uniquely dened, we must exclude the degenerate situation of redundant varibles. Theorem 2.18.1 shows that the linear projection coecient is identied (uniquely determined) under Assumptions 2.18.1. The key is invertibility of Qxx . Otherwise, there is no unique solution to the equation Qxx = Qxy . (2.33)
When Qxx is not invertible there are multiple solutions to (2.33), all of which yield an equivalent best linear predictor x0 . In this case the coecient is not identied as it does not have a unique value. Even so, the best linear predictor x0 still identied. One solution is to set E (xy ) = E xx0 where A denotes the generalized inverse of A (see Appendix A.5).
2.19
Linear Predictor Error Variance
As in the CEF model, we dene the error variance as 2 = E e2 . Setting Qyy = E y 2 and Qyx = E (yx0 ) we can write 2 as
2 2 = E y x0 = Ey 2 2E y x0 + 0 E xx0
1 = Qyy Qyx Q xx Qxy def
1 1 1 = Qyy 2Qyx Q xx Qxy + Qyx Qxx Qxx Qxx Qxy
= Qyyx .
(2.34)
1 One useful feature of this formula is that it shows that Qyyx = Qyy Qyx Q xx Qxy equals the variance of the error from the linear projection of y on x.
35
2.20
Regression Coecients
Sometimes it is useful to separate the constant from the other regressors, and write the linear projection equation in the format (2.35) y = x0 + + e where is the intercept and x does not contain a constant. Taking expectations of this equation, we nd Ey = Ex0 + E + Ee or y = 0 x + where y = Ey and x = Ex, since E (e) = 0 from (2.30). (While x does not contain a constant, the equation does so (2.30) still applies.) Rearranging, we nd = y 0 x . Subtracting this equation from (2.35) we nd y y = (x x )0 + e, (2.36)
a linear equation between the centered variables y y and x x . (They are centered at their means, so are mean-zero random variables.) Because x x is uncorrelated with e, (2.36) is also a linear projection, thus by the formula for the linear projection model, 1 E ((x x ) (y y )) = E (x x ) (x x )0 = var (x)1 cov (x, y) a function only of the covariances11 of x and y.
Theorem 2.20.1 In the linear projection model y = x0 + + e, then = y 0 x and = var (x)1 cov (x, y ) . (2.38) (2.37)
2.21
Regression Sub-Vectors

Let the regressors be partitioned as x=

11
x1 x2
(2.39)
The covariance matrix between vectors xand z is cov (x, z ) = E (x Ex) (z Ez )0 . The (co)variance 0 matrix of the vector x is var (x) = cov (x, x) = E (x Ex) (x Ex) .
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION We can write the projection of y on x as y = x0 + e

0 = x0 1 1 + x2 2 + e
36
(2.40)
E (xe) = 0. In this section we derive formula for the sub-vectors 1 and 2 . Partition Qxx comformably with x 0 E (x1 x0 Q11 Q12 1 ) E (x1 x2 ) = Qxx = 0 Q21 Q22 E (x2 x0 1 ) E (x2 x2 ) and similarly Qxy Qxy = Q1y Q2y = E (x1 y ) E (x2 y ) .
By the partitioned matrix inversion formula (A.4) 11 1 1 1 1 def Q11 Q12 Q Q12 Q Q12 Q Q 1 11 2 11 2 22 Qxx = = = . 1 1 1 Q21 Q22 Q21 Q22 Q Q 221 Q21 Q11 221
1 1 where Q112 = Q11 Q12 Q 22 Q21 and Q221 = Q22 Q21 Q11 Q12 . Thus 1 = 2 1 1 1 Q1y Q Q Q12 Q 11 2 11 2 22 = 1 1 1 Q2y Q Q 221 Q21 Q11 221 1 1 Q112 Q1y Q12 Q22 Q2y = 1 1 Q 221 Q2y Q21 Q11 Q1y 1 Q112 Q1y2 = 1 Q 221 Q2y 1 def def
(2.41)
We have shown that

1 1 = Q 112 Q1y 2 1 2 = Q 221 Q2y 1
2.22
Coecient Decomposition
In the previous section we derived formulae for the coecient sub-vectors 1 and 2 . We now use these formulae to give a useful interpretation of the coecients in terms of an iterated projection. Take equation (2.40) for the case dim(x1 ) = 1 so that 1 R. y = x1 1 + x0 2 2 + e. Now consider the projection of x1 on x2 : x1 = x0 2 2 + u1 E (x2 u1 ) = 0.
1 1 2 From (2.24) and (2.34), 2 = Q 22 Q21 and Eu1 = Q112 = Q11 Q12 Q22 Q21 . We can also calculate that 1 0 E (u1 y ) = E x1 0 2 x2 y = E (x1 y ) 2 E (x2 y ) = Q1y Q12 Q22 Q2y = Q1y 2 .
(2.42)
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION We have found that

1 1 = Q 112 Q1y2 =
37
E (u1 y) Eu2 1
the coecient from the simple regression of y on u1 . What this means is that in the multivariate projection equation (2.42), the coecient 1 equals the projection coecient from a regression of y on u1 , the error from a projection of x1 on the other regressors x2 . The error u1 can be thought of as the component of x1 which is not linearly explained by the other regressors. Thus the coecient 1 equals the linear eect of x1 on y, after stripping out the eects of the other variables. There was nothing special in the choice of the variable x1 . This derivation applies symmetrically to all coecients in a linear projection. Each coecient equals the simple regression of y on the error from a projection of that regressor on all the other regressors. Each coecient equals the linear eect of that variable on y, after linearly controlling for all the other regressors.
2.23
Omitted Variable Bias
Again, let the regressors be partitioned as in (2.39). Consider the projection of y on x1 only. Perhaps this is done because the variables x2 are not observed. This is the equation y = x0 1 1 + u E (x1 u) = 0. Notice that we have written the coecient on x1 as 1 rather than 1 and the error as u rather than e. This is because (2.43) is dierent than (2.40). Goldberger (1991) introduced the catchy labels long regression for (2.40) and short regression for (2.43) to emphasize the distinction. Typically, 1 6= 1 , except in special cases. To see this, we calculate 1 1 = E x1 x0 E (x1 y ) 1 1 0 E x1 x0 = E x1 x0 1 1 1 + x2 2 + e 1 E x1 x0 = 1 + E x1 x0 1 2 2 = 1 + 2 1 = E x1 x0 E x1 x0 1 2 (2.43)
where
is the coecient matrix from a projection of x2 on x1 . Observe that 1 = 1 + 2 6= 1 unless = 0 or 2 = 0. Thus the short and long regressions have dierent coecients on x1 . They are the same only under one of two conditions. First, if the projection of x2 on x1 yields a set of zero coecients (they are uncorrelated), or second, if the coecient on x2 in (2.40) is zero. In general, the coecient in (2.43) is 1 rather than 1 . The dierence 2 between 1 and 1 is known as omitted variable bias. It is the consequence of omission of a relevant correlated variable. To avoid omitted variables bias the standard advice is to include all potentially relevant variables in estimated models. By construction, the general model will be free of such bias. Unfortunately in many cases it is not feasible to completely follow this advice as many desired variables are not observed. In this case, the possibility of omitted variables bias should be acknowledged and discussed in the course of an empirical investigation. For example, suppose y is log wages, x1 is education, and x2 is intellectual ability. It seems reasonable to suppose that education and intellectual ability are positively correlated (highly able individuals attain higher levels of education) which means > 0. It also seems reasonable to suppose that conditional on education, individuals with higher intelligence will earn higher wages
38
on average, so that 2 > 0. This implies that 2 > 0 and 1 = 1 + 2 > 1 . Therefore, it seems reasonable to expect that in a regression of wages on education with ability omitted, the coecient on education is higher than in a regression where ability is included. In other words, in this context the omitted variable biases the regression coecient upwards.
2.24
Best Linear Approximation
The function d() is a measure of the deviation of x0 from m(x). If the two functions are identical then d() = 0, otherwise d( ) > 0. We can also view the mean-square dierence d() as a densityweighted average of the function (m(x) x0 )2 , since Z 2 m(x) x0 fx (x)dx d( ) =
Rk
There are alternative ways we could construct a linear approximation x0 to the conditional mean m(x). In this section we show that one alternative approach turns out to yield the same answer as the best linear predictor. We start by dening the mean-square approximation error of x0 to m(x) as the expected squared dierence between x0 and the conditional mean m(x) 2 (2.44) d() = E m(x) x0 .
where fx (x) is the marginal density of x. We can then dene the best linear approximation to the conditional m(x) as the function x0 obtained by selecting to minimize d() : = argmin d( ).
Rk
(2.45)
Similar to the best linear predictor we are measuring accuracy by expected squared error. The dierence is that the best linear predictor (2.21) selects to minimize the expected squared prediction error, while the best linear approximation (2.45) selects to minimize the expected squared approximation error. Despite the dierent denitions, it turns out that the best linear predictor and the best linear approximation are identical. By the same steps as in (2.18) plus an application of conditional expectations we can nd that 1 E (xm(x)) (2.46) = E xx0 0 1 E (xy) (2.47) = E xx (see Exercise 2.19). Thus (2.45) equals (2.21). We conclude that the denition (2.45) can be viewed as an alternative motivation for the linear projection coecient.
2.25
Normal Regression
Suppose the variables (y, x) are jointly normally distributed. Consider the best linear predictor of y given x y = x0 + e 1 = E xx0 E (xy) .
Since the error e is a linear transformation of the normal vector (y, x), it follows that (e, x) is jointly normal, and since they are jointly normal and uncorrelated (since E (xe) = 0) they are also independent (see Appendix B.9). Independence implies that E (e | x) = E (e) = 0
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION and E e2 | x = E e2 = 2
39
which are properties of a homoskedastic linear CEF. We have shown that when (y, x) are jointly normally distributed, they satisfy a normal linear CEF y = x0 + e where e N(0, 2 ) is independent of x. This is an alternative (and traditional) motivation for the linear CEF model. This motivation has limited merit in econometric applications since economic data is typically non-normal.
2.26
Regression to the Mean
The term regression originated in an inuential paper by Francis Galton published in 1886, where he examined the joint distribution of the stature (height) of parents and children. Eectively, he was estimating the conditional mean of childrens height given their parents height. Galton discovered that this conditional mean was approximately linear with a slope of 2/3. This implies that on average a childs height is more mediocre (average) than his or her parents height. Galton called this phenomenon regression to the mean, and the label regression has stuck to this day to describe most conditional relationships. One of Galtons fundamental insights was to recognize that if the marginal distributions of y and x are the same (e.g. the heights of children and parents in a stable environment) then the regression slope in a linear projection is always less than one. To be more precise, take the simple linear projection y = x + + e (2.48)
where y equals the height of the child and x equals the height of the parent. Assume that y and x have the same mean, so that y = x = . Then from (2.37) = (1 ) so we can write the linear projection (2.48) as P (y | x) = (1 ) + x. This shows that the projected height of the child is a weighted average of the population average height and the parents height x, with the weight equal to the regression slope . When the height distribution is stable across generations, so that var(y ) = var(x), then this slope is the simple correlation of y and x. Using (2.38) = cov (x, y ) = corr(x, y). var(x)
By the properties of correlation (e.g. equation (B.7) in the Appendix), 1 corr(x, y ) 1, with corr(x, y) = 1 only in the degenerate case y = x. Thus if we exclude degeneracy, is strictly less than 1. This means that on average a childs height is more mediocre (closer to the population average) than the parents.
40
Sir Francis Galton

Sir Francis Galton (1822-1911) of England was one of the leading gures in late 19th century statistics. In addition to inventing the concept of regression, he is credited with introducing the concepts of correlation, the standard deviation, and the bivariate normal distribution. His work on heredity made a signicant intellectual advance by examing the joint distributions of observables, allowing the application of the tools of mathematical statistics to the social sciences.
A common error known as the regression fallacy is to infer from < 1 that the population is converging, meaning that its variance is declining towards zero. This is a fallacy because we derived the implication < 1 under the assumption of constant means and variances. So certainly < 1 does not imply that the variance y is less than than the variance of x. Another way of seeing this is to examine the conditions for convergence in the context of equation (2.48). Since x and e are uncorrelated, it follows that var(y ) = 2 var(x) + var(e). Then var(y ) < var(x) if and only if 2 < 1 var(e) var(x)
which is not implied by the simple condition | | < 1. The regression fallacy arises in related empirical situations. Suppose you sort families into groups by the heights of the parents, and then plot the average heights of each subsequent generation over time. If the population is stable, the regression property implies that the plots lines will converge childrens height will be more average than their parents. The regression fallacy is to incorrectly conclude that the population is converging. A message to be learned from this example is that such plots are misleading for inferences about convergence. The regression fallacy is subtle. It is easy for intelligent economists to succumb to its temptation. A famous example is The Triumph of Mediocrity in Business by Horace Secrist, published in 1933. In this book, Secrist carefully and with great detail documented that in a sample of department stores over 1920-1930, when he divided the stores into groups based on 1920-1921 prots, and plotted the average prots of these groups for the subsequent 10 years, he found clear and persuasive evidence for convergence toward mediocrity. Of course, there was no discovery regression to the mean is a necessary feature of stable distributions.
2.27
Reverse Regression
Galton noticed another interesting feature of the bivariate distribution. There is nothing special about a regression of y on x. We can also regress x on y. (In his heredity example this is the best linear predictor of the height of parents given the height of their children.) This regression takes the form (2.49) x = y + + e . This is sometimes called the reverse regression. In this equation, the coecients , and error e are dened by linear projection. In a stable population we nd that = corr(x, y ) = = (1 ) =
41
which are exactly the same as in the projection of y on x! The intercept and slope have exactly the same values in the forward and reverse projections! While this algebraic discovery is quite simple, it is counter-intuitive. Instead, a common yet mistaken guess for the form of the reverse regression is to take the equation (2.48), divide through by and rewrite to nd the equation x=y 1 1 e (2.50)
suggesting that the projection of x on y should have a slope coecient of 1/ instead of , and intercept of / rather than . What went wrong? Equation (2.50) is perfectly valid, because it is a simple manipulation of the valid equation (2.48). The trouble is that (2.50) is neither a CEF nor a linear projection. Inverting a projection (or CEF) does not yield a projection (or CEF). Instead, (2.49) is a valid projection, not (2.50). In any event, Galtons nding was that when the variables are standardized, the slope in both projections (y on x, and x and y ) equals the correlation, and both equations exhibit regression to the mean. It is not a causal relation, but a natural feature of all joint distributions.
2.28
Limitations of the Best Linear Predictor
Lets compare the linear projection and linear CEF models. From Theorem 2.8.1.4 we know that the CEF error has the property E (xe) = 0. Thus a linear CEF is a linear projection. However, the converse is not true as the projection error does not necessarily satisfy E (e | x) = 0. Furthermore, the linear projection may be a poor approximation to the CEF. To see these points in a simple example, suppose that the true process is y = x + x2 with x N(0, 1). In this case the true CEF is m(x) = x + x2 and there is no error. Now consider the linear projection of y on x and a constant, namely the model y = x + + u. Since x N(0, 1) then x and x2 are uncorrelated the linear projection takes the form P (y | x) = x + 1. This is quite dierent from the true CEF m(x) = x + x2 . The projection error equals e = x2 1, which is a deterministic function of x, yet is uncorrelated with x. We see in this example that a projection error need not be a CEF error, and a linear projection can be a poor approximation to the CEF. Another defect of linear projection is that it is sensitive to the marginal distribution of the regressors when the conditional mean is non-linear. We illustrate the issue in Figure 2.10 for a constructed12 joint distribution of y and x. The solid line is the non-linear CEF of y given x. The data are divided in two Group 1 and Group 2 which have dierent marginal distributions for the regressor x, and Group 1 has a lower mean value of x than Group 2. The separate linear projections of y on x for these two groups are displayed in the Figure by the dashed lines. These two projections are distinct approximations to the CEF. A defect with linear projection is that it leads to the incorrect conclusion that the eect of x on y is dierent for individuals in the two groups. This conclusion is incorrect because in fact there is no dierence in the conditional mean function. The apparant dierence is a by-product of a linear approximation to a non-linear mean, combined with dierent marginal distributions for the conditioning variables.
2.29
Random Coecient Model
A model which is notationally similar to but conceptually distinct from the linear CEF model is the linear random coecient model. It takes the form y = x0
The x in Group 1 are N(2, 1) and those in Group 2 are N(4, 1), and the conditional distribution of y given x is N(m(x), 1) where m(x) = 2x x2 /6.
12
42
Figure 2.10: Conditional Mean and Two Linear Projections where the individual-specic coecient is random and independent of x. For example, if x is years of schooling and y is log wages, then is the individual-specic returns to schooling. If a person obtains an extra year of schooling, is the actual change in their wage. The random coecient model allows the returns to schooling to vary in the population. Some individuals might have a high return to education (a high ) and others a low return, possibly 0, or even negative. In the linear CEF model the regressor coecient equals the regression derivative the change in the conditional mean due to a change in the regressors, = m(x). This is not the eect on a given individual, it is the eect on the population average. In contrast, in the random coecient model, the random vector = x0 is the true causal eect the change in the response variable y itself due to a change in the regressors. It is interesting, however, to discover that the linear random coecient model implies a linear CEF. To see this, let and denote the mean and covariance matrix of : = E( ) = var ( ) and then decompose the random coecient as =+u where u is distributed independently of x with mean zero and covariance matrix . Then we can write E(y | x) = x0 E( | x) = x0 E( ) = x0 so the CEF is linear in x, and the coecients equal the mean of the random coecient . We can thus write the equation as a linear CEF y = x0 + e where e = x0 u and u = . The error is conditionally mean zero: E(e | x) = 0. (2.51)
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION Furthermore var (e | x) = x0 var ( )x = x0 x so the error is conditionally heteroskedastic with its variance a quadratic function of x.
43
Theorem 2.29.1 In the linear random coecient model y = x0 with independent of x, E kxk2 < , and E k k2 < , then var (y | x) = x0 x where = E( ) and = var ( ). E (y | x) = x0
2.30
Causal Eects
So far we have avoided the concept of causality, yet often the underlying goal of an econometric analysis is to uncover a causal relationship between variables. It is often of great interest to understand the causes and eects of decisions, actions, and policies. For example, we may be interested in the eect of class sizes on test scores, police expenditures on crime rates, climate change on economic activity, years of schooling on wages, institutional structure on growth, the eectiveness of rewards on behavior, the consequences of medical procedures for health outcomes, or any variety of possible causal relationships. In each case, the goal is to understand what is the actual eect on the outcome y due to a change in the input x. We are not just interested in the conditional mean or linear projection, we would like to know the actual change. Two inherent barriers are that the causal eect is typically specic to an individual and that it is unobserved. Consider the eect of schooling on wages. The causal eect is the actual dierence a person would receive in wages if we could change their level of education holding all else constant. This is specic to each individual as their employment outcomes in these two distinct situations is individual. The causal eect is unobserved because the most we can observe is their actual level of education and their actual wage, but not the counterfactual wage if their education had been dierent. To be even more specic, suppose that there are two individuals, Jennifer and George, and both have the possibility of being high-school graduates or college graduates, but both would have received dierent wages given their choices. For example, suppose that Jennifer would have earned $10 an hour as a high-school graduate and $20 an hour as a college graduate while George would have earned $8 as a high-school graduate and $12 as a college graduate. In this example the causal eect of schooling is $10 a hour for Jennifer and $4 an hour for George. The causal eects are specic to the individual and neither causal eect is observed. A variable x1 can be said to have a causal eect on the response variable y if the latter changes when all other inputs are held constant. To make this precise we need a mathematical formulation. We can write a full model for the response variable y as y = h (x1 , x2 , u) (2.52)
where x1 and x2 are the observed variables, u is an 1 unobserved random factor, and h is a functional relationship. This framework includes as a special case the random coecient model
44
(2.29) studied earlier. We dene the causal eect of x1 within this model as the change in y due to a change in x1 holding the other variables x2 and u constant.
Denition 2.30.1 In the model (2.52) the causal eect of x1 on y is C (x1 , x2 , u) = 1 h (x1 , x2 , u) , the change in y due to a change in x1 , holding x2 and u constant. (2.53)
To understand this concept, imagine taking a single individual. As far as our structural model is concerned, this person is described by their observables x1 and x2 and their unobservables u. In a wage regression the unobservables would include characteristics such as the persons abilities, skills, work ethic, interpersonal connections, and preferences. The causal eect of x1 (say, education) is the change in the wage as x1 changes, holding constant all other observables and unobservables. It may be helpful to understand that (2.53) is a denition, and does not necessarily describe causality in a fundamental or experimental sense. Perhaps it would be more appropriate to label (2.53) as a structural eect (the eect within the structural model). Sometimes it is useful to write this relationship as a potential outcome function y (x1 ) = h (x1 , x2 , u) where the notation implies that y(x1 ) is holding x2 and u constant. A popular example arises in the analysis of treatment eects with a binary regressor x1 . Let x1 = 1 indicate treatment (e.g. a medical procedure) and x1 = 0 indicating non-treatment. In this case y (x1 ) can be written y(0) = h (0, x2 , u) y(1) = h (1, x2 , u) . In the literature on treatment eects, it is common to refer to y(0) and y(1) as the latent outcomes associated with non-treatment and treatment, respectively. That is, for a given individual, y(0) is the health outcome if there is no treatment, and y (1) is the health outcome if there is treatment. The causal eect of treatment for the individual is the change in their health outcome due to treatment the change in y as we hold both x2 and u constant: C (x2 , u) = y (1) y (0). This is random (a function of x2 and u) as both potential outcomes y (0) and y(1) are dierent across individuals. In a sample, we cannot observe both outcomes from the same individual, we only observe the realized value y (0) if x1 = 0 y= y (1) if x1 = 1.
As the causal eect varies across individuals and is not observable, it cannot be measured on the individual level. We therefore focus on aggregate causal eects, in particular what is known as the average causal eect.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION Denition 2.30.2 In the model (2.52) the average causal eect of x1 on y conditional on x2 is ACE (x1 , x2 ) = E (C (x1 , x2 , u) | x1 , x2 ) Z = 1 h (x1 , x2 , u) f (u | x1 , x2 )du
R
45
(2.54)
where f (u | x1 , x2 ) is the conditional density of u given x1 , x2 .
We can think of the average causal eect ACE (x1 , x2 ) as the average eect in the general population. In our Jennifer & George schooling example given earlier, supposing that half of the population are Jennifers and the other half Georges, then the average causal eect of college is (10+4)/2 = $7 an hour. This is not the individual causal eect, it is the average of the causal eect across all individuals in the population. Given data on only educational attainment and wages, the ACE of $7 is the best we can hope to learn. When we conduct a regression analysis (that is, consider the regression of observed wages on educational attainment) we might hope that the regression reveals the average causal eect. Technically, that the regression derivative (the coecient on education) equals the ACE. Is this the case? In other words, what is the relationship between the average causal eect ACE (x1 , x2 ) and the regression derivative 1 m (x1 , x2 )? Equation (2.52) implies that the CEF is m(x1 , x2 ) = E (h (x1 , x2 , u) | x1 , x2 ) Z h (x1 , x2 , u) f (u | x1 , x2 )du, =
R
the average causal equation, averaged over the conditional distribution of the unobserved component u. Applying the marginal eect operator, the regression derivative is Z 1 h (x1 , x2 , u) f (u | x1 , x2 )du 1 m(x1 , x2 ) = R Z h (x1 , x2 , u) 1 f (u|x1 , x2 )du + R Z h (x1 , x2 , u) 1 f (u | x1 , x2 )du. (2.55) = ACE (x1 , x2 ) +
R
Equation (2.55) shows that in general, the regression derivative does not equal the average causal eect. The dierence is the second term on the right-hand-side of (2.55). The regression derivative and ACE equal in the special case when this term equals zero, which occurs when 1 f (u | x1 , x2 ) = 0, that is, when the conditional density of u given (x1 , x2 ) does not depend on x1 . When this condition holds then the regression derivative equals the ACE, which means that regression analysis can be interpreted causally, in the sense that it uncovers average causal eects. The condition is suciently important that it has a special name in the treatment eects literature.
Denition 2.30.3 Conditional Independence Assumption (CIA). Conditional on x2 , the random variables x1 and u are statistically independent.
46
The CIA implies f (u | x1 , x2 ) = f (u | x2 ) does not depend on x1 , and thus 1 f (u | x1 , x2 ) = 0. Thus the CIA implies that 1 m(x1 , x2 ) = ACE (x1 , x2 ), the regression derivative equals the average causal eect.
Theorem 2.30.1 In the structural model (2.52), the Conditional Independence Assumption implies 1 m(x1 , x2 ) = ACE (x1 , x2 ) the regression derivative equals the average causal eect for x1 on y conditional on x2 .
This is a fascinating result. It shows that whenever the unobservable is independent of the treatment variable (after conditioning on appropriate regressors) the regression derivative equals the average causal eect. In this case, the CEF has causal economic meaning, giving strong justication to estimation of the CEF. Our derivation also shows the critical role of the CIA. If CIA fails, then the equality of the regression derivative and ACE fails. This theorem is quite general. It applies equally to the treatment-eects model where x1 is binary or to more general settings where x1 is continuous. It is also helpful to understand that the CIA is weaker than full independence of u from the regressors (x1 , x2 ). The CIA was introduced precisely as a minimal sucient condition to obtain the desired result. Full independence implies the CIA and implies that each regression derivative equals that variables average causal eect, but full independence is not necessary in order to causally interpret a subset of the regressors. To illustrate, lets return to our education example involving a population with equal numbers of Jennifers and Georges. Recall that Jennifer earns $10 as a high-school graduate and $20 as a college graduate (and so has a causal eect of $10) while George earns $8 as a high-school graduate and $12 as a college graduate (so has a causal eect of $4). Given this information, the average causal eect of college is $7, which is what we hope to learn from a regression analysis. Now suppose that while in high school all students take an aptitude test, and if a student gets a high (H) score he or she goes to college with probability 3/4, and if a student gets a low (L) score he or she goes to college with probability 1/4. Suppose further that Jennifers get an aptitude score of H with probability 3/4, while Georges get a score of H with probability 1/4. Given this situation, 62.5% of Jennifers will go to college13 , while 37.5% of Georges will go to college14 . An econometrician who randomly samples 32 individuals and collects data on educational attainment and wages will nd the following wage distribution: $8 10 0 $10 6 0 $12 0 6 $20 0 10 Mean $8.75 $17.00
High-School Graduate College Graduate
Let college denote a dummy variable taking the value of 1 for a college graduate, otherwise 0. Thus the regression of wages on college attendence takes the form E (wage | college) = 8.25college + 8.75. The coecient on the college dummy, $8.25, is the regression derivative, and the implied wage eect of college attendence. But $8.25 overstates the average causal eect of $7. The reason is because
13 14
Pr (College|Jennif er) = Pr (College|H ) Pr (H |Jennif er) + Pr (College|L) Pr (L|Jennifer) = (3/4)2 + (1/4)2 Pr (College|George) = Pr (College|H ) Pr (H |George) + Pr (College|L) Pr (L|George) = (3/4)(1/4) + (1/4)(3/4)
47
the CIA fails. In this model the unobservable u is the individuals type (Jennifer or George) which is not independent of the regressor x1 (education), since Jennifer is more likely to go to college than George. Since Jennifers causal eect is higher than Georges, the regression derivative overstates the ACE. The coecient $8.25 is not the average benet of college attendence, rather it is the observed dierence in realized wages in a population whose decision to attend college is correlated with their individual causal eect. At the risk of repeating myself, in this example, $8.25 is the true regression derivative, it is the dierence in average wages between those with a college education and those without. It is not, however, the average causal eect of college education in the population. This does not mean that it is impossible to estimate the ACE. The key is conditioning on the appropriate variables. The CIA says that we need to nd a variable x2 such that conditional on x2 , u and x1 (type and education) are independent. In this example a variable which will achieve this is the aptitude test score. The decision to attend college was based on the test score, not on an individuals type. Thus educational attainment and type are independent once we condition on the test score. This also alters the ACE. Notice that Denition 2.30.2 is a function of x2 (the test score). Among the students who receive a high test score, 3/4 are Jennifers and 1/4 are Georges. Thus the ACE for students with a score of H is (3/4) 10 + (1/4) 4 = $8.50. Among the students who receive a low test score, 1/4 are Jennifers and 3/4 are Georges. Thus the ACE for students with a score of L is (1/4) 10 + (3/4) 4 = $5.50. The ACE varies between these two observable groups (those with high test scores and those with low test scores). Again, we would hope to be able to learn the ACE from a regression analysis, this time from a regression of wages on education and test scores. To see this in the wage distribution, suppose that the econometrician collects data on the aptitude test score as well as education and wages. Given a random sample of 32 individuals we would expect to nd the following wage distribution: High-School Graduate + High Test Score College Graduate + High Test Score High-School Graduate + Low Test Score College Graduate + Low Test Score $8 1 0 9 0 $10 3 0 3 0 $12 0 3 0 3 $20 0 9 0 1 Mean $9.50 $18.00 $8.50 $14.00
Dene the dummy variable highscore which takes the value 1 for students who received a high test score, else zero. The regression of wages on college attendence and test scores (with interactions) takes the form E (wage | college, highscore) = 1.00highscore + 5.50college + 3.00highscore college + 8.50. The cocient on college, $5.50, is the regression derivative of college attendence for those with low test scores, and the sum of this coecient with the interaction coecient, $8.50, is the regression derivative for college attendence for those with high test scores. These equal the average causal efect. In this example, by conditioning on the aptitude test score, the average causal eect of education on wages can be learned from a regression analyis. What this shows is that by conditioning on the proper variables, it may be possible to achieve the CIA, in which case regression analysis measures average causal eects.
2.31
Expectation: Mathematical Details*
We dene the mean or expectation Ey of a random variable y as follows. If y is discrete on the set {1 , 2 , ...} then X Ey = j Pr (y = j ) ,
j =1
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION and if y is continuous with density f then Ey = Z
48
yf (y)dy.
We can unify these denitions by writing the expectation as the Lebesgue integral with respect to the distribution function F Z Ey = ydF (y ). (2.56)
In the event that the integral (2.56) is not nite, separately evaluate the two integrals Z ydF (y ) I1 = 0 Z 0 I2 = ydF (y ).
(2.57) (2.58)
If I1 = and I2 < then it is typical to dene Ey = . If I1 < and I2 = then we dene Ey = . However, if both I1 = and I2 = then Ey is undened.If Z E |y | = |y | dF (y ) = I1 + I2 <
then Ey exists and is nite. In this case it is common to say that the mean Ey is well-dened. More generally, y has a nite rth moment if E |y|r < . (2.59)
By Liapunovs Inequality (B.20), (2.59) implies E |y|s < for all s r. Thus, for example, if the fourth moment is nite then the rst, second and third moments are also nite. It is common in econometric theory to assume that the variables, or certain transformations of the variables, have nite moments of a certain order. How should we interpret this assumption? How restrictive is it? One way to visualize the importance is to consider the class of Pareto densities given by f (y) = aya1 , y > 1.
This shows that if y is Pareto distributed with parameter a, then the rth moment of y is nite if and only if r < a. Higher a means higher nite moments. Equivalently, the faster the tail of the density declines to zero, the more moments are nite. This connection between tail decay and nite moments is not limited to the Pareto distribution. We can make a similar analysis using a tail bound. Suppose that y has density f (y ) which satises the bound f (y) A |y |a1 for some A < and a > 0. Since f (y) is bounded below a scale of a Pareto density, its tail behavior is similarly bounded. This means that for r < a Z Z 1 Z 2A r r < . |y | f (y )dy f (y)dy + 2A yra1 dy 1 + E |y | = a r 1 1
The parameter a of the Pareto distribution indexes the rate of decay of the tail of the density. Larger a means that the tail declines to zero more quickly. See Figure 2.11 below where we show the Pareto density for a = 1 and a = 2. The parameter a also determines which moments are nite. We can calculate that R a ra1 if r < a dy = a 1 y ar r E |y | = if r a.
49
Figure 2.11: Pareto Densities, a = 1 and a = 2 Thus if the tail of the density declines at the rate |y |a1 or faster, then y has nite moments up to (but not including) a. Broadly speaking, the restriction that y has a nite rth moment means that the tail of y s density declines to zero faster than yr1 . The faster decline of the tail means that the probability of observing an extreme value of y is a more rare event. We complete this section by adding an alternative representation of expectation in terms of the distribution function. Theorem 2.31.1 For any non-negative random variable y Z Pr (y > u) du Ey =
0
Proof of Theorem 2.31.1: Let F (x) = Pr (y > x) = 1 F (x), where F (x) is the distribution function. By integration by parts Z Z Z Z ydF (y) = ydF (y) = [yF (y )]0 + F (y )dy = Pr (y > u) du Ey =
0 0 0 0
as stated.
2.32
Existence and Uniqueness of the Conditional Expectation*
In Sections 2.3 and 2.6 we dened the conditional mean when the conditioning variables x are discrete and when the variables (y, x) have a joint density. We have explored these cases because
50
these are the situations where the conditional mean is easiest to describe and understand. However, the conditional mean exists quite generally without appealing to the properties of either discrete or continuous random variables. To justify this claim we now present a deep result from probability theory. What is says is that the conditional mean exists for all joint distributions (y, x) for which y has a nite mean.
Theorem 2.32.1 Existence of the Conditional Mean If E |y | < then there exists a function m(x) such that for all measurable sets X E (1 (x X ) y ) = E (1 (x X ) m(x)) . (2.60) The function m(x) is almost everywhere unique, in the sense that if h(x) satises (2.60), then there is a set S such that Pr(S ) = 1 and m(x) = h(x) for x S. The function m(x) is called the conditional mean and is written m(x) = E (y | x) . See, for example, Ash (1972), Theorem 6.3.3.
The conditional mean m(x) dened by (2.60) specializes to (2.7) when (y, x) have a joint density. The usefulness of denition (2.60) is that Theorem 2.32.1 shows that the conditional mean m(x) exists for all nite-mean distributions. This denition allows y to be discrete or continuous, for x to be scalar or vector-valued, and for the components of x to be discrete or continuously distributed.
2.33
Identication*
A critical and important issue in structural econometric modeling is identication, meaning that a parameter is uniquely determined by the distribution of the observed variables. It is relatively straightforward in the context of the unconditional and conditional mean, but it is worthwhile to introduce and explore the concept at this point for clarity. Let F denote the distribution of the observed data, for example the distribution of the pair (y, x). Let F be a collection of distributions F. Let be a parameter of interest (for example, the mean Ey ).
Denition 2.33.1 A parameter R is identied on F if for all F F , there is a uniquely determined value of .
Equivalently, is identied if we can write it as a mapping = g (F ) on the set F . The restriction to the set F is important. Most parameters are identied only on a strict subset of the space of all distributions. Take, for example, the mean Ey. It is uniquely odetermined if E |y | < , so it is clear that n = R is identied for the set F = F : |y | dF (y ) < . However, is also well dened when it is either positive or negative innity. Hence, dening I1 and I2 as in (2.57) and (2.58), we can deduce that is identied on the set F = {F : {I1 < } {I2 < }} . Next, consider the conditional mean. Theorem 2.32.1 demonstrates that E |y | < is a sucient condition for identication.
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION Theorem 2.33.1 Identication of the Conditional Mean If E |y| < , the conditional mean m(x) = E (y | x) is identied almost everywhere.
51
It might seem as if identication is a general property for parameters, so long as we exclude degenerate cases. This is true for moments of observed data, but not necessarily for more complicated models. As a case in point, consider the context of censoring. Let y be a random variable with distribution F. Instead of observing y, we observe y dened by the censoring rule y if y y = . if y > That is, y is capped at the value . A common example is income surveys, where income responses are top-coded, meaning that incomes above the top code are recorded as equalling the top code. The observed variable y has distribution F (u) for u F (u) = 1 for u . We are interested in features of the distribution F not the censored distribution F . For example, we are interested in the mean wage = E (y ) . The diculty is that we cannot calculate from F except in the trivial case where there is no censoring Pr (y ) = 0. Thus the mean is not generically identied from the censored distribution. A typical solution to the identication problem is to assume a parametric distribution. For example, let F be the set of normal distributions y N(, 2 ). It is possible to show that the parameters (, 2 ) are identied for all F F . That is, if we know that the uncensored distribution is normal, we can uniquely determine the parameters from the censored distribution. This is often called parametric identication as identication is restricted to a parametric class of distributions. In modern econometrics this is generally viewed as a second-best solution, as identication has been achieved only through the use of an arbitrary and unveriable parametric assumption. A pessimistic conclusion might be that it is impossible to identify parameters of interest from censored data without parametric assumptions. Interestingly, this pessimism is unwarranted. It turns out that we can identify the quantiles q of F for Pr (y ) . For example, if 20% of the distribution is censored, we can identify all quantiles for (0, 0.8). This is often called nonparametric identication as the parameters are identied without restriction to a parametric class. What we have learned from this little exercise is that in the context of censored data, moments can only be parametrically identied, while (non-censored) quantiles are nonparametrically identied. Part of the message is that a study of identication can help focus attention on what can be learned from the data distributions available.
2.34
Technical Proofs*
Proof of Theorem 2.7.1: For convenience, assume that the variables have a joint density f (y, x). Since E (y | x) is a function of the random vector x only, to calculate its expectation we integrate with respect to the density fx (x) of x, that is Z E (E (y | x)) = E (y | x) fx (x) dx.
Rk
52
Substituting in (2.7) and noting that fy|x (y|x) fx (x) = f (y, x) , we nd that the above expression equals Z Z Z Z
Rk R
yfy|x (y |x) dy fx (x) dx =
Rk
yf (y, x) dydx = E (y )
the unconditional mean of y.
Proof of Theorem 2.7.2: Again assume that the variables have a joint density. It is useful to observe that f (y, x1 , x2 ) f (x1 , x2 ) f (y |x1 , x2 ) f (x2 |x1 ) = = f (y, x2 |x1 ) , (2.61) f (x1 , x2 ) f (x1 ) the density of (y, x2 ) given x1 . Here, we have abused notation and used a single symbol f to denote the various unconditional and conditional densities to reduce notational clutter. Note that Z yf (y|x1 , x2 ) dy. (2.62) E (y | x1 , x2 ) =
R
Integrating (2.62) with respect to the conditional density of x2 given x1 , and applying (2.61) we nd that Z E (y | x1 , x2 ) f (x2 |x1 ) dx2 E (E (y | x1 , x2 ) | x1 ) = Rk2 Z Z = yf (y |x1 , x2 ) dy f (x2 |x1 ) dx2 k ZR 2 Z R = yf (y |x1 , x2 ) f (x2 |x1 ) dydx2 k ZR 2 ZR = yf (y, x2 |x1 ) dydx2
Rk2 R
= E (y | x1 )
as stated.
Proof of Theorem 2.7.3: Z Z g (x) yfy|x (y |x) dy = g (x) yfy|x (y |x) dy = g (x) E (y | x) E (g (x) y | x) =
R R
This is (2.9). The assumption that E |g (x) y | < is required for the rst equality to be welldened. Equation (2.10) follows by applying the Simple Law of Iterated Expectations to (2.9). Proof of Theorem 2.10.2: The assumption that Ey2 < implies that all the conditional expectations below exist. Set z = E(y | x1 , x2 ). By the conditional Jensens inequality (B.13), (E(z | x1 ))2 E z 2 | x1 . Taking unconditional expectations, this implies E (E(y | x1 ))2 E (E(y | x1 , x2 ))2 . (Ey )2 E (E(y | x1 ))2 E (E(y | x1 , x2 ))2 . (2.63)
Similarly,
53
The variables y, E(y | x1 ) and E(y | x1 , x2 ) all have the same mean Ey, so the inequality (2.63) implies that the variances are ranked monotonically: 0 var (E(y | x1 )) var (E(y | x1 , x2 )) . Next, for = Ey observe that E (y E(y | x)) (E(y | x) ) = E (y E(y | x)) (E(y | x) ) = 0 so the decomposition y = y E(y | x) + E(y | x) satises var (y ) = var (y E(y | x)) + var (E(y | x)) . (2.65) The monotonicity of the variances of the conditional mean (2.64) applied to the variance decomposition (2.65) implies the reverse monotonicity of the variances of the dierences, completing the proof. Proof of Theorem 2.8.1. Applying Minkowskis Inequality (B.19) to e = y m(x), (E |e|r )1/r = (E |y m(x)|r )1/r (E |y |r )1/r + (E |m(x)|r )1/r < , where the two parts on the right-hand are nite since E |y |r < by assumption and E |m(x)|r < by the Conditional Expectation Inequality (B.14). The fact that (E |e|r )1/r < implies E |e|r < . Proof of Theorem 2.18.1. For part 1, by the Expectation Inequality (B.15), (A.9) and Assumption 2.18.1, 0 E xx E xx0 = E kxk2 < . Similarly, using the Expectation Inequality (B.15), the Cauchy-Schwarz Inequality (B.17) and Assumption 2.18.1, 1/2 1/2 Ey2 < . kE (xy)k E kxyk = E kxk2 (2.64)
Thus the moments E (xy ) and E (xx0 ) are nite and well dened. For part 2, the coecient = (E (xx0 ))1 E (xy ) is well dened since (E (xx0 ))1 exists under Assumption 2.18.1. Part 3 follows from Denition 2.18.1 and part 2. For part 4, rst note that 2 Ee2 = E y x0 = Ey 2 2E y x0 + 0 E xx0 1 E (xy ) = Ey 2 2E y x0 E xx0 Ey 2 < . The rst inequality holds because E (y x0 ) (E (xx0 ))1 E (xy ) is a quadratic form and therefore necessarily non-negative. Second, by the Expectation Inequality (B.15), the Cauchy-Schwarz Inequality (B.17) and Assumption 2.18.1, 1/2 1/2 < . Ee2 kE (xe)k E kxek = E kxk2
It follows that the expectation E (xe) is nite, and is zero by the calculation (2.28).
CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION For part 6, Applying Minkowskis Inequality (B.19) to e = y x0 , r 1/r (E |e|r )1/r = E y x0 r 1/r (E |y |r )1/r + E x0 < ,
54
(E |y |r )1/r + (E kxkr )1/r kk
the nal inequality by assumption.
55
Exercises
Exercise 2.1 Find E (E (E (y | x1 , x2 , x3 ) | x1 , x2 ) | x1 ) . Exercise 2.2 If E (y | x) = a + bx, nd E (yx) as a function of moments of x. Exercise 2.3 Prove Theorem 2.8.1.4 using the law of iterated expectations. Exercise 2.4 Suppose that the random variables y and x only take the values 0 and 1, and have the following joint probability distribution x=0 .1 .4 x=1 .2 .3
y=0 y=1
Find E (y | x) , E y 2 | x and var (y | x) for x = 0 and x = 1.
Exercise 2.5 Show that 2 (x) is the best predictor of e2 given x: (a) Write down the mean-squared error of a predictor h(x) for e2 . (b) What does it mean to be predicting e2 ? (c) Show that 2 (x) minimizes the mean-squared error and is thus the best predictor.
Exercise 2.6 Use y = m(x) + e to show that var (y ) = var (m(x)) + 2 Exercise 2.7 Show that the conditional variance can be written as 2 (x) = E y2 | x (E (y | x))2 .
Exercise 2.8 Suppose that y is discrete-valued, taking values only on the non-negative integers, and the conditional distribution of y given x is Poisson: Pr (y = j | x) = exp (x0 ) (x0 )j , j! j = 0, 1, 2, ...
Compute E (y | x) and var (y | x) . Does this justify a linear regression model of the form y = x0 + e? )j , then Ey = and var(y ) = . Hint: If Pr (y = j ) = exp( j! Exercise 2.9 Suppose you have two regressors: x1 is binary (takes values 0 and 1) and x2 is categorical with 3 categories (A, B, C ). Write E (y | x1 , x2 ) as a linear regression.
Exercise 2.12 True or False. If y = x0 + e and E (e | x) = 0, then e is independent of x. Exercise 2.13 True or False. If y = x0 + e and E(xe) = 0, then E (e | x) = 0.
Exercise 2.11 True or False. If y = x + e, x R, and E (xe) = 0, then E x2 e = 0.
Exercise 2.10 True or False. If y = x + e, x R, and E (e | x) = 0, then E x2 e = 0.
56
Exercise 2.14 True or False. If y = x0 + e, E (e | x) = 0, and E e2 | x = 2 , a constant, then e is independent of x. Exercise 2.15 Consider the intercept-only model y = + e dened as the best linear predictor. Show that = E(y ). 2 2 on 0 x 1, 0 y 1. + y Exercise 2.16 Let x and y have the joint density f (x, y) = 3 x 2 Compute the coecients of the best linear predictor y = + x + e. Compute the conditional mean m(x) = E (y | x) . Are the best linear predictor and conditional mean dierent? Exercise 2.17 Let x be a random variable with = Ex and 2 = var(x). Dene x 2 . g x | , = (x )2 2 Show that Eg (x | m, s) = 0 if and only if m = and s = 2 . 1 x = x2 x3 Exercise 2.18 Suppose that
and x3 = 1 + 2 x2 is a linear function of x2 .
(a) Show that Qxx = E (xx0 ) is not invertible. (b) Use a linear transformation of x to nd an expression for the best linear predictor of y given x. (Be explicit, do not just use the generalized inverse formula.) Exercise 2.19 Show (2.46)-(2.47), namely that for 2 d() = E m(x) x0 = argmin d() 1 E (xm(x)) = E xx0 0 1 E (xy) . = E xx
Rk
then
Hint: To show E (xm(x)) = E (xy) use the law of iterated expectations. Exercise 2.20 Verify that (2.60) holds with m(x) dened in (2.7) when (y, x) have a joint density f (y, x).
Chapter 3
The Algebra of Least Squares

3.1 Introduction
In this chapter we introduce the popular least-squares estimator. Most of the discussion will be algebraic, with questions of distribution and inference defered to later chapters.
3.2
Random Samples
We can estimate from observational data which includes joint measurements on the variables (y, x) . For example, supposing we are interested in estimating a wage equation, we would use a dataset with observations on wages (or weekly earnings), education, experience (or age), and demographic characteristics (gender, race, location). One possible dataset is the Current Population Survey (CPS), a survey of U.S. households which includes questions on employment, income, education, and demographic characteristics. Notationally we wish to emphasize when we are discussing observations. Typically in econometrics we denote observations by appending a subscript i which runs from 1 to n, thus the ith observation is (yi , xi ), and n denotes the sample size. The dataset is then {(yi , xi ); i = 1, ..., n}. From the viewpoint of empirical analysis, a dataset is a array of numbers often organized as a table, where the columns of the table correspond to distinct variables and the rows correspond to distinct observations. For empirical analysis, the dataset and observations are xed in the sense that they are numbers presented to the researcher. For statistical analysis we need to view the dataset as random, or more precisely as a realization of a random process. For cross-sectional studies, the most common approach is to treat the individual observations as independent draws from an underlying population F. When the observations are realizations of independent and identically distributed random variables, we say that the data is a random sample. Assumption 3.2.1 The observations {(y1 , x1 ), ..., (yi , xi ), ..., (yn , xn )} are a random sample.
In Section 2.18 we derived and discussed the best linear predictor of y given x for a pair of random variables (y, x) R Rk , and called this the linear projection model. We are now interested in estimating the parameters of this model, in particular the projection coecient 1 E (xy) . = E xx0
With a random sample, the ordering of the data is irrelevant. There is nothing special about any specic observation or ordering. You can permute the order of the observations and no information is gained or lost. 57
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES
58
As most economic data sets are not literally the result of a random experiment, the random sampling framework is best viewed as an approximation rather than being literally true. The linear projection model applies to the random observations (yi , xi ) . This means that the probability model for the observations is the same as that described in Section 2.18. We can write the model as (3.1) yi = x0 i + ei where the linear projection coecient is dened as = argmin S (),
Rk
(3.2)
the minimizer of the expected squared error 2 S () = E yi x0 i , (3.3) (3.4)
and has the explicit solution
1 E (xi yi ) . = E xi x0 i
3.3
Sample Means
Consider the intercept-only model yi = + ei E(ei ) = 0. In this case the regression parameter is the unconditional mean = E(yi ). The standard estimator of a population mean is the sample mean, namely b= 1X yi . n
i=1 n
The sample mean is the empirical analog of the population mean, and is the conventional estimator in the lack of other information about or the distribution of y. We call b the moment estimator for . Indeed, whenever we have a parameter which can be written as the expectation of a function of random variables, a natural estimator of the parameter is the moment estimator, which is the sample 2 ) the moment mean of the corresponding function of the observations. For example, for 2 = E(yi 1 Pn 2 b 1 Pn y1i y2i . estimator is b2 = i=1 yi , and for = E(y1i y2i ) the moment estimator is = n n i=1
3.4
Least Squares Estimator
The linear projection coecient is dened in (3.2) as the minimizer of the expected squared error S ( ) dened in (3.3). For given , the expected squared error is the expectation of the 2 squared error (yi x0 i ) . The moment estimator of S ( ) is the sample average: Sn () = 2 1 X yi x0 i n
i=1 n
(3.5)
1 = SSEn () n where
n X 2 SSEn () = yi x0 i i=1
59
Figure 3.1: Sum-of-Squared Errors Function is called the sum-of-squared-errors function. Since Sn () is a sample average, we can interpret it as an estimator of the expected squared error S (). Examining Sn () as a function of therefore is informative about how S ( ) varies with . The projection coecient coeent minimizes S (), an analog estimator minimizes (3.5): b = argmin Sn ().
Rk
b as the minAlternatively, as Sn () is a scale multiple of SSEn (), we may equivalently dene b is commonly called the least-squares (LS) (or ordinary least imizer of SSEn (). Hence squares (OLS)) estimator of . Here, as is common in econometrics, we put a hat ^ over the b is a sample estimate of . This is a helpful convention, as just by parameter to indicate that b seeing the symbol we can immediately interpret it as an estimator (because of the hat), and as an estimator of a parameter labelled . Sometimes when we want to be explicit about the estimation b ols to signify that it is the OLS estimator. It is also common to see the method, we will write b notation n , where the subscript n indicates that the estimator depends on the sample size n. It is important to understand the distinction between population parameters such as and b . The population parameter is a non-random feature of the population sample estimates such as b is a random feature of a random sample. is xed, while b varies while the sample estimate across samples. To visualize the quadratic function Sn ( ), Figure 3.1 displays an example sum-of-squared erb is the the pair ( b1 , b2 ) rors function SSEn ( ) for the case k = 2. The least-squares estimator minimizing this function.
3.5
Solving for Least Squares with One Regressor
For simplicity, we start by considering the case k = 1 so that the coecient is a scalar. Then the sum of squared errors is a simple quadratic
60
SSEn ( ) = =
n X i=1
b minimizes this function. From elementary algebra we know that the minimizer The OLS estimator of the quadratic function a 2bx + cx2 is x = b/c. Thus the minimizer of SSEn ( ) is Pn i=1 xi yi b (3.6) = P n 2 . i=1 xi The intercept-only model is the special case xi = 1. In this case we nd Pn n 1yi 1X i =1 b yi = y, = Pn 2 = n i=1 1
i=1
n X
i=1
(yi xi )2
2 yi
n X
i=1
xi yi
n X
i=1
x2 i
(3.7)
the sample mean of yi . Here, as is common, we put a bar over y to indicate that the quantity is a sample mean. This calculation shows that the OLS estimator in the intercept-only model is the sample mean.
3.6
Solving for Least Squares with Multiple Regressors
This is a quadratic expression in the vector argument . The rst-order-condition for minimization of SSEn () is n n X X b ) = 2 b SSEn ( 0= xi yi + 2 xi x0 (3.8) i .
i=1 i=1
We now consider the case with k 1 so that the coecient is a vector. b , expand the SSE function to nd To solve for n n n X X X 2 0 0 SSEn () = yi 2 xi yi + xi x0 i .
i=1 i=1 i=1
This is the natural estimator of the best linear projection coecient dened in (3.2), and can also be called the linear projection estimator. We see that (3.9) simplies to the expression (3.6) when k = 1. The expression (3.9) is a notationally simple generalization but requires a careful attention to vector and matrix manipulations. Alternatively, equation (3.4) writes the projection coecient as an explicit function of the population moments Qxy and Qxx . Their moment estimators are the sample moments X b xy = 1 xi yi Q n
i=1 i=1 n
We have written this using a single expression, but it is actually a system of k equations with k b ). unknowns (the elements of b The solution for may be found by solving the system of k equations in can write P(3.8). We 0 we nd an x x this solution compactly using matrix algebra. Inverting the k k matrix n i=1 i i explicit formula for the least-squares estimator n !1 n ! X X b= xi x0 xi yi . (3.9)
i i=1 i=1
b xx Q
n 1X = xi x0 i. n
61
The moment estimator of replaces the population moments in (3.4) with the sample moments:
1 b=Q b b xx Qxy n !1 n ! 1X 1X 0 = xi xi xi yi n n i=1 i=1 n !1 n ! X X = xi x0 xi yi i i=1 i=1
which is identical with (3.9).
Least Squares Estimation b is Denition 3.6.1 The least-squares estimator where b = argmin Sn ()
Rk n
Sn () =
and has the solution n !1 n ! X X b= xi x0 xi yi .

i i=1 i=1
2 1 X yi x0 i n
i=1
Adrien-Marie Legendre
The method of least-squares was rst published in 1805 by the French mathematician Adrien-Marie Legendre (1752-1833). Legendre proposed leastsquares as a solution to the algebraic problem of solving a system of equations when the number of equations exceeded the number of unknowns. This was a vexing and common problem in astronomical measurement. As viewed by Legendre, (3.1) is a set of n equations with k unknowns. As the equations cannot be solved exactly, Legendres goal was to select to make the set of errors as small as possible. He proposed the sum of squared error criterion, and derived the algebraic solution presented above. As he noted, the rstorder conditions (3.8) is a system of k equations with k unknowns, which can be solved by ordinary methods. Hence the method became known as Ordinary Least Squares and to this day we still use the abbreviation OLS to refer to Legendres estimation method.
62
3.7
Illustration
We illustrate the least-squares estimator in practice with the data set used to generate the estimates from Chapter 2. This is the March 2009 Current Population Survey, which has extensive information on the U.S. population. This data set is described in more detail in Section 3.19. For this illustration, we use the sub-sample of married (spouse present) black female wages earners with 12 years potential work experience. This sub-sample has 20 observations. Let yi be log wages and xi be years of education and an intercept. Then
n X i=1 n X i=1
xi yi =
995.86 62.64
xi x0 i !1
5010 314 314 20
, (3.10)
and
Thus
n X
i=1
xi x0 i
0.0125 0.196 0.196 3.124 995.86 62.64
b= =
0.0125 0.196 0.196 3.124 0.155 0.698 .
We often write the estimated equation using the format \ log( W age) = 0.155 education + 0.698. (3.11)
An interpretation of the estimated equation is that each year of education is associated with an 16% increase in mean wages. Equation (3.11) is called a bivariate regression as there are only two variables. A multivariate regression has two or more regressors, and allows a more detailed investigation. Lets take an example similar to (3.11) but include all levels of experience. This time, we use the sub-sample of single (never married) asian men, which has 268 observations. Including as regressors years of potential work experience (experience ) and its square (experience 2 /100) (we divide by 100 to simplify reporting), we obtain the estimates \ log( W age) = 0.143 education + 0.036 experience 0.071 experience2 /100 + 0.575. (3.12)
These estimates suggest a 14% increase in mean wages per year of education, holding experience constant.
3.8
Least Squares Residuals

b y i = x0 i
As a by-product of estimation, we dene the tted value
and the residual
b i = yi x0 e i = yi y i .
(3.13)
63
i is a Sometimes y i is called the predicted value, but this is a misleading label. The tted value y function of the entire sample, including yi , and thus cannot be interpreted as a valid prediction of i as a tted rather than a predicted value. yi . It is thus more accurate to describe y i + e i and Note that yi = y b i . (3.14) yi = x0 i + e We make a distinction between the error ei and the residual e i . The error ei is unobservable while the residual e i is a by-product of estimation. These two variables are frequently mislabeled, which can cause confusion. Equation (3.8) implies that n X xi e i = 0. (3.15)
i=1
To see this by a direct calculation, using (3.13) and (3.9),

n X i=1
xi e i = =
n X i=1 n X i=1 n X i=1 n X i=1
b xi yi x0 i xi yi xi yi xi yi
n X i=1 n X i=1 n X i=1
b xi x0 i xi x0 i xi yi
= =
n X
i=1
xi x0 i
!1 n X
i=1
xi yi
= 0.
When xi contains a constant, an implication of (3.15) is 1X e i = 0. n

i=1 n
(3.16)
Thus the residuals have a sample mean of zero and the sample correlation between the regressors and the residual is zero. These are algebraic results, and hold true for all linear regression estimates.
3.9
Model in Matrix Notation
For many purposes, including computation, it is convenient to write the model and statistics in matrix notation. The linear equation (2.26) is a system of n equations, one for each observation. We can stack these n equations together as y1 = x0 1 + e1 y2 = x0 2 + e2 . . . yn = x0 n + en . Now dene y1 y2 . . . yn x0 1 x0 2 . . . x0 n e1 e2 . . . en
y=
X=
e=
64
Observe that y and e are n 1 vectors, and X is an n k matrix. Then the system of n equations can be compactly written in the single equation y = X + e. Sample sums can be written in matrix notation. For example
n X 0 xi x0 i =X X
(3.17)
i=1 n X i=1
xi yi = X 0 y .
The matrix version of (3.14) and estimated version of (3.17) is b +e y = X , b. e = y X = 0. X 0e
Therefore the least-squares estimator can be written as b = X 0 X 1 X 0 y .
(3.18)
or equivalently the residual vector is
Using the residual vector, we can write (3.15) as
(3.20)
Using matrix notation we have simple expressions for most estimators. This is particularly convenient for computer programming, as most languages allow matrix notation and manipulation.
Important Matrix Expressions y = X + e b = X 0 X 1 X 0 y b e = y X
X 0e = 0.
Early Use of Matrices

The earliest known treatment of the use of matrix methods to solve simultaneous systems is found in Chapter 8 of the Chinese text The Nine Chapters on the Mathematical Art, written by several generations of scholars from the 10th to 2nd century BCE.
65
3.10
Projection Matrix
1 0 X. P = X X 0X
Dene the matrix Observe that
This is a property of a projection matrix. More generally, for any matrix Z which can be written as Z = X for some matrix (we say that Z lies in the range space of X ), then 1 0 P Z = P X = X X 0X X X = X = Z. X = [X 1 X 2] ,
1 0 P X = X X 0X X X = X.
As an important example, if we partition the matrix X into two matrices X 1 and X 2 so that
then P X 1 = X 1 . (See Exercise 3.7.) The matrix P is symmetric and idempotent1 . To see that it is symmetric, 1 0 0 X P 0 = X X 0X 0 0 1 0 XX = X0 (X )0 0 1 0 = X X 0X X 0 1 0 = X (X )0 X 0 X = P. To establish that it is idempotent, the fact that P X = X implies that 1 0 P P = P X X 0X X 0 1 0 =X XX X = P. The matrix P has the property that it creates the tted values in a least-squares regression: 1 0 b=y X y = X . P y = X X 0X
Because of this property, P is also known as the hat matrix. A special example of a projection matrix occurs when X = 1 is an n-vector of ones. Then 1 0 1 P 1 = 1 10 1 1 = 110 . n Note that 1 0 1y P 1 y = 1 10 1 = 1y creates an n-vector whose elements are the sample mean y of yi .
1
A matrix P is symmetric if P 0 = P . A matrix P is idempotent if P P = P . See Appendix A.8.
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES The ith diagonal element of P = X (X 0 X )

1
66
X 0 is (3.21)
which is called the leverage of the ith observation. Some useful properties of the the matrix P and the leverage values hii are now summarized.
0 1 xi hii = x0 i X X
Theorem 3.10.1
and
n X i=1
hii = tr P = k
(3.22)
0 hii 1.
(3.23)
To show (3.22), 1 0 tr P = tr X X 0 X X 1 0 XX = tr X 0 X = tr (I k ) = k. See Appendix A.4 for denition and properties of the trace operator. The proof of (3.23) is defered to Section 3.21.
3.11
Orthogonal Projection
Dene M = In P 1 0 X = I n X X 0X
where I n is the n n identity matrix. Note that
M X = (I n P ) X = X P X = X X = 0. Thus M and X are orthogonal. We call M an orthogonal projection matrix or an annihilator matrix due to the property that for any matrix Z in the range space of X then M Z = Z P Z = 0. For example, M X 1 = 0 for any subcomponent X 1 of X , and M P = 0 (see Exercise 3.7). The orthogonal projection matrix M has many similar properties with P , including that M is symmetric (M 0 = M ) and idempotent (M M = M ). Similarly to (3.22) we can calculate tr M = n k. (See Exercise 3.9.) While P creates tted values, M creates least-squares residuals: b=b M y = y P y = y X e. (3.25) (3.24)
67
As discussed in the previous section, a special example of a projection matrix occurs when X = 1 is an n-vector of ones, so that P 1 = 1 (10 1)1 10 . Similarly, set M 1 = In P 1 1 0 = I n 1 10 1 1. M 1 y = y 1y . For simplicity we will often write the right-hand-side as y y . The ith element is yi y , the demeaned value of yi . We can also use (3.25) to write an alternative expression for the residual vector. Substituting y = X + e into e = M y and using M X = 0 we nd e = M y = M (X + e) = M e which is free of dependence on the regression coecient . (3.26)
While P 1 creates a vector of sample means, M 1 creates demeaned values:
3.12
Estimation of Error Variance
The error variance 2 = Ee2 i is a moment, so a natural estimator is a moment estimator. If ei were observed we would estimate 2 by 1X 2 ei . = n
2 i=1 n
(3.27)
However, this is infeasible as ei is not observed. In this case it is common to take a two-step i approach to estimation. The residuals e i are calculated in the rst step, and then we substitute e for ei in expression (3.27) to obtain the feasible estimator 2 = 1X 2 e i . n
i=1 n
(3.28)
In matrix notation, we can write (3.27) and (3.28) as 2 = n1 e0 e and Recall the expressions b e = M y = M e from (3.25) and (3.26). Applied to (3.29) we nd 2 = n1 e 0 e = n1 y 0 M M y = n1 y 0 M y = n1 e0 M e the third equality since M M = M . An interesting implication is that 2 2 = n1 e0 e n1 e0 M e = n1 e0 P e 0. The nal inequality holds because P is positive semi-denite and e0 P e is a quadratic form. This shows that the feasible estimator 2 is numerically smaller than the idealized estimator (3.27). 2 = n1 b e0 b e. (3.29)
68
3.13
Analysis of Variance
Another way of writing (3.25) is y = P y + My = y +e . This decomposition is orthogonal, that is y 0 e = (P y )0 (M y ) = y 0 P M y = 0. It follows that y0 y = y 0 y + 2y 0 e + e 0 e = y 0 y +e 0 e or
n X i=1 2 yi n X i=1 2 y i n X i=1
(3.30)
e 2 i.
Subtracting y from both sizes of (3.30) we obtain
y 1y =y 1y +e This decomposition is also orthogonal when X contains a constant, as = y 0 e y 10 e = 0 (y 1y )0 e under (3.16). It follows that (y 1y )0 (y 1y ) = (y 1y )0 (y 1y ) + e 0 e or
n X i=1
(yi y )2 =
It is often described as the fraction of the sample variance of yi which is explained by the leastsquares t. R2 is a crude measure of regression t. We have better measures of t, but these require a statistical (not just algebraic) analysis and we will return to these issues later. One deciency with R2 is that it increases when regressors are added to a regression (see Exercise 3.16) so the t can be always increased by increasing the number of regressors.
This is commonly called the analysis-of-variance formula for least squares regression. A commonly reported statistic is the coecient of determination or R-squared: Pn Pn 2 yi y )2 i 2 i=1 ( i=1 e . R = Pn 2 = 1 Pn ) )2 i=1 (yi y i=1 (yi y
n X i=1
( yi y )2 +
n X i=1
e 2 i.
3.14
Regression Components
X = [X 1 X 2] .
Partition and = Then the regression model can be rewritten as y = X 1 1 + X 2 2 + e. (3.31)
1 2
69
b and b . We are interested in algebraic expressions for 1 2 The algebra for the estimator is identical as that for the population coecients as presented in Section 2.21. b xy as b xx and Q Partition Q b xx = Q and similarly Qxy b 12 b 11 Q Q b 21 Q b 22 Q 1 X0 X1 n 1 = 1 0 X X1 n 2 1 0 X 1X 2 n 1 0 X 2X 2 n
0 0 The OLS estimator of = (0 1 , 2 ) is obtained by regression of y on X = [X 1 X 2 ] and can be written as b + X 2 b +e b +e . (3.32) y = X = X 1 1 2
1 b 0 X y Q1y n 1 b xy = . = Q 1 b 0 Q2y X y n 2 By the partitioned matrix inversion formula (A.4) 11 1 1 1 b b b b 12 1 b 12 b 12 Q b b 11 Q Q Q Q Q Q Q 11 2 11 2 22 1 b def Q = = xx = 21 22 1 1 1 b b b b b b b b Q21 Q22 Q Q Q221 Q221 Q21 Q11
1 b 112 = Q b b b 11 Q b 12 Q b b b b 1 b where Q 22 Q21 and Q221 = Q22 Q21 Q11 Q12 . Thus ! b1 b= b2 " #" 1 1 1 b b b b Q Q Q Q 112 112 12 22 = b 21 Q b 1 b 1 b 1 Q Q Q 221 11 221 1 ! b b Q 112 Q1y 2 = 1 b 221 Q b 2y1 Q
(3.33)
b 1y Q b 2y Q
Now
1 b 11 Q b 12 Q b 112 = Q b b Q 22 Q21
1 1 0 = X0 1X 1 X 1X 2 n n 1 0 = X 1M 2X 1 n
1 0 X X2 n 2
1 0 X X1 n 2
where
b 221 = 1 X 0 M 1 X 2 where is the orthogonal projection matrix for X 2 . Similarly Q n 2 1 0 X1 M 1 = In X1 X0 1X 1
1 0 X2 M 2 = In X2 X0 2X 2
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES is the orthogonal projection matrix for X 1 . Also
1 b 1y Q b 12 Q b 1y2 = Q b b Q 22 Q2y 1 1 0 1 0 1 0 1 0 X 2X 2 X y = X 1y X 1X 2 n n n n 2 1 M 2y = X0 n 1
70
b 2y1 = 1 X 0 M 1y. and Q n 2 Therefore and
These are algebraic expressions for the sub-coecient estimates from (3.32).
b = X 0 M 1 X 2 1 X 0 M 1 y . 2 2 2
b = X 0 M 2 X 1 1 X 0 M 2 y 1 1 1
(3.34) (3.35)
3.15
Residual Regression
As rst recognized by Frisch and Waugh (1933), expressions (3.34) and (3.35) can be used to b can be found by a two-step regression procedure. b and show that the least-squares estimators 1 2 Take (3.35). Since M 1 is idempotent, M 1 = M 1 M 1 and thus b = X 0 M 1 X 2 1 X 0 M 1 y 2 2 2 1 0 = X0 X 2M 1M 1y 2M 1M 1X 2 0 1 0 f2 X f2 f2 e = X 1 X f2 = M 1 X 2 X e 1 = M 1 y .
where and
b is algebraically equal to the least-squares regression of e Thus the coecient estimate 1 on 2 f X 2 . Notice that these two are y and X 2 , respectively, premultiplied by M 1 . But we know that 1 is simply the multiplication by M 1 is equivalent to creating least-squares residuals. Therefore e f2 are the least-squares least-squares residual from a regression of y on X 1 , and the columns of X residuals from the regressions of the columns of X 2 on X 1 . We have proven the following theorem.
Theorem 3.15.1 Frisch-Waugh-Lovell In the model (3.31), the OLS estimator of 2 and the OLS residuals e may be equivalently computed by either the OLS regression (3.32) or via the following algorithm: 1 ; 1. Regress y on X 1 , obtain residuals e f2 ; 2. Regress X 2 on X 1 , obtain residuals X
b 2 and residuals e f2 , obtain OLS estimates 3. Regress e 1 on X .
71
In some contexts, the FWL theorem can be used to speed computation, but in most cases there is little computational advantage to using the two-step algorithm. This result is a direct analogy of the coecient representation obtained in Section 2.22. The result obtained in that section concerned the population projection coecients, the result obtained here concern the least-squares estimates. The key message is the same. In the least-squares b 2 numerically equals the regression of y on the regressors regression (3.32), the estimated coecient X 2 , only after the regressors X 1 have been linearly projected out. Similarly, the coecient estimate b numerically equals the regression of y on the regressors X 1 , after the regressors X 2 have been 1 linearly projected out. This result can be very insightful when intrepreting regression coecients. A common application of the FWL theorem, which you may have seen in an introductory econometrics course, is the demeaning formula for regression. Partition X = [X 1 X 2 ] where X 1 = 1 is a vector of ones and X 2 is a matrix of observed regressors. In this case, 1 0 M 1 = I n 1 10 1 1. Observe that and f2 = M 1 X 2 = X 2 X 2 X y = M 1y = y y
b 2 is the OLS estimate from a regression are the demeaned variables. The FWL theorem says that of yi y on x2i x2 : n ! !1 n X X 0 b = (x2i x2 ) (x2i x2 ) (x2i x2 ) (yi y ) .
2 i=1 i=1
Thus the OLS estimator for the slope coecients is a regression with demeaned data.
Ragnar Frisch
Ragnar Frisch (1895-1973) was co-winner with Jan Tinbergen of the rst Nobel Memorial Prize in Economic Sciences in 1969 for their work in developing and applying dynamic models for the analysis of economic problems. Frisch made a number of foundational contributions to modern economics beyond the Frisch-Waugh-Lovell Theorem, including formalizing consumer theory, production theory, and business cycle theory.
3.16
Prediction Errors
The least-squares residual e i are not true prediction errors, as they are constructed based on the full sample including yi . A proper prediction for yi should be based on estimates constructed using only the other observations. We can do this by dening the leave-one-out OLS estimator of as that obtained from the sample of n 1 observations excluding the ith observation: b (i) = 1 n1 X
j 6=i
1 X (i) y (i) . = X0 (i) X (i)
xj x0 j
1 n1
X
j 6=i
xj yj
(3.36)
72
Here, X (i) and y (i) are the data matrices omitting the ith row. The leave-one-out predicted value for yi is b y i = x0 i (i) , and the leave-one-out residual or prediction error or prediction residual is e i = yi y i . b A convenient alternative expression for (i) (derived in Section 3.21) is where hii are the leverage values as dened in (3.21). Using (3.37) we can simplify the expression for the prediction error: b e i = yi x0 i (i)
1 0 0 1 = yi x0 xi X X xi e i i + (1 hii ) 1 0 1 b b xi e i XX (i) = (1 hii )
(3.37)
=e i + (1 hii )1 hii e i = (1 hii )1 e i .
(3.38)
To write this in vector notation, dene M = (I n diag{h11 , .., hnn })1 Then (3.38) is equivalent to A convenient feature of this expression is that it shows that computation of the full vector of prediction errors e e is based on a simple linear operation, and does not really require n separate estimations. One use of the prediction errors is to estimate the out-of-sample mean squared error 1X 2 e i = n
2 i=1 i=1 n
= diag{(1 h11 )1 , .., (1 hnn )1 }. e e. e = M b
(3.39) (3.40)
n 1X = (1 hii )2 e 2 i. n
(3.41) 2 is
This is also known as the sample mean squared prediction error. Its square root = the prediction standard error.
3.17
Inuential Observations
Another use of the leave-one-out estimator is to investigate the impact of inuential observations, sometimes called outliers. We say that observation i is inuential if its omission from the sample induces a substantial change in a parameter estimate of interest. For illustration, consider Figure 3.2 which shows a scatter plot of random variables (yi , xi ). The 25 observations shown with the open circles are generated by xi U [1, 10] and yi N (xi , 4). The 26th observation shown with the lled circle is x26 = 9, y26 = 0. (Imagine that y26 = 0 was incorrectly recorded due to a mistaken key entry.) The Figure shows both the least-squares tted line from the full sample and that obtained after deletion of the 26th observation from the sample. In this example we can see how the 26th observation (the outlier) greatly tilts the least-squares
73
10
leaveoneout OLS 8 6
OLS y 0 2 4
4 x
10
Figure 3.2: Impact of an inuential observation on the least-squares estimator tted line towards the 26th observation. In fact, the slope coecient decreases from 0.97 (which is close to the true value of 1.00) to 0.56, which is substantially reduced. Neither y26 nor x26 are unusual values relative to their marginal distributions, so this outlier would not have been detected from examination of the marginal distributions of the data. The change in the slope coecient of 0.41 is meaningful and should raise concern to an applied economist. From (3.37)-(3.38) we know that b b (i) = (1 hii )1 X 0 X 1 xi e i 0 1 = XX xi e i .
(3.42)
By direct calculation of this quantity for each observation i, we can directly discover if a specic observation i is inuential for a coecient estimate of interest. For a general assessment, we can focus on the predicted values. The dierence between the full-sample and leave-one-out predicted values is
0b b i = x0 y i y i xi (i) 0 1 = x0 xi e i i X X
= hii e i
which is a simple function of the leverage values hii and prediction errors e i . Observation i is i | is large, which requires that both hii and |e i | are large. inuential for the predicted value if |hii e One way to think about this is that a large leverage value hii gives the potential for observation i to be inuential. A large hii means that observation i is unusual in the sense that the regressor xi is far from its sample mean. We call an observation with large hii a leverage point. A leverage point is not necessarily inuential as the latter also requires that the prediction error e i is large. To determine if any individual observations are inuential in this sense, several diagnostics have been proposed (some names include DFITS, Cooks Distance, and Welsch Distance). Unfortunately, from a statistical perspective it is dicult to recommend these diagnostics for applications as they are not based on statistical theory. Probably the most relevant measure is the change in the coecient estimates given in (3.42). The ratio of these changes to the coecients standard error is called its DFBETA, and is a postestimation diagnostic available in STATA. While there is no magic threshold, the concern is whether or not an individual observation meaningfully changes an
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES estimated coecient of interest. A simple diagnostic for inuential observations is to calculate i y i | = max |hii e i | . Inf luence = max |y
1in 1in
74
This is the largest (absolute) change in the predicted value due to a single observation. If this diagnostic is large relative to the distribution of yi , it may indicate that that observation is inuential. If an observation is determined to be inuential, what should be done? As a common cause of inuential observations is data entry error, the inuential observations should be examined for evidence that the observation was mis-recorded. Perhaps the observation falls outside of permitted ranges, or some observables are inconsistent (for example, a person is listed as having a job but receives earnings of $0). If it is determined that an observation is incorrectly recorded, then the observation is typically deleted from the sample. This process is often called cleaning the data. The decisions made in this process involve an fair amount of individual judgement. When this is done it is proper empirical practice to document such choices. (It is useful to keep the source data in its original form, a revised data le after cleaning, and a record describing the revision process. This is especially useful when revising empirical work at a later date.) It is also possible that an observation is correctly measured, but unusual and inuential. In this case it is unclear how to proceed. Some researchers will try to alter the specication to properly model the inuential observation. Other researchers will delete the observation from the sample. The motivation for this choice is to prevent the results from being skewed or determined by individual observations, but this practice is viewed skeptically by many researchers who believe it reduces the integrity of reported empirical results. For an empirical illustration, consider the log wage regression (3.12) for single asian males. This regression, which has 268 observations, has Inf luence = 0.29. This means that the most inuential observation, when deleted, changes the predicted (tted) value of the dependent variable log(W age) by 0.29, or equivalently the wage by 29%. This is a meaningful change and suggests further investigation. We examine the inuential observation, and nd that its leverage hii is 0.33, which is disturbingly large. (Rcall that the leverage values are all positive and sum to one. One third of the leverage in this sample of 268 observations is contained in just this single observation!) Examining further, we nd that this individual is 65 years old with 8 years education, so that his potential experience is 51 years. This is the highest experience in the subsample the next highest is 41 years. The large leverage is due to to his unusual characteristics (very low education and very high experience) within this sample. Essentially, regression (3.12) is attempting to estimate the conditional mean at experience = 51 with only one observation, so it is not surprising that this observation determines the t and is thus inuential. A reasonable conclusion is the regression function can only be estimated over a smaller range of experience. We restrict the sample to individuals with less than 45 years experience, re-estimate, and obtain the following estimates. \ log( W age) = 0.144 education + 0.043 experience 0.095 experience2 /100 + 0.531. (3.43)
For this regression, we calculate that Inf luence = 0.11, which is greatly reduced relative to the regression (3.12). Comparing (3.43) with (3.12), the slope coecient for education is essentially unchanged, but the coecients in experience and its square have slightly increased. By eliminating the inuential observation, equation (3.43) can be viewed as a more robust estimate of the conditional mean for most levels of experience. Whether to report (3.12) or (3.43) in an application is largely a matter of judgment.
3.18
Normal Regression Model
The normal regression model is the linear regression model under the restriction that the error ei is independent of xi and has the distribution N 0, 2 . We can write this as ei | xi N 0, 2 .
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES This assumption implies 2 yi | xi N x0 i , .
75
Normal regression is a parametric model, where likelihood methods can be used for estimation, testing, and distribution theory. The log-likelihood function for the normal regression model is ! n X 2 1 1 2 0 log exp 2 yi xi log L(, ) = 2 (22 )1/2 i=1 1 n n (3.44) = log (2 ) log 2 2 SSEn (). 2 2 2
2 ) maximizes log L( , 2 ). Since the latter b mle , The maximum likelihood estimator (MLE) ( mle is a function of only through the sum of squared errors SSEn (), maximizing the likelihood is identical to minimizing SSEn (). Hence
b the MLE for equals the OLS estimator. Due to this equivalence, the least squares estimator ols is often called the MLE. b into the log-likelihood we obtain We can also nd the MLE for 2 . Plugging Maximization with respect to 2 yields the rst-order condition b b , 2 = n log (2 ) n log 2 SSEn (mle ) . log L mle 2 2 2 2 1 n 2 b , b ) = 0. log L SSEn ( = 2 + mle mle 2 2 2 ( 2 )2
2 mle n b ) 1X 2 SSEn ( mle = = e i n n i=1
b b mle = ols ,
Solving for 2 yields the MLE for 2
which is the same as the moment estimator (3.28). Plugging the estimates into (3.44) we obtain the maximized log-likelihood 2 n n 2 b mle , = (log (2 ) + 1) log mle . mle log L 2 2
(3.45)
The log-likelihood (or the negative log-likelihood) is typically reported as a measure of t. b mle is numerically equal to the OLS estimator, despite It may seem surprising that the MLE emerging from quite dierent motivations. It is not completely accidental. The least-squares estimator minimizes a particular sample loss function the sum of squared error criterion and most loss functions are equivalent to the likelihood of a specic parametric distribution, in this case the normal regression model. In this sense it is not surprising that the least-squares estimator can be motivated as either the minimizer of a sample loss function or as the maximizer of a likelihood function.
76
Carl Friedrich Gauss

The mathematician Carl Friedrich Gauss (1777-1855) proposed the normal regression model, and derived the least squares estimator as the maximum likelihood estimator for this model. He claimed to have discovered the method in 1795 at the age of eighteen, but did not publish the result until 1809. Interest in Gausss approach was reinforced by Laplaces simultaneous discovery of the central limit theorem, which provided a justication for viewing random disturbances as approximately normal.
3.19
CPS Data Set
In this section we describe the data set used in the empirical illustrations. The Current Population Survey (CPS) is a monthly survey of about 57,000 U.S. households conducted by the Bureau of the Census of the Bureau of Labor Statistics. The CPS is the primary source of information on the labor force characteristics of the U.S. population. The survey covers employment, earnings, educational attainment, income, poverty, health insurance coverage, job experience, voting and registration, computer usage, veteran status, and other variables. Details can be found at www.census.gov/cps and dataferrett.census.gov. From the March 2009 survey we extracted the individuals with non-allocated variables who were full-time employed (dened as those who had worked at least 36 hours per week for at least 48 weeks the past year), and excluded those in the military. This sample has 50,742 individuals. We extracted 14 variables from the CPS on these individuals and created the data les cps09mar.dta (Stata format) and cps98mar.txt (text format). The variables are: 1. age: years, capped at 85 2. female: 1 if female, 0 otherwise 3. hisp: 1 if Spanish, Hispanic, or Latino, 0 otherwise 4. education 0 4 6 8 9 10 11 12 13 14 16 18 20 Less than 1st grade 1st, 2nd, 3rd, or 4th grade 5th or 6th grade 7th or 8th grade 9th grade 10th grade 11th grade or 12th grade with no high school diploma High school graduate, high school diploma or equivalent Some college but no degree Associate degree in college, including occupation/vocation programs Bachelors degree or equivalent (BA, AB, BS) Masters degree (MA, MS MENG, MED, MSW, MBA) Professional degree or Doctorate degree (MD, DDS, DVM, LLB, JD, PHD, EDD)
5. earnings: total annual wage and salary earnings
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES 6. hours: number of hours worked per week 7. week: number of weeks worked per year 8. union: 1 for member of a labor union, 0 otherwise 9. uncov: 1 if covered by a union or employee association contract, 0 otherwise 10. region 1 2 3 4 11. Race 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 White only Black only American Indian, Alaskan Native (AI) only Asian only Hawaiian/Pacic Islander (HP) only White-Black White-AI White-Asian White-HP Black-AI Black-Asian Black-HP AI-Asian Asian-HP White-Black-AI White-Black-Asian White-AI-Asian White-Asian-HP White-Black-AI-Asian 2 or 3 races 4 or 5 races Northeast Midwest South West
77
12. marital 1 2 3 4 5 6 7 Married - civilian spouse present Married - Armed Forces spouse present Married - spouse absent (except separated) Widowed Divorced Separated Never married
78
3.20
Programming
Most packages allow both interactive programming (where you enter commands one-by-one) and batch programming (where you run a pre-written sequence of commands from a le). Interactive programming can be useful for exploratory analysis, but eventually all work should be executed in batch mode. This is the best way to control and document your work. Batch programs are text les where each line executes a single command. For Stata, this le needs to have the lename extension .do, and for Matlab .m, while for Gauss and R there are no specic naming requirements. To execute a program le, you type a command within the program. Stata: do chapter3 executes the le chapter3.do Gauss: run chapter3.prg executes the le chapter3.prg Matlab: run chapter3 executes the le chapter3.m R: source(chapter3.r) executes the le chatper3.r When writing batch les, it is useful to include comments for documentation and readability. We illustrate programming les for Stata, Gauss, R, and Matlab, which execute a portion of the empirical illustrations from Sections 3.7 and 3.17.
Stata do File * Clear memory and load the data clear use cps09mar.dta * Generate transformations gen wage=ln(earnings/(hours*week)) gen experience = age - education - 6 gen exp2 = (experience^2)/100 * Create indicator for subsamples gen mbf = (race == 2) & (marital <= 2) & (female == 1) gen sam = (race == 4) & (marital == 7) & (female == 0) * Regressions reg wage education if (mbf == 1) & (experience == 12) reg wage education experience exp2 if sam == 1 * Leverage and inuence predict leverage,hat predict e,residual gen d=e*leverage/(1-leverage) summarize d if mnwf ==1
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES Gauss Program File /* Load the data and create subsamples */ load dat[50742,12]=cps09mar.txt; experience=dat[.,1]-dat[.,4]-6; mbf=(dat[.,11].==2).*(dat[.,12].<=2).*(dat[.,2].==1).*(experience.==12); sam=(dat[.,11].==4).*(dat[.,12].==7).*(dat[.,2].==0); dat1=selif(dat,mbf); dat2=selif(dat,sam); /* First regression */ y=ln(dat1[.,5]./(dat1[.,6].*dat1[.,7])); x=dat1[.,4]~ones(rows(dat1),1); beta=invpd(xx)*(xy); beta; /* Second regression */ y=ln(dat2[.,5]./(dat2[.,6].*dat2[.,7])); experience=dat2[.,1]-dat2[.,4]-6; exp2 = (experience.^2)/100; x=dat2[.,4]~experience~exp2~ones(rows(dat2),1); beta=invpd(xx)*(xy); beta; /* Create leverage and inuence */ e=y-x*beta; leverage=sumc((x.*(x*invpd(xx)))); d=leverage.*e./(1-leverage); "Inuence " maxc(abs(d));
79
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES R Program File # Load the data and create subsamples dat <- read.table("cps09mar.txt") experience <- dat[,1]-dat[,4]-6 mbf <- (dat[,11]==2)&(dat[,12]<=2)&(dat[,2]==1)&(experience==12) sam <- (dat[,11]==4)&(dat[,12]==7)&(dat[,2]==0) dat1 <- dat[mbf,] dat2 <- dat[sam,] # First regression y <- as.matrix(log(dat1[,5]/(dat1[,6]*dat1[,7]))) x <- cbind(dat1[,4],matrix(1,nrow(dat1),1)) beta <- solve(t(x)%*%x,t(x)%*%y) print(beta) # Second regression y <- as.matrix(log(dat2[,5]/(dat2[,6]*dat2[,7]))) experience <- dat2[,1]-dat2[,4]-6 exp2 <- (experience^2)/100 x <- cbind(dat2[,4],experience,exp2,matrix(1,nrow(dat2),1)) beta <- solve(t(x)%*%x,t(x)%*%y) print(beta) # Create leverage and inuence e <- y-x%*%beta leverage <- rowSums(x*(x%*%solve(t(x)%*%x))) r <- e/(1-leverage) d <- leverage*e/(1-leverage) print(max(abs(d)))
80
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES Matlab Program File % Load the data and create subsamples load cps09mar.txt; dat=cps09mar; experience=dat(:,1)-dat(:,4)-6; mbf = (dat(:,11)==2)&(dat(:,12)<=2)&(dat(:,2)==1)&(experience==12); sam = (dat(:,11)==4)&(dat(:,12)==7)&(dat(:,2)==0); dat1=dat(mbf,:); dat2=dat(sam,:); % First regression y=log(dat1(:,5)./(dat1(:,6).*dat1(:,7))); x=[dat1(:,4),ones(length(dat1),1)]; beta=inv(x*x)*(x*y); display(beta); % Second regression y=log(dat2(:,5)./(dat2(:,6).*dat2(:,7))); experience=dat2(:,1)-dat2(:,4)-6; exp2 = (experience.^2)/100; x=[dat2(:,4),experience,exp2,ones(length(dat2),1)]; beta=inv(x*x)*(x*y); display(beta); % Create leverage and inuence e=y-x*beta; leverage=sum((x.*(x*inv(x*x)))); d=leverage.*e./(1-leverage); inuence=max(abs(d)); display(inuence);
81
82
3.21
Technical Proofs*
1
0 Proof of Theorem 3.10.1, equation (3.23): First, hii = x0 xi 0 since it is a i (X X ) 0 quadratic form and X X > 0. Next, since hii is the ith diagonal element of the projection matrix 1 P = X (X 0 X ) X, then hii = s0 P s
where
is a unit vector with a 1 in the ith place (and zeros elsewhere). By the spectral decomposition of the idempotent matrix P (see equation (A.5)) Ik 0 0 P =B B 0 0
0 . . . s= 1 . . . 0
where B 0 B = I n . Thus letting b = Bs denote the ith column of B , and partitioning b0 = b0 1 then Ik 0 0 0 Bs hii = s B 0 0 Ik 0 0 = b1 b1 0 0 = b0 1 b1 b0 b
b0 2
=1
the nal equality since b is the ith column of B and B 0 B = I n . We have shown that hii 1, establishing (3.23). Proof of Equation (3.37). The ShermanMorrison formula (A.3) from Appendix A.5 states that for nonsingular A and vector b 1 1 0 1 1 = A1 + 1 b0 A1 b A bb A . A bb0 This implies and thus X 0 X xi x0 i 1 1 1 0 1 = X 0X + (1 hii )1 X 0 X xi x0 i X X
b = (X 0 X )1 X 0 y and hii = x0 (X 0 X )1 xi , and the the third equality making the substitutions i remainder collecting terms.
0 0 0 1 b X y xi yi (i) = X X xi xi 1 0 1 = X 0X X y X 0X xi yi 1 0 1 0 + (1 hii )1 X 0 X xi x0 X y xi yi i X X b X 0 X 1 xi yi + (1 hii )1 X 0 X 1 xi x0 b hii yi = i b (1 hii )1 X 0 X 1 xi (1 hii ) yi x0 b + hii yi = i 0 1 1 b (1 hii ) = xi e i XX
83
Exercises
Exercise 3.1 Let y be a random variable with = Ey and 2 = var(y). Dene y 2 . g y, , = (y )2 2 Pn
i=1 g (yi , m, s) .
Let ( , 2 ) be the values such that g n ( , 2 ) = 0 where g n (m, s) = n1 2 and are the sample mean and variance.
Show that
Exercise 3.2 Consider the OLS regression of the n 1 vector y on the n k matrix X . Consider an alternative set of regressors Z = XC , where C is a k k non-singular matrix. Thus, each column of Z is a mixture of some of the columns of X . Compare the OLS estimates and residuals from the regression of y on X to the OLS estimates from the regression of y on Z . = 0. Exercise 3.3 Using matrix algebra, show X 0 e Exercise 3.4 Let e be the OLS residual from a regression of y on X = [X 1 X 2 ]. Find X 0 . 2e Exercise 3.5 Let e be the OLS residual from a regression of y on X . Find the OLS coecient from a regression of e on X . on X . Exercise 3.6 Let y = X (X 0 X )1 X 0 y . Find the OLS coecient from a regression of y Exercise 3.7 Show that if X = [X 1 X 2 ] then P X 1 = X 1 and M X 1 = 0. Exercise 3.8 Show that M is idempotent: M M = M . Exercise 3.9 Show that tr M = n k. Exercise 3.10 Show that if X = [X 1 X 2 ] and X 0 1 X 2 = 0 then P = P 1 + P 2 . Exercise 3.11 Show that when X contains a constant, 1 Pn y i = y . n i=1
Exercise 3.12 A dummy variable takes on only the values 0 and 1. It is used for categorical data, such as an individuals gender. Let d1 and d2 be vectors of 1s and 0s, with the i0 th element of d1 equaling 1 and that of d2 equaling 0 if the person is a man, and the reverse if the person is a woman. Suppose that there are n1 men and n2 women in the sample. Consider tting the following three equations by OLS y = + d1 1 + d2 2 + e y = d1 1 + d2 2 + e y = + d1 + e Can all three equations (3.46), (3.47), and (3.48) be estimated by OLS? Explain if not. (a) Compare regressions (3.47) and (3.48). Is one more general than the other? Explain the relationship between the parameters in (3.47) and (3.48). (b) Compute 0 d1 and 0 d2 , where is an n 1 is a vector of ones. (c) Letting = (1 2 )0 , write equation (3.47) as y = X +e. Consider the assumption E(xi ei ) = 0. Is there any content to this assumption in this setting? (3.46) (3.47) (3.48)
CHAPTER 3. THE ALGEBRA OF LEAST SQUARES Exercise 3.13 Let d1 and d2 be dened as in the previous exercise. (a) In the OLS regression y = d1 1 + d2 2 + u ,
84
show that 1 is the sample mean of the dependent variable among the men of the sample 2 is the sample mean among the women (y 2 ). (y 1 ), and that (b) Let X (n k) be an additional matrix of regressions. Describe in words the transformations
0 X = X d1 x0 1 d2 x2
y = y d1 y 1 d2 y2
where x1 and x2 are the k 1 means of the regressors for men and women, respectively. e from the OLS regresion (c) Compare b from the OLS regression with e +e y = X
b n = (X 0 X n )1 X 0 y n denote the OLS estimate when y n is n 1 and X n is Exercise 3.14 Let n n n k. A new observation (yn+1 , xn+1 ) becomes available. Prove that the OLS estimate computed using this additional observation is 0 1 1 0 b + b b y = X x x X n+1 n+1 n+1 n n n n+1 n . 1 0 1 + x0 xn+1 n+1 (X n X n ) Exercise 3.15 Prove that R2 is the square of the sample correlation between y and y . Exercise 3.16 Consider two least-squares regressions e1 + e y = X 1
b +e y = d1 1 + d2 2 + X .
and
2 and R2 be the R-squared from the two regressions. Show that R2 R2 . Is there a case Let R1 2 2 1 2 = R2 ? (explain) when there is equality R2 1
b + X 2 b +e y = X 1 . 1 2
2 . Is equality possible? Exercise 3.17 Show that 2 b b Exercise 3.18 For which observations will (i) = ?
Exercise 3.19 Use the data set from Section 3.19 and the sub-sample used for equation (3.43) (see Section 3.20) for data construction) 1. Estimate equation (3.43) and compute the equation R2 and sum of squared errors. 2. Re-estimate the slope on education using the residual regression approach. Regress log(Wage) on experience and its square, regress education on experience and its square, and the residuals on the residuals. Report the estimates from this nal regression, along with the equation R2 and sum of squared errors. Does the slope coecient equal the value in (3.43)? Explain. 3. Do the R2 and sum-of-squared errors from parts 1 and 2 equal? Explain.
85
Exercise 3.20 Estimate equation (3.43) as in part 1 of the previous question. Let e i be the OLS residual, y i the predicted value from the regression, x1i be education and x2i be experience. Numerically calculate the following: (a) (b) (c) (d) (e) (f) (g) Pn Pn Pn Pn Pn Pn Pn i i=1 e i i=1 x1i e i i=1 x2i e
2 i i=1 x1i e 2 i i=1 x2i e
i e i i=1 y 2 i i=1 e
Are these calculations consistent with the theoretical properties of OLS? Explain. Exercise 3.21 Use the data set from Section 3.19. 1. Estimate a log wage regression for the subsample of white male Hispanics. In addition to education, experience, and its square, include a set of binary variables for regions and marital status. For regions, you create dummy variables for Northeast, South and West so that Midwest is the excluded group. For marital status, create variables for married, windowed or divorced, and separated, so that single (never married) is the excluded group. 2. Repeat this estimation using a dierent econometric package. Compare your results. Do they agree?
Chapter 4
Least Squares Regression

4.1 Introduction
In this chapter we investigate some nite-sample properties of least-squares applied to a random sample in the the linear regression model. In particular, we calculate the nite-sample mean and covariance matrix and propose standard errors for the coecient estimates.
4.2
Sample Mean
To start with the simplest setting, we rst consider the intercept-only model yi = + ei E (ei ) = 0. which is equivalent to the regression model with k = 1 and xi = 1. In the intercept model, = E (yi ) b = y equals the sample mean is the mean of yi . (See Exercise 2.15.) The least-squares estimator as shown in equation (3.7). We now calculate the mean and variance of the estimator y . Since the sample mean is a linear function of the observations, its expectation is simple to calculate n ! n 1X 1X yi = Eyi = . E (y ) = E n n
i=1 i=1
This shows that the expected value of least-squares estimator (the sample mean) equals the projection coecient (the population mean). An estimator with the property that its expectation equals the parameter it is estimating is called unbiased.
b for is unbiased if E b = . Denition 4.2.1 An estimator We next calculate the variance of the estimator y. Making the substitution yi = + ei we nd y= 1X ei . n
i=1 n
86
CHAPTER 4. LEAST SQUARES REGRESSION Then var (y) = E (y )2 n ! n X X 1 1 =E ei ej n n

i=1 j =1
87
1 n2 1 n2
n n X X i=1 j =1 n X 2
E (ei ej )
= =
i=1
1 2 . n
The second-to-last inequality is because E (ei ej ) = 2 for i = j yet E (ei ej ) = 0 for i 6= j due to independence. 1 2 . This is the familiar formula for the variance of the sample We have shown that var (y ) = n mean.
4.3
Linear Regression Model
We now consider the linear regression model. Throughout the remainder of this chapter we maintain the following.
Assumption 4.3.1 Linear Regression Model The observations (yi , xi ) come from a random sample and satisfy the linear regression equation yi = x0 i + ei E (ei | xi ) = 0. The variables have nite second moments
2 < , Eyi
(4.1) (4.2)
E kxi k2 < , and an invertible design matrix Qxx = E xi x0 i > 0. We will consider both the general case of heteroskedastic regression, where the conditional variance 2 2 E e2 i | xi = (xi ) = i is unrestricted, and the specialized case of homoskedastic regression, where the conditional variance is constant. In the latter case we add the following assumption.
CHAPTER 4. LEAST SQUARES REGRESSION Assumption 4.3.2 Homoskedastic Linear Regression Model In addition to Assumption 4.3.1, 2 2 (4.3) E e2 i | xi = (xi ) = is independent of xi .
88
4.4
Mean of Least-Squares Estimator
In this section we show that the OLS estimator is unbiased in the linear regression model. This calculation can be done using either summation notation or matrix notation. We will use both. First take summation notation. Observe that under (4.1)-(4.2) E (yi | X ) = E (yi | xi ) = x0 i. (4.4)
The rst equality states that the conditional expectation of yi given {x1 , ..., xn } only depends on xi , since the observations are independent across i. The second equality is the assumption of a linear conditional mean. Using denition (3.9), the conditioning theorem, the linearity of expectations, (4.4), and properties of the matrix inverse, !1 n ! n X X b | X = E xi x0 xi yi | X E i
i=1 i=1
n X n X
i=1 i=1 i=1
xi x0 i xi x0 i xi x0 i xi x0 i
!1 !1 !1 !1
n X n X
i=1
n X i=1 n X i=1 n X i=1
n X
i=1
xi yi
|X
E (xi yi | X ) xi E (yi | X ) xi x0 i
= .
Similarly
Now lets show the same result using matrix notation. (4.4) implies . . . . . 0. = x = X . E (y | X ) = | X ) E ( y i i . . . . . . = E (ei | xi ) = 0. E (e | X ) = E ( e | X ) i . . . . . . . . . . . .
(4.5)
(4.6)
CHAPTER 4. LEAST SQUARES REGRESSION
89
Using denition (3.18), the conditioning theorem, the linearity of expectations, (4.5), and the properties of the matrix inverse, b | X = E X 0 X 1 X 0 y | X E 1 0 = X 0X X E (y | X ) 0 1 0 X X = XX = . At the risk of belaboring the derivation, another way to calculate the same result is as follows. b to obtain Insert y = X + e into the formula (3.18) for b = X 0 X 1 X 0 (X + e) 1 0 1 0 = X 0X Xe X X + X 0X 0 1 0 X e. (4.7) =+ XX
b into the true parameter and the stochastic This is a useful linear decomposition of the estimator 1 0 0 component (X X ) X e. Once again, we can calculate that b | X = E X 0 X 1 X 0 e | X E 1 0 X E (e | X ) = X 0X = 0.
b | X = . Applying the law of iterated Regardless of the method, we have shown that E expectations, we nd that b =E E b|X E = . We have shown the following theorem. Theorem 4.4.1 Mean of Least-Squares Estimator In the linear regression model (Assumption 4.3.1) b|X = E b ) = . E(
(4.8)
and
(4.9)
b is unbiased for , meaning that the distribution of Equation (4.9) says that the estimator b is centered at . Equation (4.8) says that the estimator is conditionally unbiased, which is a b is unbiased for any realization of the regressor matrix X . stronger result. It says that
4.5
Variance of Least Squares Estimator
In this section we calculate the conditional variance of the OLS estimator. For any r 1 random vector Z dene the r r covariance matrix var(Z ) = E (Z EZ ) (Z EZ )0 = EZZ 0 (EZ ) (EZ )0
CHAPTER 4. LEAST SQUARES REGRESSION and for any pair (Z, X) dene the conditional covariance matrix var(Z | X ) = E (Z E (Z | X )) (Z E (Z | X ))0 | X . def b V = var | X
90
We dene
The ith diagonal element of D is
the conditional covariance matrix of the regression coecient estimates. We now derive its form. The conditional covariance matrix of the n 1 regression error e is the n n matrix D = E ee0 | X . 2 2 E e2 i | X = E ei | xi = i
while the ij 0 th o-diagonal element of D is
E (ei ej | X ) = E (ei | xi ) E (ej | xj ) = 0. where the rst equality uses independence of the observations (Assumption 1.5.1) and the second 2: is (4.2). Thus D is a diagonal matrix with ith diagonal element i 2 1 0 0 2 2 0 2 0 2 (4.10) D = diag 1 , ..., n = . . . . . . .. . . . . . 2 0 0 n In the special case of the linear homoskedastic regression model (4.3), then 2 2 E e2 i | xi = i = D = I n 2 . In general, however, D need not necessarily take this simplied form. For any matrix n r matrix A = A(X ), var(A0 y | X ) = var(A0 e | X ) = A0 DA. b = A0 y where A = X (X 0 X )1 and thus In particular, we can write It is useful to note that
n X i=1
and we have the simplication
(4.11)
0 1 0 1 0 b X DX X 0 X . V = var( | X ) = A DA = X X X 0 DX =
2 xi x0 i i ,
a weighted version of X 0 X . In the special case of the linear homoskedastic regression model, D = I n 2 , so X 0 DX = 0 X X 2 , and the variance matrix simplies to 0 1 2 . V = X X
CHAPTER 4. LEAST SQUARES REGRESSION Theorem 4.5.1 Variance of Least-Squares Estimator In the linear regression model (Assumption 4.3.1) b V = var | X 1 0 1 = X 0X X DX X 0 X 0 1 2 . V = X X
91
(4.12)
where D is dened in (4.10). In the homoskedastic linear regression model (Assumption 4.3.2)
4.6
Gauss-Markov Theorem
Now consider the class of estimators of which are linear functions of the vector y , and thus can be written as e = A0 y
where A is an n k function of X . As noted before, the least-squares estimator is the special case obtained by setting A = X (X 0 X )1 . What is the best choice of A? The Gauss-Markov theorem, which we now present, says that the least-squares estimator is the best choice among linear unbiased estimators when the errors are homoskedastic, in the sense that the least-squares estimator has the smallest variance among all unbiased linear estimators. e = A0 y we have To see this, since E (y | X ) = X , then for any linear estimator e | X = A0 E (y | X ) = A0 X , E e is unbiased if (and only if) A0 X = I k . Furthermore, we saw in (4.11) that so e | X = var A0 y | X = A0 DA = A0 A 2 var
the last equality using the homoskedasticity assumption D = I n 2 . The best unbiased linear 0 estimator is obtained by nding the matrix A0 satisfying A0 0 X = I k such that A0 A0 is minimized in the positive denite sense, in that for any other matrix A satisfying A0 X = I k , then A0 A A0 0 A0 is positive semi-denite.
Theorem 4.6.1 Gauss-Markov 1. In the homoskedastic linear regression model (Assumption 4.3.2), the best (minimum-variance) unbiased linear estimator is the leastsquares estimator b = X 0 X 1 X 0 y
2. In the linear regression model (Assumption 4.3.1), the best unbiased linear estimator is e = X 0 D1 X 1 X 0 D1 y (4.13)
92
The rst part of the Gauss-Markov theorem is a limited eciency justication for the leastsquares estimator. The justication is limited because the class of models is restricted to homoskedastic linear regression and the class of potential estimators is restricted to linear unbiased estimators. This latter restriction is particularly unsatisfactory as the theorem leaves open the possibility that a non-linear or biased estimator could have lower mean squared error than the least-squares estimator. The second part of the theorem shows that in the (heteroskedastic) linear regression model, within the class of linear unbiased estimators the best estimator is not least-squares but is (4.13). This is called the Generalized Least Squares (GLS) estimator. The GLS estimator is infeasible as the matrix D is unknown. This result does not suggest a practical alternative to least-squares. We return to the issue of feasible implementation of GLS in Section 9.2. We give a proof of the rst part of the theorem below, and leave the proof of the second part for Exercise 4.3. Proof of Theorem 4.6.1.1. Let A be any n k function of X such that A0 X = I k . The variance 1 of the least-squares estimator is (X 0 X ) 2 and that of A0 y is A0 A 2 . It is sucient to show 1 1 that the dierence A0 A (X 0 X ) is positive semi-denite. Set C = A X (X 0 X ) . Note that X 0 C = 0. Then we calculate that 1 1 0 1 0 1 = C + X X 0X A0 A X 0 X C + X X 0X XX 1 1 = C 0C + C 0X X 0X + X 0X X 0C 0 1 0 0 1 0 1 XX XX XX + XX = C 0C . The matrix C 0 C is positive semi-denite (see Appendix A.8) as required.
4.7
Residuals
b b What are some properties of the residuals e i = yi x0 i = yi x0 i and prediction errors e i (i) , at least in the context of the linear regression model? Recall from (3.26) that we can write the residuals in vector notation as e = Me where M = I n X (X 0 X ) conditional expectation
1
X 0 is the orthogonal projection matrix. Using the properties of
E (e | X ) = E (M e | X ) = M E (e | X ) = 0 and var (e | X ) = var (M e | X ) = M var (e | X ) M = M DM where D is dened in (4.10). We can simplify this expression under the assumption of conditional homoskedasticity 2 E e2 i | xi = . In this case (4.14) simplies to var (e | X ) = M 2 . (4.15) (4.14)
93
As this variance is a function of hii and hence xi , the residuals e i are heteroskedastic even if the errors ei are homoskedastic. i can be written in Similarly, recall from (3.40) that the prediction errors e i = (1 hii )1 e vector notation as e = M e where M is a diagonal matrix with ith diagonal element (1 hii )1 . Thus e = M M e. We can calculate that E (e | X ) = M M E (e | X ) = 0 and var (e | X ) = M M var (e | X ) M M = M M DM M which simplies under homoskedasticity to var (e | X ) = M M M M 2 = M M M 2 . The variance of the ith prediction error is then 2 i | X var ( ei | X ) = E e
In particular, for a single observation i, we can nd the (conditional) variance of e i by taking the ith diagonal element of (4.16). Since the ith diagonal element of M is 1 hii as dened in (3.21) we obtain 2 i | X = (1 hii ) 2 . (4.16) var ( ei | X ) = E e
= (1 hii )1 2 .
= (1 hii )1 (1 hii ) (1 hii )1 2
A residual with constant conditional variance can be obtained by rescaling. The standardized residuals are i , (4.17) e i = (1 hii )1/2 e and in vector notation e = ( e1 , ..., e n )0 = M 1/2 M e. From our above calculations, under homoskedasticity, var (e | X ) = M 1/2 M M 1/2 2 and 2 i | X = 2 var ( ei | X ) = E e (4.18)
and thus these standardized residuals have the same bias and variance as the original errors when the latter are homoskedastic.
4.8
Estimation of Error Variance
The error variance 2 = Ee2 i can be a parameter of interest, even in a heteroskedastic regression or a projection model. 2 measures the variation in the unexplained part of the regression. Its method of moments estimator (MME) is the sample average of the squared residuals: 1X 2 e i = n
2 i=1 n
and equals the MLE in the normal regression model (3.28).
94
In the linear regression model we can calculate the mean of 2 . From (3.26), the properties of projection matrices and the trace operator, observe that 2 = Then 1 1 0 1 1 1 e e = e0 M M e = e0 M e = tr e0 M e = tr M ee0 . n n n n n
1 2 E | X = tr E M ee0 | X n 1 = tr M E ee0 | X n 1 (4.19) = tr (M D) . n 2 2 Adding the assumption of conditional homoskedasticity E e2 i | xi = , so that D = I n , then (4.19) simplies to 1 2 E | X = tr M 2 n 2 nk = , n the nal equality by (3.24). This calculation shows that 2 is biased towards zero. The order of the bias depends on k/n, the ratio of the number of estimated coecients to the sample size. Another way to see this is to use (4.16). Note that
n 1X 2 2 E e i | X E |X = n
= =
1 n
nk n
i=1 n X i=1
(1 hii ) 2 2 (4.20)
By the above calculation, so
the last equality using Theorem 3.10.1. Since the bias takes a scale form, a classic method to obtain an unbiased estimator is by rescaling the estimator. Dene n 1 X 2 2 e i . (4.21) s = nk
i=1
E s2 | X = 2 E s2 = 2
(4.22)
and the estimator s2 is unbiased for 2 . Consequently, s2 is known as the bias-corrected estimator for 2 and in empirical practice s2 is the most widely used estimator for 2 . Interestingly, this is not the only method to construct an unbiased estimator for 2 . An estimator constructed with the standardized residuals e i from (4.17) is 2 = 1X 2 1X e i = (1 hii )1 e 2 i. n n
i=1 i=1 n n
(4.23)
CHAPTER 4. LEAST SQUARES REGRESSION You can show (see Exercise 4.6) that 2 E | X = 2
95
(4.24)
and thus 2 is unbiased for 2 (in the homoskedastic linear regression model). When k/n is small (typically, this occurs when n is large), the estimators 2 , s2 and 2 are 2 2 2 are generally preferred to . Consequently it is likely to be close. However, if not then s and best to use one of the bias-corrected variance estimators in applications.
4.9
Mean-Square Forecast Error
A major purpose of estimated regressions is to predict out-of-sample values. Consider an outof-sample observation (yn+1 , xn+1 ) where xn+1 will be observed but not yn+1 . Given the coecient b the standard point estimate of E (yn+1 | xn+1 ) = x0 is y b estimate n+1 = x0 n+1 n+1 . The forecast n+1 = yn+1 y n+1 . error is the dierence between the actual value yn+1 and the point forecast, e The mean-squared forecast error (MSFE) is 2 M SF En = Ee n+1 . b , so = en+1 x0 n+1
n+1
In the linear regression model, e n+1
The rst term in (4.25) is 2 . The second term in (4.25) is zero since en+1 x0 n+1 is independent b of and both are mean zero. The third term in (4.25) is 0 b b tr E xn+1 x0 E n+1 0 0 b b = tr E xn+1 xn+1 E = tr E xn+1 x0 n+1 EV = E tr xn+1 x0 n+1 V = E x0 (4.26) xn+1 n+1 V b and use the face V where we use the fact that xn+1 is independent of Thus V x M SF En = 2 + E x0 n+1 . n+1
0 b M SF En = Ee2 n+1 2E en+1 xn+1 0 b b xn+1 . + E x0
(4.25)
0 b b |X . =E
Under conditional homoskedasticity, this simplies to 0 1 M SF En = 2 1 + E x0 xn+1 . n+1 X X 1X 2 e i = n

2 i=1 n
A simple estimator for the MSFE is obtained by averaging the squared prediction errors (3.41)
CHAPTER 4. LEAST SQUARES REGRESSION b i (1 hii )1 . Indeed, we can calculate that where e i = yi x0 i (i) = e
96
This is the MSFE based on a sample of size n 1, rather than size n. The dierence arises because the in-sample prediction errors e i for i n are calculated using an eective sample size of n1, while the out-of sample prediction error e n+1 is calculated from a sample with the full n observations. Unless n is very small we should expect M SF En1 (the MSFE based on n 1 observations) to 2 is a reasonable estimator for be close to M SF En (the MSFE based on n observations). Thus M SF En . Theorem 4.9.1 MSFE In the linear regression model (Assumption 4.3.1) 2 0 2 = + E x V x M SF En = Ee n+1 n+1 n+1
By the same calculations as in (4.26) we nd E 2 = 2 + E x0 iV
2 E 2 = Ee i 2 b = E ei x0 i (i) 0 b b = 2 + E x0 xi . (i) (i) i

(i)
xi = M SF En1 .
b where V = var | X . Furthermore, 2 dened in (3.41) is an unbiased estimator of M SF En1 : E 2 = M SF En1
4.10
Covariance Matrix Estimation Under Homoskedasticity
which is known up to the unknown scale 2 . In Section 4.8 we discussed three estimators of 2 . The most commonly used choice is s2 , leading to the classic covariance matrix estimator 0 1 2 b0 s . (4.27) V = X X b0 Since s2 is conditionally unbiased for 2 , it is simple to calculate that V is conditionally unbiased for V under the assumption of homoskedasticity: 0 1 2 0 b E V E s |X |X = X X 1 2 = X 0X =V .
For inference, we need an estimate of the covariance matrix V of the least-squares estimator. In this section we consider the homoskedastic regression model (Assumption 4.3.2). Under homoskedasticity, the covariance matrix takes the relatively simple form 0 1 2 V = X X
97
2 = x2 implies 2 = E 2 = Ex2 .) The constant is the (Notice that we use the fact that i i i i , and can be any number greater than standardized forth moment (or kurtosis) of the regressor x i 2 one. For example, if xi N 0, then = 3, so the true variance V is three times larger
This estimator was the dominant covariance matrix estimator in applied econometrics for many years, and is still the default method in most regression packages. If the estimator (4.27) is used, but the regression error is heteroskedastic, it is possible for 0 1 1 0 b (X 0 DX ) (X 0 X ) . For V to be quite biased for the correct covariance matrix V = (X X ) 2 = x2 with Ex = 0. The ratio of the true variance of the least-squares example, suppose k = 1 and i i i estimator to the expectation of the variance estimator is Pn 4 V Ex4 i i=1 xi 0 = 2 P ' 2 = . n 2 2 x b Exi i=1 i E V |X
b0 than the expected homoskedastic estimator V . But can be much larger. Suppose, for example, 1 . In this case = 15 , so that the true variance V that xi 2 is fteen times larger than 1 b0 the expected homoskedastic estimator V . While this is an extreme and constructed example, the point is that the classic covariance matrix estimator (4.27) may be quite biased when the homoskedasticity assumption fails.
4.11
Covariance Matrix Estimation Under Heteroskedasticity
In the previous section we showed that that the classic covariance matrix estimator can be highly biased if homoskedasticity fails. In this section we show how to contruct covariance matrix estimators which do not require homoskedasticity. Recall that the general form for the covariance matrix is 0 1 0 1 X DX X 0 X . V = X X
= E (D0 | X ) 2 where D0 = diag e1 , ..., e2 n . Thus D 0 is a conditionally unbiased estimator for D . If the squared 2 errors ei were observable, we could construct the unbiased estimator 1 0 1 b ideal = X 0X X D0X X 0 X V n ! 0 1 0 1 X 2 XX = XX xi x0 . i ei
i=1
This depends on the unknown matrix D which we can write as 2 2 , ..., n D = diag 1 = E ee0 | X
Indeed, ! n ideal X 0 1 1 2 b E V | X = X 0X xi x0 XX i E ei | X = X 0X =V 1 0 1 = X 0X X DX X 0 X n 1 X
i=1 i=1 2 xi x0 i i
X 0X
CHAPTER 4. LEAST SQUARES REGRESSION b ideal is unbiased for V verifying that V
98
While the scaling by n/(n k) is ad hoc, it is recommended over the unscaled estimator (4.28). Alternatively, we could use the prediction errors e i or the standardized residuals e i , yielding the estimators n ! 0 1 X 0 1 e = X X xi x0 2 XX V ie i
2 We know, however, that e 2 i is biased towards zero. To estimate the variance the unbiased 2 2 by n/(n k) . Making the same adjustment we obtain estimator s scales the moment estimator the estimator n ! 0 1 0 1 X n b = xi x0 2 . (4.29) V XX XX ie i nk i=1
b ideal Since the errors e2 is not a feasible estimator. However, we can replace i are unobserved, V i . Making this substitution we obtain the estimator the errors ei with the least-squares residuals e n ! 0 1 X 0 1 W b xi x0 2 . (4.28) V XX = X X ie i
i=1
= X 0X
n 1 X
i=1
i=1
(1 hii )2 xi x0 2 ie i
0 1 XX
(4.30)
and V n ! 0 1 X 0 1 0 2 = XX xi xi e i XX
i=1 i=1
n ! 0 1 X 0 1 1 0 2 = XX (1 hii ) xi xi e i . XX
(4.31)
by Andrews (1991) based on the principle of leave-one-out cross-validation (and is implemented using the vce(hc3) option in Stata). The estimator V was introduced by Horn, Horn and Duncan (1975) (and is implemented using the vce(hc2) option in Stata). Since (1 hii )2 > (1 hii )1 > 1 it is straightforward to show that e bW V <V <V (4.32)
rst developed by Eicker (1963) and introduced to econometrics by White (1980), and is sometimes b is the called the Eicker-White or White covariance matrix estimator. The scaled estimator V e was introduced default robust covariance matrix estimator implemented in Stata. The estimator V
b ,V b , V e , and V are collectively called robust, heteroskedasticityThe four estimators V b was consistent, or heteroskedasticity-robust covariance matrix estimators. The estimator V
(See Exercise 4.7). The inequality A < B when applied to matrices means that the matrix B A is positive denite. In general, the bias of the covariance matrix estimators is quite complicated, but they greatly
CHAPTER 4. LEAST SQUARES REGRESSION simplify under the assumption of homoskedasticity (4.3). For example, using (4.16), n ! W 1 X 2 0 1 0 0 b xi xi E e i | X XX E V |X = X X = X 0X n ! 0 1 2 0 1 X 0 1 2 = XX XX xi x0 XX i hii 1 2 X 0X =V .
W i=1
99
n 1 X
i=1
i=1
2 xi x0 i (1 hii )
0 1 XX
b is biased towards zero. This calculation shows that V e is biased away from zero, Similarly, (again under homoskedasticity) we can calculate that V specically e | X X 0 X 1 2 E V (4.33) while the estimator V is unbiased 1 2 | X = X 0X . E V (4.34)
b0 should you use? The classic estimator V is typically a poor choice, as it is only valid under the unlikely homoskedasticity restriction. For this reason it is not typically used in contemporary b0 econometric research. Unfortunately, standard regression packages set their default choice as V , so users must intentionally select a robust covariance matrix estimator. b are the most commonly used, and in particular V b bW Of the four robust estimators, V and V e is the default robust covariance matrix option in Stata. However, V and V are preferred based

(See Exercise 4.8.) It might seem rather odd to compare the bias of heteroskedasticity-robust estimators under the assumption of homoskedasticity, but it does give us a baseline for comparison. bW b , V e , and V . Which b0 We have introduced ve covariance matrix estimators, V , V , V
e and V are simple to implement, this should not be a barrier. on their improved bias. As V
Halbert L. White
Hal White (1950-2012) of the United States was an inuential econometrician of recent years. His 1980 paper on heteroskedasticity-consistent covariance matrix estimation for many years has been the most cited paper in economics. His research was central to the movement to view econometric models as approximations, and to the drive for increased mathematical rigor in the discipline. In addition to being a highly prolic and inuential scholar, he also co-founded the economic consulting rm Bates White.
100
4.12
Standard Errors
b. A b is an estimate of the variance of the distribution of A variance estimator such as V more easily interpretable measure of spread is its square root the standard deviation. This is so important when discussing the distribution of parameter estimates, we have a special name for estimates of their standard deviation. b) for a real-valued estimator b Denition 4.12.1 A standard error s( b is an estimate of the standard deviation of the distribution of . b and covariance matrix estimate V b , standard errors for When is a vector with estimate b . That is, individual elements are the square roots of the diagonal elements of V rh q i j ) = V b = b . V s( j
jj
As we discussed in the previous section, there are multiple possible covariance matrix estimators, so standard errors are not unique. It is therefore important to understand what formula and method is used by an author when studying their work. It is also important to understand that a particular standard error may be relevant under one set of model assumptions, but not under another set of assumptions. To illustrate, we return to the log wage regression (3.11) of Section 3.7. We calculate that 2 s = 0.160. Therefore the homoskedastic covariance matrix estimate is 1 5010 314 0.002 0.031 b0 = 0 . 215 = . V 314 20 0.031 0.499 We also calculate that
n X i=1
(1 hii )
xi x0 2 ie i
763.26 48.51 48.51 3.11
Therefore the Horn-Horn-Duncan covariance matrix estimate is 1 1 763.26 48.51 5010 314 5010 314 V = 48.51 3.11 314 20 314 20 0.001 0.015 = . 0.015 0.243
(4.35)
The standard errors are the square roots of the diagonal elements of these matrices. A conventional format to write the estimated equation with standard errors is \ log( W age) = 0.155 Education + 0.698 . (0.031) (0.493)
Alternatively, standard errors could be calculated using the other formulae. We report the dierent standard errors in the following table. Homoskedastic (4.27) White (4.28) Scaled White (4.29) Andrews (4.30) Horn-Horn-Duncan (4.31) Education 0.045 0.029 0.030 0.033 0.031 Intercept 0.707 0.461 0.486 0.527 0.493
101
The homoskedastic standard errors are noticably dierent (larger, in this case) than the others, but the four robust standard errors are quite close to one another.
4.13
Computation
We illustrate methods to compute standard errors for equation (3.12) extending the code of Section 3.20.
Stata do File (continued) * reg * reg * reg * reg Homoskedastic formula (4.27): wage education experience exp2 if (mnwf Scaled White formula (4.29): wage education experience exp2 if (mnwf Andrews formula (4.30): wage education experience exp2 if (mnwf Horn-Horn-Duncan formula (4.31): wage education experience exp2 if (mnwf
== 1) == 1), r == 1), vce(hc3) == 1), vce(hc2)
Gauss Program File (continued) n=rows(y); k=cols(x); a=n/(n-k); sig2=(ee)/(n-k); u1=x.*e; u2=x.*(e./(1-leverage)); u3=x.*(e./sqrt(1-leverage)); xx=inv(xx); v0=xx*sig2; v1=xx*(u1u1)*xx; v1a=a*xx*(u1u1)*xx; v2=xx*(u2u2)*xx; v3=xx*(u3u3)*xx s0=sqrt(diag(v0)); @ Homoskedastic formula @ s1=sqrt(diag(v1)); @ White formula @ s1a=sqrt(diag(v1a)); @ Scaled White formula @ s2=sqrt(diag(v2)); @ Andrews formula @ s3=sqrt(diag(v3)); @ Horn-Horn-Duncan formula @
CHAPTER 4. LEAST SQUARES REGRESSION R Program File (continued) n <- nrow(y) k <- ncol(x) a <- n/(n-k) sig2 <- (t(e) %*% e)/(n-k) u1 <- x*(e%*%matrix(1,1,k)) u2 <- x*((e/(1-leverage))%*%matrix(1,1,k)) u3 <- x*((e/sqrt(1-leverage))%*%matrix(1,1,k)) v0 <- xx*sig2 xx <- solve(t(x)%*%x) v1 <- xx %*% (t(u1)%*%u1) %*% xx v1a <- a * xx %*% (t(u1)%*%u1) %*% xx v2 <- xx %*% (t(u2)%*%u2) %*% xx v3 <- xx %*% (t(u3)%*%u3) %*% xx s0 <- sqrt(diag(v0)) # Homoskedastic formula s1 <- sqrt(diag(v1)) # White formula s1a <- sqrt(diag(v1a)) # Scaled White formula s2 <- sqrt(diag(v2)) # Andrews formula s3 <- sqrt(diag(v3)) # Horn-Horn-Duncan formula
102
Matlab Program File (continued) [n,k]=size(x); a=n/(n-k); sig2=(e*e)/(n-k); u1=x.*(e*ones(1,k)); u2=x.*((e./(1-leverage))*ones(1,k)); u3=x.*((e./sqrt(1-leverage))*ones(1,k)); xx=inv(x*x); v0=xx*sig2; v1=xx*(u1*u1)*xx; v1a=a*xx*(u1*u1)*xx; v2=xx*(u2*u2)*xx; v3=xx*(u3*u3)*xx; s0=sqrt(diag(v0)); # Homoskedastic formula s1=sqrt(diag(v1)); # White formula s1a=sqrt(diag(v1a)); # Scaled White formula s2=sqrt(diag(v2)); # Andrews formula s3=sqrt(diag(v3)); # Horn-Horn-Duncan formula
4.14
Measures of Fit
As we described in the previous chapter, a commonly reported measure of regression t is the regression R2 dened as Pn 2 e 2 2 R = 1 Pn i=1 i 2 = 1 2 . y ) i=1 (yi y

2 = n1 where y
103
Pn
i=1 (yi
y)2 . R2 can be viewed as an estimator of the population parameter 2 = 2 var (x0 i ) = 1 2. var(yi ) y
2 are biased estimators. Theil (1961) proposed replacing these by the unbiy However, 2 and Pn 2 2 = (n 1)1 ased versions s2 and y i=1 (yi y ) yielding what is known as R-bar-squared or adjusted R-squared: P (n 1) n e 2 s2 2 R =1 2 =1 Pn i=1 i 2 . y (n k) i=1 (yi y )
where e i are the prediction errors (3.38) and 2 is the MSPE from (3.41). As described in Section e2 is a good (4.9), 2 is a good estimator of the out-of-sample mean-squared forecast error, so R estimator of the percentage of the forecast variance which is explained by the regression forecast. e2 is a good measure of t. In this sense, R 2 e2 , is that R2 One problem with R2 , which is partially corrected by R and fully corrected by R necessarily increases when regressors are added to a regression model. This occurs because R2 is a negative function of the sum of squared residuals which cannot increase when a regressor is added. 2 e2 are non-monotonic in the number of regressors. R e2 can even be negative, In contrast, R and R which occurs when an estimated model predicts worse than a constant-only model. In the statistical literature the MSPE 2 is known as the leave-one-out cross validation criterion, and is popular for model comparison and selection, especially in high-dimensional (none2 or 2 to compare and select models. Models with parametric) contexts. It is equivalent to use R e2 (or low 2 ) are better models in terms of expected out of sample squared error. In contrast, high R 2 R cannot be used for model selection, as it necessarily increases when regressors are added to a 2 regression model. R is also an inappropriate choice for model selection (it tends to select models with too many parameters), though a justication of this assertion requires a study of the theory 2 of model selection. Unfortunately, R is routinely used by some economists, possibly as a hold-over from previous generations. e2 and/or 2 in regression analysis, In summary, it is recommended to calculate and report R 2 and omit R2 and R .
While R is an improvement on R2 , a much better improvement is Pn 2 e 2 2 e R = 1 Pn i=1 i 2 = 1 2 y ) i=1 (yi y
Henri Theil
Henri Theil (1924-2000) of Holland invented R and two-stage least squares, both of which are routinely seen in applied econometrics. He also wrote an early inuential advanced textbook on econometrics (Theil, 1971).
2
4.15
Empirical Example
We again return to our wage equation, but use a much larger sample of all individuals with at least 12 years of education. For regressors we include years of education, potential work experience, experience squared, and dummy variable indicators for the following: female, female union member,
104
male union member, married female1 , married male, formerly married female2 , formerly married male, hispanic, black, American Indian, Asian, and mixed race3 . The available sample is 46,943 so the parameter estimates are quite precise and reported in Table 4.1. For standard errors we use the unbiased Horn-Horn-Duncan formula. Table 4.1 displays the parameter estimates in a standard tabular format. The table clearly states the estimation method (OLS), the dependent variable (log(Wage)), and the regressors are clearly labeled. Both parameter estimates and standard errors are reported for all coecients. In addition to the coecient estimates, the table also reports the estimated error standard deviation and the sample size. These are useful summary measures of t which aid readers. Table 4.1 OLS Estimates of Linear Equation for Log(Wage) ) s( Education 0.117 0.001 Experience 0.033 0.001 -0.056 0.002 Experience2 /100 Female -0.098 0.011 Female Union Member 0.023 0.020 Male Union Member 0.095 0.020 Married Female 0.016 0.010 Married Male 0.211 0.010 Formerly Married Female -0.006 0.012 Formerly Married Male 0.083 0.015 Hispanic -0.108 0.008 Black -0.096 0.008 American Indian -0.137 0.027 Asian -0.038 0.013 Mixed Race -0.041 0.021 Intercept 0.909 0.021 0.565 Sample Size 46,943 Note: Standard errors are heteroskedasticity-consistent (Horn-Horn-Duncan formula)
As a general rule, it is advisable to always report standard errors along with parameter estimates. This allows readers to assess the precision of the parameter estimates, and as we will discuss in later chapters, form condence intervals and t-tests for individual coecients if desired. The results in Table 4.1 conrm our earlier ndings that the return to a year of education is approximately 12%, the return to experience is concave, that single women earn approximately 10% less then single men, and blacks earn about 10% less than whites. In addition, we see that Hispanics earn about 11% less than whites, American Indians 14% less, and Asians and Mixed races about 4% less. We also see there are wage premiums for men who are members of a labor union (about 10%), married (about 21%) or formerly married (about 8%), but no similar premiums are apparant for women.
1 2
Dening married as marital code 1, 2, or 3. Dening formerly married as marital code 4, 5, or 6. 3 Race code 6 or higher.
105
4.16
Multicollinearity
1 b are not dened. This situation is called strict If X 0 X is singular, then (X 0 X ) and multicollinearity, as the columns of X are linearly dependent, i.e., there is some 6= 0 such that X = 0. Most commonly, this arises when sets of regressors are included which are identically related. For example, if X includes both the logs of two prices and the log of the relative prices, log(p1 ), log(p2 ) and log(p1 /p2 ), for then X 0 X will necessarily be singular. When this happens, the applied researcher quickly discovers the error as the statistical software will be unable to construct (X 0 X )1 . Since the error is discovered quickly, this is rarely a problem for applied econometric practice. The more relevant situation is near multicollinearity, which is often called multicollinearity for brevity. This is the situation when the X 0 X matrix is near singular, when the columns of X are close to linearly dependent. This denition is not precise, because we have not said what it means for a matrix to be near singular. This is one diculty with the denition and interpretation of multicollinearity. One potential complication of near singularity of matrices is that the numerical reliability of the calculations may be reduced. In practice this is rarely an important concern, except when the number of regressors is very large. A more relevant implication of near multicollinearity is that individual coecient estimates will be imprecise. We can see this most simply in a homoskedastic linear regression model with two regressors yi = x1i 1 + x2i 2 + ei ,
and
1 0 XX= n
1 1
In this case
The correlation indexes collinearity, since as approaches 1 the matrix becomes singular. We can see the of collinearity on precision by observing that the variance of a coecient esti eect 1 approaches innity as approaches 1. Thus the more collinear are the mate 2 n 1 2 regressors, the worse the precision of the individual coecient estimates. What is happening is that when the regressors are highly dependent, it is statistically dicult to disentangle the impact of 1 from that of 2 . As a consequence, the precision of individual estimates are reduced. The imprecision, however, will be reected by large standard errors, so there is no distortion in inference. Some earlier textbooks overemphasized a concern about multicollinearity. A very amusing parody of these texts appeared in Chapter 23.3 of Goldbergers A Course in Econometrics (1991), which is reprinted below. 1 To understand his basic point, you should notice how the estimation 2 2 depends equally and symmetrically on the correlation and the sample variance n 1 size n.
2 1 1 2 1 b . = var | X = 1 n n (1 2 ) 1
106
Arthur S. Goldberger
Art Goldberger (1930-2009) was one of the most distinguished members of the Department of Economics at the University of Wisconsin. His PhD thesis developed an early macroeconometric forecasting model (known as the Klein-Goldberger model) but most of his career focused on microeconometric issues. He was the leading pioneer of what has been called the Wisconsin Tradition of empirical work a combination of formal econometric theory with a careful critical analysis of empirical work. Goldberger wrote a series of highly regarded and inuential graduate econometric textbooks, including including Econometric Theory (1964), Topics in Regression Analysis (1968), and A Course in Econometrics (1991).
CHAPTER 4. LEAST SQUARES REGRESSION Micronumerosity Arthur S. Goldberger A Course in Econometrics (1991), Chapter 23.3 Econometrics texts devote many pages to the problem of multicollinearity in multiple regression, but they say little about the closely analogous problem of small sample size in estimating a univariate mean. Perhaps that imbalance is attributable to the lack of an exotic polysyllabic name for small sample size. If so, we can remove that impediment by introducing the term micronumerosity. Suppose an econometrician set out to write a chapter about small sample size in sampling from a univariate population. Judging from what is now written about multicollinearity, the chapter might look like this: 1. Micronumerosity The extreme case, exact micronumerosity, arises when n = 0, in which case the sample estimate of is not unique. (Technically, there is a violation of the rank condition n > 0 : the matrix 0 is singular.) The extreme case is easy enough to recognize. Near micronumerosity is more subtle, and yet very serious. It arises when the rank condition n > 0 is barely satised. Near micronumerosity is very prevalent in empirical economics. 2. Consequences of micronumerosity The consequences of micronumerosity are serious. Precision of estimation is reduced. There are two aspects of this reduction: estimates of may have large errors, and not only that, but Vy will be large. Investigators will sometimes be led to accept the hypothesis = 0 because y / y is small, even though the true situation may be not that = 0 but simply that the sample data have not enabled us to pick up. The estimate of will be very sensitive to sample data, and the addition of a few more observations can sometimes produce drastic shifts in the sample mean. The true may be suciently large for the null hypothesis = 0 to be 2 rejected, even though Vy = /n is large because of micronumerosity. But if the true is small (although nonzero) the hypothesis = 0 may mistakenly be accepted.
107
108
3. Testing for micronumerosity Tests for the presence of micronumerosity require the judicious use of various ngers. Some researchers prefer a single nger, others use their toes, still others let their thumbs rule. A generally reliable guide may be obtained by counting the number of observations. Most of the time in econometric analysis, when n is close to zero, it is also far from innity. Several test procedures develop critical values n , such that micronumerosity is a problem only if n is smaller than n . But those procedures are questionable. 4. Remedies for micronumerosity If micronumerosity proves serious in the sense that the estimate of has an unsatisfactorily low degree of precision, we are in the statistical position of not being able to make bricks without straw. The remedy lies essentially in the acquisition, if possible, of larger samples from the same population. But more data are no remedy for micronumerosity if the additional data are simply more of the same. So obtaining lots of small samples from the same population will not help.
4.17
Normal Regression Model
In the special case of the normal linear regression model introduced in Section 3.18, we can derive exact sampling distributions for the least-squares estimator, residuals, and variance estimator. In particular, under the normality assumption ei | xi N 0, 2 then we have the multivariate implication e | X N 0, I n 2 .
That is, the error vector e is independent of X and is normally distributed. Since linear functions of normals are also normal, this implies that conditional on X 2 1 1 b 0 (X 0 X ) (X 0 X ) X 0 e N 0, = M e 0 2 M
1
where M = I n X (X 0 X ) X 0 . Since uncorrelated normal variables are independent, it follows b is independent of any function of the OLS residuals including the estimated error variance that 2 2 or prediction errors e . s or The spectral decomposition (see equation (A.5)) of M yields I nk 0 H0 M =H 0 0
CHAPTER 4. LEAST SQUARES REGRESSION where H 0 H = I n . Let u = 1 H 0 e N (0, H 0 H ) N (0, I n ) . Then n 2 (n k ) s2 = 2 2 1 0 = 2e e 1 = 2 e0 M e 1 I nk 0 = 2 e0 H H 0e 0 0 I nk 0 0 =u u 0 0 2 nk ,
109
a t distribution with n k degrees of freedom.
a chi-square distribution with n k degrees of freedom. Furthermore, if standard errors are calculated using the homoskedastic formula (4.27) h i 1 0 2 N 0, (X X ) bj j bj j N (0, 1) jj rh = q 2 tnk = rh q i i j ) nk 1 1 s( 0 0 2 2 (X X ) s (X X ) nk nk nk
jj jj
Theorem 4.17.1 Normal Regression In the linear regression (Assumption 4.3.1) if ei is independent of model xi and distributed N 0, 2 then b N 0, 2 (X 0 X )1
n 2 2
(nk)s2 2
2 nk
j j j ) s(
tnk
These are the exact nite-sample distributions of the least-squares estimator and variance estimators, and are the basis for traditional inference in linear regression. While elegant, the diculty in applying Theorem 4.17.1 is that the normality assumption is too restrictive to be empirical plausible, and therefore inference based on Theorem 4.17.1 has no guarantee of accuracy. We develop an alternative inference theory based on large sample (asymptotic) approximations in the following chapter.
William Gosset
William S. Gosset (1876-1937) of England is most famous for his derivation of the students t distribution, published in the paper The probable error of a mean in 1908. At the time, Gosset worked at Guiness Brewery, which prohibited its employees from publishing in order to prevent the possible loss of trade secrets. To circumvent this barrier, Gosset published under the pseudonym Student. Consequently, this famous distribution is known as the students t rather than Gossets t!
110
Exercises
Exercise 4.1 Explain the dierence between
1 n
Exercise 4.2 True or False. If yi =P xi + ei , xi R, E(ei | xi ) = 0, and e i is the OLS residual n 2 i = 0. from the regression of yi on xi , then i=1 xi e Exercise 4.3 Prove Theorem 4.6.1.2. Exercise 4.4 In a linear model y = X + e, E(e | X ) = 0, var (e | X ) = 2
Pn
0 i=1 xi xi
and E (xi x0 i) .
with a known function of X , the GLS estimator is e = X 0 1 X 1 X 0 1 y , s2 = 1 e 0 1 e . nk
e , and an estimate of 2 is the residual vector is e = y X e|X . (a) Find E
1 0 1 X . (c) Prove that e = M 1 e, where M 1 = I X X 0 1 X 1 0 1 1 1 (d) Prove that M 0 1 X X 0 1 X X . 1 M 1 = 2 (e) Find E s | X . (f) Is s2 a reasonable estimator for 2 ? Exercise 4.5 Let (yi , xi ) be a random sample with E(y | X ) = X . Consider the Weighted Least Squares (WLS) estimator of e = X 0 W X 1 X 0 W y
e|X . (b) Find var
2 where W = diag (w1 , ..., wn ) and wi = x ji , where xji is one of the xi .
e would perform better than (b) Using your intuition, in which situations would you expect that OLS? Exercise 4.6 Show (4.24) in the homoskedastic regression model. Exercise 4.7 Prove (4.32). Exercise 4.8 Show (4.33) and (4.34) in the homoskedastic regression model. Exercise 4.9 Let = E (yi ) , 2 = E (yi )2 and 3 = E (yi )3 and consider the sample mean P n 3 1 2 y=n i=1 yi . Find E (y ) as a function of , , 3 and n.
e be a good estimator? (a) In which contexts would
111
| xi ) = 0. De ne Exercise 4.10 Take the simple regression model yi = xi + ei , xi R, E(ei 3 2 = E(e | x ) and = E(e3 | x ) and consider the OLS coecient . b Find E b | X . i i i 3i i i Exercise 4.11 Continue the empirical analysis in Exercise 3.19. 1. Calculate standard errors using the homoskedasticity formula and using the four covariance matrices from Section 4.11. 2. Repeat in your second programming language. Are they identical? Exercise 4.12 Continue the empirical analysis in Exercise 3.21. Calculate standard errors using the Horn-Horn-Duncan method. Repeat in your second programming language. Are they identical?
Chapter 5
An Introduction to Large Sample Asymptotics

5.1 Introduction
In Chapter 4 we derived the mean and variance of the least-squares estimator in the context of the linear regression model, but this is not a complete description of the sampling distribution, nor sucient for inference (condence intervals and hypothesis testing) on the unknown parameters. Furthermore, the theory does not apply in the context of the linear projection model, which is more relevant for empirical applications. To illustrate the situation with an example, let yi and xi be drawn from the joint density 1 1 1 2 2 exp (log y log x) exp (log x) f (x, y ) = 2xy 2 2 be the slope coecient estimate from a least-squares regression of yi on xi and a constant. and let was computed and plotted in Figure 5.1 for Using simulation methods, the density function of sample sizes of n = 25, n = 100 and n = 800. The vertical line marks the true projection coecient. From the gure we can see that the density functions are dispersed and highly non-normal. As the sample size increases the density becomes more concentrated about the population coecient. ? Is there a simple way to characterize the sampling distribution of In principle the sampling distribution of is a function of the joint distribution of (yi , xi ) and the sample size n, but in practice this function is extremely complicated so it is not feasible to except in very special cases. Therefore we typically analytically calculate the exact distribution of rely on approximation methods. The most widely used and versatile method is asymptotic theory, which approximates sampling distributions by taking the limit of the nite sample distribution as the sample size n tends to innity. It is important to understand that this is an approximation technique, as the asymptotic distributions are used to assess the nite sample distributions of our estimators in actual practical samples. The primary tools of asymptotic theory are the weak law of large numbers (WLLN), central limit theorem (CLT), and continuous mapping theorem (CMT). With these tools we can approximate the sampling distributions of most econometric estimators. In this chapter we provide a concise summary. It will be useful for most students to review this material, even if most is familiar.
5.2
Asymptotic Limits
Asymptotic analysis is a method of approximation obtained by taking a suitable limit. There is more than one method to take limits, but the most common is to take the limit of the sequence 112
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS
113
Figure 5.1: Sampling Density of of sampling distributions as the sample size tends to positive innity, written as n . It is not meant to be interpreted literally, but rather as an approximating device. The rst building block for asymptotic analysis is the concept of a limit of a sequence.
Denition 5.2.1 A sequence an has the limit a, written an a as n , or alternatively as limn an = a, if for all > 0 there is some n < such that for all n n , |an a| .
In words, an has the limit a if the sequence gets closer and closer to a as n gets larger. If a sequence has a limit, that limit is unique (a sequence cannot have two distinct limits). If an has the limit a, we also say that an converges to a as n . Not all sequences have limits. For example, the sequence {1, 2, 1, 2, 1, 2, ...} does not have a limit. It is therefore sometimes useful to have a more general denition of limits which always exist, and these are the limit superior and limit inferior of sequence
Denition 5.2.2 lim inf n an = limn inf mn an Denition 5.2.3 lim supn an = limn supmn an
def
def
The limit inferior and limit superior always exist, and equal when the limit exists. In the example given earlier, the limit inferior of {1, 2, 1, 2, 1, 2, ...} is 1, and the limit superior is 2.
114
5.3
Convergence in Probability
A sequence of numbers may converge to a limit, Pn but what about a sequence of random variables? 1 For example, consider a sample mean y = n i=1 yi based on an random sample of n observations. As n increases, the distribution of y changes. In what sense can we describe the limit of y ? In what sense does it converge? Since y is a random variable, we cannot directly apply the deterministic concept of a sequence of numbers. Instead, we require a denition of convergence which is appropriate for random variables. There are more than one such denition, but the most commonly used is called convergence in probability.
Denition 5.3.1 A random variable zn R converges in probability p to z as n , denoted zn z, or alternatively plimn zn = z , if for all > 0, (5.1) lim Pr (|zn z | ) = 1.
n
We call z the probability limit (or plim) of zn .
The denition looks quite abstract, but it formalizes the concept of a sequence of random variables concentrating about a point. The event {|zn z | } occurs when zn is within of the point z. Pr (|zn z | ) is the probability of this event that zn is within of the point z . Equation (5.1) states that this probability approaches 1 as the sample size n increases. The denition of convergence in probability requires that this holds for any . So for any small interval about z the distribution of zn concentrates within this interval for large n. You may notice that the denition concerns the distribution of the random variables zn , not their realizations. Furthermore, notice that the denition uses the concept of a conventional (deterministic) limit, but the latter is applied to a sequence of probabilities, not directly to the random variables zn or their realizations. Two comments about the notation are worth mentioning. First, it is conventional to write the p convergence symbol as where the p above the arrow indicates that the convergence is in probability. You should try and adhere to this notation, and not simply write zn z . Second, it is important to include the phrase as n to be specic about how the limit is obtained. A common mistake to confuse convergence in probability with convergence in expectation: Ezn Ez. (5.2)
They are related but distinct concepts. Neither (5.1) nor (5.2) implies the other. To see the distinction it might be helpful to think through a stylized example. Consider a discrete random variable zn which takes the value 0 with probability 1 n1 and the value an 6= 0 with probability n1 , or Pr (zn = 0) = 1 Pr (zn = an ) = 1 . n 1 n (5.3)
In this example the probability distribution of zn concentrates at zero as n increases, regardless of p the sequence an . You can check that zn 0 as n . In this example we can also calculate that the expectation of zn is Ezn = an . n
115
Despite the fact that zn converges in probability to zero, its expectation will not decrease to zero unless an /n 0. If an diverges to innity at a rate equal to n (or faster) then Ezn will not converge p to zero. For example, if an = n, then Ezn = 1 for all n, even though zn 0. This example might seem a bit articial, but the point is that the concepts of convergence in probability and convergence in expectation are distinct, so it is important not to confuse one with the other. Another common source of confusion with the notation surrounding probability limits is that p the expression to the right of the arrow must be free of dependence on the sample size n. p Thus expressions of the form zn cn are notationally meaningless and should not be used.
5.4
Weak Law of Large Numbers
In large samples we expect parameter estimates to be close to the population values. For example, in Section 4.2 we saw that the sample mean y is unbiased for = Ey and has variance 2 /n. As n gets large its variance decreases and thus the distribution of y concentrates about the population mean . It turns out that this implies that the sample mean converges in probability to the population mean. When y has a nite variance there is a fairly straightforward proof by applying Chebyshevs inequality. Theorem 5.4.1 Chebyshevs Inequality. For any random variable zn and constant > 0 Pr (|zn Ezn | > ) var(zn ) . 2
Chebyshevs inequality is terrically important in asymptotic theory. While its proof is a technical exercise in probability theory, it is quite simple so we discuss it forthwith. Let Fn (u) denote the distribution of zn Ezn . Then Z 2 2 Pr (|zn Ezn | > ) = Pr (zn Ezn ) > = dFn (u).
{u2 > 2 }
u2 The integral is over the event u2 > 2 , so that the inequality 1 2 holds throughout. Thus Z Z Z u2 u2 E (zn Ezn )2 var(zn ) dFn (u) dF ( u ) dF ( u ) = = , n n 2 2 2 2 {u2 > 2 } {u2 > 2 }
which establishes the desired inequality. Applied to the sample mean y, Chebyshevs inequality shows that for any > 0 Pr (|y Ey | > ) 2 . n 2
For xed 2 and , the bound on the right-hand-side shrinks to zero as n . Thus the probability that y is within of Ey = approaches 1 as n gets large, or
n
lim Pr (|y | ) = 1.
This means that y converges in probability to as n . This result is called the weak law of large numbers. Our derivation assumed that y has a nite variance, but with a more careful proof all that is necessary is a nite mean.
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS Theorem 5.4.2 Weak Law of Large Numbers (WLLN) If yi are independent and identically distributed and E |y | < , then as n , n 1X p y= yi E(y ). n
i=1
116
The proof of Theorem 5.4.2 is presented in Section 5.14. The WLLN shows that the estimator y converges in probability to the true population mean . In general, an estimator which converges in probability to the population value is called consistent.
of a parameter is consistent if Denition 5.4.1 An estimator as n .
Theorem 5.4.3 If yi are independent and identically distributed and E |y | < , then b = y is consistent for the population mean . Consistency is a good property for an estimator to possess. It means that for any given data will be arbitrarily distribution, there is a sample size n suciently large such that the estimator close to the true value with high probability. The theorem does not tell us, however, how large this n has to be. Thus the theorem does not give practical guidance for empirical practice. Still, it is a minimal property for an estimator to be considered a good estimator, and provides a foundation for more useful approximations.
5.5
Almost Sure Convergence and the Strong Law*
Convergence in probability is sometimes called weak convergence. A related concept is almost sure convergence, also known as strong convergence. (In probability theory the term almost sure means with probability equal to one. An event which is random but occurs with probability equal to one is said to be almost sure.)
Denition 5.5.1 A random variable zn R converges almost surely a.s. to z as n , denoted zn z, if for every > 0 Pr lim |zn z | = 1. (5.4)
n
The convergence (5.4) is stronger than (5.1) because it computes the probability of a limit rather than the limit of a probability. Almost sure convergence is stronger than convergence in p a.s. probability in the sense that zn z implies zn z .
117
In the example (5.3) of Section 5.3, the sequence zn converges in probability to zero for any sequence an , but this is not sucient for zn to converge almost surely. In order for zn to converge to zero almost surely, it is necessary that an 0. In the random sampling context the sample mean can be shown to converge almost surely to the population mean. This is called the strong law of large numbers.
Theorem 5.5.1 Strong Law of Large Numbers (SLLN) If yi are independent and identically distributed and E |y | < , then as n , n 1 X a.s. y= yi E(y ). n
i=1
The proof of the SLLN is technically quite advanced so is not presented here. For a proof see Billingsley (1995, Section 22) or Ash (1972, Theorem 7.2.5). The WLLN is sucient for most purposes in econometrics, so we will not use the SLLN in this text.
5.6
Vector-Valued Moments
Our preceding discussion focused on the case where y is real-valued (a scalar), but nothing important changes if we generalize to the case where y Rm is a vector. To x notation, the elements of y are y1 y2 y = . . . . ym The population mean of y is just the vector of marginal means E (y1 ) E (y2 ) = E(y ) = . . . . E (ym )
When working with random vectors y it is convenient to measure their magnitude by their Euclidean length or Euclidean norm 2 2 1/2 ky k = y1 + + ym . ky k2 = y 0 y .
In vector notation we have
It turns out that it is equivalent to describe niteness of moments in terms of the Euclidean norm of a vector or all individual components.
Theorem 5.6.1 For y Rm , E ky k < if and only if E |yj | < for j = 1, ..., m.
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS The m m variance matrix of y is
118
V is often called a variance-covariance matrix. You can show that the elements of V are nite if E ky k2 < . A random sample {y 1 , ..., y n } consists of n observations of independent and identically distributed draws from the distribution of y . (Each draw is an m-vector.) The vector sample mean y1 n y 1X 2 y= yi = . . n . i=1 ym
V = var (y ) = E (y ) (y )0 .
is the vector of sample means of the individual variables. Convergence in probability of a vector can be dened as convergence in probability of all elep p ments in the vector. Thus y if and only if yj j for j = 1, ..., m. Since the latter holds if E |yj | < for j = 1, ..., m, or equivalently E ky k < , we can state this formally as follows. Theorem 5.6.2 Weak Law of Large Numbers (WLLN) for random vectors If y i are independent and identically distributed and E ky k < , then as n , n 1X p y= y i E(y ). n
i=1
5.7
Convergence in Distribution
The WLLN is a useful rst step, but does not give an approximation to the distribution of an estimator. A large-sample or asymptotic approximation can be obtained using the concept of convergence in distribution.
denoted z n z , if for all u at which F (u) = Pr (z u) is continuous, Fn (u) F (u) as n .
Denition 5.7.1 Let z n be a random vector with distribution Fn (u) = Pr (z n u) . We say that z n converges in distribution to z as n ,
d
When z n z , it is common to refer to z as the asymptotic distribution or limit distribution of z n . When the limit distribution z is degenerate (that is, Pr (z = c) = 1 for some c) we can write the convergence as z n c, which is equivalent to convergence in probability, z n c. The typical path to establishing convergence in distribution is through the central limit theorem (CLT), which states that a standardized sample average converges in distribution to a normal random vector.
d p
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS Theorem 5.7.1 LindebergLvy Central Limit Theorem (CLT). If y i are independent and identically distributed and E ky k2 < , then as n n 1 X d n (y ) = (y i ) N (0, V ) n i=1 where = Ey and V = E (y ) (y )0 .
119
The standardized sum z n = n (y n ) has mean zero and variance V . What the CLT adds is that the variable z n is also approximately normally distributed, and that the normal approximation improves as n increases. The CLT is one of the most powerful and mysterious results in statistical theory. It shows that the simple process of averaging induces normality. The rst version of the CLT (for the number of heads resulting from many tosses of a fair coin) was established by the French mathematician Abraham de Moivre in an article published in 1733. This was extended to cover an approximation to the binomial distribution in 1812 by Pierre-Simon Laplace in his book Thorie Analytique des Probabilits, and the most general statements are credited to articles by the Russian mathematician Aleksandr Lyapunov (1901) and the Finnish mathematician Jarl Waldemar Lindeberg (1920, 1922). The above statement is known as the classic (or Lindeberg-Lvy) CLT due to contributions by Lindeberg (1920) and the French mathematician Paul Pierre Lvy. A more general version which does not require the restriction to identical distributions was provided by Lindeberg (1922).
Theorem 5.7.2 Lindeberg Central Limit Theorem (CLT). Suppose nite that yi are independent but not necessarily identically distributed Pn with 2 2 2 2 means i = Eyi and variances i = E (yi i ) . Set n = i=1 i . If for all > 0 n 1 X E (yi i )2 1 (|yi i | n ) = 0 (5.5) lim 2 n n
i=1
then
n 1 X d (yi i ) N (0, 1) . n i=1
Equation (5.5) is known as Lindebergs condition. A standard method to verify (5.5) is via Lyapunovs condition: For some > 0 lim
n 1 X i=1
2+ n
E (yi i )2+ = 0.
(5.6)
It is easy to verify that (5.6) implies (5.5), and (5.6) is often easy to verify. For example, if 2 c > 0 then supi E (yi i )3 < and inf i i
n 1 X n E (yi i )3 0 3 n (nc)3/2 i=1
so (5.6) is satised.
120
5.8
Higher Moments
Often we want to estimate a parameter which is the expected value of a transformation of a random vector y . That is, can be written as = Eh (y ) for some function h : Rm Rk . For example, the second moment of y is Ey2 , the kth is Eyk , the moment generating function is E exp (ty) , and the distribution function is E1 {y x} . Estimating parameters of this form ts into our previous analysis by dening the random variable z = h (y ) for then = Ez is just a simple moment of z . This suggests the moment estimator n n 1X 1X b= zi = h (y i ) . n n i=1 i=1 P k generating function For example, the moment estimator of Eyk is n1 n i=1 yi , that of the moment P Pn n 1 1 exp ( ty ) , and for the distribution function the estimator is n is n i i=1 i=1 1 {yi x} b is a sample average, and transformations of iid variables are also iid, the asymptotic Since results of the previous sections immediately apply. Theorem 5.8.1 If y i are independent and identically distributed, = 1 Pn b = n i=1 h (y i ) , as n , Eh (y ) , and E kh (y )k < , then for p b .
d b ) N (0, V ) n ( where V = E (h (y ) ) (h (y ) )0 .
Theorem 5.8.2 If y i are independent and P identically distributed, = 2 1 b=n n Eh (y ) , and E kh (y )k < , then for i=1 h (y i ) , as n ,
b is consistent for and asymptotically Theorems 5.8.1 and 5.8.2 show that the estimate normally distributed, so long as the stated moment conditions hold. A word of caution. Theorems 5.8.1 and 5.8.2 give the impression that it is possible to estimate any moment of y. Technically this is the case so long as that moment is nite. What is hidden by the notation, however, is that estimates of highP order momnets can be quite imprecise. For 1 th 8 b8 = n n example, consider the sample 8 moment i=1 yi , and suppose for simplicity that y is 1 1 b8 ) = n 645, 015, which is huge, even for large n! In N(0, 1). Then we can calculate that var ( general, higher-order moments are challenging to estimate because their variance depends upon even higher moments which can be quite large in some cases.
1
2, 027, 025 and Ey 8 = 7!! = 105 where k!! = k(k 2) is the double factorial for odd k.
2 . Since y is N(0, 1), Ey 16 = 15!! = By the formula for the variance of a mean var ( 8 ) = n1 Ey 16 Ey 8
121
5.9
Functions of Moments
We now expand our investigation and consider estimation of parameters which can be written as a continuous function of = Eh (y ). That is, the parameter of interest can be written as = g () = g (Eh (y )) for some functions g : Rk R and h : Rm Rk . As one example, the geometric mean of wages w is = exp (E (log (w))) . This is (5.7) with g (u) = exp (u) and h(w) = log(w). A simple yet common example is the variance 2 = E (w Ew)2 This is (5.7) with h(w) = and g (1 , 2 ) = 2 2 1. Similarly, the skewness of the wage distribution is E (w Ew)3 sk = 3/2 . E (w Ew)2 w h(w) = w2 w3 g (1 , 2 , 3 ) = 3 32 1 + 23 1 3/2 . 2 2 1 (5.9) (5.8) (5.7)
= Ew2 (Ew)2 . w w2
This is (5.7) with
and
The parameter = g () is not a population moment, so it does not have a direct moment estimator. Instead, it is common to use a plug-in estimate formed by replacing the unknown b and then plugging this into the expression for . The rst step is with its point estimate and the second step is 1X b= h (y i ) n
i=1 n
b is a sample estimate of . Again, the hat ^ indicates that For example, the plug-in estimate of the geometric mean of the wage distribution from (5.8) is b = exp( b) with 1X log (wagei ) . b= n
i=1 n
b = g ( b) .
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS The plug-in estimate of the variance is 1X 2 = wi n
2 i=1 i=1 n
122
and that for the skewness is
n 1X = (wi w)2 . n
1X wi n
i=1
!2
where
b2 b1 + 2 b3 b3 3 1 c= sk 3/2 2 b1 b2 3 1 Pn i=1 (wi w) n = 3/2 2 1 Pn ( w w ) i i=1 n bj = 1X j wi . n

i=1 n
A useful property is that continuous functions are limit-preserving. Theorem 5.9.1 Continuous Mapping Theorem (CMT). If z n c p as n and g () is continuous at c, then g (z n ) g (c) as n . The proof of Theorem 5.9.1 is given in Section 5.14. p For example, if zn c as n then zn + a c + a azn ac as the functions g (u) = u + a, g (u) = au, and g (u) = u2 are continuous. Also a p a zn c if c 6= 0. The condition c 6= 0 is important as the function g (u) = a/u is not continuous at u = 0. If y i are independent and identically distributed, = Eh (y ) , and E kh (y )k < , then for p p 1 Pn b = g ( b=n b . Applying the CMT, b ) g () = . i=1 h (y i ) , as n , Theorem 5.9.2 If y i are independent and identically distributed, = g (Eh (y )) , E kh (y )k < , and g (u) is continuous at u = , then for p b b = g 1 Pn h (y ) , as n , . i i=1 n
2 zn c2 p p p p
To apply Theorem 5.9.2 it is necessary to check if the function g is continuous at . In our rst example g (u) = exp (u) is continuous everywhere. It therefore follows from Theorem 5.6.2 and p Theorem 5.9.2 that if E |log (wage)| < then as n , b . In the example of the variance, g is continuous for all . Thus if Ew2 < then as n , p 2 b 2 . In our third example g dened in (5.9) is continuous for all such that var(w) = 2 2 1 > 0, which holds unless w has a degenerate distribution. Thus if E |w|3 < and var(w) > 0 then as p c n , sk sk.
123
5.10
Delta Method
In this section we introduce two tools an extended version of the CMT and the Delta Method b. which allow us to calculate the asymptotic distribution of the parameter estimate We rst present an extended version of the continuous mapping theorem which allows convergence in distribution. Theorem 5.10.1 Continuous Mapping Theorem d If z n z as n and g : Rm Rk has the set of discontinuity points d Dg such that Pr (z Dg ) = 0, then g(z n ) g (z ) as n .
For a proof of Theorem 5.10.1 see Theorem 2.3 of van der Vaart (1998). It was rst proved by Mann and Wald (1943) and is therefore sometimes referred to as the Mann-Wald Theorem. Theorem 5.10.1 allows the function g to be discontinuous only if the probability at being at a discontinuity point is zero. For example, the function g (u) = u1 is discontinuous at u = 0, but if
1 z 1 . zn z N (0, 1) then Pr (z = 0) = 0 so zn A special case of the Continuous Mapping Theorem is known as Slutskys Theorem. d d
Theorem 5.10.2 Slutskys Theorem p d If zn z and cn c as n , then 1. zn + cn z + c 2. zn cn zc 3.

d d
zn d z if c 6= 0 cn c
Even though Slutskys Theorem is a special case of the CMT, it is a useful statement as it focuses on the most common applications addition, multiplication, and division. b is a function of b for which we have an asymptotic Despite the fact that the plug-in estimator b . This is distribution, Theorem 5.10.1 does not directly give us an asymptotic distribution for b = g ( b ) . b ) is written as a function of b , not of the standardized sequence n ( because We need an intermediate step a rst order Taylor series expansion. This step is so critical to statistical theory that it has its own name The Delta Method.
Theorem 5.10.3 Delta Method: d b ) , where g (u) is continuously dierentiable in a neighIf n ( borhood of then as n where G(u) = as n
0 u g (u)
d b ) g ()) N 0, G0 V G . n (g (
and G = G(). In particular, if N (0, V ) then (5.11)
d b ) g ()) G0 n (g (
(5.10)
124
The Delta Method allows us to complete our derivation of the asymptotic distribution of the b of . estimator By combining Theorems 5.8.2 and 5.10.3 we can nd the asymptotic distribution of the plug-in b. estimator Theorem 5.10.4 If y i are independent and identically distributed, = g (u)0 is continuous Eh (y ), = g () , E kh (y )k2 < , and G (u) = u 1 Pn b=g in a neighborhood of , then for i=1 h (y i ) , as n n where V = E (h (y ) ) (h (y ) )0 and G = G () . b for , and Theorem 5.10.4 established its asympTheorem 5.9.2 established the consistency of totic normality. It is instructive to compare the conditions required for these results. Consistency required that h (y ) have a nite mean, while asymptotic normality requires that this variable have a nite variance. Consistency required that g (u) be continuous, while asymptotic normality required that g (u) be continuously dierentiable. d b n N 0, G0 V G
5.11
Stochastic Order Symbols
It is convenient to have simple symbols for random variables and vectors which converge in probability to zero or are stochastically bounded. In this section we introduce some of the most commonly found notation. It might be useful to review the common notation for non-random convergence and boundedness. Let xn and an , n = 1, 2, ..., be a non-random sequences. The notation xn = o(1) (pronounced small oh-one) is equivalent to xn 0 as n . The notation xn = o(an )
1 is equivalent to a n xn 0 as n . The notation
xn = O(1) (pronounced big oh-one) means that xn is bounded uniformly in n : there exists an M < such that |xn | M for all n. The notation xn = O(an )
1 is equivalent to a n xn = O (1). We now introduce similar concepts for sequences of random variables. Let zn and an , n = 1, 2, ... be sequences of random variables. (In most applications, an is non-random.) The notation
zn = op (1) (small oh-P-one) means that zn 0 as n . We also write zn = op (an )

p

1 b if a n zn = op (1). For example, for any consistent estimator for we can write
125
Similarly, the notation zn = Op (1) (big oh-P-one) means that zn is bounded in probability. Precisely, for any > 0 there is a constant M < such that lim sup Pr (|zn | > M ) .
n
b = + op (1).
Furthermore, we write zn = Op (an )

1 if a n zn = Op (1). Op (1) is weaker than op (1) in the sense that zn = op (1) implies zn = Op (1) but not the reverse. However, if zn = Op (an ) then zn = op (bn ) for any bn such that an /bn 0. d
If a random vector converges in distribution z n z (for example, if z N (0, V )) then b which satisfy the convergence of Theorem 5.10.4 then z n = Op (1). It follows that for estimators we can write b = + Op (n1/2 ).
b equals the true coecient plus a random In words, this statement says that the estimator component which is shrinking to zero at the rate n1/2 . Another useful observation is that a random sequence with a bounded moment is stochastically bounded. Theorem 5.11.1 If z n is a random vector which satises E kz n k = O (an ) for some sequence an and > 0, then
/ z n = Op (a1 n ).
Similarly, E kz n k = o (an ) implies z n = op (an ).
1/
This can be shown using Markovs inequality (B.21). The assumptions imply that there is some 1/ M . Then M < such that E kz n k M an for all n. For any set B = M an 1/ Pr an kz n k > B = Pr kz n k > E kz n k M an
as required. There are many simple rules for manipulating op (1) and Op (1) sequences which can be deduced from the continuous mapping theorem or Slutskys Theorem. For example, op (1) + op (1) = op (1) op (1) + Op (1) = Op (1) Op (1) + Op (1) = Op (1) op (1)op (1) = op (1) op (1)Op (1) = op (1) Op (1)Op (1) = Op (1)
126
5.12
Uniform Stochastic Bounds*
For some applications it can be useful to obtain the stochastic order of the random variable
1in
max |yi | .
This is the magnitude of the largest observation in the sample {y1 , ..., yn }. If the support of the distribution of yi is unbounded, then as the sample size n increases, the largest observation will also tend to increase. It turns out that there is a simple characterization.
Theorem 5.12.1 Assume y i are independent and identically distributed. If E |y |r < , then as n n1/r max |yi | 0.
1in p
(5.12)
If E exp(ty ) < for all t < , then (log n)1 max |yi | 0.
1in p
(5.13)
The proof of Theorem 5.12.1 is presented in Section 5.14. Equivalently, (5.12) can be written as
1in
max |yi | = op (n1/r )
(5.14)
and (5.13) as
1in
max |yi | = op (log n).
(5.15)
Equation (5.12) says that if y has r nite moments, then the largest observation will diverge at a rate slower than n1/r . As r increases this rate decreases. Equation (5.13) shows that if we strengthen this to y having all nite moments and a nite moment generating function (for example, if y is normally distributed) then the largest observation will diverge slower than log n. Thus the higher the moments, the slower the rate of divergence. To simplify the notation, we write (5.14) as yi = op (n1/r ) uniformly in 1 i n, and similarly (5.15) as yi = op (log n), uniformly in 1 i n. It is important to understand when the Op or op symbols are applied to subscript i random variables whether the convergence is pointwise in i, or is uniform in i in the sense of (5.14)-(5.15). Theorem 5.12.1 applies to random vectors. If E ky kr < then
1in
max ky i k = op (n1/r ),
and if E exp(t0 y ) < for all ktk < then (log n)1 max ky i k 0.
1in p
(5.16)
127
5.13
Semiparametric Eciency
b = g ( b and plug-in estimator b ) are ecient In this section we argue that the sample mean estimators of the parameters and . Our demonstration is based on the rich but technically challenging theory of semiparametric eciency bounds. An excellent accessible review has been provided by Newey (1990). We will also appeal to the asymptotic theory of maximum likelihood estimation (see Section B.11). b will follow from b , for the asymptotic eciency of We start by examining the sample mean b. that of Recall, we know that if E ky k2 < then the sample mean has the asymptotic distribution d b ) N (0, V ) . We want to know if b is the best feasible estimator, or if there is another n ( estimator with a smaller asymptotic variance. While it seems intuitively unlikely that another estimator could have a smaller asymptotic variance, how do we know that this is not the case? b is the best estimator, we need to be clear about the class of models the class When we ask if of permissible distributions. For estimation of the mean of the distribution of y the broadest purposes, as conceivable class is L1 = {F : E ky k < } . This class is too broad n for our current n o b is not asymptotically N (0, V ) for all F L1 . A more realistic choice is L2 = F : E ky k2 < the class of nite-variance distributions. When we seek an ecient estimator of the mean in the class of models L2 what we are seeking is the best estimator, given that all we know is that F L2 . To show that the answer is not immediately obvious, it might be helpful to review a setting where the sample mean cient. is ine Suppose that y R has the double exponential den 1 / 2 exp |y | 2 . Since var (y ) = 1 we see that the sample mean satsity f (y | ) = 2 d ) N (0, 1). In this model the maximum likelihood estimator (MLE) for ises n ( is the sample median. Recall from the theory of maximum likelhood that the MLE satises 1 d 2 where S = n ( ) N 0, ES log f (y | ) = 2 sgn (y ) is the score. We can d calculate that ES 2 = 2 and thus conclude that n ( ) N (0, 1/2) . The asymptotic variance of the MLE is one-half that of the sample mean. Thus when the true density is known to be double exponential the sample mean is inecient. But the estimator which achieves this improved eciency the sample median is not generically consistent for the population mean. It is inconsistent if the density is asymmetric or skewed. So the improvement comes at a great cost. Another way of looking at this is that the sample median is ecient in the class of densities f (y | ) = 21/2 exp |y | 2 but unless it is known that this is the correct distribution class this knowledge is not very useful. The relevant question is whether or not the sample mean is ecient when the form of the distribution is unknown. We call this setting semiparametric as the parameter of interest (the mean) is nite dimensional while the remaining features of the distribution are unspecied. In the semiparametric context an estimator is called semiparametrically ecient if it has the smallest asymptotic variance among all semiparametric estimators. The mathematical trick is to reduce the semiparametric model to a set of parametric submodels. The Cramer-Rao variance bound can be found for each parametric submodel. The variance bound for the semiparametric model (the union of the submodels) is then dened as the supremum of the individual variance bounds. R Formally, suppose that the true density of y is the unknown function f (y ) with mean = Ey = y f (y )dy . A parametric submodel for f (y ) is a density f (y | ) which is a smooth function of a parameter , and there is a true value 0 such that f (y | 0 ) = f (y ). The index indicates the submodels. The equality f (y | 0 ) = f (y ) means that the submodel class passes through the true density, so the submodel is a true model. The class of submodels and R parameter 0 depend on the true density f. In the submodel f (y | ) , the mean is () = y f (y | ) dy which varies with the parameter . Let be the class of all submodels for f.
128
Since each submodel is parametric we can calculate the eciency bound for estimation of within this submodel. Specically, given the density f (y | ) its likelihood score is log f (y | 0 ) , 1 0 . Dening M = (0 )0 , so the Cramer-Rao lower bound for estimation of is ES S by Theorem B.11.5 the Cramer-Rao lower bound for estimation of within the submodel is 1 0 M . V = M0 ES S As V is the eciency bound for the submodel class f (y | ) , no estimator can have an asymptotic variance smaller than V for any density f (y | ) in the submodel class, including the true density f . This is true for all submodels . Thus the asymptotic variance of any semiparametric estimator cannot be smaller than V for any conceivable submodel. Taking the supremum of the Cramer-Rao bounds lower from all conceivable submodels we dene2 S = V = sup V .
The asymptotic variance of any semiparametric estimator cannot be smaller than V , since it cannot be smaller than any individual V . We call V the semiparametric asymptotic variance bound or semiparametric eciency bound for estimation of , as it is a lower bound on the asymptotic variance for any semiparametric estimator. If the asymptotic variance of a specic semiparametric estimator equals the bound V we say that the estimator is semiparametrically ecient. For many statistical problems it is quite challenging to calculate the semiparametric variance bound. However, in some cases there is a simple method to nd the solution. Suppose that we can nd a submodel 0 whose Cramer-Rao lower bound satises V 0 = V where V is the asymptotic variance of a known semiparametric estimator. In this case, we can deduce that V = V 0 = V . Otherwise there would exist another submodel 1 whose Cramer-Rao lower bound satises V 0 < V 1 but this would imply V < V 1 which contradicts the Cramer-Rao Theorem. b . Our goal is to nd a parametric submodel We now nd this submodel for the sample mean whose Cramer-Rao bound for is V . This can be done by creating a tilted version of the true density. Consider the parametric submodel (5.17) f (y | ) = f (y ) 1 + 0 V 1 (y ) where f (y ) is the true density and = Ey . Note that Z Z Z 0 1 f (y | ) dy = f (y )dy + V f (y ) (y ) dy = 1
and for all close to zero f (y | ) 0. Thus f (y | ) is a valid density function. It is a parametric submodel since f (y | 0 ) = f (y ) when 0 = 0. This parametric submodel has the mean Z () = y f (y | ) dy Z Z = y f (y )dy + f (y )y (y )0 V 1 dy =+ which is a smooth function of .
It is not obvious that this supremum exists, as V is a matrix so there is not a unique ordering of matrices. However, in many cases (including the ones we study) the supremum exists and is unique.
2
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS Since
129
it follows that the score function for is S =
V 1 (y ) log f (y | ) = log 1 + 0 V 1 (y ) = 1 + 0 V 1 (y ) log f (y | 0 ) = V 1 (y ) . (5.18)
By Theorem B.11.3 the Cramer-Rao lower bound for is 1 1 1 E S S0 = V E (y ) (y )0 V 1 =V. (5.19)
The Cramer-Rao lower bound for () = + is also V , and this equals the asymptotic variance b . This was what we set out to show. of the moment estimator In summary, we have shown that in the submodel (5.17) the Cramer-Rao lower bound for estimation of is V which equals the asymptotic variance of the sample mean. This establishes the following result.
Proposition 5.13.1 In the class of distributions F L2 , the semiparametric variance bound for estimation of is V = var(yi ), and the sample b is a semiparametrically ecient estimator of the population mean mean . We call this result a proposition rather than a theorem as we have not attended to the regularity conditions. b = g ( b ). We know from It is a simple matter to extend this result to the plug-in estimator 2 di erentiable at u = then the plugTheorem 5.10.4 that if E ky k < and g (u) is continuously b d 0 in estimator has the asymptotic distribution n N (0, G V G) . We therefore consider 2 < , Ey 2 < , and Ey 6= 0 . For example, if = 1 /2 where 1 = Ey1 and 2 = Ey2 then L2 (g ) = F : Ey1 2 2 For any submodel the Cramer-Rao lower bound for estimation of = g () is G0 V G by Theorem B.11.5. For the submodel (5.17) this bound is G0 V G which equals the asymptotic variance b from Theorem 5.10.4. Thus b is semiparametrically ecient. of Proposition 5.13.2 In the class of distributions F L2 (g ) the semiparametric variance bound for estimation of = g () is G0 V G, and the b = g ( b ) is a semiparametrically ecient estimator of plug-in estimator . The result in Proposition 5.13.2 is quite general. Smooth functions of sample moments are ecient estimators for their population counterparts. This is a very powerful result, as most econometric estimators can be written (or approximated) as smooth functions of sample means.
the class of distributions n o L2 (g) = F : E ky k2 < , g (u) is continuously dierentiable at u = Ey .
130
5.14
Technical Proofs*
In this section we provide proofs of some of the more technical points in the chapter. These proofs may only be of interest to more mathematically inclined. Proof of Theorem 5.4.2: Without loss of generality, we can assume E(yi ) = 0 by recentering yi on its expectation. We need to show that for all > 0 and > 0 there is some N < so that for all n N, Pr (|y | > ) . Fix and . Set = /3. Pick C < large enough so that E (|yi | 1 (|yi | > C )) (5.20)
(where 1 () is the indicator function) which is possible since E |yi | < . Dene the random variables wi = yi 1 (|yi | C ) E (yi 1 (|yi | C )) zi = yi 1 (|yi | > C ) E (yi 1 (|yi | > C )) y =w+z and E |y | E |w| + E |z | . (5.21) We now show that sum of the expectations on the right-hand-side can be bounded below 3. First, by the Triangle Inequality (A.12) and the Expectation Inequality (B.15), E |zi | = E |yi 1 (|yi | > C ) E (yi 1 (|yi | > C ))| 2E |yi 1 (|yi | > C )| so that
E |yi 1 (|yi | > C )| + |E (yi 1 (|yi | > C ))| (5.22)
2,
and thus by the Triangle Inequality (A.12) and (5.22) n n 1 X 1 X E |z | = E zi E |zi | 2. n n

i=1 i=1
(5.23)
Second, by a similar argument
|wi | = |yi 1 (|yi | C ) E (yi 1 (|yi | C ))| 2 |yi 1 (|yi | C )|
|yi 1 (|yi | C )| + |E (yi 1 (|yi | C ))| 2C (5.24)
where the nal inequality is (5.20). Then by Jensens Inequality (B.12), the fact that the wi are iid and mean zero, and (5.24), (E |w|)2 E |w|2 =
2 4C 2 Ewi = 2 n n
(5.25)
the nal inequality holding for n 4C 2 /2 = 36C 2 / 2 2 . Equations (5.21), (5.23) and (5.25) together show that (5.26) E |y | 32 as desired.
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS Finally, by Markovs Inequality (B.21) and (5.26),
131
3 E |y | = , the nal equality by the denition of . We have shown that for any > 0 and > 0 then for all n 36C 2 / 2 2 , Pr (|y | > ) , as needed. Pr (|y| > ) Proof of Theorem 5.6.1: By Loves cr Inequality (A.19) 1/2 m m X X 2 yj |yj | . ky k =
j =1 j =1
Thus if E |yj | < for j = 1, ..., m, then
E ky k
m X j =1
E |yj | < .
For the reverse inequality, the Euclidean norm of a vector is larger than the length of any individual component, so for any j, |yj | ky k . Thus, if E ky k < , then E |yj | < for j = 1, ..., m. Proof of Theorem 5.7.1: The moment bound Ey 0 i y i < is sucient to guarantee that the elements of and V are well dened and nite. Without loss of generality, it is sucient to consider the case = 0. Our proof method is to calculate the characteristic function of ny n and show that it converges pointwise to the characteristic function of N (0, V ) . By Lvys Continuity Theorem (see Van der Vaart (2008) Theorem 2.13) this is sucient to established that ny n converges in distribution to N (0, V ) . For Rm , let C () = E exp i0 y i denote the characteristic function of y i and set c () = log C (). Since y i has two nite moments the rst and second derivatives of C () are continuous in . They are C () = iE y i exp i0 y i 0 2 2 0 . 0 C () = i E y i y i exp i y i C (0) = 1 C (0) = iE (y i ) = 0 2 0 0 C (0) = E y i y i = V .
When evaluated at = 0
Furthermore, c () =
c() = C ()1 C () 2 2 1 c () = C () C () C ()2 C () 0 c() = C () 0 0 so when evaluated at = 0 c(0) = 0 c (0) = 0 c (0) = V .
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS By a second-order Taylor series expansion of c() about = 0, 1 1 c() = c(0) + c (0)0 + 0 c ( ) = 0 c ( ) 2 2
132
(5.27)
0 and where lies on the line segment joining . We now compute Cn () = E exp i0 ny n , the characteristic function of ny n . By the properties of the exponential function, the independence of the y i , the denition of c() and (5.27) ! n 1 X 0 log Cn () = log E exp i yi n i=1 n Y 1 0 exp i y i = log E n i=1 n Y 1 0 E exp i y i = log n i=1 n X 1 0 log E exp i y i = n i=1 = nc n 1 = 0 c (n ) 2 where n lies on the line segment joining 0 and / n. Since n 0 and c () is continuous, c (n ) c (0) = V. We thus nd that as n , 1 log Cn () 0 V 2 and 1 0 Cn () exp V 2
which is the characteristic function of the N (0, V ) distribution. This completes the proof.
Proof of Theorem 5.9.1: Since g is continuous at c, for all > 0 we can nd a > 0 such that if kz n ck < then kg (z n ) g (c)k . Recall that A B implies Pr(A) Pr(B ). Thus p Pr (kg (z n ) g (c)k ) Pr (kz n ck < ) 1 as n by the assumption that z n c. p Hence g (z n ) g (c) as n . Proof of Theorem 5.10.3: By a vector Taylor series expansion, for each element of g , gj (n ) = gj () + gj ( jn ) ( n ) where nj lies on the line segment between n and and therefore converges in probability to . It follows that ajn = gj ( jn ) gj 0. Stacking across elements of g , we nd d n (g (n ) g ()) = (G + an )0 n (n ) G0 . (5.28)
p
d d The convergence is by Theorem 5.10.1, as G + an G, n (n ) , and their product is continuous. This establishes (5.10) When N (0, V ) , the right-hand-side of (5.28) equals G0 = G0 N (0, V ) = N 0, G0 V G
CHAPTER 5. AN INTRODUCTION TO LARGE SAMPLE ASYMPTOTICS establishing (5.11).
133
Proof of Theorem 5.12.1: First consider (5.12). Take any > 0. The event max1in |yi | > n1/r S 1/r means that at least the |yi | exceeds n1/r , which is the same as the event n i=1 |yi | > n Sn one of or equivalently i=1 {|yi |r > r n} . Since the probability of the union of events is smaller than the sum of the probabilities, n ! [ r 1/r r max |yi | > = Pr {|yi | > n} Pr n
1in i=1
n X i=1
Pr (|yi |r > n r )
1 = r E (|yi |r 1 (|yi |r > n r ))
n 1 X E (|yi |r 1 (|yi |r > n r )) n r i=1
where the second inequality is the strong form of Markovs inequality (Theorem B.22) and the nal equality is since the yi are iid. Since E |y |r < this nal expectation converges to zero as n . This is because Z r E |yi | = |y |r dF (y ) < implies E (|yi | 1 (|yi | > c)) =
r r
|y| >c
|y|r dF (y) 0
(5.29)
as c . This establishes (5.12). Now consider (5.13). Take any > 0 and set t = 1/. By a similar calculation n ! [ 1 {exp |tyi | > exp (t log n)} Pr (log n) max |yi | > = Pr
1in i=1
E (exp |ty | 1 (exp |ty| > n)) where the second line uses exp (t log n) = exp (log n) = n. The assumption E exp(ty ) < means E (exp |ty | 1 (exp |ty | > n)) 0 as n by the same argument as in (5.29). This establishes (5.13).
n X i=1
Pr (exp |tyi | > n)
134
Exercises
Exercise 5.1 For the following sequences, nd the liminf, limsup and limit (if it exists) as n 1. an = 1/n n 2. an = sin 2 1 n 3. an = sin n 2
Exercise 5.2 A weighted sample mean takes the form y = P n 1 constants wi satisfying n i=1 wi = 1. Assume yi is iid. 1. Show that y is unbiased for = Eyi . 2. Calculate var(y ). 3. Show that a sucient condition for y is that
p 1 n2
1 n
Pn
i=1 wi yi
for some non-negative
4. Show that a sucient condition for the condition in part 3 is maxin wi = o(n). Exercise 5.3 Take a random variable Z such that EZ = 0 and var(Z ) = 1. Use Chebyshevs inequality to nd a such that Pr (|Z | > ) 0.05. Contrast this with the exact which solves Pr (|Z | > ) = 0.05 when Z N (0, 1) . Comment on the dierence.
3 and show that Exercise 5.4 Find the moment estimator b3 of 3 = Eyi 2 2 for some v . Write v as a function of the moments of yi . p p
Pn
2 i=1 wi
0.
2 c2 as n without using the Exercise 5.5 Suppose zn c as n . Show that zn denition of convergence in probability, without using the CMT.
d n ( b3 3 ) N 0, v 2
Exercise 5.6 Suppose
1. Use the Delta Method to obtain an asymptotic distribution for
d b= n ( b ) N 0, v2 and set = 2 and b2 .
2. Now suppose = 0. Describe what happens to the asymptotic distribution from the previous part. 3. Improve on the previous answer. Under the assumption = 0, nd the asymptotic distribub = n tion for n b2 .
b n .
4. Comment on the dierences between the answers in parts 1 and 3.
Chapter 6
Asymptotic Theory for Least Squares

6.1 Introduction
It turns out that the asymptotic theory of least-squares estimation applies equally to the projection model and the linear CEF model, and therefore the results in this chapter will be stated for the broader projection model described in Section 2.18. Recall that the model is yi = x0 i + ei for i = 1, ..., n, where the linear projection is 1 = E xi x0 E (xi yi ) . i
Some of the results of this section hold under random sampling (Assumption 1.5.1) and nite second moments (Assumption 2.18.1). We restate this condition here for clarity.
Assumption 6.1.1 1. The observations (yi , xi ), i = 1, ..., n, are independent and identically distributed. 2. Ey 2 < . 3. E kxk2 < . 4. Qxx = E (xx0 ) is positive denite.
Some of the results will require a strengthening to nite fourth moments.
4 < and Assumption 6.1.2 In addition to Assumption 6.1.1, Eyi E kxi k4 < .
135
CHAPTER 6. ASYMPTOTIC THEORY FOR LEAST SQUARES
136
6.2
Consistency of Least-Squares Estimation
b xx = 1 Pn xi x0 and Q b xy = 1 Pn xi yi . is a function of the sample moments Q i i=1 i=1 n n Second, by an application of the WLLN these sample moments converge in probability to the population moments. Specically, the fact that (yi , xi ) are mutually independent and identically distributed implies that any function of (yi , xi ) is iid, including xi x0 i and xi yi . These variables also have nite expectations by Theorem 2.18.1.1. Under these conditions, the WLLN (Theorem 5.6.2) implies that as n , n X p 0 b xx = 1 xi x0 (6.1) Q i E xi xi = Qxx n
i=1 n
In this section we use the weak law of large numbers (WLLN, Theorem 5.4.2 and Theorem 5.6.2) and continuous mapping theorem (CMT, Theorem 5.9.1) to show that the least-squares estimator b is consistent for the projection coecient . This derivation is based on three key components. First, the OLS estimator can be written as a continuous function of a set of sample moments. Second, the WLLN shows that sample moments converge in probability to population moments. And third, the CMT states that continuous functions preserve convergence in probability. We now explain each step in brief and then in greater detail. First, observe that the OLS estimator !1 n ! n X X 1 1 1 b= b b xi x0 xi yi = Q i xx Qxy n n
i=1 i=1
and
b converges Third, the CMT ( Theorem 5.9.1) allows us to combine these equations to show that in probability to . Specically, as n ,
1 b=Q b b xx Qxy p 1 Q xx Qxy
X p b xy = 1 Q xi yi E (xi yi ) = Qxy . n
i=1
(6.2)
= .
(6.3)
p b We have shown that , as n . In words, the OLS estimator converges in probability to the projection coecient vector as the sample size n gets large. To fully understand the application of the CMT we walk through it in detail. We can write b=g Q b xy b xx , Q
where g (A, b) = A1 b is a function of A and b. The function g (A, b) is a continuous function of 1 A and b at all values of the arguments such that A1 exists. Assumption 2.18.1 implies that Q xx exists and thus g (A, b) is continuous at A = Qxx . This justies the application of the CMT in (6.3). For a slightly dierent demonstration of (6.3), recall that (4.7) implies that b =Q b b 1 Q xx xe
n i=1
(6.4)
where
X b xe = 1 xi ei . Q n
CHAPTER 6. ASYMPTOTIC THEORY FOR LEAST SQUARES The WLLN and (2.27) imply Therefore
p b xe E (xi ei ) = 0. Q 1 b =Q b b xx Qxe
137
(6.5)
=0
1 Q xx 0
p b which is the same as .
Theorem 6.2.1 Consistency of Least-Squares 1 p p p 1 b xy b b xx Qxx , Q Qxy , Q Under Assumption 6.1.1, Q xx Qxx , p p b as n . b xe 0, and Q b converges in probability to as n increases, Theorem 6.2.1 states that the OLS estimator b and thus is consistent for . In the stochastic order notation, Theorem 6.2.1 can be equivalently written as b = + op (1). (6.6) ln(W agei ) = 1 Educationi + 2 Experiencei + 3 Experience2 i + 4 + ei . We use the sample of 24,344white men from the March 2009 CPS. Randomly sorting the observations, and sequentially estimating the model by least-squares, starting with the rst 5 observations, and continuing until the full sample is used, the sequence of estimates are displayed in Figure 6.1. You can see how the least-squares estimate changes with the sample size, but as the number of 1 = 0.114. observations increases it settles down to the full-sample estimate
To illustrate the eect of sample size on the least-squares estimator consider the least-squares regression
6.3
Asymptotic Normality
We started this chapter discussing the need for an approximation to the distribution of the OLS b . In Section 6.2 we showed that b converges in probability to . Consistency is a good estimator rst step, but in itself does not describe the distribution of the estimator. In this section we derive an approximation typically called the asymptotic distribution. The derivation starts by writing the estimator as a function of sample moments. One of the moments must be written as a sum of zero-mean random vectors and normalized so that the central limit theorem can be applied. The steps are as follows. Take equation (6.4) and multiply it by n. This yields the expression n !1 ! n b 1X 1 X 0 n = xi xi xi ei . (6.7) n n
i=1 i=1
b is a function of the sample This shows that the normalized and centered estimator n 1 Pn 1 Pn 0 average n i=1 xi xi and the normalized sample average n i=1 xi ei . Furthermore, the latter has mean zero so the central limit theorem (CLT, Theorem 5.7.1) applies.
138
OLS Estimation
0.08
0.09
0.10
0.11
0.12
0.13
0.14
0.15
5000
10000
15000
20000
Number of Observations
1 as a function of sample size n Figure 6.1: The least-squares estimator The product xi ei is iid (since the observations are iid) and mean zero (since E (xi ei ) = 0). Dene the k k covariance matrix 2 (6.8) = E xi x0 i ei . We require the elements of to be nite, written < . By the Expectation Inequality (B.15), 2 2 2 2 kk E xi x0 i ei = E kxi ei k = E kxi k ei
which is nite if xi and ei have nite fourth moments. As ei is a linear combination of yi and xi , it is sucient that the observables have nite fourth moments (Theorem 2.18.1.6). We can then apply the CLT (Theorem 5.7.1).
or equivalently that E kxi ei k2 < . Usingkxi ei k2 = kxi k2 e2 i and the Cauchy-Schwarz Inequality (B.17), 1/2 1/2 2 2 2 4 2 Ee4 E k x e e k = E k x k e k (6.9) = E k x kk E xi x0 i i i i i i i i
Theorem 6.3.1 Under Assumption 6.1.2, kk E kxi ei k2 < and 1 X d xi ei N (0, ) n

i=1 n
(6.10)
(6.11)
as n .
Putting together (6.1), (6.7), and (6.11), d 1 b n Q xx N (0, ) 1 1 = N 0, Q xx Qxx
139
as n , where the nal equality follows from the property that linear combinations of normal vectors are also normal (Theorem B.9.1). We have derived the asymptotic normal approximation to the distribution of the least-squares estimator.
Theorem 6.3.2 Asymptotic Normality of Least-Squares Estimator Under Assumption 6.1.2, as n d b n N (0, V ) where Qxx
1 1 V = Q xx Qxx , 0 2 = E (xi x0 i ) , and = E xi xi ei .
(6.12)
In the stochastic order notation, Theorem 6.3.2 implies that b = + Op (n1/2 ) (6.13)
which is stronger than (6.6). b 1 1 The matrix V = Q Q is the variance of the asymptotic distribution of n . xx xx b . The expression Consequently, V is often referred to as the asymptotic covariance matrix of 1 1 V = Qxx Qxx is called a sandwich form, as the matrix is sandwiched between two copies of 1 Q xx . It is useful to compare the variance of the asymptotic distribution given in (6.12) and the nite-sample conditional variance in the CEF model as given in (4.12): 1 0 1 b X DX X 0 X = var | X = X 0X . (6.14) V b and V is the asymptotic variance of is the exact conditional variance of Notice that V b n . Thus V should be (roughly) n times as large as V , or V nV . Indeed, multiplying (6.14) by n and distributing, we nd nV = 1 0 XX n 1 1 0 X DX n 1 0 XX n 1
which looks like an estimator of V . Indeed, as n nV V .

p
The expression V is useful for practical inference (such as computation of standard errors and b , while V is useful for asymptotic theory as it tests) since it is the variance of the estimator is well dened in the limit as n goes to innity. We will make use of both symbols and it will be advisable to adhere to this convention. There is a special case where and V simplify. We say that ei is a Homoskedastic Projection Error when 2 (6.15) cov(xi x0 i , ei ) = 0.
140
Figure 6.2: Density of Normalized OLS estimator with Double Pareto Error Condition (6.15) holds in the homoskedastic linear regression model, but is somewhat broader. Under (6.15) the asymptotic variance formulae simplify as 2 2 (6.16) = E xi x0 i E ei = Qxx
1 1 1 2 0 V = Q xx Qxx = Qxx V
(6.17)
1 2 0 In (6.17) we dene V 0 = Qxx whether (6.15) is true or false. When (6.15) is true then V = V , 0 otherwise V 6= V 0 . We call V the homoskedastic asymptotic covariance matrix. Theorem 6.3.2 states that the sampling distribution of the least-squares estimator, after rescaling, is approximately normal when the sample size n is suciently large. This holds true for all joint distributions of (yi , xi ) which satisfy the conditions of Assumption 6.1.2, and is therefore broadly applicable. Consequently, asymptotic normality is routinely used to approximate the nite sample b . distribution of n b can be arbitrarily far from the A diculty is that for any xed n the sampling distribution of normal distribution. In Figure 5.1 we have already seen a simple example where the least-squares estimate is quite asymmetric and non-normal even for reasonably large sample sizes. The normal approximation improves as n increases, but how large should n be in order for the approximation to be useful? Unfortunately, there is no simple answer to this reasonable question. The trouble is that no matter how large is the sample size, the normal approximation is arbitrarily poor for some data distribution satisfying the assumptions. We illustrate this problem using a simulation. Let yi = 1 xi + 2 + ei where xi is N (0, 1) , and ei is independent of xi with the Double Pareto 1 , |e| 1. If > 2 the error ei has zero mean and variance /( 2). density f (e) = 2 |e| As approaches 2, however, variance diverges to innity. In this context the normalized leastq its 2 squares slope estimator n 1 1 has the N(0, 1) asymptotic distibution for any > 2. q 2 In Figure 6.2 we display the nite sample densities of the normalized estimator n 1 1 , setting n = 100 and varying the parameter . For = 3.0 the density is very close to the N(0, 1) density. As diminishes the density changes signicantly, concentrating most of the probability mass around zero.
141
Figure 6.3: Density of Normalized OLS estimator with error process (6.18) Another example is shown in Figure 6.3. Here the model is yi = + ei where k uk i E ui ei = 1/2 k E uk 2 E u2 i i
(6.18)
b and ui N(0, 1). We show the sampling distribution of n setting n = 100, for k = 1, 4, 6 and 8. As k increases, the sampling distribution becomes highly skewed and non-normal. The lesson from Figures 6.2 and 6.3 is that the N(0, 1) asymptotic approximation is never guaranteed to be accurate.
6.4
Joint Distribution
Theorem 6.3.2 gives the joint asymptotic distribution of the coecient estimates. We can use the result to study the covariance between the coecient estimates. For example, suppose k = 2 2 ). For simplicity suppose that the regressors are mean zero. Then 1 , and write the estimates as ( we can write 2 1 1 2 Qxx = 2 1 2 2
2 and 2 are the variances of x where 1 1i and x2i , and is their correlation. If the error is ho2 2 ) is V 0 = Q1 2 . By the formula for 1 , moskedastic, then the asymptotic variance matrix for ( xx inversion of a 2 2 matrix, 2 1 2 1 2 1 Qxx = 2 2 . 2 1 1 2 (1 2 ) 1 2
1 and 2 are negatively correlated (and Thus if x1i and x2i are positively correlated ( > 0) then vice-versa). For illustration, Figure 6.4 displays the probability contours of the joint asymptotic distribution 2 2 when 1 = 2 = 0, 2 = 2 = 2 = 1, and = 0.5. The coecient estimates of 1 1 and 1 2 1 is are negatively correlated since the regressors are positively correlated. This means that if
142
1 , 2 ), homoskedastic case Figure 6.4: Contours of Joint Distribution of ( 2 is unusually positive, or conversely. It is also unlikely that unusually negative, it is likely that 2 unusually large and of the same sign. 1 and we will observe both This nding that the correlation of the regressors is of opposite sign of the correlation of the coefcient estimates is sensitive to the assumption of homoskedasticity. If the errors are heteroskedastic then this relationship is not guaranteed. This can be seen through a simple constructed example. Suppose that x1i and x2i only take the values {1, +1}, symmetrically, with Pr (x1i = x2i = 1) = Pr (x1i = x2i = 1) = 3/8, and Pr (x1i = 1, x2i = 1) = Pr (x1i = 1, x2i = 1) = 1/8. You can check that the regressors are mean zero, unit variance and correlation 0.5, which is identical with the setting displayed in Figure 6.4. Now suppose that the error is heteroskedastic. Specically, suppose that E e2 i | x1i = x2i = 1 5 2 = E x2 e2 = 1 and and E e2 . You can check that E e2 = 1, E x2 i | x1i 6= x2i = i 1i ei 2i i 4 4 7 2 E x1i x2i ei = . Therefore 8
1 1 V = Q xx Qxx 1 9 1 2 1 = 7 16 1 1 2 8
7 1 8 1 1 2
1 2 1
1 and 2 are positively correlated (their correlation is 1/4.) The Thus the coecient estimates joint probability contours of their asymptotic distribution is displayed in Figure 6.5. We can see how the two estimates are positively associated. What we found through this example is that in the presence of heteroskedasticity there is no simple relationship between the correlation of the regressors and the correlation of the parameter
4 1 3 1 4
1 4 . 1
143
1 and 2 , heteroskedastic case Figure 6.5: Contours of Joint Distribution of estimates. We can extend the above analysis to study the covariance between coecient sub-vectors. For 0 0 0 0 0 0 example, partitioning xi = (x1i , x2i ) and = 1 , 2 , we can write the general model as
0 yi = x0 1i 1 + x2i 2 + ei 0 0 b0 = b , b . Make the partitions and the coecient estimates as 1 2
Qxx =
Q11 Q12 Q21 Q22
11 12 21 22
(6.19)
From (2.41)
1 Q xx
1 1 where Q112 = Q11 Q12 Q 22 Q21 and Q221 = Q22 Q21 Q11 Q12 . Thus when the error is homoskedastic, b , b = 2 Q1 Q Q1 cov 1 2 112 12 22
1 1 1 Q Q 112 112 Q12 Q22 1 1 1 Q221 Q21 Q11 Q221
which is a matrix generalization of the two-regressor case. In the general case, you can show that (Exercise 6.5) V 11 V 12 V = V 21 V 22
(6.20)
where 1 1 1 1 1 1 V 11 = Q 112 11 Q12 Q22 21 12 Q22 Q21 + Q12 Q22 22 Q22 Q21 Q112 1 1 1 1 1 1 V 21 = Q 221 21 Q21 Q11 11 22 Q22 Q21 + Q21 Q11 12 Q22 Q21 Q112 1 1 1 1 1 1 V 22 = Q 221 22 Q21 Q11 12 21 Q11 Q12 + Q21 Q11 11 Q11 Q12 Q221 (6.21) (6.22) (6.23)
Unfortunately, these expressions are not easily interpretable.
144
6.5
Consistency of Error Variance Estimators
1 Pn 2 Using the methods of Section 6.2 we can show that the estimators 2 = n 2 i and s = i=1 e P n 1 2 2 i are consistent for . i=1 e nk The trick is to write the residual e i as equal to the error ei plus a deviation term
b e i = yi x0 i
Thus the squared residual equals the squared error plus a deviation 0 2 2 0 b 0 b b e i = ei 2ei xi + xi xi .
0b = ei + x0 i xi b . = ei x0 i
(6.24)
Indeed, the WLLN shows that
So when we take the average of the squared residuals we obtain the average of the squared errors, plus two terms which are (hopefully) asymptotically negligible. n ! n ! n 0 1 X X X 1 1 0 0 b + b b . 2 = (6.25) e2 2 e x x x i i i i i n n n
i=1 i=1 i=1
1X 2 p 2 ei n
i=1 i=1
n 0 1X p ei x0 i E ei xi = 0 n i=1
n 1X p 0 xi x0 i E xi xi = Qxx n p b and Theorem 6.2.1 shows that . Hence (6.25) converges in probability to 2 , as desired. Finally, since n/(n k) 1 as n , it follows that n p 2 s = 2 2 . nk
Thus both estimators are consistent.
Theorem 6.5.1 Under Assumption 6.1.1, 2 2 and s2 2 as n .
6.6
Homoskedastic Covariance Matrix Estimation
b Theorem 6.3.2 shows that n is asymptotically normal with with asymptotic covariance matrix V . For asymptotic inference (condence intervals and tests) we need a consistent 1 2 estimate of V . Under homoskedasticity, V simplies to V 0 = Qxx , and in this section we consider the simplied problem of estimating V 0 .
145
b xx dened in (6.1), and thus an estimator for Q1 The standard moment estimator of Qxx is Q xx 1 b xx . Also, the standard estimator of 2 is the unbiased estimator s2 dened in (4.21). Thus a is Q 1 2 b0 b 1 2 natural plug-in estimator for V 0 = Qxx is V = Qxx s . b 0 for V 0 follows from consistency of the moment estimates Q b xx and s2 , and Consistency of V an application of the continuous mapping theorem. Specically, Theorem 6.2.1 established that p p 1 2 b xx Qxx , and Theorem 6.5.1 established s2 2 . The function V 0 Q = Qxx is a continuous function of Qxx and 2 so long as Qxx > 0, which holds true under Assumption 6.1.1.4. It follows by the CMT that 1 2 0 b 1 2 p b0 V = Qxx s Qxx = V
0 b0 so that V is consistent for V , as desired. p 0 b0 Theorem 6.6.1 Under Assumption 6.1.1, V V as n .
It is instructive to notice that Theorem 6.6.1 does not require the assumption of homoskedasticb 0 is consistent for V 0 regardless if the regression is homoskedastic or heteroskedastic. ity. That is, V 0 b b0 However, V = V = avar() only under homoskedasticity. Thus in the general case, V is consistent for a well-dened but non-useful object.
6.7
Heteroskedastic Covariance Matrix Estimation

b n is V =
Theorems 6.3.2 established that the asymptotic covariance matrix of
1 1 Q xx Qxx . We now consider estimation of this covariance matrix without imposing homoskedasticity. The standard approach is to use a plug-in estimator which replaces the unknowns with sample moments. 1 b 1 b As described in the previous section, a natural estimator for Q xx is Qxx , where Qxx dened in (6.1). The moment estimator for is n X b = 1 xi x0 2 (6.26) ie i, n i=1
leading to the plug-in covariance matrix estimator
b W = nV bW bW You can check that V where V is the White covariance matrix estimator introduced in (4.28). 1 p 1 b b As shown in Theorem 6.2.1, Q xx Qxx , so we just need to verify the consistency of . 2 2 The key is to replace the squared residual e i with the squared error ei , and then show that the dierence is asymptotically negligible. Specically, observe that X b = 1 xi x0 2 ie i n
i=1 i=1 n
b 1 b b 1 bW V = Qxx Qxx .
(6.27)
n n 2 1X 1X 2 i e2 = xi x0 e + xi x0 i i i e i . n n i=1
(6.28)
146
2 The rst term is an average of the iid random variables xi x0 i ei , and therefore by the WLLN converges in probability to its expectation, namely,
1X 2 p 0 2 xi x0 i ei E xi xi ei = . n
i=1
Technically, this requires that has nite elements, which was shown in (6.10). b is consistent for it remains to show that So to establish that 2 p 1X xi x0 i e2 i e i 0. n
i=1 n
(6.29)
There are multiple ways to do this. A reasonable straightforward yet slightly tedious derivation is to start by applying the Triangle Inequality (A.12) n n 1 X 2 1 X 2 0 2 xi x0 i ei i e2 xi xi e i e i n n
i=1 i=1 n 2 1X = kxi k2 e i e2 i . n i=1
(6.30)
Then recalling the expression for the squared residual (6.24), apply the Triangle Inequality and then the Schwarz Inequality (A.10) twice 0 2 0 b 0 b b e + x x x 2 e i e2 i i i i i 2 0 0 b b = 2 |ei | xi + xi 2 b 2 b 2 |ei | kxi k + kxi k . (6.31) Combining (6.30) and (6.31), we nd n n ! n ! 2 1 X 2 1X 1X b 3 4 b 0 2 + e 2 x x e k x k | e | k x k i i i i i i i n n n
i=1 i=1 i=1
= op (1). (6.32) p b 0 and both averages in parenthesis are averages of The expression is op (1) because random variables with nite mean under Assumption 6.1.2 (and are thus Op (1)). Indeed, by Hlders Inequality (B.16) 4/3 3/4 3/4 1/4 1/4 3 3 4 = E k x k < . E kxi k |ei | E kxi k Ee4 Ee4 i i i
We have established (6.29), as desired.
p b Theorem 6.7.1 Under Assumption 6.1.2, as n , and W p b V . V
For an alternative proof of this result, see Section 6.21.
147
6.8
Summary of Covariance Matrix Notation
The notation we have introduced may be somewhat confusing so it is helpful to write it down b b in one place. The exact variance of and the asymptotic variance of n are 1 0 1 0 b V X DX X 0 X = var | X = X X b = Q1 Q1 V = avar n xx xx
The White estimates of these two covariance matrices are n ! 0 1 X 0 1 W b xi x0 2 V XX = X X ie i

i=1
and satisfy the simple relationship
bW V
b b 1 b 1 Q xx Qxx
Similarly, under the assumption of homoskedasticity the exact and asymptotic variances simplify to 0 1 2 V0 = X X
1 2 V0 = Qxx
b W = nV bW V .
and their standard estimators are 0 1 2 b0 s V = X X b 0 = nV b0 V .
which also satisfy the relationship
b 1 2 b0 V = Qxx s
The exact formula and estimates are useful when constructing test statistics and standard errors, b , not in the asymptotic as in practice we are interested in the covariance matrix of the estimates variance. However, for theoretical purposes the asymptotic formula (variances and their estimates) are more useful, as these retain non-generate limits as the sample sizes diverge. That is why both sets of notation are useful.
6.9
Alternative Covariance Matrix Estimators*
b = nV b , V e = nV e , and V = nV . To do so we introduce their scaled versions, e.g. V These are (alternative) estimates of the asymptotic covariance matrix V . bW b = nV b = n V bW b . Notice that V First, consider V nk where V was dened in (6.27) and n shown consistent for V in Theorem 6.7.1. If k is xed as n , then n k 1 and thus
p b = (1 + o(1))V b W V V .
bW bW bW In Section 6.7 we introduced V from as an estimator of V . V is a scaled version of V Section 4.11, where we also introduced the alternative heteroskedasticity-robust covariance matrix e and V . We now discuss the consistency properties of these estimators. b , V estimators V

b is consistent for V . Thus V
CHAPTER 6. ASYMPTOTIC THEORY FOR LEAST SQUARES b replaced by e and V take the form (6.27) but with The alternative estimators V X e = 1 (1 hii )2 xi x0 2 ie i n
i=1 n n
148
and
b , it is sucient respectively. To show that these estimators also consistent for V , given e b and b converge in probability to zero as n . to show that the dierences The trick is to use the fact that the leverage values are asymptotically negligible: h n = max hii = op (1).
1in
1X = (1 hii )1 xi x0 2 ie i, n
i=1
(6.33)
(See Theorem 6.22.1 in Section 6.22).) Then using the Triangle Inequality
n 1X 2 1 b xi x0 i (1 hii ) 1 i e n i=1 n ! 1X 1 (1 h kxi k2 e 2 ) 1 . i n n i=1
The sum in parenthesis can be shown to be Op (1) under Assumption 6.1.2 by the same argument asin in the proof of Theorem 6.7.1. (In fact, it can be shown to converge in probability to 2 2 E kxi k ei .) The term in absolute values is op (1) by (6.33). Thus the product is op (1), which b + op (1) . means that = Similarly,
n 2 e b 1 X 2 xi x0 (1 h e ) 1 ii i i n i=1 n ! 1X 2 2 2 kxi k e i (1 hn ) 1 n i=1
= op (1).
p p e Theorem 6.9.1 Under Assumption 6.1.2, as n , , , p p p b e V V , V V , and V V .
Theorem 6.9.1 shows that the alternative covariance matrix estimators are also consistent for the asymptotic covariance matrix.
6.10
Functions of Parameters
Sometimes we are interested in a transformation of the coecient vector = (1 , ..., k ). For example, we may be interested in a single coecient j , or a ratio j /l . In these cases we can write
149
p b we can deduce that By the continuous mapping theorem (Theorem 5.9.1) and the fact b is consistent for .
the transformation as a function of the coecients, e.g. = r() for some function r : Rk Rq . The estimate of is b = r( b ).
Theorem 6.10.1 Under Assumption 6.1.1, if r() is continuous at the p b true value of , then as n , . Furthermore, if the transformation is suciently smooth, by the Delta Method (Theorem 5.10.3) b is asymptotically normal. we can show that Assumption 6.10.1 r() : Rk Rq is continuously dierentiable at the 0 true value of and R = r ( ) has rank q.
where
Theorem 6.10.2 Asymptotic Distribution of Functions of Parameters Under Assumptions 6.1.2 and 6.10.1, as n , d b n N (0, V ) (6.34) V = R0 V R (6.35)
In many cases, the function r() is linear: r() = R0 for some k q matrix R. In particular, if R is a selector matrix I R= 0
0 0 0 0 0 0 then we can conformably partition = (0 1 , 2 ) so that R = 1 for = ( 1 , 2 ) . Then I = V 11 , V = I 0 V 0
(6.36)
b are approximately normal with variances given by the comformable subcomThat is, subsets of ponents of V .
the upper-left sub-matrix of V 11 given in (6.21). In this case (6.34) states that d b n N (0, V 11 ) .
1 1
150
so
To illustrate the case of a nonlinear transformation, take the example = j /l for j 6= l. Then 0 . . . 1/l . . (6.37) R= . j / 2 l . . . 0
2 V = V jj /l2 + V ll j /l4 2V jl j /l3
where V ab denotes the abth element of V . For inference we need an estimate of the asymptotic variance matrix V = R0 V R, and for this it is typical to use a plug-in estimator. The natural estimator of R is the derivative evaluated at the point estimates b )0 . b = r( (6.38) R The derivative in (6.38) may be calculated analytically or numerically. By analytically, we mean working out for the formula for the derivative and replacing the unknowns by point estimates. For example, if = j /l , then r ( ) is (6.37). However in some cases the function r ( ) may be extremely complicated and a formula for the analytic derivative may not be easily available. In this case calculation by numerical dierentiation may be preferable. Let l = (0 1 0)0 be the b is unit vector with the 1 in the lth place. Then the jlth element of a numerical derivative R b b b jl = rj ( + l ) rj () R b 0V b = R b R b. V
for some small . The estimate of V is
(6.39)
Given (6.38), (6.39) and (6.40) are simple to calculate using matrix operations. b is the asymptotic approximation (6.34), V b is often called As the primary justication for V an asymptotic covariance matrix estimator. p b b is consistent for V under the conditions of Theorem 6.10.2 since V V The estimator V by Theorem 6.7.1, and p b )0 b = r( r()0 = R R
p b and the function since 0 r ( )
e b b0 Alternatively, V , V or V may be used in place of V . For example, the homoskedastic covariance matrix estimator is b0 = R b 0R b =R b 0Q b 1 R b 2 b 0V V (6.40) xx s
is continuous.
Theorem 6.10.3 Under Assumptions 6.1.2 and 6.10.1, as n ,

p b V V .
151
b is consistent for V and thus may be used for asymptotic Theorem 6.10.3 shows that V inference. In practice, we may set b , or substitute an alternative covariance estimator such as V . as an estimate of the variance of b 0V b = R b R b = n1 R b 0V b R b V (6.41)
6.11
Asymptotic Standard Errors
As described in Section 4.12, a standard error is an estimate of the standard deviation of the b , then standard b is an estimate of the covariance matrix of distribution of an estimator. Thus if V errors are the square roots of the diagonal elements of this matrix. These take the form rh q i b b V s(j ) = V = . j
jj
) an asymptotic standard j ) or s( When the justication is based on asymptotic theory we call s( j or . When reporting your results, it is good practice to report standard errors for error for each reported estimate, and this includesfunctions and transformations of your parameter estimates. This helps users of the work (including yourself) assess the estimation precision. We illustrate using the log wage regression log(W age) = 1 education + 2 experience + 3 experience2 /100 + 4 + e. Consider the following three parameters of interest. 1. Percentage return to education: 1 = 1001
b are constructed similarly. Supposing that q = 1 (so h() is real-valued), then Standard errors for is the square root of (6.41) the standard error for r q 0 b b b b 0V b R b. s() = R V R = n1 R
(100 times the partial derivative of the conditional expectation of log wages with respect to education.) 2. Percentage return to experience for individuals with 10 years of experience 2 = 1002 + 23 (100 times the partial derivative of the conditional expectation of log wages with respect to experience, evaluated at experience = 10) 3. Experience level which maximizes expected log wages: 3 = 502 /3 (The level of experience at which the partial derivative of the conditional expectation of log wages with respect to experience equals 0.)
CHAPTER 6. ASYMPTOTIC THEORY FOR LEAST SQUARES The 4 1 vector R for these three examples is 100 0 0 100 R= 0 , 2 , 0 0
152
respectively. We use the subsample of married black women (all experience levels), which has 982 observations. The point estimates and standard errors are 0.118 education + 0.016 experience 0.022 experience2 /100 + 0.947 . (0.157) (0.008) (0.006) (0.012) (6.42) The standard errors are the square roots of the Horn-Horn-Duncan covariance matrix estimate 0.632 0.131 0.143 11.1 0.131 0.390 0.731 6.25 104 . V (6.43) = 0.143 0.731 1.48 9.43 11.1 6.25 9.43 246 We calculate that b1 b1 = 100 = 11.8 b1 ) = s( \ log( W age) =
0 50/3 2 , 502 /3 0
= 100 0.118
= 0.8
p 1002 0.632 104
b2 + 2 b3 b2 = 100 = 1.12 s
= 100 0.016 2 0.022
b2 ) = s(
100 2
0.390 0.731 0.731 1.48
100 2
104
= 0.40
b2 / b3 b3 = 50 = 35.2
= 50 0.016/0.022
v ! u 0.390 0.731 u b3 50 / t 2 b3 ) = b2 / b b3 50 s( 104 50/ 3 b2 / b2 0.731 1.48 50 3 = 7.0 The calculations show that the estimate of the percentage return to education is about 12% per year, with a standard of 0.8. The estimate of the percentage return to experience for those with 10 years of experience is 1.2% per year, with a standard error of 0.4. And the estimate of the experience level which maximizes expected log wages is 35 years, with a standard error of 7.
153
6.12
t statistic
Let = r() : Rk R be any parameter of interest (for example, could be a single element b its estimate and s( b) its asymptotic standard error. Consider the statistic of ), tn () = b . b) s(
(6.44)
Dierent writers have called (6.44) a t-statistic, a t-ratio, a z-statistic or a studentized statistic, sometimes using the dierent labels to distinguish between nite-sample and asymptotic inference. As the statistics themselves are always (6.44) we wont make such as distinction, and will simply refer to tn () as a t-statistic or a t-ratio. We also often suppress the parameter dependence, writing it as tn . The t-statistic is a simple function of the estimate, its standard error, and the parameter. b p d b N (0, V ) and V V . Thus By Theorems 6.10.2 and 6.10.3, n tn () = b b) s( b n q = b V
d
N (0, V ) V = Z N (0, 1) . The last equality is by the property that linear scales of normal distributions are normal. Thus the asymptotic distribution of the t-ratio tn () is the standard normal. Since this distribution does not depend on the parameters, we say that tn () is asymptotically pivotal. In special cases (such as the normal regression model, see Section 3.18), the statistic tn has an exact t distribution, and is therefore exactly free of unknowns. In this case, we say that tn is exactly pivotal. In general, however, pivotal statistics are unavailable and we must rely on asymptotically pivotal statistics. As we will see in the next section, it is also useful to consider the distribution of the absolute d d t-ratio |tn ()| . Since tn () Z, the continuous mapping theorem yields |tn ()| |Z| . Letting (u) = Pr (Z u) denote the standard normal distribution function, we can calculate that the distribution function of |Z| is Pr (|Z| u) = Pr (u Z u) = (u) (u)
def
= Pr (Z u) Pr (Z < u) = 2(u) 1 = (u).
(6.45)
N (0, 1) and |tn ()| |Z| .
Theorem 6.12.1 Under Assumptions 6.1.2 and 6.10.1, tn () Z

d
The asymptotic normality of Theorem 6.12.1 is used to justify condence intervals and tests for the parameters.
154
6.13
Condence Intervals
b is a point estimate for , meaning that b is a single value in Rk . A The OLS estimate broader concept is a set estimate Cn which is a collection of values in Rk . When the parameter is real-valued then it is common to focus on intervals Cn = [Ln , Un ] and which is called an interval estimate for . The goal of an interval estimate Cn is to contain the true value, e.g. Cn , with high probability. The interval estimate Cn is a function of the data and hence is random. The coverage probability of the interval Cn = [Ln , Un ] is Pr ( Cn ). The randomness comes from Cn as the parameter is treated as xed. Interval estimates Cn are typically called condence intervals as the goal is typically to set the coverage probability to equal a pre-specied target, typically 90% or 95%. Cn is called a (1 )% condence interval if inf Pr ( Cn ) = 1 . There is not a unique method to construct condence intervals. For example, a simple (yet silly) interval is R with probability 1 Cn = b with probability b has a continuous distribution, Pr( Cn ) = 1 , so this condence interval By construction, if has perfect coverage, but Cn is uninformative about and is therefore not useful. b with standard error s( b), the When we have an asymptotically normal parameter estimate standard condence interval for takes the form h i b c s( b), b + c s( b) (6.46) Cn = where c > 0 is a pre-specied constant. This condence interval is symmetric about the point b and its length is proportional to the standard error s( b). estimate , Equivalently, Cn is the set of parameter values for such that the t-statistic tn () is smaller (in absolute value) than c, that is ( ) b Cn = { : |tn ()| c} = : c c . b) s( The coverage probability of this condence interval is Pr ( Cn ) = Pr (|tn ()| c) which is generally unknown. We can approximate the coverage probability by taking the asymptotic limit as n . Since |tn ()| is asymptotically |Z| (Theorem 6.12.1), it follows that as n that Pr ( Cn ) Pr (|Z| c) = (c) where (u) is given in (6.45). We call this the asymptotic coverage probability. Since the tratio is asymptotically pivotal, the asymptotic coverage probability is independent of the parameter , and is only a function of c. As we mentioned before, an ideal condence interval has a pre-specied probability coverage 1 , typically 90% or 95%. This means selecting the constant c so that (c) = 1 . Eectively, this makes c a function of , and can be backed out of a normal distribution table. For example, = 0.05 (a 95% interval) implies c = 1.96 and = 0.1 (a 90% interval) implies c = 1.645. Rounding 1.96 to 2, we obtain the most commonly used condence interval in applied econometric practice h i b 2s( b), b + 2s( b) . (6.47) Cn =
155
This is a useful rule-of thumb. This asymptotic 95% condence interval Cn is simple to compute and can be roughly calculated from tables of coecient estimates and standard errors. (Technically, it is an asymptotic 95.4% interval, due to the substitution of 2.0 for 1.96, but this distinction is overly precise.)
Theorem 6.13.1 Under Assumptions 6.1.2 and 6.10.1, for Cn dened in (6.46), Pr ( Cn ) (c). For c = 1.96, Pr ( Cn ) 0.95.
Condence intervals are a simple yet eective tool to assess estimation uncertainty. When reading a set of empirical results, look at the estimated coecient estimates and the standard errors. For a parameter of interest, compute the condence interval Cn and consider the meaning of the spread of the suggested values. If the range of values in the condence interval are too wide to learn about , then do not jump to a conclusion about based on the point estimate alone. For illustration, consider the three examples presented in Section 6.11 based on the log wage regression for single asian men. Percentage return to education. A 95% asymptotic condence interval is 11.81.960.8 = [10.2, 13.3]. Percentage return to experience for individuals with 12 years experience. A 90% asymptotic condence interval is 1.1 1.645 0.4 = [0.5, 1.8]. Experience level which maximizes expected log wages. An 80% asymptotic condence interval is 35 1.28 7 = [26, 44].
6.14
Regression Intervals
m(x) = E (yi | xi = x) = x0 .
In the linear regression model the conditional mean of yi given xi = x is
In some cases, we want to estimate m(x) at a particular point x. Notice that this is a (linear) 0 b = x0 b and R = x, so b (x) = function qof . Letting r() = x and = r(), we see that m b) = x0 V b x. Thus an asymptotic 95% condence interval for m(x) is s(
It is interesting to observe that if this is viewed as a function of x, the width of the condence set is dependent on x. To illustrate, we return to the log wage regression (3.11) of Section 3.7. The estimated regression equation is \ b = 0.155x + 0.698 log( W age) = x0 where x = education. The covariance matrix estimate from (4.35) is 0.001 0.015 b = . V 0.015 0.243
q 0b 0 b x . x 1.96 x V
Thus the 95% condence interval for the regression takes the form p 0.155x + 0.698 1.96 0.001x2 0.015x + 0.243.
156
Figure 6.6: Wage on Education Regression Intervals The estimated regression and 95% intervals are shown in Figure 6.6. Notice that the condence bands take a hyperbolic shape. This means that the regression line is less precisely estimated for very large and very small values of education. Plots of the estimated regression line and condence intervals are especially useful when the regression includes nonlinear terms. To illustrate, consider the log wage regression (6.42) which includes experience and its square, with covariance matrix (6.43). We are interested in plotting the regression estimate and regression intervals as a function of experience. Since the regression also includes education, to plot the estimates in a simple graph we need to x education at a specic value. We select education =12. This only aects the level of the estimated regression, since education enters without an interaction. Dene the points of evaluation 12 x z (x) = x2 /100 1 where x =experience. Thus the 95% regression interval for education=12, as a function of x =experience is 0.118 12 + 0.016 x 0.022 x2 /100 + 0.947 v u 0.632 0.131 0.143 11.1 u u 0.131 u 0.390 0.731 6.25 1.96uz (x)0 t 0.143 0.731 1.48 9.43 11.1 6.25 9.43 246
The estimated regression and 95% intervals are shown in Figure 6.7. The regression interval widens greatly for small and large values of experience, indicating considerable uncertainty about the eect of experience on mean wages for this population. The condence bands take a more complicated shape than in Figure 6.6 due to the nonlinear specication.
= 0.016 x .00022 x2 + 2.36 p 0.0196 70.608 9.356 x + 0.54428 x2 0.01462 x3 + 0.000148 x4
z (x) 104
157
Figure 6.7: Wage on Experience Regression Intervals
6.15
Forecast Intervals
Suppose we a given a value of the regressor vector xn+1 for an individual outside the sample, and we want to forecast (guess) yn+1 for this individual. This is equivalent to forecasting yn+1 given xn+1 = x, which will generally be a function of x. A reasonable forecasting rule is the conditional mean m(x) as it is the mean-square-minimizing forecast. A point forecast is the estimated b . We would also like a measure of uncertainty for the forecast. conditional mean m b (x) = x0 b . As the out-of-sample error en+1 b (x) = en+1 x0 The forecast error is e n+1 = yn+1 m b , this has variance is independent of the in-sample estimate 0 2 0 b b x = E e | x = x + x E Ee 2 n +1 n+1 n+1 = 2 (x) + x0 V x.
2 b x, so a standard 2 + x0 V Assuming E e2 n+1 | xn+1 = , the natural estimate of this variance is q 2 0 b +xV error for the forecast is s (x) = x. Notice that this is dierent from the standard error for the conditional mean. The conventional 95% forecast interval for yn+1 uses a normal approximation and sets h i b 2 x0 s(x) .
The diculty is that the equation error en+1 is generally non-normal, and asymptotic theory cannot be applied to a single observation. The only special exception is the case where en+1 has the exact distribution N(0, 2 ), which is generally invalid. To get an accurate forecast interval, we need to estimate the conditional distribution of en+1 given xn+1 = x, which is a much more dicult task. Perhaps due to this diculty, many applied
It is dicult, however, to fully justify this choice. It would be correct if we have a normal approximation to the ratio b en+1 x0 . s (x)
158
i h b 2 s(x) despite the lack of a convincing forecasters use the simple approximate interval x0 justication.
6.16
Wald Statistic
b = nV b . When q = 1, then Wn () = tn ()2 is the square of the t-ratio. When q > 1, where V Wn () is typically called a Wald statistic. We are interested in its sampling distribution. The asymptotic distribution of Wn () is simple to derive given Theorem 6.10.2 and Theorem 6.10.3, which show that d b n Z N (0, V ) and It follows that Wn () = 0 1 b d 1 b b n n V Z0 V Z
p b V . V
b its estimate and V b its Let = r() : Rk Rq be any parameter vector of interest, covariance matrix estimator. Consider the quadratic form 0 1 0 1 b =n b V b . b V b b (6.48) Wn () =
(6.49)
a quadratic in the normal random vector Z. Here we can appeal to a useful result from probability theory. (See Theorem B.9.3 in the Appendix.) Theorem 6.16.1 If Z N (0, A) with A > 0, q q, then Z 0 A1 Z 2 q, a chi-square random variable with q degrees of freedom.
The asymptotic distribution in (6.49) takes exactly this form. Note that V > 0 since R is full rank under Assumption 6.10.1 It follows that Wn () converges in distribution to a chi-square random variable. Theorem 6.16.2 Under Assumptions 6.1.2 and 6.10.1, as n , Wn () 2 q.
d
Theorem 6.16.2 is used to justify multivariate condence regions and mutivariate hypothesis tests.
6.17
Homoskedastic Wald Statistic
2 Under the conditional homoskedasticity assumption E e2 i | xi = we can construct the Wald 0 b dened in (6.40). This yields a statistic using the homoskedastic covariance matrix estimator V homoskedastic Wald statistic 0 0 1 0 0 1 0 b =n b b . b b b V V (6.50) () = Wn Under the additional assumption of conditional homoskedasticity, it has the same asymptotic distribution as Wn ().
CHAPTER 6. ASYMPTOTIC THEORY FOR LEAST SQUARES Theorem 6.17.1 Under Assumptions 6.1.2 and 6.10.1, and E e2 i | xi = 2 , as n , d 0 () 2 Wn q.
159
6.18
Condence Regions
which shows that Cn has asymptotic coverage (1 )%. To illustrate the construction of a condence region, consider the estimated regression (6.42) of the model \ log( W age) = 1 education + 2 experience + 3 experience2 /100 + 4 . Suppose that the two parameters of interest are the percentage return to education 1 = 1001 and the percentage return to experience for individuals with 10 years experience 2 = 1002 + 203 . These two parameters are a linear transformation of the regression parameters with point estimates 0 100 0 0 11.8 b b = = , 0 0 100 20 1.2 and have the covariance matrix estimate 0 0 0 100 0 0 b 100 0 b V V = 0 100 0 0 100 20 0 20 0.632 0.103 = 0.103 0.157 with inverse
1 b V =
with c1 the 1 th quantile of the 2 q distribution. (Thus Fq (c1 ) = 1 .) These quantiles critical value table. can be found from the 2 q Theorem 6.16.2 implies Pr ( Cn ) Pr 2 q c1 = 1
A condence region Cn is a set estimator for Rq when q > 1. A condence region Cn is a set in Rq intended to cover the true parameter value with a pre-selected probability 1 . Thus an ideal condence region has the coverage probability Pr( Cn ) = 1 . In practice it is typically not possible to construct a region with exact coverage, but we can calculate its asymptotic coverage. When the parameter estimate satises the conditions of Theorem 6.16.2, a good choice for a condence region is the ellipse Cn = { : Wn () c1 } .
Thus the Wald statistic is 0 1 b b V b Wn () = 0 1.77 1.16 11.8 1 11.8 1 = 1.2 2 1.2 2 1.16 7.13
1.77 1.16 1.16 7.13
= 1.77 (11.8 1 )2 2.32 (11.8 1 ) (1.2 2 ) + 7.13 (1.2 2 )2 .
160
Figure 6.8: Condence Region for Return to Experience and Return to Education
2 The 90% quantile of the 2 2 distribution is 4.605 (we use the 2 distribution as the dimension of is two), so an asymptotic 90% condence region for the two parameters is the interior of the ellipse Wn () = 4.605 which is displayed in Figure 6.8. Since the estimated correlation of the two coecient estimates is modest (about 0.3) the region is modestly elliptical.
6.19
Semiparametric Eciency in the Projection Model
b satises the central limit theorem. the set of distributions for which
In Section 4.6 we presented the Gauss-Markov theorem, which stated that in the homoskedastic CEF model, in the class of linear unbiased estimators the one with the smallest variance is leastsquares. As we noted in that section, the restriction to linear unbiased estimators is unsatisfactory as it leaves open the possibility that an alternative (non-linear) estimator could have a smaller asymptotic variance. In addition, the restriction to the homoskedastic CEF model is also unsatisfactory as the projection model is more relevant for empirical application. The question remains: what is the most ecient estimator of the projection coecient (or functions = h()) in the projection model? It turns out that it is straightforward to show that the projection model falls in the estimator class considered in Proposition 5.13.2. It follows that the least-squares estimator is semiparametrically ecient in the sense that it has the smallest asymptotic variance in the class of semiparametric estimators of . This is a more powerful and interesting result than the Gauss-Markov theorem. To see this, it is worth rephrasing Proposition 5.13.2 with amended notation. Suppose that a pa1 Pn b = n i=1 rameter of interest is = g( ) where = Ez i , for which the moment estimators are n o zi 2 b b ). Let L2 (g ) = F : E kz k < , g (u) is continuously dierentiable at u = Ez be and = g ( b is semiProposition 6.19.1 In the class of distributions F L2 (g ), parametrically ecient for in the sense that its asymptotic variance equals the semiparametric eciency bound.
161
b is asymptotically normal, Proposition 6.19.1 says that under the minimal conditions in which b. then no semiparametric estimator can have a smaller asymptotic variance than To show that an estimator is semiparametrically ecient it is sucient to show that it falls in the class covered by this Proposition. To show that the projection model falls in this class, we 1 0 write = Q xx Qxy = g () where = Ez i and z i = (xi xi , xi yi ) . The class L2 (g ) equals the class of distributions n o > 0 . L4 ( ) = F : Ey4 < , E kxk4 < , Exi x0 i Proposition 6.19.2 In the class of distributions F L4 (), the leastb is semiparametrically ecient for . squares estimator The least-squares estimator is an asymptotically ecient estimator of the projection coecient because the latter is a smooth function of sample moments and the model implies no further restrictions. However, if the class of permissible distributions is restricted to a strict subset of L4 () then least-squares can be inecient. For example, the linear CEF model with heteroskedastic errors is a strict subset of L4 (), and the GLS estimator has a smaller asymptotic variance than OLS. In this case, the knowledge that true conditional mean is linear allows for more ecient estimation of the unknown parameter. b = h( b ) are semiparametFrom Proposition 6.19.1 we can also deduce that plug-in estimators rically ecient estimators of = h() when h is continuously dierentiable. We can also deduce that other parameters estimators are semiparametrically ecient, such as 2 for 2 . To see this, note that we can write 2 2 = E yi x0 i 2 0 0 = Eyi 2E yi x0 i + E xi xi
1 = Qyy Qyx Q xx Qxy
which is a smooth function of the moments Qyy , Qyx and Qxx . Similarly the estimator 2 equals 1X 2 = e i n
2 i=1 n
2 , y x0 and x x0 all have nite variances when F L ( ), the conditions of Since the variables yi i i i i 4 Proposition 6.19.1 are satised. We conclude:
b Q b b yx Q byy Q =Q xx xy
Proposition 6.19.3 In the class of distributions F L4 (), 2 is semi2 parametrically ecient for .
162
6.20
Semiparametric Eciency in the Homoskedastic Regression Model*
In Section 6.19 we showed that the OLS estimator is semiparametrically ecient in the projection model. What if we restrict attention to the classical homoskedastic regression model? Is OLS still ecient in this class? In this section we derive the asymptotic semiparametric eciency bound for this model, and show that it is the same as that obtained by the OLS estimator. Therefore it turns out that least-squares is ecient in this class as well. Recall that in the homoskedastic regression model the asymptotic variance of the OLS estimator b for is V 0 = Q1 2 . Therefore, as described in Section 5.13, it is sucient to nd a parametric xx 0 submodel whose Cramer-Rao bound for estimation of is V 0 . This would establish that V is b is semiparametrically ecient for . the semiparametric variance bound and the OLS estimator Let the joint density of y and x be written as f (y, x) = f1 (y | x) f2 (x) , the product of the conditional density of y given x and the marginal density of x. Now consider the parametric submodel (6.51) f (y, x | ) = f1 (y | x) 1 + y x0 x0 / 2 f2 (x) .
You can check that in thissubmodel the marginal density of x is f2 (x) and the conditional density 0 0 2 the latter is a valid conditional of y given x is f1 (y | x) 1 + (y x ) (x ) / . To see that R density, observe that the regression assumption implies that yf1 (y | x) dy = x0 and therefore Z f1 (y | x) 1 + y x0 x0 / 2 dy Z Z = f1 (y | x) dy + f1 (y | x) y x0 dy x0 /2 = 1.
In this parametric submodel the conditional mean of y given x is Z E (y | x) = yf1 (y | x) 1 + y x0 x0 / 2 dy Z Z = yf1 (y | x) dy + yf1 (y | x) y x0 x0 / 2 dy Z Z 2 y x0 f1 (y | x) x0 / 2 dy = yf1 (y | x) dy + Z + y x0 f1 (y | x) dy x0 x0 / 2 = x0 ( + ) , R using the homoskedasticity assumption (y x0 )2 f1 (y | x) dy = 2 . This means that in this parametric submodel, the conditional mean is linear in x and the regression coecient is () = + . We now calculate the score for estimation of . Since x (y x0 ) / 2 log f (y, x | ) = log 1 + y x0 x0 /2 = 1 + (y x0 ) (x0 ) / 2 s=
the score is
log f (y, x | 0 ) = xe/2 . The Cramer-Rao bound for estimation of (and therefore () as well) is 1 0 1 4 1 0 = E (xe) (xe)0 = 2 Q E ss xx = V .
163
We have shown that there is a parametric submodel (6.51) whose Cramer-Rao bound for estimation of is identical to the asymptotic variance of the least-squares estimator, which therefore is the semiparametric variance bound.
Theorem 6.20.1 In the homoskedastic regression model, the semipara1 metric variance bound for estimation of is V 0 = 2 Q xx and the OLS estimator is semiparametrically ecient.
This result is similar to the Gauss-Markov theorem, in that it asserts the eciency of the leastsquares estimator in the context of the homoskedastic regression model. The dierence is that the Gauss-Markov theorem states that OLS has the smallest variance among the set of unbiased linear estimators, while Theorem 6.20.1 states that OLS has the smallest asymptotic variance among all regular estimators. This is a much more powerful statement.
6.21
Uniformly Consistent Residuals*
It seems natural to view the residuals e i as estimates of the unknown errors ei . Are they consistent estimates? In this section we develop an appropriate convergence result. This is not a widely-used technique, and can safely be skipped by most readers. Notice that we can write the residual as b e i = yi x0 i
b = Op (n1/2 ) from Theorem 6.3.2, but we also need to To bound (6.53) we can use bound the random variable kxi k. If the regressor is bounded, that is, kxi k B < , then b |e i ei | B = Op (n1/2 ). However if the regressor does not have bounded support then we have to be more careful. The key is Theorem 5.12.1 which shows that E kxi kr < implies xi = op n1/r uniformly in i, or p n1/r max kxi k 0.
1in
b 0 it seems reasonable to guess that e i will be close to ei if n is large. Since We can bound the dierence in (6.52) using the Schwarz inequality (A.10) to nd b b k x k |e i ei | = x0 . i i
0b = ei + x0 i xi b . = ei x0 i
(6.52)
(6.53)
Applied to (6.53) we obtain

1in
b i ei | max kxi k max |e = op (n1/2+1/r ).

1in
We have shown the following.
CHAPTER 6. ASYMPTOTIC THEORY FOR LEAST SQUARES Theorem 6.21.1 Under Assumption 6.1.2 and E kxi kr < , then uniformly in 1 i n (6.54) e i = ei + op (n1/2+1/r ).
164
The rate of convergence in (6.54) depends on r. Assumption 6.1.2 requires r 4, so the rate of convergence is at least op (n1/4 ). As r increases, the rate improves. As a limiting case, from Theorem 5.12.1 we see if E exp( t0 xi ) < for all ktk < then xi = op (log n) uniformly in i, that and thus e i = ei + op n1/2 log n . We mentioned in Section 6.7 that there are multiple ways to prove the consistent of the cob . We now show that Theorem 6.21.1 provides one simple method to variance matrix estimator i ei | = op (n1/4 ). Since establish (6.32) and thus Theorem 6.7.1. Let qn = max1in |e
2 ei ei ) + ( ei ei )2 , e 2 i ei = 2ei (
then n n 1 X 2 2 1 X 0 2 xi x0 i e2 xi xi e i ei i e i n n
i=1 i=1 i=1
n n 2X 1X kxi k2 |ei | |e i ei | + kxi k2 |e i ei |2 n n i=1 n n 2X 1X 2 kxi k2 |ei | qn + kxi k2 qn n n
op (n
i=1 1/4
i=1
).
6.22
Asymptotic Leverage*
These are the diagonal elements of the projection matrix P and appear in the formula for leaveone-out prediction errors and several covariance matrix estimators. We can show that under iid sampling the leverage values are uniformly asymptotically small. Let min (A) and max (A) denote the smallest and largest eigenvalues of a symmetric square matrix A, and note that max (A1 ) = (min (A))1 . 1 0 p p 1 Since n X 0 X Qxx > 0 then by the CMT, min n X X min (Qxx ) > 0. (The latter is positive since Qxx is positive denite and thus all its eigenvalues are positive.) Then by the Trace Inequality (A.13) 0 1 hii = x0 xi i X X ! 1 1 0 1 0 XX xi xi = tr n n 1 ! 1 0 1 max XX xi x0 tr i n n 1 1 1 0 XX kxi k2 = min n n 1 (min (Qxx ) + op (1))1 max kxi k2 . (6.55) n 1in
Recall the denition of leverage from (3.21) 0 1 hii = x0 xi . i X X
165
Theorem 5.12.1 shows that E kxi kr < implies max1in kxi k = op n1/r and thus (6.55) is op n2/r1 . Theorem 6.22.1 If xi is independent and identically distributed and Ek xi kr < for some r 2, then uniformly in 1 i n, hii = op n2/r1 . For any r 2 then hii = op (1) (uniformly in i n). Larger r implies a stronger rate of convergence, for example r = 4 implies hii = op n1/2 . Theorem (6.22.1) implies that under random sampling with nite variances and large samples, no individual observation should have a large leverage value. Consequently individual observations should not be inuential, unless one of these conditions is violated.
166
Exercises
0 Exercise 6.1 Take the model yi = x0 1i 1 + x2i 2 + ei with Exi ei = 0. Suppose that 1 is estimated by regressing yi on x1i only. Find the probability limit of this estimator. In general, is it consistent for 1 ? If not, under what conditions is this estimator consistent for 1 ?
Exercise 6.2 Let y be n 1, X be n k (rank k). y = X + e with E(xi ei ) = 0. Dene the ridge regression estimator !1 n ! n X X b= xi x0 xi yi (6.56) i + I k
i=1 i=1
b as n . Is b consistent for ? where > 0 is a xed constant. Find the probability limit of
Exercise 6.3 For the ridge regression estimator (6.56), set = cn where c > 0 is xed as n . b as n . Find the probability limit of Pr (x1i = x2i = 1) = Pr (x1i = x2i = 1) = 3/8
Exercise 6.4 Verify some of the calculations reported in Section 6.4. Specically, suppose that x1i and x2i only take the values {1, +1}, symmetrically, with Pr (x1i = 1, x2i = 1) = Pr (x1i = 1, x2i = 1) = 1/8 5 E e2 i | x1i = x2i = 4 2 1 E ei | x1i 6= x2i = . 4
Verify the following: (a) Ex1i = 0 (b) Ex2 1i = 1 (c) Ex1i x2i = 1 2
(d) E e2 i =1 2 2 (e) E x1i ei = 1
Exercise 6.5 Show (6.20)-(6.23). Exercise 6.6 The model is yi = x0 i + ei E (xi ei ) = 0 2 = E xi x0 i ei .
7 (f) E x1i x2i e2 i = . 8
b, b ) for (, ) . Find the method of moments estimators ( (b) If so, in what sense are they ecient?
b, b ) ecient estimators of (, )? (a) In this model, are (
167
, y , x ) only the pair (y , x ) are observed. In this case, we say Exercise 6.7 Of the variables (yi i i i i that yi is a latent variable. Suppose = x0 yi i + ei
E (xi ei ) = 0
yi = yi + ui
where ui is a measurement error satisfying E (xi ui ) = 0

ui ) = 0 E (yi
b denote the OLS coecient from the regression of yi on xi . Let (a) Is the coecient from the linear projection of yi on xi ? b consistent for as n ? (b) Is b n as n .
(c) Find the asymptotic distribution of
Exercise 6.8 Find the asymptotic distribution of Exercise 6.9 The model is
2 n 2 as n .
yi = xi + ei E (ei | xi ) = 0 where xi R. Consider the two estimators Pn i=1 xi yi b = P n 2 i=1 xi n X yi e= 1 . n xi

i=1
(a) Under the stated assumptions, are both estimators consistent for ? (b) Are there conditions under which either estimator is ecient? Exercise 6.10 In the homoskedastic regression model y = X + e with E(ei | xi ) = 0 and 2 b b , based E(e2 i | xi ) = , suppose is the OLS estimate of with covariance matrix estimate V on a sample of size n. Let 2 be the estimate of 2 . You wish to forecast an out-of-sample value of yn+1 given that xn+1 = x. Thus the available information is the sample (y , X ), the estimates b, V b , ( 2 ), the residuals e , and the out-of-sample value of the regressors, xn+1 . (a) Find a point forecast of yn+1 . (b) Find an estimate of the variance of this forecast Exercise 6.11 As in Exercise 3.21, use the CPS dataset and the subsample of white male Hispanics. Estimate the regression \ log( W age) = 1 education + 2 experience + 3 experience2 /100 + 4 . (a) Report the coecients and robust standard errors.
168
(b) Let be the ratio of the return to one year of education to the return to one year of experib from the ence. Write as a function of the regression coecients and variables. Compute estimated model.
(d) Construct a 90% asymptotic condence interval for from the estimated model.
b as a function of the covariance (c) Write out the formula for the asymptotic standard error for b b matrix for . Compute s () from the estimated model. (e) Compute the regression function at edu = 12 and experience=20. Compute a 95% condence interval for the regression function at this point (f) Consider an out-of-sample individual with 16 years of education and 5 years experience. Construct an 80% forecast interval for their log wage and wage. [To obtain the forecast interval for the wage, apply the exponential function to both endpoints.]
Chapter 7
Restricted Estimation
7.1 Introduction
yi = x0 i + ei E (xi ei ) = 0 a common task is to impose on the coecient vector . For example, partitioning a constraint 0 , x0 ) and 0 = 0 , 0 , a typical constraint is an exclusion restriction of the form = ( x x0 1 2 i 1i 2i 2 = 0. In this case the constrained model is yi = x0 1i 1 + ei E (xi ei ) = 0 At rst glance this appears the same as the linear projection model, but there is one important 0 0 dierence: the error ei is uncorrelated with the entire regressor vector x0 i = (x1i , x2i ) not just the included regressor x1i . In general, a set of q linear constraints on takes the form R0 = c (7.1)
In the linear projection model
where R is k q, rank(R) = q < k and c is q 1. The assumption that R is full rank means that the constraints are linearly independent (there are no redundant or contradictory constraints). The constraint 2 = 0 discussed above is a special case of the constraint (7.1) with 0 R= , (7.2) I a selector matrix, and c = 0. Another common restriction is that a set of coecients sum to a known constant, i.e. 1 +2 = 1. This constraint arises in a constant-return-to-scale production function. Other common restrictions include the equality of coecients 1 = 2 , and equal and osetting coecients 1 = 2 . A typical reason to impose a constraint is that we believe (or have information) that the constraint is true. By imposing the constraint we hope to improve estimation eciency. The goal is to obtain consistent estimates with reduced variance relative to the unconstrained estimator. The questions then arise: How should we estimate the coecient vector imposing the linear restriction (7.1)? If we impose such constraints, what is the sampling distribution of the resulting estimator? How should we calculate standard errors? These are the questions explored in this chapter. 169
CHAPTER 7. RESTRICTED ESTIMATION
170
7.2
Constrained Least Squares
An intuitively appealing method to estimate a constrained linear projection is to minimize the least-squares criterion subject to the constraint R0 = c. The constrained least-squares estimator is e = argmin SSEn ( ) cls
R0 =c n X 2 = y 0 y 2y 0 X + 0 X 0 X . yi x0 SSEn ( ) = i i=1
(7.3)
where
(7.4)
e minimizes the sum of squared errors over all such that B R , or equivalently The estimator cls e the constrained least-squares (CLS) estimator. such that the restriction (7.1) holds. We call cls e is a restricted We follow the convention of using a tilde ~ rather than a hat ^ to indicate that cls b e cls to be clear estimator in contrast to the unrestricted least-squares estimator , and write it as that the estimation method is CLS. One method to nd the solution to (7.3) uses the technique of Lagrange multipliers. The problem (7.3) is equivalent to the minimization of the Lagrangian 1 L(, ) = SSEn () + 0 R0 c 2 (7.5)
over (, ), where is an s 1 vector of Lagrange multipliers. The rst-order conditions for minimization of (7.5) are e , e cls ) = X 0 y + X 0 X e + R e cls = 0 L( cls cls e , e cls ) = R0 e c = 0. L( cls
1
(7.6)
and
(7.7)
Premultiplying (7.6) by R0 (X 0 X )
we obtain
e cls c = 0 from b = (X 0 X ) X 0 y is the unrestricted least-squares estimator. Imposing R0 where e (7.7) and solving for cls we nd
1
e cls = 0 b + R0 e + R0 X 0 X 1 R R0 cls h i1 e cls = R0 X 0 X 1 R b c . R0
(7.8)
e cls we nd the solution to the constrained Substuting this expression into (7.6) and solving for minimization problem (7.3) h i1 0b e = b X 0 X 1 R R0 X 0 X 1 R c . R cls (7.9)
(See Exercise 7.4 to verify that (3.4) satises (7.1).) This is a general formula for the CLS estimator. It also can be written as 1 1 0 b 1 0b b Q e = b R R R R c . Q cls xx xx e the residuals are Given cls e e i = yi x0 i cls .
(7.10)
(7.11)
CHAPTER 7. RESTRICTED ESTIMATION The moment estimator of 2 is

2 cls
171
A bias-corrected version of
2 cls
is s2 cls =
1X 2 = e i . n
i=1
2 so that s2 cls is unbiased for .
You can show (See Exercise 7.6) that in the homoskedastic linear regression model under (7.1), 2 (7.12) E s2 cls | X =
n X 1 e 2 i. nk+q i=1
7.3
Exclusion Restriction
While (7.9) is a general formula for the CLS estimator, in most cases the estimator can be found by applying least-squares to a reparameterized equation. To illustrate, let us return to the rst example presented at the beginning of the chapter a simple exclusion restriction. Recall the unconstrained model is 0 (7.13) yi = x0 1i 1 + x2i 2 + ei the exclusion restriction is 2 = 0, and the constrained equation is yi = x0 1i 1 + ei . (7.14)
It is not immediately obvious, but (7.9) and (7.16) are algebraically (and numerically) equivalent. To see this, the rst component of (7.9) with (7.2) is " # 1 1 0 1 0 b Q b . e = I 0 b b 0 I Q 0 I 1 xx xx I I Using (3.33) this equals 22 1 12 b b Q e = b b Q 1 1 2
0 The CLS estimator of the entire vector = 0 1 , 2 is e 1 e= . 0

0
In this setting the CLS estimator is OLS of yi on x1i . (See Exercise 7.1.) We can write this as n !1 n ! X X 0 e = x1i x1i x1i yi . (7.15) 1
i=1 i=1
(7.16)
b +Q b b b 1 b b 1 Q = 1 112 12 Q22 Q221 2 b 1y Q b 1 Q b 2y b 1 Q b 12 Q =Q 112 22 1 1 1 1 b 112 Q b b 12 Q b 22 Q b 221 Q b b b b +Q Q Q Q Q 2y 21 11 1y 221 1 1 1 b b b b b b b =Q 112 Q1y Q12 Q22 Q21 Q11 Q1y 1 1 1 b b b b b b b Q =Q Q Q Q 11 12 22 21 Q11 Q1y 112
1 b b =Q 11 Q1y
which is (7.16) as originally claimed.
172
7.4
Minimum Distance
b and . Jn () is small if is close to b, This is a (squared) weighted Euclidean distance between b e and is minimized at zero only if = . A minimum distance estimator md for minimizes Jn () subject to the constraint (7.1), that is, e = argmin Jn () . md
R0 =c
A minimum distance estimator tries to nd a parameter value which satises the constraint b be the unconstrained leastwhich is as close as possible to the unconstrained estimate. Let squares estimator, and for some k k positive denite weight matrix W n > 0 dene the quadratic criterion function 0 b b Wn (7.17) Jn ( ) = n
(7.18)
To see the equality of CLS and minimum distance, rewrite the least-squares criterion as follows. b i and substitute this equation Write the unconstrained least-squares tted equation as yi = x0 i + e into SSEn ( ) to obtain
n X 2 yi x0 SSEn () = i i=1 i=1 n X i=1
b xx , and we write this criterion function as The CLS estimator is the special case when W n = Q 0 0 b Q b b xx () = n Jn (7.19)
n 2 X 0 b +e = x x0 i i i
e 2 i
0 = n 2 + Jn ()
n ! 0 X 0 b b + xi xi
i=1
(7.20)
P P 0 b where the third equality uses the fact that n i = 0, and the last line uses n i=1 xi e i=1 xi xi = nQxx 0 . The expression (7.20) only depends on through Jn () . Thus minimization of SSEn () and 0 ( ) are equivalent, and hence e = e when W n = Q b xx . Jn md cls e We can solve for md explicitly by the method of Lagrange multipliers. The Lagrangian is which is minimized over ( , ). The solution is 1 L( , ) = Jn (, W n ) + 0 R0 c 2
e md specializes to e cls when we (See Exercise 7.7.) Comparing (7.22) with (7.10) we can see that b set W n = Qxx . An obvious question is which weight matrix W n is best. We will address this question after we derive the asymptotic distribution for a general weight matrix.
e md = n R0 W 1 R 1 R0 b c n e md = b c . b W 1 R R0 W 1 R 1 R0 n n
(7.21) (7.22)
173
7.5
Asymptotic Distribution
We rst show that the class of minimum distance estimators are consistent for the population parameters when the constraints are valid.
Assumption 7.5.1 R0 = c where R is k q with rank(R) = q.
Assumption 7.5.2 W n W > 0.
Theorem 7.5.1 Consistency p e as n . Under Assumptions 6.1.1, 7.5.1, and 7.5.2, md For a proof, see Exercise 7.8. Theorem 7.5.1 shows that consistency holds for any weight matrix with a positive denite limit, so the result includes the CLS estimator. Similarly, the constrained estimators are asymptotically normally distributed.
as n , where
Theorem 7.5.2 Asymptotic Normality Under Assumptions 6.1.2, 7.5.1, and 7.5.2, d e md n N (0, V (W )) 1 0 RV V (W ) = V W 1 R R0 W 1 R 0 1 1 0 1 V R R W R RW 1 0 1 1 +W 1 R R0 W 1 R R0 V R R0 W 1 R RW
(7.23)
(7.24)
1 1 and V = Q xx Qxx .
For a proof, see Exercise 7.9. Theorem 7.5.2 shows that the minimum distance estimator is asymptotically normal for all positive denite weight matrices. The asymptotic variance depends on W . The theorem includes the CLS estimator as a special case by setting W = Qxx .
CHAPTER 7. RESTRICTED ESTIMATION Theorem 7.5.3 Asymptotic Distribution of CLS Estimator Under Assumptions 6.1.2 and 7.5.1, as n d e cls n N (0, V cls ) where 0 1 1 0 1 RV V cls = V Q xx R R Qxx R 0 1 1 0 1 V R R Qxx R R Qxx 0 1 1 0 1 0 1 1 1 + Qxx R R Qxx R R V R R0 Q R Qxx xx R For a proof, see Exercise 7.10.
174
7.6
Ecient Minimum Distance Estimator
b 1 the ecient minimum distance We call the minimum distance estimator setting W n = V estimator and takes the form 1 b V b c . e emd = b R b R R0 V R0 (7.25)
Theorem 7.5.2 shows that the minimum distance estimators, which include CLS as a special case, are asymptotically normal with an asymptotic covariance matrix which depends on the weight matrix W . The asymptotically optimal weight matrix is the one which minimizes the asymptotic 1 1 variance V (W ). This turns out to be W = V as is shown in Theorem 7.6.1 below. Since V 1 is unknown this weight matrix cannot be used for a feasible estimator, but we can replace V 1 b with a consistent estimate V and the asymptotic distribution (and eciency) are unchanged.
The asymptotic distribution of (7.25) can be deduced from Theorem 7.5.2. (See Exercises 7.11 and 7.12.)
Theorem 7.6.1 Ecient Minimum Distance Estimator Under Assumptions 6.1.2 and 7.5.1, d e emd n N 0, V as n , where 0 1 0 V R V . = V V R R V R V V (7.26)
Since
(7.27)
the estimator (7.25) has lower asymptotic variance than the unrestricted estimator. Furthermore, for any W , V V (W ) (7.28)
so (7.25) is asymptotically ecient in the class of minimum distance estimators.
175
Theorem 7.6.1 shows that the minimum distance estimator with the smallest asymptotic variance is (7.25). One implication is that the constrained least squares estimator is generally inecient. The interesting exception is the case of conditional homoskedasticity, in which case the 1 so in this case CLS is an ecient minimum distance estimator. optimal weight matrix is W = V 0 Otherwise when the error is conditionally heteroskedastic, there are asymptotic eciency gains by using minimum distance rather than least squares. The fact that CLS is generally inecient is counter-intuitive and requires some reection to understand. Standard intuition suggests to apply the same estimation method (least squares) to the unconstrained and constrained models, and this is the most common empirical practice. But Theorem 7.6.1 shows that this is not the ecient estimation method. Instead, the ecient minimum distance estimator has a smaller asymptotic variance. Why? The reason is that the least-squares estimator does not make use of the regressor x2i . It ignores the information E (x2i ei ) = 0. This information is relevant when the error is heteroskedastic and the excluded regressors are correlated with the included regressors. e Inequality (7.27) shows that the ecient minimum distance estimator emd has a smaller asb ymptotic variance than the unrestricted least squares estimator . This means that estimation is more ecient by imposing correct restrictions when we use the minimum distance method.
7.7
Exclusion Restriction Revisited

0 yi = x0 1i 1 + x2i 2 + ei
We return to the example of estimation with a simple exclusion restriction. The model is
with the exclusion restriction 2 = 0. We have introduced three estimators of 1 . The rst is unconstrained least-squares applied to (7.13), which can be written as
1 b =Q b b 1 112 Q1y 2 .
From Theorem 6.34 and equation (6.21) its asymptotic variance is b ) = Q1 11 Q Q1 21 12 Q1 Q + Q Q1 22 Q1 Q Q1 . avar( 1 12 22 21 12 22 21 112 22 22 112 The second estimator of 1 is the CLS estimator, which can be written as e b 1 b 1,cls = Q11 Q1y .
Its asymptotic variance can be deduced from Theorem 7.5.3, but it is simpler to apply the CLT directly to show that 1 1 e (7.29) avar( 1,cls ) = Q11 11 Q11 . The third estimator of 1 is the ecient minimum distance estimator. Applying (7.25), it equals e 1,md = b1 V b b 1 b 12 V 22 2 b = V " b 11 V b 12 V b 21 V b 22 V # . (7.30)
where we have partitioned
From Theorem 7.6.1 its asymptotic variance is
See Exercise 7.13 to verify equations (7.29), (7.30), and (7.31).
1 e avar( 1,md ) = V 11 V 12 V 22 V 21 .
(7.31)
176
In general, the three estimators are dierent, and they have dierent asymptotic variances. It is quite instructive to compare the asymptotic variances of the CLS and unconstrained leastsquares estimators to assess whether or not the constrained estimator is necessarily more ecient than the unconstrained estimator. First, consider the case of conditional homoskedasticity. In this case the two covariance matrices simplify to b 1 ) = 2 Q1 avar( 112 and If Q12 = 0 (so x1i and x2i are orthogonal) then these two variance matrices equal and the two 1 estimators have equal asymptotic eciency. Otherwise, since Q12 Q 22 Q21 0, then Q11 Q11 1 Q12 Q22 Q21 , and consequently 1 2 1 2 1 . Q 11 Q11 Q12 Q22 Q21
2 1 e avar( 1,cls ) = Q11 .
e This means that under conditional homoskedasticity, 1,cls has a lower asymptotic variance matrix b than 1 . Therefore in this context, constrained least-squares is more ecient than unconstrained least-squares. This is consistent with our intuition that imposing a correct restriction (excluding an irrelevant regressor) improves estimation eciency. However, in the general case of conditional heteroskedasticity this ranking is not guaranteed. In fact what is really amazing is that the variance ranking can be reversed. The CLS estimator can have a larger asymptotic variance than the unconstrained least squares estimator. To see this lets use the simple heteroskedastic example from Section 6.4. In that example, 1 7 Q11 = Q22 = 1, Q12 = , 11 = 22 = 1, and 12 = . We can calculate (see Exercise 7.14) that 2 8 3 Q112 = and 4 2 3 e 1,cls ) = 1 avar( b )= avar( 1 (7.32) (7.33)
e 1,md ) = 5 . avar( (7.34) 8 e Thus the restricted least-squares estimator 1,cls has a larger variance than the unrestricted leastb ! The minimum distance estimator has the smallest variance of the three, as squares estimator 1 expected. What we have found is that when the estimation method is least-squares, deleting the irrelevant variable x2i can actually increase estimation variance, or equivalently, adding an irrelevant variable can actually decrease the estimation variance. To repeat this unexpected nding, we have shown in a very simple example that it is possible for least-squares applied to the short regression (7.14) to be less ecient for estimation of 1 than least-squares applied to the long regression (7.13), even though the constraint 2 = 0 is valid! This result is strongly counter-intuitive. It seems to contradict our initial motivation for pursuing constrained estimation to improve estimation eciency. It turns out that a more rened answer is appropriate. Constrained estimation is desirable, but not constrained least-squares estimation. While least-squares is asymptotically ecient for estimation of the unconstrained projection model, it is not an ecient estimator of the constrained projection model.
177
7.8
Variance and Standard Error Estimation
The asymptotic covariance matrix (7.26) may be estimated by replacing V with a consistent b . This variance estimator is estimates such as V 1 b = V b R b . b V b R R0 V V R0 V (7.35)
7.9
e so long as h does not lie in We can calculate standard errors for any linear combination h0 0e the range space of R. A standard error for h is 1/2 e ) = n1 h0 V b s(h0 h . (7.36)
Misspecication
e if the constraint (7.1) is incorrect? What are the consequences for a constrained estimator To be specic, suppose that R0 = c where c is not necessarily equal to c. This situation is a generalization of the analysis of omitted variable bias from Section 2.23, where we found that the short regression (e.g. (7.15)) is estimating a dierent projection coecient than the long regression (e.g. (7.13)). One mechanical answer is that we can use the formula (7.22) for the minimum distance estimator to nd that 1 p 1 e md R R0 W 1 R (c c) . (7.37) md = W 1 (c c), shows that imposing an incorrect constraint leads The second term, W 1 R R0 W 1 R to inconsistency an asymptotic bias. We can call the limiting value md the minimum-distance projection coecient or the pseudo-true value implied by the restriction. However, we can say more. For example, we can describe some characteristics of the approximating projections. The CLS estimator projection coecient has the representation 2 0 cls = argmin E yi xi ,
R0 =c
the best linear predictor subject to the constraint (7.1). The minimum distance estimator converges to 0 md = argmin ( 0 ) W ( 0 )
R0 =c
where 0 is the true coecient. That is, md is the coecient vector satisfying (7.1) closest to the true value n the weighted Euclidean norm. These calculations show that the constrained estimators are still reasonable in the sense that they produce good approximations to the true coecient, conditional on being required to satisfy the constraint. e We can also show that md has an asymptotic normal distribution. The trick is to dene the pseudo-true value 0 1 1 1 (c c) . (7.38) n = Wn R R Wn R (Note that (7.37) and (7.38) are dierent!) Then e md = n b W 1 R R0 W 1 R 1 n R0 b c n n n n 1 1 0 1 b R0 n = I W n R R Wn R 1 0 d I W 1 R R0 W 1 R R N (0, V ) = N (0, V (W )) .
(7.39)
CHAPTER 7. RESTRICTED ESTIMATION In particular d e n emd n N 0, V .
178
This means that even when the constraint (7.1) is misspecied, the conventional covariance matrix estimator (7.35) and standard errors (7.36) are appropriate measures of the sampling variance, though the distributions are centered at the pseudo-true values (or projections) n rather than . The fact that the estimators are biased is an unavoidable consequence of misspecication. An alternative approach to the asymptotic distribution theory under misspecication uses the concept of local alternatives. It is a technical device which might seem a bit articial, but it is a powerful method to derive useful distributional approximations in a wide variety of contexts. The idea is to index the true coecient n by n via the relationship R0 n = c + n1/2 . (7.40)
Equation (7.40) species that n violates (7.1) and thus the constraint is misspecied. However, the constraint is close to correct, as the dierence R0 n c = n1/2 is small in the sense that it decreases with the sample size n. We call (7.40) local misspecication. The asymptotic theory is then derived as n under the sequence of probability distributions with the coecients n . The way to think about this is that the true value of the parameter is n , and it is close to satisfying (7.1). The reason why the deviation is proportional to n1/2 is because this is the only choice under which the localizing parameter appears in the asymptotic distribution but does not dominate it. The best way to see this is to work through the asymptotic approximation. Since n is the true coecient value, then yi = x0 i n +ei and we have the standard representation for the unconstrained estimator, namely n !1 ! n X X 1 1 0 b = n xi xi xi ei n n n
i=1 i=1
N (0, V ) .
(7.41)
There is no dierence under xed (classical) or local asymptotics, since the right-hand-side is independent of the coecient n . A dierence arises for the constrained estimator. Using (7.40), c = R0 n n1/2 , so b + n1/2 b c = R0 R0 n and b W 1 R R0 W 1 R 1 R0 b c e md = n n b n + W 1 R R0 W 1 R 1 n1/2 . b W 1 R R0 W 1 R 1 R0 = n n n n 0 1 1 0 e 1 b n md n = I W R n n n R R Wn R 0 1 1 1 . + W n R R Wn R
It follows that
The rst term is asymptotically normal (from 7.41)). The second term converges in probability to a constant. This is because the n1/2 local scaling in (7.40) is exactly balanced by the n scaling of the estimator. No alternative rate would have produced this result. Consequently, we nd that the asymptotic distribution equals 1 d e md n n N (0, V ) + W 1 R R0 W 1 R = N ( , V (W )) (7.42)
CHAPTER 7. RESTRICTED ESTIMATION where 1 . = W 1 R R0 W 1 R
179
The asymptotic distribution (7.42) is an approximation of the sampling distribution of the restricted estimator under misspecication. The distribution (7.42) contains an asymptotic bias component . The approximation is not fundamentally dierent from (7.39) they both have the same asymptotic variances, and both reect the bias due to misspecication. The dierence is that (7.39) puts the bias on the left-side of the convergence arrow, while (7.42) has the bias on the right-side. There is no substantive dierence between the two, but (7.42) is more convenient for some purposes, such as the analysis of the power of tests, as we will explore in the next chapter.
7.10
Nonlinear Constraints
In some cases it is desirable to impose nonlinear constraints on the parameter vector . They can be written as r( ) = 0 (7.43) where r : Rk Rq . This includes the linear constraints (7.1) as a special case. An example of (7.43) which cannot be written as (7.1) is 1 2 = 1, which is (7.43) with r() = 1 2 1. The constrained least-squares and minimum distance estimators of subject to (7.43) solve the minimization problems e = argmin SSEn ( ) (7.44) cls
r()=0
where SSEn () and Jn () are dened in (7.4) and (7.17), respectively. The solutions minimize the Lagrangians 1 L(, ) = SSEn () + 0 r() (7.46) 2 or 1 L(, ) = Jn () + 0 r() (7.47) 2 over (, ). Computationally, there is in general no explicit expression for the solutions so they must be found numerically. Algorithms to numerically solve (7.44) and (7.45) are known as constrained optimization methods, and are available in programming languages including Matlab, Gauss and R.
e = argmin Jn ( ) md
r()=0
(7.45)
Assumption 7.10.1 r() = 0 with rank(R) = q, where R =
r()0 .
The asymptotic distribution is a simple generalization of the case of a linear constraint, but the proof is more delicate.
CHAPTER 7. RESTRICTED ESTIMATION e = Theorem 7.10.1 Under Assumptions 6.1.2, 7.10.1, and 7.5.2, for e e e md and = cls dened in (7.44) and (7.45), d e n N (0, V (W ))
180
e , W = Q and as n , where V (W ) is dened in (7.24). For cls xx V (W ) = V cls as dened in Theorem 7.5.3. V (W ) is minimized with 1 W =V , in which case the asymptotic variance is 0 1 0 R V . V = V V R R V R The asymptotic variance matrix for the ecient minimum distance estimator can be estimated by 0 1 0 b R b V b R b b V b b = V b V b R R V e )0 . b = r( R md (7.48)
where
e b Standard errors for the elements of = md are the square roots of the diagonal elements of V 1 b n V .
7.11
Inequality Restrictions
Inequality constraints on the parameter vector take the form r( ) 0 for some function r : Rk Rq . The most common example is a non-negative constraint 1 0. The constrained least-squares and minimum distance estimators can be written as e = argmin SSEn ( ) cls
r()0
(7.49)
(7.50)
and
Except in special cases the constrained estimators do not have simple algebraic solutions. An important exception is when there is a single non-negativity constraint, e.g. 1 0 with q = 1. In this case the constrained estimator can be found by two-step approach. First compute the e= b . Second, if b1 < 0 then impose 1 = 0 (eliminate b . If b1 0 then uncontrained estimator the regressor X1 ) and re-estimate. This yields the constrained least-squares estimator. While this method works when there is a single non-negativity constraint, it does not immediately generalize to other contexts. The computational problems (7.50) and (7.51) are examples of quadratic programming problems. Quick and easy computer algorithms are available in programming languages including Matlab, Gauss and R.
e = argmin Jn () . md
r()0
(7.51)
181
Inference on inequality-constrained estimators is unfortunately quite challenging. The conventional asymptotic theory gives rise to the following dichotomy. If the true parameter satises the strict inequality r() > 0, then asymptotically the estimator is not subject to the constraint and the inequality-constrained estimator has an asymptotic distribution equal to the unconstrained case. However if the true parameter is on the boundary, e.g. r() = 0, then the estimator has a trun cated structure. This is easiest to see in the one-dimensional case. If we have an estimator which d = max[, 0] satises n Z = N (0, V ) and = 0, then the constrained estimator d will have the asymptotic distribution n max[Z, 0], a half-normal distribution.
7.12
Constrained MLE
1 n log L( , 2 ) = log 2 2 2 SSEn (). 2 2
Recall that the log-likelihood function (3.44) for the normal regression model is
2 b cmle ) maximizes log L(, 2 ) The constrained maximum likelihood estimator (CMLE) ( cmle , subject to the constraint (7.43) Since log L( , 2 ) is a function of only through the sum of squared b b errors SSEn ( ), maximizing the likelihood is identical to minimizing SSEn (). Hence cmle = cls 2 2 and cmle = cls .
7.13
Technical Proofs*
Proof of Theorem 7.6.1, Equation (7.28). Let R be a full rank k (k q ) matrix satisfying R0 V R = 0 and then set C = [R, R ] which is full rank and invertible. Then we can calculate that R0 V R0 V 0 R R C V C = 0 R0 V R R V R 0 0 = 0 0 R V R and C 0 V (W )C R0 V R0 V (W )R (W )R = 0 R0 V (W )R R V (W )R 0 0 = . 1 1 0 0 0 R0 R0 V R (R0 W R) R0 W R V R + R W R (R W R) Thus C 0 V (W ) V C = C 0 V (W )C C 0 V C 0 0 = 1 0 1 0 0 R0 R V R (R0 W R) R0 W R W R (R W R) 0
Since C is invertible it follows that V (W ) V 0 which is (7.28).
e= e , as Proof of Theorem 7.10.1. We show the result for the minimum distance estimator md the proof for the constrained least-squares estimator is similar. For simplicity we assume that the
182
p e constrained estimator is consistent . This can be shown with more eort, but requires a deeper treatment than appropriate for this textbook. For each element rj () of the q -vector r(), by the mean value theorem there exists a j on e the line segment joining and such that
Let R n be the k q matrix R n =
e . e ) = rj () + rj ( )0 rj ( j r1 ( 1) r2 ( 2)
(7.52)
rq ( q)
p p p e it follows that Since j , and by the CMT, Rn R. Stacking the (7.52), we obtain
e ) = 0 by construction and r() = 0 by Assumption 7.5.1, this implies Since r( 0 e . 0 = R n The rst-order condition for (7.47) is b e =R e. b Wn
e . e ) = r() + R0 r( n
(7.53)
b is dened in (7.48). where R 1 Premultiplying by R0 W n , inverting, and using (7.53), we nd
Thus
1 1 0 b 0 1 b 0 b e e = R0 W 1 R b R R n n n = Rn W n R n . e = 1 1 b 0 1 b 0 b . I W n R Rn W n R Rn (7.54)
From Theorem 6.3.2 and Theorem 6.7.1 we nd 1 1 b 0 1 e 0 e b n = I W n R Rn W n R Rn n 1 0 d I W 1 R R0 W 1 R R N (0, V ) = N (0, V (W )) .
183
Exercises
Exercise 7.1 In the model y = X 1 1 + X 2 2 + e, show directly from denition (7.3) that the CLS estimate of = ( 1 , 2 ) subject to the constraint that 2 = 0 is the OLS regression of y on X 1. Exercise 7.2 In the model y = X 1 1 + X 2 2 + e, show directly from denition (7.3) that the CLS estimate of = (1 , 2 ), subject to the constraint that 1 = c (where c is some given vector) is the OLS regression of y X 1 c on X 2 . Exercise 7.3 In the model y = X 1 1 + X 2 2 + e, with X 1 and X 2 each n k, nd the CLS estimate of = ( 1 , 2 ), subject to the constraint that 1 = 2 . e cls dened in (7.9) that R0 e cls = c. Exercise 7.4 Verify that for b c = R0 (X 0 X ) (a) R0
1
Exercise 7.5 Let e e be the vector of constrained least-squares residuals (7.11). Show that under (7.1), X 0e 1 1 e = (X 0 X )1 X 0 e (X 0 X )1 R R0 (X 0 X )1 R (b) R0 (X 0 X ) X 0 e cls (d) Show that A is symmetric and idempotent, tr A = q, and P A = 0.
2 Exercise 7.6 Show (7.12), that is, E s2 cls | X = , under the assumptions of the homoskedastic regression model and (7.1). Hint: Use the results of Exercise 7.5 e md with W n = Exercise 7.7 Verify (7.21) and (7.22), and that the minimum distance estimator b Qxx equals the CLS estimator.
1 (c) e e = (I P + A) e for P = X (X 0 X ) X 0 and some matrix A (nd this matrix A).
Exercise 7.8 Prove Theorem 7.5.1. Exercise 7.9 Prove Theorem 7.5.2.
Exercise 7.10 Prove Theorem 7.5.3. (Hint: Use that CLS is a special case of Theorem 7.5.2.)
1 Exercise 7.11 Verify that (7.26) is V (W ) with W = V .
Exercise 7.12 Prove (7.27). Hint: Use (7.26). Exercise 7.13 Verify (7.29), (7.30) and (7.31) Exercise 7.14 Verify (7.32), (7.33), and (7.34).
184
Exercise 7.15 As in Exercise 6.11 and 3.21, use the CPS dataset and the subsample of white male Hispanics. (a) Estimate the regression \ log( W age) = 1 education + 2 experience + 3 experience2 /100 + 4 M arried1 + 5 M arried2 + 6 M arried3 + 7 W idowed + 8 Divorced + 9 Separated + 10 where M arried1 , M arried2 , and M arried3 are the rst three marital status codes as listed in Section 3.19. (b) Estimate the equation using constrained least-squares, imposing the constraints 4 = 7 and 8 = 9 , and report the estimates and standard errors (c) Estimate the equation using ecient minimum distance, imposing the same constraints, and report the estimates and standard errors (d) Under what constraint on the coecients is the wage equation non-decreasing in experience for experience up to 50? (e) Estimate the equation imposing 4 = 7 , 8 = 9 , and the inequality from part (d).
Chapter 8
Hypothesis Testing
8.1 Hypotheses
In Chapter 7 we discussed estimation subject to restrictions, including linear restrictions (7.1), nonlinear restrictions (7.43), and inequality restrictions (7.49). In this chapter we discuss tests of such restrictions. Hypothesis tests attempt to assess whether there is evidence to contradict a proposed parametric restriction. Let = r() be a q 1 parameter of interest where r : Rk Rq is some transformation. For example, may be a single coecient, e.g. = j , the dierence between two coecients, e.g. = j , or the ratio of two coecients, e.g. = j / . A point hypothesis concerning is a proposed restriction such as = 0 (8.1)
where 0 is a hypothesized value. More generally, letting B Rk be the parameter space, a hypothesis is a restriction B 0 where B 0 is a proper subset of B . This specializes to (8.1) by setting B 0 = { B : r() = 0 } . In this chapter we will focus primarily on point hypotheses of the form (8.1) as they are the most common and relatively simple to handle. The hypothesis to be tested is called the null hypothesis.
Denition 1 The null hypothesis, written H0 , is the restriction = 0 or B 0 .
We often write the null hypothesis as H0 : = 0 or H0 : r() = 0 . The complement of the null hypothesis (the collection of parameter values which do not satisfy the null hypothesis) is called the alternative hypothesis.
Denition 2 The alternative hypothesis, written H1 , is the set / B0} . { : 6= 0 } or { B :
185
CHAPTER 8. HYPOTHESIS TESTING
186
We often write the alternative hypothesis as H1 : 6= 0 or H0 : r() 6= 0 . For simplicity, we often refer to the hypotheses as the null and the alternative. In hypothesis testing, we assume that there is a true (but unknown) value of and this value either satises H0 or does not satisfy H0 . The goal of hypothesis testing is to assess whether or not H0 is true, by asking if H0 is consistent with the observed data. To be specic, take our example of wage determination and consider the question: Does union membership aects wages? We can turn this into a hypothesis test by specifying the null as the restriction that a coecient on union membership is zero in a wage regression. Consider, for example, the estimates reported in Table 4.1. The coecient for Male Union Member is 0.095 (a wage premium of 9.5%) and the coecient for Female Union Member is 0.022 (a wage premium of 2.2%). These are estimates, not the true values. The question is: Are the estimates close enough to zero that they are compatible with the hypothesis that the true values of the coecients are zero? Or are the estimates suciently dierent from zero that this hypothesis must be false?
8.2
Acceptance and Rejection
A hypothesis test either accepts the null hypothesis, or rejects the null hypothesis in favor of the alternative hypothesis. We can describe these two decisions as Accept H0 and Reject H0 . In the example given in the previous section, the decision would be either to accept the hypothesis that union membership does not aect wages, or to reject the hypothesis in favor of the alternative that union membership does aect wages. The decision is based on the data, and so is a mapping from the sample space to the decision set. This splits the sample space into two regions S0 and S1 such that if the observed sample falls into S0 we accept H0 , while if the sample falls into S1 we reject H0 . The set S0 can be called the acceptance region and the set S1 the rejection or critical region. It is convenient to express this mapping as a real-valued function called a test statistic Tn = Tn ((y1 , x1 ) , ..., (yn , xn )) relative to a critical value c. The hypothesis test then consists of the decision rule 1. Accept H0 if Tn c, 2. Reject H0 if Tn > c. A test statistic Tn should be designed so that small values are likely when H0 is true and large values are likely when H1 is true. There is a well developed statistical theory concerning the design of optimal tests. We will not review that theory here, but instead refer the reader to Lehmann and Romano (2005). In this chapter we will summarize the main approaches to the design of test statistics. The most commonly used test statistic is the absolute value of the t-statistic tn = |tn (0 )| where tn () = b b) s( (8.2)
(8.3)
b is a point estimate and s( b) its standard error. tn is an is the t-statistic from (6.44), where appropriate statistic when testing hypotheses on individual coecients or real-valued parameters = h() and 0 is the hypothesized value. Quite typically, 0 = 0, as interest focuses on whether or not a coecient equals zero, but this is not the only possibility. For example, interest may focus on whether an elasticity equals 1, in which case we may wish to test H0 : = 1.
187
8.3
Type I Error
A false rejection of the null hypothesis H0 (rejecting H0 when H0 is true) is called a Type I error. The probability of a Type I error is Pr (Reject H0 | H0 true) = Pr (Tn > c | H0 true) . (8.4)
The nite sample size of the test is dened as the supremum of (8.4) across all data distributions which satisfy H0 . A primary goal of test construction is to limit the incidence of Type I error by bounding the size of the test. For the reasons discussed in Chapter 6, in typical econometric models the exact sampling distributions of estimators and test statistics are unknown and hence we cannot explicitly calculate (8.4). Instead, we typically rely on asymptotic approximations. We start by assuming that the test statistic has an asymptotic distribution under H0 . That is, when H0 is true Tn T
d
(8.5)
as n for some continuously-distributed random variable T . This is not a substantive restriction, as most conventional econometric tests satisfy (8.5). Let G(u) = Pr (T u) denote the distribution of T . We call T (or G) the asymptotic null distribution. It is generally desirable to design test statistics Tn whose asymptotic null distribution G is known and does not depend on unknown parameters. In this case we say that the statistic Tn is asymptotically pivotal. For example, if Tn = tn , the absolute t-statistic from (8.2), then we know from Theorem 6.12.1 that if = 0 (that is, the null hypothesis holds), then tn |Z| as n where Z N(0, 1). This means that G(u) = Pr (T |Z|) = (u), the symmetrized normal distribution function dened in (6.45). This distribution does not depend on unknowns and is pivotal. We dene the asymptotic size of the test as the asymptotic probability of a Type I error:
n d
lim Pr (Tn > c | H0 true) = Pr (T > c) = 1 G(c).
We see that the asymptotic size of the test is a simple function of the asymptotic null distribution G and the critical value c. For example, the asymptotic size of a test based on the absolute t-statistic with critical value c is 1 (c). In the dominant approach to hypothesis testing, the researcher pre-selects a signicance level (0, 1) and then selects c so that the (asymptotic) size is no larger than . When the asymptotic null distribution G is pivotal, we can accomplish this by setting c equal to the (1 )th quantile of the distribution G. (If the distribution G is not pivotal, more complicated methods must be used, pointing out the great convenience of using asymptotically pivotal test statistics.) We call c the asymptotic critical value because it has been selected from the asymptotic null distribution. For example, since (1.96) = 0.95, it follows that the 5% asymptotic critical value for the absolute t-statistic is c = 1.96.
8.4
t tests
As we mentioned earlier, the most common test of the one-dimensional hypothesis H0 : = 0 against the alternative H1 : 6= 0 (8.7) (8.6)
188
is the absolute value of the t-statistic (8.3). We now formally state its asymptotic null distribution, which is a simple application of Theorem 6.12.1.
Theorem 8.4.1 Under Assumptions 6.1.2, 6.10.1, and H0 : = 0 , tn (0 ) Z. For c satisfying = 2 (1 (c)) , Pr (|tn (0 )| > c | H0 ) so the test Reject H0 if |tn (0 )| > c has asymptotic size .
d
The theorem shows that asymptotic critical values can be taken from the normal distribution table. The alternative hypothesis (8.7) is sometimes called a two-sided alternative. Sometimes we are interested in testing for one-sided alternatives such as H1 : > 0 or H1 : < 0 . (8.9) Tests of (8.6) against (8.8) or (8.9) are based on the signed t-statistic tn = tn (0 ). The hypothesis (8.6) is rejected in favor of (8.8) if tn > c where c satises = 1 (c). Negative values of tn are b less than 0 do not point to (8.8). Since the not taken as evidence against H0 , as point estimates critical values are taken from the single tail of the normal distribution, they are smaller than for two-sided tests. Specically, the asymptotic 5% critical value is = 1.645. Thus, we reject (8.6) in favor of (8.8) if tn > 1.645. Conversely, tests of (8.6) against (8.9) reject H0 for negative t-statistics, e.g. if tn c. For this alternative large positive values of tn are not evidence against H0 . An asymptotic 5% test rejects if tn < 1.645. There seems to be an ambiguity. Should we use the two-sided critical value 1.96 or the onesided critical value 1.645? The answer is that we should use one-sided tests and critical values only when the parameter space is known to satisfy a one-sided restriction such as 0 . This is when the test of (8.6) against (8.8) makes sense. If the restriction 0 is not known a priori, then imposing this restriction to test (8.6) against (8.8) does not makes sense. Since linear regression coecients typically do not have a priori sign restrictions, we conclude that two-sided tests are generally appropriate. (8.8)
8.5
Type II Error and Power
A false acceptance of the null hypothesis H0 (accepting H0 when H1 is true) is called a Type II error. The rejection probability under the alternative hypothesis is called the power of the test, and equals 1 minus the probability of a Type II error: n () = Pr (Reject H0 | H1 true) = Pr (Tn > c | H1 true) . We call n () the power function and is written as a function of to indicate its dependence on the true value of the parameter .
189
In the dominant approach to hypothesis testing, the goal of test construction is to have high power, subject to the constraint that the size of the test is lower than the pre-specied signicance level. Generally, the power of a test depends on the true value of the parameter , and for a well behaved test the power is increasing both as moves away from the null hypothesis 0 and as the sample size n increases. Given the two possible states of the world (H0 or H1 ) and the two possible decisions (Accept H0 or Reject H0 ), there are four possible pairings of states and decisions as is depicted in the following chart. Hypothesis Testing Decisions H0 true H1 true Accept H0 Correct Decision Type II Error Reject H0 Type I Error Correct Decision
Given a test statistic Tn , increasing the critical value c increases the acceptance region S0 while decreasing the rejection region S1 . This decreases the likelihood of a Type I error (decreases the size) but increases the likelihood of a Type II error (decreases the power). Thus the choice of c involves a trade-o between size and the power. This is why the signicance level of the test cannot be set arbitrarily small. (Otherwise the test will not have meaningful power.) It is important to consider the power of a test when interpreting hypothesis tests, as an overly narrow focus on size can lead to poor decisions. For example, it is trivial to design a test which has perfect size yet has trivial power. Specically, for any hypothesis we can use the following test: Generate a random variable U U [0, 1] and reject H0 if U < . This test has exact size of . Yet the test also has power precisely equal to . When the power of a test equals the size, we say that the test has trivial power. Nothing is learned from such a test.
8.6
Statistical Signicance
Testing requires a pre-selected choice of signicance level , yet there is no objective scientic basis for choice of . Never-the-less, the common practice is to set = 0.05 (5%). Alternative values are = 0.10 (10%) and = 0.01 (1%). These choices are somewhat the by-product of traditional tables of critical values and statistical software. The informal reasoning behind the choice of a 5% critical value is to ensure that Type I errors should be relatively unlikely that the decision Reject H0 has scientic strength yet the test retains power against reasonable alternatives. The decision Reject H0 means that the evidence is inconsistent with the null hypothesis, in the sense that it is relatively unlikely (1 in 20) that data generated by the null hypothesis would yield the observed test result. In contrast, the decision Accept H0 is not a strong statement. It does not mean that the evidence supports H0 , only that there is insucient evidence to reject H0 . Because of this, it is more accurate to use the label Do not Reject H0 instead of Accept H0 . When a test rejects H0 at the 5% signicance level it is common to say that the statistic is statistically signicant and if the test accepts H0 it is common to say that the statistic is not statistically signicant or that that it is statistically insignicant. It is helpful to remember that this is simply a way of saying Using the statistic Tn , the hypothesis H0 can [cannot] be rejected at the asymptotic 5% level. When the null hypothesis H0 : = 0 is rejected it is common to say that the coecient is statistically signicant, because the test has rejected the hypothesis that the coecient is equal to zero. Let us return to the example about the union wage premium as measured in Table 4.1. The absolute t-statistic for the coecient on Male Union Member is 0.095/0.020 = 4.75, which is
190
greater than the 5% asymptotic critical value of 1.96. Therefore we reject the hypothesis that union membership does not aect wages for men. In this case, we can say that union membership is statistically signicant for men. However, the absolute t-statistic for the coecient on Female Union Member is 0.022/0.020 = 1.10, which is less than 1.96 and therefore we do not reject the hypothesis that union membership does not aect wages for women. In this case we nd that membership for women is not statistically signicant. When a test accepts a null hypothesis (when a test is not statistically signicant), a common misinterpretation is that this is evidence that the null hypothesis is true. This is incorrect. Failure to reject is by itself not evidence. Without an analysis of power, we do not know the likelihood of making a Type II error, and thus are uncertain. In our wage example, it would be a mistake to write that the regression nds that female union membership has no eect on wages. This is an incorrect and most unfortunate interpretation. The test has failed to reject the hypothesis that the coecient is zero, but that does not mean that the coecient is actually zero. When a test rejects a null hypothesis (when a test is statistically signicant) it does not mean that the null hypothesis is false (as we could be making a Type I error) but we know that this event is unlikely. Thus it is appropriate to interpret the rejection as an evidential statement: The null hypothesis appears to be incorrect. However, while we can conclude that the true value is numerically dierent than the hypothesized value 0 , the test alone does not tell us that the true value is meaingfully dierent than 0 . That is, whether or not the deviation of from 0 is meaningful with respect to the interpretation of the coecient. This is where an examination of condence intervals can be quite helpful.
8.7
P-Values
Continuing with the wage regression estimates reported in Table 4.1, consider another question: Does marriage status aect wages? To test the hypothesis that marriage status has no eect on wages, we examine the t-statistics for the coecients on Married Male and Married Female in Table 4.1, which are 0.180/0.008 = 22.5 and 0.016/0.008 = 2.0, respectively. Both exceed the asymptotic 5% critical value of 1.96, so we reject the hypothesis for both men and women. But the statistic for men is exceptionally high, and that for women is only slightly above the critical value. Suppose in contrast that the t-statistic had been 1.9, which is less than the critical value. This would lead to the decision Accept H0 rather than Reject H0 . Should we really be making a dierent decision if the t-statistic is 1.9 rather than 2.0? The dierence in values is small, shouldnt the dierence in the decision be also small? Thinking through these examples it seems unsatisfactory to simply report Accept H0 or Reject H0 . These two decisions do not summarize the evidence. Instead, the magnitude of the statistic Tn suggests a degree of evidence against H0 . How can we take this into account? The answer is to report what is known as the asymptotic p-value pn = 1 G(Tn ). Since the distribution function G is monotonically increasing, the p-value is a monotonically decreasing function of Tn and is an equivalent test statistic. Instead of rejecting H0 at the signicance level if Tn > c, we can reject H0 if pn < . Thus it is sucient to report pn , and let the reader decide. In is instructive to interpret pn as the marginal signicance level: the largest value of for which the test Tn rejects the null hypothesis. That is, pn = 0.11 means that Tn rejects H0 for all signicance levels greater than 0.11, but fails to reject H0 for signicance levels less than 0.11. Furthermore, the asymptotic p-value has a very convenient asymptotic null distribution. Since
CHAPTER 8. HYPOTHESIS TESTING Tn T under H0 , then pn = 1 G(Tn ) 1 G(T ), which has the distribution Pr (1 G(T ) u) = Pr (1 u G(T )) = 1 Pr T G1 (1 u) = 1 G G1 (1 u) = u, = 1 (1 u)
d d d
191
which is the uniform distribution on [0, 1]. Thus pn U[0, 1]. This means that the unusualness of pn is easier to interpret than the unusualness of Tn . An important caveat is that the p-value pn should not be interpreted as the probability that either hypothesis is true. For example, a common mis-interpretation is that pn is the probability that the null hypothesis is false. This is incorrect. Rather, pn is a measure of the strength of information against the null hypothesis. Returing to our empirical example, for the test that the coecient on Married Male is zero, the p-value is 0.000. This means that it would be highly unlikely to observe a t-statistic as large as 22.5 when the true value of the coecient is zero, and thus we can reject that the true value is zero. When presented with such evidence we can say that we strongly reject the null hypothesis, that the test is highly signicant, or that the test rejects at any conventional critical value. In contrast, the p-value for the coecient on Married Female is 0.046. In this context it is typical to say that the test is marginally signicant, meaning that the test statistic is close to the asymptotic 5% critical value. A related (but somewhat inferior) empirical practice is to append asterisks (*) to coecient estimates or test statistics to indicate the level of signicance. A common practice to to append a single asterisk (*) for an estimate or test statistic which exceeds the 10% critical value (i.e., is signicant at the 10% level), append a double asterisk (**) for a test which exceeds the 5% critical value, or append a trip asterisk (***) for a test which which exceeds the 1% critical value. Such a practice can be better than a table of raw test statistics as the asterisks permit a quick interpretation of signicance. On the other hand, asterisks are inferior to p-values, which are also easy and quick to interpret. The goal is essentially the same; it seems wiser to report p-values whenever possible and avoid the use of asterisks. Our recommendation is that the best empirical practice is to compute and report the asymptotic p-value pn rather than simply the test statistic Tn , the binary decision Accept/Reject, or appending asterisks. The p-value is a simple statistic, easy to interpret, and contains more information than the other choices. We now summarize the main features of hypothesis testing. 1. Select a signicance level . 2. Select a test statistic Tn with asymptotic distribution Tn T under H0 . 3. Set the asymptotic critical value c so that 1 G(c) = , where G is the distribution function of T. 4. Calculate the asymptotic p-value pn = 1 G(Tn ). 5. Reject H0 if Tn > c, or equivalently pn < . 6. Accept H0 if Tn c, or equivalently pn . 7. Report pn to summarize the evidence concerning H0 versus H1 .
d
192
8.8
t-ratios and the Abuse of Testing
and In Section 4.15, we argued that a good applied practice is to report coecient estimates ) for all coecients of interest in estimated models. With standard errors s( and s() the reader ) for hypotheses 2s( )] and t-statistics 0 /s( can easily construct condence intervals [ of interest. ( ), and t-ratios tn = /s Some applied papers (especially older ones) instead report estimates not standard errors. Reporting t-ratios instead of standard errors is poor econometric practice. While the same information is being reported (you can back out standard errors by division, e.g. ) = /t n ), standard errors are generally more helpful to readers than t-ratios. Standard errors s( help the reader focus on the estimation precision and condence intervals, while t-ratios focus attention on statistical signicance. While statistical signicance is important, it is less important that the parameter estimates themselves and their condence intervals. The focus should be on the meaning of the parameter estimates, their magnitudes, and their interpretation, not on listing which variables have signicant (e.g. non-zero) coecients. In many modern applications, sample sizes are very large so standard errors can be very small. Consequently t-ratios can be large even if the coecient estimates are economically small. In such contexts it may not be interesting to announce The coecient is non-zero! Instead, what is interesting to announce is that The coecient estimate is economically interesting! In particular, some applied papers report coecient estimates and t-ratios, and limit their discussion of the results to describing which variables are signicant (meaning that their t-ratios exceed 2) and the signs of the coecient estimates. This is very poor empirical work, and should be studiously avoided. It is also a receipe for banishment of your work to lower tier economics journals. Fundamentally, the common t-ratio is a test for the hypothesis that a coecient equals zero. This should be reported and discussed when this is an interesting economic hypothesis of interest. But if this is not the case, it is distracting. In general, when a coecient is of interest, it is constructive to focus on the point estimate, its standard error, and its condence interval. The point estimate gives our best guess for the value. The standard error is a measure of precision. The condence interval gives us the range of values consistent with the data. If the standard error is large then the point estimate is not a good summary about . The endpoints of the condence interval describe the bounds on the likely possibilities. If the condence interval embraces too broad a set of values for , then the dataset is not suciently informative to render useful inferences about . On the other hand if the condence interval is tight, then the data have produced an accurate estimate, and the focus should be on the value and interpretation of this estimate. In contrast, the statement the t-ratio is highly signicant has little interpretive value. The above discussion requires that the researcher knows what the coecient means (in terms of the economic problem) and can interpret values and magnitudes, not just signs. This is critical for good applied econometric practice. For example, consider the question about the eect of marriage status on mean log wages. We had found that the eect is highly signicant for men and marginally signicant for women. Now, lets construct asymptotic condence intervals for the coecients. The one for men is [0.16, 0.20] and that for women is [0.00, 0.03]. This shows that average wages for married men are about 16-20% higher than for unmarried men, which is very substantial, while the dierence for women is about 0-3%, which is small. These magnitudes are more informative than the results of the hypothesis tests.
193
8.9
Wald Tests
The t-test is appropriate when the null hypothesis is a real-valued restriction. More generally, there may be multiple restrictions on the coecient vector . Suppose that we have q > 1 restrictions which can written in the form (8.1). It is natural to estimate = r( ) by the plug-in estimate b . b = r( b ). To test H0 : = 0 one approach is to measure the magnitude of the discrepancy As this is a vector, there is more than one measure of its length. One simple measure is the weighted quadratic form known as the Wald statistic. This is (6.48) evaluated at the null hypothesis 0 1 b 0 V b 0 b Wn = Wn (0 ) = (8.10) b )0 . Notice that we can write Wn b R b is an estimate of V and R b 0V b = r( b = R where V alternatively as 0 1 b 0 b 0 V b Wn = n
b as b , or we can write it directly as a function of using the asymptotic variance estimate V 1 0 0 b ) 0 . b ) 0 b R b b V Wn = r( R r ( (8.11) Also, when r() = R0 is a linear function of , then the Wald statistic simplies to 0 1 0b b 0 b R R0 V R . Wn = R0 0
b 0 . When The Wald statistic Wn is a weighted Euclidean measure of the length of the vector q = 1 then Wn = t2 n , the square of the t-statistic, so hypothesis tests based on Wn and |tn | are equivalent. The Wald statistic (8.10) is a generalization of the t-statistic to the case of multiple restrictions. d As shown in Theorem 6.16.2, when satises r() = 0 then Wn 2 q , a chi-square random distribution function. For a given variable with q degrees of freedom. Let Gq (u) denote the 2 q signicance level , the asymptotic critical value c satises = 1 Gq (c) and can be found from the chi-square distribution table. For example, the 5% critical values for q = 1, q = 2, and q = 3 are 3.84, 5.99, and 7.82, respectively. An asymptotic test rejects H0 in favor of H1 if Wn > c. As with t-tests, it is conventional to describe a Wald test as signicant if Wn exceeds the 5% asymptotic critical value.
Theorem 8.9.1 Under Assumptions 6.1.2 and 6.10.1, and H0 : = 0 , then d Wn 2 q, and for c satisfying = 1 Gq (c), Pr (Wn > c | H0 ) so the test Reject H0 if Wn > c has asymptotic size .
Notice that the asymptotic distribution in Theorem 8.9.1 depends solely on q, the number of restrictions being tested. It does not depend on k, the number of parameters estimated. The asymptotic p-value for Wn is pn = 1 Gq (Wn ). It is particularly useful to report p-values instead of the Wald statistic. For example, if you write that a Wald test on eight restrictions
194
(q = 8) has the value Wn = 11.2, it is dicult for a reader to assess the magnitude of this statistic without the time-consuming and cumbersome process of looking up the critical values from a table. Instead, if you write that the p-value is pn = 0.19 (as is the case for Wn = 11.2 and q = 8) then it is simple for a reader to intrepret its magnitude as insignicant. For example, consider the empirical results presented in Table 4.1. The hypothesis Union membership does not aect wages is the joint restriction that both coecients on Male Union Member and Female Union Member are zero. We calculate the Wald statistic (8.10) for this joint hypothesis and nd Wn = 23.14 with a p-value of pn = 0.000. Thus we reject the hypothesis in favor of the alternative that at least one of the coecients is non-zero. This does not mean that both coecients are non-zero, just that one of the two is non-zero. Therefore examining the joint Wald statistic and the individual t-statistics is useful for interpretation. The Wald statistic is named after the statistician Abraham Wald, who showed that Wn has optimal weighted average power in certain settings.
8.10
Homoskedastic Wald Tests
If the error is known to be homoskedastic, then it is appropriate to use the homoskedastic Wald b0 b with the homoskedastic estimate V statistic (6.50) which replaces V . This statistic equals 0 0 1 0 b 0 b 0 b = Wn V 0 0 1 b ) 0 /s2 . b ) 0 b b X 0 X 1 R r( = r( R (8.12)
We call (8.12) the homoskedastic Wald statistic as it is an appropriate test when the errors are conditionally homoskedastic. 0 = t2 , the square of the t-statistic, where the latter is computed As for Wn , when q = 1 then Wn n with a homoskedastic standard error. In the case of linear hypotheses H0 : R0 = 0 the homoskedastic Wald statistic equals 0 0 1 0 b 0 b 0 /s2 . b X 0 X 1 R R R0 = R0 Wn 2 Theorem 8.10.1 Under Assumptions 6.1.2 and 6.10.1, E e2 i | xi = , and H0 : = 0 , then 0 d 2 Wn q, and for c satisfying = 1 Gq (c), 0 Pr Wn > c | H0 (8.13)
0 > c has asymptotic size . so the test Reject H0 if Wn
8.11
Criterion-Based Tests
b 0 : the discrepancy between the The Wald statistic is based on the length of the vector b b estimate = r() and the hypothesized value 0 . An alternative class of tests is based on the discrepancy between the criterion function minimized with and without the restriction.
195
Criterion-based testing applies when we have a criterion function, say Jn () with B , which is minimized for estimation, and the goal is to test H0 : B 0 versus H0 : / B0 where B 0 B . Minimizing the criterion function over B and B 0 we obtain the unrestricted and restricted estimators b = argmin Jn ()
B B 0
The criterion-based statistic for H0 versus H1 is proportional to Jn = min Jn () min Jn () e ) Jn ( b ). = Jn (

B 0 B
e = argmin Jn () .
The criterion-based statistic Jn is sometimes called a distance statistic, a minimum-distance statistic, or a likelihood-ratio-like statistic. e ) Jn ( b ) and thus Jn 0. The statistic Jn measures the cost Since B 0 is a subset of B , Jn ( (on the criterion) of imposing the null restriction B 0 .
8.12
Minimum Distance Tests
The minimum distance test is a criterion-based test where Jn () is the minimum distance criterion (7.17) 0 b b Wn (8.14) Jn ( ) = n
Consider the class of linear hypotheses H0 : R0 = 0 . In this case we know from (7.25) that 0 e the ecient minimum distance estimator emd subject to the constraint R = 0 is and thus 1 b 0 . b e emd = V b R b R R0 V R0 1 0b 0b b V e b R = R R R V 0 emd
e b 1 The ecient minimum distance estimator emd is obtained by setting W n = V in (8.14) and (8.15). The ecient minimum distance statistic for H0 : B 0 is therefore 0 1 b e b e b Jn (8.16) =n V emd emd .
b the unrestricted (LS) estimator. The restricted estimator e with md minimizes (8.14) subject to b B 0 . Observing that Jn () = 0, the minimum distance statistic simplies to 0 b e e )=n b e Jn = Jn ( (8.15) W n md md md .
Substituting into (8.16) we nd
1 1 0 1 0b 0b b 0 b R b V b b Jn = n R0 R0 V R R R R0 V R V V 0 0 1 b 0 b 0 b R R0 V R0 = n R0 = Wn ,
(8.17)
196
which is the Wald statistic (8.10). is identical Thus for linear hypotheses H0 : R0 = 0 , the ecient minimum distance statistic Jn to the Wald statistic (8.10). For non-linear hypotheses, however, the Wald and minimum distance statistics are dierent. for linear and nonNewey and West (1987) established the asymptotic null distribution of Jn linear hypotheses. Theorem 8.12.1 Under Assumptions 6.1.2 and 6.10.1, and H0 : = 0 , d 2 then Jn q.
is similar to testing using the Wald statistic Testing using the minimum distance statistic Jn Wn . Critical values and p-values are computed using the 2 q distribution. H0 is rejected in favor of exceeds the level critical value. The asymptotic p-value is p = 1 F (J ). H1 if Jn n q n
8.13
Minimum Distance Tests Under Homoskedasticity
A minimum distance statistic for H0 : B 0 is
b xx in (8.14) we obtain the criterion (7.19) If we set W n = Q 0 0 b . b Q b xx Jn () = n

0 0 Jn = min Jn () /s2 . B 0
Notice that we have scaled the criterion by the unbiased variance estimator s2 from (4.21) for reasons which will become clear momentarily. Equation (7.20) showed that
0 2 + Jn () SSEn () = n 0 ( ) are identical. Thus the constrained minimizer of and so the minimizers of SSEn () and Jn 0 Jn () is constrained least-squares
and therefore
e cls = argmin J 0 () = argmin SSEn () n

B 0 B 0 0 0 e Jn = Jn (cls )/s2 0 b e cls /s2 . b e cls Q b xx =n
(8.18)
In the special case of linear hypotheses H0 : R0 = 0 , the constrained least-squares estimator subject to R0 = 0 has the solution (7.10) 1 1 0 b 1 b 0 b Q e = b R0 cls xx R R Qxx R and solving we nd 0 1 1 0 0b 2 0 b 0 b Jn R0 Q R = n R0 R 0 /s = Wn . xx (8.19)
This is the homoskedastic Wald statistic (8.13). Thus for testing linear hypotheses, homoskedastic minimum distance and Wald statistics agree. For nonlinear hypotheses they disagree, but have the same null asymptotic distribution.
CHAPTER 8. HYPOTHESIS TESTING 2 Theorem 8.13.1 Under Assumptions 6.1.2 and 6.10.1, E e2 i | xi = ,
d
197
0 2 . and H0 : = 0 , then Jn q
8.14
F Tests
where
The F statistic for testing H0 : B 0 is e ) SSEn ( b ) /q SSEn ( cls Fn = b )/(n k ) SSEn ( SSEn () =
n X 2 yi x0 i i=1
(8.20)
e is the constrained least-squares estimator (8.18), b is the unconis the sum-of-squared errors, cls strained least-squares estimator, q is the number of restrictions, and k is the number of unconstrained coecients. b )/(n k), we can also write (8.20) as Noting that s2 = SSEn ( Fn = which is a scale of the dierence of sum-of-squared errors, and is thus a criterion-based statistic. Using (7.20) we can also write the statistic as
0 /q, Fn = Jn
e ) SSEn ( b) SSEn ( cls qs2
so the F stastistic is identical to the homoskedastic minimum distance statistic divided by the number of restrictions q. Another useful way of writing (8.20) is 2 2 nk (8.21) Fn = q 2 where 2 = b) 1X 2 SSEn ( = e i n n
i=1 n
is the residual variance estimate under H1 and 2 =
n e ) 1X 2 SSEn ( cls = e i n n i=1
e with e i = yi x0 i cls is the residual variance estimate under H0 . As we discussed in the previous section, in the special case of linear hypotheses H0 : R0 = 0 , 0 0 . It follows that in this case F = W 0 /q . Thus the F statistic equals the homoskedastic Jn = Wn n n n Wald statistic divided by q. It follows that they are equivalent tests for H0 against H1 . In many statistical packages, linear hypothesis tests are reported as F statistics rather than Wald statistics. While they are equivalent, it is important to know which is being reported to know which critical values to use. (If p-values are directly reported this is not an issue.)
198
When reporting an Fn statistic it is conventional to calculate critical values and p-values using the F (q, n k) distribution instead of the asymptotic 2 q /q distribution. This is a prudent small sample adjustment, as the F distribution is exact when the errors are independent of the regressors and normally distributed. However, when the degrees of freedom n k are large then the dierence is negligible. More relevantly, if n k is small enough to make a dierence, probably we shouldnt be trusting the asymptotic approximation anyway! An elegant feature about (8.20) or (8.21) is that they are directly computable from the standard output from two simple OLS regressions, as the sum of squared errors (or regression variance) is a typical printed output from statistical packages, and is often reported in applied tables. Thus Fn can be calculated by hand from standard reported statistics even if you dont have the original data (or if you are sitting in a seminar and listening to a presentation!). If you are presented with an Fn statistic (or a Wald statistic, as you can just divide by q ) but dont have access to critical values, a useful rule of thumb is to know that for large n, the 5% asymptotic critical value is decreasing as q increases, and is less than 2 for q 7. In many statistical packages, when an OLS regression is estimated an F-statistic is automatically reported, even though hypothesis test is requested. This is the F statistic Fn where H0 restricts all coecients except the intercept to be zero. This was a popular statistic in the early days of econometric reporting, when sample sizes were very small and researchers wanted to know if there was any explanatory power to their regression. This is rarely an issue today, as sample sizes are typically suciently large that this F statistic is nearly always highly signicant. While there are special cases where this F statistic is useful, these cases are atypical. As a general rule, there is no reason to report this F statistic. The Fn statistic is named after the statistician Ronald Fisher, one of the founders of modern statistical theory.
8.15
Likelihood Ratio Test
b and e are the unrestricted and constrained MLE. where In the normal linear model the maximized log likelihood (3.45) at the unrestricted and restricted estimates are 2 n n 2 b , log log L mle mle = (log (2 ) + 1) 2 2 and 2 n n 2 e cmle , = (log (2 ) + 1) log cmle log L 2 2 respectively. Thus the LR statistic is 2 2 LRn = n log log 2 = n log 2 2 . Recall that the F statistic (8.21) is also a monotonic which is a monotonic function of 2 / 2 2 . Thus LRn and Fn are fundamentally the same statistic and have the same function of / information about H0 versus H1 .
For a model with parameter and likelihood function Ln () the likelihood ratio statistic for H0 : 0 versus H1 : c 0 is LRn = 2 max log Ln () max log Ln () 0 b) log Ln ( e) = 2 log Ln (
CHAPTER 8. HYPOTHESIS TESTING Furthermore, by a rst-order Taylor series approximation 2 2 n n 1 ' Fn . LRn /q = log 1 + 2 1 ' q q 2
199
This shows that the two statistics (LRn and Fn ) will be numerically close. It also shows that the F statistic and the homoskedastic Wald statistic for linear hypotheses can also be interpreted as approximate likelihood ratio statistics under normality.
8.16
Problems with Tests of NonLinear Hypotheses
While the t and Wald tests work well when the hypothesis is a linear restriction on , they can work quite poorly when the restrictions are nonlinear. This can be seen by a simple example introduced by Lafontaine and White (1986). Take the model yi = + ei ei N(0, 2 ) and consider the hypothesis H0 : = 1. and Let 2 be the sample mean and variance of yi . The standard Wald test for H0 is 2 1 . Wn = n 2 Now notice that H0 is equivalent to the hypothesis H0 (s) : s = 1 for any positive integer s. Letting r( ) = s , and noting R = s s1 , we nd that the standard Wald test for H0 (s) is 2 s 1 . Wn (s) = n 2s2 2 s2 While the hypothesis s = 1 is unaected by the choice of s, the statistic Wn (s) varies with s. This is an unfortunate feature of the Wald statistic. To demonstrate this eect, we have plotted in Figure 8.1 the Wald statistic Wn (s) as a function = 0.8. The decreasing dashed of s, setting n/ 2 = 10. The increasing solid line is for the case = 1.6. It is easy to see that in each case there are values of s for which the line is for the case test statistic is signicant relative to asymptotic critical values, while there are other values of s for which the test statistic is insignicant. This is distressing since the choice of s is arbitrary and irrelevant to the actual hypothesis. d Our rst-order asymptotic theory is not useful to help pick s, as Wn (s) 2 1 under H0 for any s. This is a context where Monte Carlo simulation can be quite useful as a tool to study and compare the exact distributions of statistical procedures in nite samples. The method uses random simulation to create articial datasets, to which we apply the statistical tools of interest. This produces random draws from the statistics sampling distribution. Through repetition, features of this distribution can be calculated. In the present context of the Wald statistic, one feature of importance is the Type I error of the test using the asymptotic 5% critical value 3.84 the probability of a false rejection, Pr (Wn (s) > 3.84 | = 1) . Given the simplicity of the model, this probability depends only on s, n, and 2 . In Table 8.1 we report the results of a Monte Carlo simulation where we vary these
200
Figure 8.1: Wald Statistic as a function of s three parameters. The value of s is varied from 1 to 10, n is varied among 20, 100 and 500, and is varied among 1 and 3. The Table reports the simulation estimate of the Type I error probability from 50,000 random samples. Each row of the table corresponds to a dierent value of s and thus corresponds to a particular choice of test statistic. The second through seventh columns contain the Type I error probabilities for dierent combinations of n and . These probabilities are calculated as the percentage of the 50,000 simulated Wald statistics Wn (s) which are larger than 3.84. The null hypothesis s = 1 is true, so these probabilities are Type I error. To interpret the table, remember that the ideal Type I error probability is 5% (.05) with deviations indicating distortion. Type I error rates between 3% and 8% are considered reasonable. Error rates above 10% are considered excessive. Rates above 20% are unacceptable. When comparing statistical procedures, we compare the rates row by row, looking for tests for which rejection rates are close to 5% and rarely fall outside of the 3%-8% range. For this particular example the only test which meets this criterion is the conventional Wn = Wn (1) test. Any other choice of s leads to a test with unacceptable Type I error probabilities. Table 8.1 Type I Error Probability of Asymptotic 5% Wn (s) Test =1 =3 s n = 20 n = 100 n = 500 n = 20 n = 100 n = 500 1 .06 .05 .05 .07 .05 .05 2 .08 .06 .05 .15 .08 .06 3 .10 .06 .05 .21 .12 .07 4 .13 .07 .06 .25 .15 .08 5 .15 .08 .06 .28 .18 .10 6 .17 .09 .06 .30 .20 .11 7 .19 .10 .06 .31 .22 .13 8 .20 .12 .07 .33 .24 .14 9 .22 .13 .07 .34 .25 .15 10 .23 .14 .08 .35 .26 .16 Note: Rejection frequencies from 50,000 simulated random samples
201
In Table 8.1 you can also see the impact of variation in sample size. In each case, the Type I error probability improves towards 5% as the sample size n increases. There is, however, no magic choice of n for which all tests perform uniformly well. Test performance deteriorates as s increases, which is not surprising given the dependence of Wn (s) on s as shown in Figure 8.1. In this example it is not surprising that the choice s = 1 yields the best test statistic. Other choices are arbitrary and would not be used in practice. While this is clear in this particular example, in other examples natural choices are not always obvious and the best choices may in fact appear counter-intuitive at rst. This point can be illustrated through another example which is similar to one developed in Gregory and Veall (1985). Take the model yi = 0 + x1i 1 + x2i 2 + ei E (xi ei ) = 0 and the hypothesis H0 : 1 = 0 2 (8.22)
0 1/2 is s( ) = R b 1V b R b so that the standard error for . In this case a t-statistic for H0 is
1
where 0 is a known constant. Equivalently, dene = 1 /2 , so the hypothesis can be stated as H0 : = 0 . 1 , 2 ) be the least-squares estimates of (8.22), let V b = ( 0 , b be an estimate of the Let b covariance matrix for and set = 1 /2 . Dene 0 1 2 b1 = R 1 2
2
t1n =
1 2
) s(
An alternative statistic can be constructed through reformulating the null hypothesis as H0 : 1 0 2 = 0. A t-statistic based on this formulation of the hypothesis is 2 1 0 t2n = 1/2 . 0 b R2 V R
2
where
To compare t1n and t2n we perform another simple Monte Carlo simulation. We let x1i and x2i be mutually independent N(0, 1) variables, ei be an independent N(0, 2 ) draw with = 3, and normalize 0 = 0 and 1 = 1. This leaves 2 as a free parameter, along with sample size n. We vary 2 among .1, .25, .50, .75, and 1.0 and n among 100 and 500.
0 R2 = 1 . 0
CHAPTER 8. HYPOTHESIS TESTING Table 8.2 Type I Error Probability of Asymptotic 5% t-tests n = 100 n = 500 Pr (tn < 1.645) Pr (tn > 1.645) Pr (tn < 1.645) Pr (tn > 1.645) t1n t2n t1n t2n t1n t2n t1n t2n .47 .06 .00 .06 .28 .05 .00 .05 .26 .06 .00 .06 .15 .05 .00 .05 .15 .06 .00 .06 .10 .05 .00 .05 .12 .06 .00 .06 .09 .05 .00 .05 .10 .06 .00 .06 .07 .05 .02 .05
202
2 .10 .25 .50 .75 1.00
The one-sided Type I error probabilities Pr (tn < 1.645) and Pr (tn > 1.645) are calculated from 50,000 simulated samples. The results are presented in Table 8.2. Ideally, the entries in the table should be 0.05. However, the rejection rates for the t1n statistic diverge greatly from this value, especially for small values of 2 . The left tail probabilities Pr (t1n < 1.645) greatly exceed 5%, while the right tail probabilities Pr (t1n > 1.645) are close to zero in most cases. In contrast, the rejection rates for the linear t2n statistic are invariant to the value of 2 , and are close to the ideal 5% rate for both sample sizes. The implication of Table 4.2 is that the two t-ratios have dramatically dierent sampling behavior. The common message from both examples is that Wald statistics are sensitive to the algebraic formulation of the null hypothesis. A simple solution is to use the minimum distance statistic Jn , which equals Wn with r = 1 in the rst example, and |t2n | in the second example. The minimum distance statistic is invariant to the algebraic formulation of the null hypothesis, so is immune to this problem. Whenever possible, the Wald statistic should not be used to test nonlinear hypotheses.
8.17
Monte Carlo Simulation
In the Section 8.16 we introduced the method of Monte Carlo simulation to illustrate the small sample problems with tests of nonlinear hypotheses. In this section we describe the method in more detail. Recall, our data consist of observations (yi , xi ) which are random draws from a population distribution F. Let be a parameter and let Tn = Tn ((y1 , x1 ) , ..., (yn , xn ) , ) be a statistic of or a t-statistic ( )/s( ). The exact distribution of Tn is interest, for example an estimator Gn (u, F ) = Pr (Tn u | F ) . While the asymptotic distribution of Tn might be known, the exact (nite sample) distribution Gn is generally unknown. Monte Carlo simulation uses numerical simulation to compute Gn (u, F ) for selected choices of F. This is useful to investigate the performance of the statistic Tn in reasonable situations and sample sizes. The basic idea is that for any given F, the distribution function Gn (u, F ) can be calculated numerically through simulation. The name Monte Carlo derives from the famous Mediterranean gambling resort where games of chance are played. The method of Monte Carlo is quite simple to describe. The researcher chooses F (the distribution of the data) and the sample size n. A true value of is implied by this choice, or equivalently the value is selected directly by the researcher which implies restrictions on F . Then the following experiment is conducted by computer simulation:
, x ) , i = 1, ..., n, are drawn from the distribution F using 1. n independent random pairs (yi i the computers random number generator. , x ) , ..., (y , x ) , ) is calculated on this pseudo data. 2. The statistic Tn = Tn ((y1 n n 1
203
For step 1, most computer packages have built-in procedures for generating U[0, 1] and N(0, 1) random numbers, and from these most random variables can be constructed. (For example, a chi-square can be generated by sums of squares of normals.) For step 2, it is important that the statistic be evaluated at the true value of corresponding to the choice of F. The above experiment creates one random draw from the distribution Gn (u, F ). This is one observation from an unknown distribution. Clearly, from one observation very little can be said. So the researcher repeats the experiment B times, where B is a large number. Typically, we set B = 1000 or B = 5000. We will discuss this choice later. Notationally, let the bth experiment result in the draw Tnb , b = 1, ..., B. These results are stored. After all B experiments have been calculated, these results constitute a random sample of size B from the distribution of Gn (u, F ) = Pr (Tnb u) = Pr (Tn u | F ) . From a random sample, we can estimate any feature of interest using (typically) a method of moments estimator. We now describe some specic examples. Suppose we are interested in the bias, mean-squared error (MSE), and/or variance of the dis , run the above experiment, and calculate . We then set Tn = tribution of
B B X 1 X \ ) = 1 Tnb = b Bias ( B B
\ ) = 1 M SE ( B
B 2 1 X b (Tnb ) = B b=1 b=1 2 \ \ \ var() = M SE () Bias() 2
b=1 B X
b=1
Suppose we are interested in the Type I error associated with an asymptotic 5% two-sided t-test. ) and calculate We would then set Tn = /s(
B X = 1 P 1 (Tnb 1.96) , B b=1
(8.23)
the percentage of the simulated t-ratios which exceed the asymptotic 5% critical value. or Tn = /s( ) We Suppose we are interested in the 5% and 95% quantile of Tn = then compute the 5% and 95% sample quantiles of the sample {Tnb }. The % sample quantile is a number q such that % of the sample are less than q . A simple way to compute sample quantiles is to sort the sample {Tnb } from low to high. Then q is the N th number in this ordered sequence, where N = (B + 1). It is therefore convenient to pick B so that N is an integer. For example, if we set B = 999, then the 5% sample quantile is 50th sorted value and the 95% sample quantile is the 950th sorted value. The typical purpose of a Monte Carlo simulation is to investigate the performance of a statistical procedure (estimator or test) in realistic settings. Generally, the performance will depend on n and F. In many cases, an estimator or test may perform wonderfully for some values, and poorly for others. It is therefore useful to conduct a variety of experiments, for a selection of choices of n and F. As discussed above, the researcher must select the number of experiments, B. Often this is called the number of replications. Quite simply, a larger B results in more precise estimates of the features of interest of Gn , but requires more computational time. In practice, therefore, the choice of B is often guided by the computational demands of the statistical procedure. Since the results of a Monte Carlo experiment are estimates computed from a random sample of size B, it is straightforward to calculate standard errors for any quantity of interest. If the standard error is too large to make a reliable inference, then B will have to be increased.
204
In particular, it is simple to make inferences about rejection probabilities from statistical tests, such as the percentage estimate reported in (8.23). The random variable 1 (Tnb 1.96) is iid 1.96) . The average (8.23) is therefore an Bernoulli, equalling 1 with probability p = E1 (Tnbp unbiased estimator of p with standard error s ( p) = p (1 p) /B . As p is unknown, this may be approximated by replacing p with p or with an hypothesized value. For example, if we are assessing p an asymptotic 5% test, then we can set s ( p) = (.05) (.95) /B ' .22/ B. Hence, standard errors for B = 100, 1000, and 5000, are, respectively, s ( p) = .022, .007, and .003. Most papers in econometric methods, and some empirical papers, include the results of Monte Carlo simulations to illustrate the performance of their methods. When extending existing results, it is good practice to start by replicating existing (published) results. This is not exactly possible in the case of simulation results, as they are inherently random. For example suppose a paper investigates a statistical test, and reports a simulated rejection probability of 0.07 based on a simulation with B = 100 replications. Suppose you attempt to replicate this result, and nd a rejection probability of 0.03 (again using B = 100 simulation replications). Should you conclude that you have failed in your attempt? Absolutely not! Under the hypothesis that both simulations = 0.03, of a common probability are identical, you have two independent estimates, p 1 = 0.07 and p 2 d p1 p 2 ) N(0, 2p(1 p)), so p. The asymptotic (as B ) distribution pof their dierence is B ( 2 = 0.04 is s = 2p(1 p)/B ' 0.03, using the estimate p = ( p1 + p 2 )/2. a standard error for p 1 p Since the t-ratio 0.04/0.03 = 1.3 is not statistically signicant, it is incorrect to reject the null hypothesis that the two simulations are identical. The dierence between the results p 1 = 0.07 and p 2 = 0.03 is consistent with random variation. What should be done? The rst mistake was to copy the previous papers choice of B = 100. 1 p 2 = 0.03 and Instead, suppose you setp B = 5000. Suppose you now obtain p 2 = 0.04. Then p a standard error is s = p(1 p) (1/100 + 1/5000) ' 0.02. Still we cannot reject the hypothesis that the two simulations are dierent. Even though the estimates (0.07 and 0.04) appear to be quite dierent, the diculty is that the original simulation used a very small number of replications (B = 100) so the reported estimate is quite imprecise. In this case, it is appropriate to conclude that your results replicate the previous study, as there is no statistical evidence to reject the hypothesis that they are equivalent. Most journals have policies requiring authors to make available their data sets and computer programs required for empirical results. They do not have similar policies regarding simulations. Never-the-less, it is good professional practice to make your simulations available. The best practice is to post your simulation code on your webpage. This invites others to build on and use your results, leading to possible collaboration, citation, and/or advancement.
8.18
Condence Intervals by Test Inversion
There is a close relationship between hypothesis tests and condence intervals. We observed in Section 6.13 that the standard 95% asymptotic condence interval for a parameter is i h b 1.96 s( b), b + 1.96 s( b) (8.24) Cn = = { : |tn ()| 1.96} . That is, we can describe Cn as The point estimate plus or minus 2 standard errors or The set of parameter values not rejected by a two-sided t-test. The second denition, known as test statistic inversion is a general method for nding condence intervals, and typically produces condence intervals with excellent properties. Given a test statistic Tn () and critical value c, the acceptance region Accept if Tn () c is identical to the condence interval Cn = { : Tn () c}. Since the regions are identical, the probability of coverage Pr ( Cn ) equals the probability of correct acceptance Pr (Accept|) which
205
interval is inverting the t-test based on the nonlinear hypothesis r() = . The trouble is that in Section 8.16 we learned that there is no unique t-statistic for tests of nonlinear hypotheses and that the choice of parameterization matters greatly. For example, if = 1 /2 then the coverage probability of the standard interval (8.24) is 1 minus the probability of the Type I error, which as shown in Table 8.2 can be far from the nominal 5%. In this example a good solution is the same as discussed in Section 8.16 to rewrite the hypothesis as a linear restriction. The hypothesis = 1 /2 is the same as 2 = 1 . The tstatistic for this restriction is 2 1 tn () = 1/2 b R R0 V where R= 1
is exactly 1 minus the Type I error probability. Thus inverting a test with good Type I error probabilities yields a condence interval with good coverage probabilities. Now suppose that the parameter of interest = r() is a nonlinear function of the coecient = r( b) vector . In this case the standardq condence interval for is the set Cn as in (8.24) where ) = R b 0V b R b is the delta method standard error. This condence is the point estimate and s(
2 ). A 95% condence interval for = 1 /2 is the set of 1 b is the covariance matrix for ( and V values of such that |tn ()| 1.96. Since appears in both the numerator and denominator, tn () is a non-linear function of so the easiest method to nd the condence set is by grid search over . For example, in the wage equation log(W age) = 1 Experience + 2 Experience2 /100 + the highest expected wage occurs at Experience = 501 /2 . From Table 4.1 we have the point = 29.8 and we can calculate the standard error s( ) = 0.022 for a 95% condence interval estimate [29.8, 29.9]. However, if we instead invert the linear form of the test we can numerically nd the interval [29.1, 30.6] which is much larger. From the evidence presented in Section 8.16 we know the rst interval can be quite inaccurate and the second interval is greatly preferred.
8.19
Power and Test Consistency
The power of a test is the probability of rejecting H0 when H1 is true. For simplicity suppose that yi is i.i.d. N (, 2 ) with 2 known, consider the t-statistic tn () = n ( y ) /, and tests of H0 : = 0 against H1 : > 0. We reject H0 if tn = tn (0) > c. Note that tn = tn () + n/ and tn () = Z has an exact N(0, 1) distribution. This is because tn () is centered at the true mean , while the test statistic tn (0) is centered at the (false) hypothesized mean of 0. The power of the test is Pr (tn > c | ) = Pr Z + n/ > c = 1 c n/ .
This function is monotonically increasing in and n, and decreasing in and c. Notice that for any c and 6= 0, the power increases to 1 as n . This means that for H1 , the test will reject H0 with probability approaching 1 as the sample size gets large. We call this property test consistency.
CHAPTER 8. HYPOTHESIS TESTING Denition 8.19.1 A test of H0 : 0 is consistent against xed alternatives if for all 1 , Pr (Reject H0 | ) 1 as n .
206
For tests of the form Reject H0 if Tn > c, a sucient condition for test consistency is that the Tn diverges to positive innity with probability one for all 1 . Denition 8.19.2 Tn as n if for all M < , p Pr (Tn M ) 0 as n . Similarly, Tn as n if for all M < , Pr (Tn M ) 0 as n .
p
In general, t-test and Wald tests are consistent against xed alternatives. Take a t-statistic for a test of H0 : = 0 b 0 tn = b) s( q b) = n1 V . Note that where 0 is a known value and s( tn = b n ( 0 ) q . + b) s( V
The rst term on the right-hand-side converges in distribution to N(0, 1). The second term on the right-hand-side equals zero if = 0 , converges in probability to + if > 0 , and converges in probability to if < 0 . Thus the two-sided t-test is consistent against H1 : 6= 0 , and one-sided t-tests are consistent against the alternatives for which they are designed.
Theorem 8.19.1 Under Assumptions 6.1.2 and 6.10.1, for = r( ) 6= 0 p and q = 1, then |tn | , so for any c < the test Reject H0 if |tn | > c is consistent against xed alternatives.
under H1 , Wn . Again, this implies that Wald tests are consistent tests.
p b 6= 0 . Thus Under H1 , p
The Wald statistic for H0 : = r() = 0 against H1 : 6= 0 is 0 1 b 0 . b 0 V b Wn = n
0 1 p 1 b 0 b 0 V b ( 0 )0 V ( 0 ) > 0. Hence
Theorem 8.19.2 Under Assumptions 6.1.2 and 6.10.1, for = r( ) 6= p 0 , then Wn , so for any c < the test Reject H0 if Wn > c is consistent against xed alternatives.
207
8.20
Asymptotic Local Power
Consistency is a good property for a test, but does not give a useful approximation to the power of a test. To approximate the power function we need a distributional approximation. The standard asymptotic method for power analysis uses what are called local alternatives. This is similar to our analysis of restriction estimation under misspecication (Section 7.9). The technique is to index the parameter by sample size so that the asymptotic distribution of the statistic is continuous in a localizing parameter. In this section we consider t-tests on real-valued parameters and in the next section consider Wald tests. Specically, we consider parameter vectors n which are indexed by sample size n and satisfy the real-valued relationship n = r(n ) = 0 + n1/2 h (8.25)
where the second equality is (8.25). The rst term is asymptotically normal: b d p n n V Z. where Z N(0, 1). Therefore d p b 0 n V Z + h
where the scalar h is is called a localizing parameter. We index n and n by sample size to indicate their dependence on n. The way to think of (8.25) is that the true value of the parameters are n and n . The parameter n is close to the hypothesized value 0 , with deviation n1/2 h. The specication (8.25) states that for any xed h , n approaches 0 as n gets large. Thus n is close or local to 0 . The concept of a localizing sequence (8.25) might seem odd at rst as in the actual world the sample size cannot mechanically aect the value of the parameter. Thus (8.25) should not be interpreted literally. Instead, it should be interpreted as a technical device which allows the asymptotic distribution of the test statistic to be continuous in the alternative hypothesis. To evaluate the asymptotic distribution of the test statistic we start by examining the scaled estimate centered at the hypothesized value 0 . Breaking it into a term centered at the true value n and a remainder we nd b 0 = n b n + n (n 0 ) n b n + h = n
or N(h, V ). This is a continuous asymptotic distribution, and depends continuously on the localing parameter h. Applied to the t statistic we nd tn = b 0 b) s( V Z + h d V Z+
(8.26)
where = h/ V . This generalizes Theorem 8.4.1 (which assumes H0 is true) to allow for local alternatives of the form (8.25). Consider a t-test of H0 against the one-sided alternative H1 : > 0 which rejects H0 for tn > c where (c ) = 1 . The asymptotic local power of this test is the limit (as the sample size
208
Figure 8.2: Asymptotic Local Power Function of One-Sided t Test diverges) of the rejection probability under the local alternative (8.25)
n
lim Pr (Reject H0 ) = lim Pr (tn > c ) = Pr (Z + > c ) = 1 (c )

def n
= ( c ) = ( ).
We call ( ) the local power function. In Figure 8.2 we plot the local power function ( ) as a function of [1, 4] for tests of asymptotic size = 0.10, = 0.05, and = 0.01. = 0 corresponds to the null hypothesis so . Note that the power is lower ( ) = . The power functions are monotonically increasing in than for < 0 due to the one-sided nature of the test. We can see that the three power functions are ranked by so that the test with = 0.10 has higher power than the test with = 0.01. This is the inherent trade-o between size and power. Decreasing size induces a decrease in power, and conversely. The coecient can be interpreted as the parameter deviation measured as a multiple of the q ). To see this, recall that s( ) = n1/2 V ' n1/2 V and then note that standard error s( n1/2 h h n 0 = ' = . ) ) V s( s(
). Thus as Thus equals the deviation n 0 expressed as multiples of the standard error s( we examine Figure 8.2, we can interpret the power function at = 1 (e.g. 26% for a 5% size test) as the power when the parameter n is one standard error above the hypothesized value. For example, from Table 4.1 the standard error for the coecient on Married Female is 0.008. Thus in this example, = 1 corresonds to n = 0.008 or an 0.8% wage premium for married females. Our calculations show that the asymptotic power of a one-sided 5% test against this alternative is about 26%.
209
The dierence between power functions can be measured either vertically or horizontally. For example, in Figure 8.2 there is a vertical dotted line at = 1, showing that the asymptotic local power function ( ) equals 39% for = 0.10, equals 26% for = 0.05 and equals 9% for = 0.01. This is the dierence in power across tests of diering size, holding xed the parameter in the alternative A horizontal comparison can also be illuminating. To illustrate, in Figure 8.2 there is a horizontal dotted line at 50% power. 50% power is a useful benchmark, as it is the point where the test has equal odds of rejection and acceptance. The dotted line crosses the three power curves at = 1.29 ( = 0.10), = 1.65 ( = 0.05), and = 2.33 ( = 0.01). This means that the parameter must be at least 1.65 standard errors above the hypothesized value for the one-sided test to have 50% (approximate) power. As discussed agove, these values can be interpreted as the multiple of the standard error needed for a coecient to obtain power equal to 50%. The ratio of these values (e.g. 1.65/1.29 = 1.28 for the asymptotic 5% versus 10% tests) measures the relative parameter magnitude needed to achieve the same power. (Thus, for a 5% size test to achieve 50% power, the parameter must be 28% larger than for a 10% size test.) Even more interesting, the square of this ratio (e.g. (1.65/1.29)2 = 1.64) can be interpreted as the increase in sample size needed to achieve the same power under xed parameters. That is, to achieve 50% power, a 5% size test needs 64% more observations than a 10% size test. This interpretation follows by the following informal argument. By denition and (8.25) = h/ V = n (n 0 ) / V . Thus holding and V xed, we can see that 2 is proportional to n. The analysis of a two-sided t test is similar. (8.26) implies that b 0 d tn = |Z + | b) s( and thus the local power of a two-sided t test is
n
lim Pr (Reject H0 ) = lim Pr (tn > c ) = ( c ) ( c ) = Pr (|Z + | > c )

n
which is monotonically increasing in | |. Theorem 8.20.1 Under Assumptions 6.1.2 and 6.10.1, and n = r(n ) = r0 + n1/2 h, then b 0 d t(0 ) = Z + b) s( where Z N(0, 1) and = h/ V . For c such that Pr (Z > c ) = , Pr (t(0 ) > c ) ( c ) . Furthermore, for c such that Pr (|Z| > c ) = , Pr (|t(0 )| > c ) ( c ) ( c ) .
210
8.21
Asymptotic Local Power, Vector Case
In this section we extend the local power analysis of the previous section to the case of vectorvalued alternatives. We generalize (8.25) to allow n to be vector-valued. The local parameterization takes the form (8.27) n = r(n ) = 0 + n1/2 h where h is q 1. Under (8.27), b 0 = n b n + h n
d
Zh N(h, V ),
a normal random vector with mean h and variance matrix V . Applied to the Wald statistic we nd 0 1 b 0 b 0 V b Wn = n
1 Z0 h V Zh . d
(8.28)
The asymptotic distribution (8.28) is a quadratic in the normal random vector Zh , similar to that found for the asymptotic null distribution of the Wald statistic. The important dierence, however, is that Zh has a mean of h. The distribution of the quadratic form (8.28) is a close relative of the chi-square distribution.
1 Theorem 8.21.1 If Zh N(h, V ) with V > 0, q q, then Z0 Zh hV 2 q (), a non-central chi-square random variable with q degrees of freedom and non-centrality parameter = h0 V 1 h.
The convergence (8.28) shows that under the local alternatives (8.27), Wn 2 q (). This generalizes the null asymptotic distribution which obtains as the special case = 0. We can use this result to obtain a continuous asymptotic approximation to the power function. For any signicance 2 level > 0 set the asymptotic critical value c so that Pr q > c = . Then as n , The asymptotic local power function ,q () depends only on , q, and . def Pr (Wn > c ) Pr 2 q () > c = ,q ().
Theorem 8.21.2 Under Assumptions 6.1.2 and 6.10.1, and n r(n ) = 0 + n1/2 h, then Wn 2 q ()
d
2 1 where = h0 V h. Furthermore, for c such that Pr q > c = , Pr (Wn > c ) Pr 2 q () > c .
211
Figure 8.3: Asymptotic Local Power Function, Varying q The non-central chi-square distribution is a generalization of the chi-square, with 2 q () special2 2 2 2 izing to q when = 0. In the case q = 1, q () = |Z + | with = , and thus Theorem 8.21.2 generalizes Theorem 8.20.1 from q = 1 to q 1. Figure 8.3 plots 0.05,q () (the power of asymptotic 5% tests) as a function of for q = 1, q = 2, and q = 3. The power functions are monotonically increasing in and asymptote to one. Figure 8.3 also shows the power loss for xed non-centrality parameter as the dimensionality of the test increases. The power curves shift to the right as q increases, resulting in a decrease in power. This is illustrated by the dotted line at 50% power. The dotted line crosses the three power curves at = 3.85 (q = 1), = 4.96 (q = 2), and = 5.77 (q = 3). The ratio of these values correspond to the relative sample sizes needed to obtain the same power. Thus increasing the dimension of the test from q = 1 to q = 2 requires a 28% increase in sample size, or an increase from q = 1 to q = 3 requires a 50% increase in sample size, to obtain a test with 50% power.
8.22
Technical Proofs*
Proof of Theorem 8.12. The conditions of Theorem 7.10.1 hold, since H0 implies Assumption 1 b 1 7.5.1. From (7.54) with W n = V , we see that 1 0 b b 0 b e b b b = V n R n R R V R emd n n 0 1 0 d V R R V R R N(0, V ) = V R Z. where Z N(0, (R0 V R)
1
). Thus 0 1 b e b e b Jn =n emd V emd

1 Z0 R0 V V V R Z = Z0 R0 V R Z d
= 2 q.
212
1 Zh depends only on Proof of Theorem 8.21.1. We show that the random variable Q = Z0 hV q and . 1 First, let G be a square root of V so that V = GG0 . Dene Z h = G Zh N(h , I q ) with 0 0 h = G1 h. Note that = h V 1 h = h h . Second, construct an orthogonal q q matrix H = [H 1 , H 2 ] whose rst column equals H 1 = 0 1/2 1/2 and H 0 h = 0. Dene Z = H 0 Z N(h , I ) where . Note that H 0 h h h q 1h = 2 h h h
=H =
H0 1h H0 2h
1/2 0
1 . q1
Note that the distribution of Z h is only a function of and q. Finally, observe that
1 Zh Q = Z0 hV 0 = Z h Zh 0 = Z h Zh
which is only a function of Z h and thus its distribution only depends on and q.
213
Exercises
Exercise 8.1 Prove that if an additional regressor X k+1 is added to X , Theils adjusted R k+1 /s( k+1 ) is the t-ratio for k+1 and increases if and only if |tk+1 | > 1, where tk+1 = k+1 ) = s2 [(X 0 X )1 ]k+1,k+1 1/2 s(
2
is the homoskedasticity-formula standard error.
Exercise 8.2 You have two independent samples (y 1 , X 1 ) and (y 2 , X 2 ) which satisfy y 1 = X 1 1 + e1 and y 2 = X 2 2 + e2 , where E (x1i e1i ) = 0 and E (x2i e2i ) = 0, and both X 1 and X 2 have k b be the OLS estimates of and . For simplicity, you may assume that b and columns. Let 1 2 1 2 both samples have the same number of observations n. b b ( ) as n . (a) Find the asymptotic distribution of n 2 1 2 1 (b) Find an appropriate test statistic for H0 : 2 = 1 . (c) Find the asymptotic distribution of this statistic under H0 . Exercise 8.3 The data set invest.dat contains data on 565 U.S. rms extracted from Compustat for the year 1987. The variables, in order, are Ii Qi Ci Di Investment to Capital Ratio (multiplied by 100). Total Market Value to Asset Ratio (Tobins Q). Cash Flow to Asset Ratio. Long Term Debt to Asset Ratio.
The ow variables are annual sums for 1987. The stock variables are beginning of year. (a) Estimate a linear regression of Ii on the other variables. Calculate appropriate standard errors. (b) Calculate asymptotic condence intervals for the coecients. (c) This regression is related to Tobins q theory of investment, which suggests that investment should be predicted solely by Qi . Thus the coecient on Qi should be positive and the others should be zero. Test the joint hypothesis that the coecients on Ci and Di are zero. Test the hypothesis that the coecient on Qi is zero. Are the results consistent with the predictions of the theory?
2 2 (d) Now try a non-linear (quadratic) specication. Regress Ii on Qi , Ci , Di , Q2 i , Ci , Di , Qi Ci , Qi Di , Ci Di . Test the joint hypothesis that the six interaction and quadratic coecients are zero.
Exercise 8.4 In a paper in 1963, Marc Nerlove analyzed a cost function for 145 American electric companies. (The problem is discussed in Example 8.3 of Greene, section 1.7 of Hayashi, and the empirical exercise in Chapter 1 of Hayashi). The data le nerlov.dat contains his data. The variables are described on page 77 of Hayashi. Nerlov was interested in estimating a cost function : T C = f (Q, P L, P F, P K ).
CHAPTER 8. HYPOTHESIS TESTING (a) First estimate an unrestricted Cobb-Douglass specication log T Ci = 1 + 2 log Qi + 3 log P Li + 4 log P Ki + 5 log P Fi + ei .
214
(8.29)
Report parameter estimates and standard errors. You should obtain the same OLS estimates as in Hayashis equation (1.7.7), but your standard errors may dier. (b) What is the economic meaning of the restriction H0 : 3 + 4 + 5 = 1? (c) Estimate (8.29) by constrained least-squares imposing 3 +4 +5 = 1. Report your parameter estimates and standard errors. (d) Estimate (8.29) by ecient minimum distance imposing 3 + 4 + 5 = 1. Report your parameter estimates and standard errors. (e) Test H0 : 3 + 4 + 5 = 1 using a Wald statistic (f) Test H0 : 3 + 4 + 5 = 1 using a minimum distance statistic
Chapter 9
Regression Extensions
9.1 NonLinear Least Squares
In some cases we might use a parametric regression function m (x, ) = E (yi | xi = x) which is a non-linear function of the parameters . We describe this setting as non-linear regression. Example 9.1.1 Exponential Link Regression m (x, ) = exp x0
The exponential link function is strictly positive, so this choice can be useful when it is desired to constrain the mean to be strictly positive. Example 9.1.2 Logistic Link Regression m (x, ) = x0 (u) = (1 + exp(u))1 (9.1)
where
is the Logistic distribution function. Since the logistic link function lies in [0, 1], this choice can be useful when the conditional mean is bounded between 0 and 1. Example 9.1.3 Exponentially Transformed Regressors m (x, ) = 1 + 2 exp(3 x) Example 9.1.4 Power Transformation m (x, ) = 1 + 2 x3 with x > 0. Example 9.1.5 Box-Cox Transformed Regressors m (x, ) = 1 + 2 x(3 ) where x() x 1 , if > 0 = log( x), if = 0
(9.2)
and x > 0. The function (9.2) is called the Box-Cox Transformation and was introduced by Box and Cox (1964). The function nests linearity ( = 1) and logarithmic ( = 0) transformations continuously. 215
CHAPTER 9. REGRESSION EXTENSIONS Example 9.1.6 Continuous Threshold Regression m (x, ) = 1 + 2 x + 3 (x 4 ) 1 (x > 4 ) Example 9.1.7 Threshold Regression 0 0 x1 1 (x2 < 3 ) + 2 x1 1 (x2 3 ) m (x, ) = 1 Example 9.1.8 Smooth Transition m (x, ) = where (u) is the logit function (9.1).
0 1 x1
216
0 + 2 x1
x2 3 4
What dierentiates these examples from the linear regression model is that the conditional mean cannot be written as a linear function of the parameter vector . Nonlinear regression is sometimes adopted because the functional form m (x, ) is suggested by an economic model. In other cases, it is adopted as a exible approximation to an unknown regression function. b minimizes the normalized sum-of-squared-errors The least squares estimator 1X (yi m (xi , ))2 . Sn () = n
i=1 n
When the regression function is nonlinear, wecall this the nonlinear least squares (NLLS) b estimator. The NLLS residuals are e i = yi m xi , . One motivation for the choice of NLLS as the estimation method is that the parameter is the solution to the population problem min E (yi m (xi , ))2 b must be found by numerical Since sum-of-squared-errors function Sn () is not quadratic, methods. See Appendix E. When m(x, ) is dierentiable, then the FOC for minimization are 0= where
n X i=1
b e m xi , i
(9.3)
m (x, ) =
m (x, ) .
Theorem 9.1.1 Asymptotic Distribution of NLLS Estimator If the model is identied and m (x, ) is dierentiable with respect to , b d n N (0, V ) where mi = m (xi , 0 ). 1 1 2 E mi m0 E mi m0 V = E mi m0 i i ei i
CHAPTER 9. REGRESSION EXTENSIONS Based on Theorem 9.1.1, an estimate of the asymptotic variance V is n !1 n ! n !1 X X X 1 1 1 b = m i m 0 m i m 0 2 m i m 0 V i i i e i n n n
i=1 i=1 i=1
217
b) and e b). where m i = m (xi , i = yi m(xi , Identication is often tricky in nonlinear regression models. Suppose that
0 m(xi , ) = 0 1 z i + 2 xi ( )
where xi ( ) is a function of xi and the unknown parameter . Examples include xi ( ) = x i, xi ( ) = exp (xi ) , and xi ( ) = xi 1 (g (xi ) > ). The model is linear when 2 = 0, and this is often a useful hypothesis (sub-model) to consider. Thus we want to test H0 : 2 = 0. However, under H0 , the model is yi = 0 1 z i + ei
and both 2 and have dropped out. This means that under H0 , is not identied. This renders the distribution theory presented in the previous section invalid. Thus when the truth is that 2 = 0, the parameter estimates are not asymptotically normally distributed. Furthermore, tests of H0 do not have asymptotic normal or chi-square distributions. The asymptotic theory of such tests have been worked out by Andrews and Ploberger (1994) and B. E. Hansen (1996). In particular, Hansen shows how to use simulation (similar to the bootstrap) to construct the asymptotic critical values (or p-values) in a given application. Proof of Theorem 9.1.1 (Sketch). NLLS estimation falls in the class of optimization estimators. For this theory, it is useful to denote the true value of the parameter as 0 . p 0 . Proving that nonlinear estimators are consistent is more The rst step is to show that minimizes challenging than for linear estimators. We sketch the main argument. The idea is that the sample criterion function Sn (), which (for any ) converges in probability to the mean-squared will converge in error function E (yi m (xi , ))2 . Thus it seems reasonable that the minimizer 2 probability to 0 , the minimizer of E (yi m (xi , )) . It turns out that to show this rigorously, we need to show that Sn () converges uniformly to its expectation E (yi m (xi , ))2 , which means that the maximum discrepancy must converge in probability to zero, to exclude the possibility that Sn () is excessively wiggly in . Proving uniform convergence is technically challenging, but it can be shown to hold broadly for relevant nonlinear regression models, especially if the regression function m (xi , ) is dierentiabel in . For a complete treatment of the theory of optimization estimators see Newey and McFadden (1994). p is close to 0 for n large, so the minimization of Sn () only needs to be 0 , Since examined for close to 0 . Let 0 = ei + m0 yi i 0 . For close to the true value 0 , by a rst-order Taylor series approximation, m (xi , ) ' m (xi , 0 ) + m0 i ( 0 ) . Thus yi m (xi , ) ' (ei + m (xi , 0 )) m (xi , 0 ) + m0 i ( 0 )
0 = yi m0 i .
= ei m0 i ( 0 )
CHAPTER 9. REGRESSION EXTENSIONS Hence the sum of squared errors function is Sn () =

n X i=1
218
(yi m (xi , ))2 '
n X 2 0 yi m0 i i=1
0 on m . Thus the NLLS and the right-hand-side is the SSE function for a linear regression of yi i 0 on m , b estimator has the same asymptotic distribution as the (infeasible) OLS regression of yi i which is that stated in the theorem.
9.2
Generalized Least Squares
In the projection model, we know that the least-squares estimator is semi-parametrically ecient for the projection coecient. However, in the linear regression model yi = x0 i + ei E (ei | xi ) = 0, the least-squares estimator is inecient. The theory of Chamberlain (1987) can be used to show that in this model the semiparametric eciency bound is obtained by the Generalized Least Squares (GLS) estimator (4.13) introduced in Section 4.6.1. The GLS estimator is sometimes called the Aitken estimator. The GLS estimator (9.2) is infeasible since the matrix D is unknown. 2 , ..., 2 }. = diag{ n A feasible GLS (FGLS) estimator replaces the unknown D with an estimate D 1 We now discuss this estimation problem. Suppose that we model the conditional variance using the parametric form
2 = 0 + z 0 i 1i 1
= 0 z i , where z 1i is some q 1 function of xi . Typically, z 1i are squares (and perhaps levels) of some (or all) elements of xi . Often the functional form is kept simple for parsimony. Let i = e2 i . Then E (i | xi ) = 0 + z 0 1i 1 and we have the regression equation i = 0 + z 0 1i 1 + i E (i | xi ) = 0. This regression error i is generally heteroskedastic and has the conditional variance var (i | xi ) = var e2 i | xi 2 2 E e | x | x = E e2 i i i i 4 2 2 . = E ei | xi E ei | xi Suppose ei (and thus i ) were observed. Then we could estimate by OLS: 1 0 p b = Z 0Z Z (9.4)
and
d b ) N (0, V ) n (
CHAPTER 9. REGRESSION EXTENSIONS where 1 0 2 0 1 E z i z i i E z i z i . V = E zi z0 i

2 =e 2 i ei
219
(9.5)
0 b b While ei is not observed, we have the OLS residual e i = yi x0 i = ei xi ( ). Thus
i i i
And then
b + ( b )0 xi x0 ( b ). = 2ei x0 i i
n n n X 2 X 1 X b + 1 b )0 xi x0 ( b ) n z i i = z i ei x0 n z i ( i i n n n i=1 i=1 i=1
Let
be from OLS regression of i on z i . Then 1 1/2 0 e ) = n ( b ) + n1 Z 0 Z n ( n Z N (0, V )

d
1 0 e = Z 0Z Z
(9.6)
(9.7)
Thus the fact that i is replaced with i is asymptotically irrelevant. We call (9.6) the skedastic regression, as it is estimating the conditional variance of the regression of yi on xi . We have shown 2 = z 0 by that is consistently estimated by a simple procedure, and hence we can estimate i i
2 i = 0 zi . 2 > 0 for all i. Then set Suppose that i 2 2 e = diag{ D 1 , ..., n }
(9.8)
and
This is the feasible GLS, or FGLS, estimator of . Since there is not a unique specication for the conditional variance the FGLS estimator is not unique, and will depend on the model (and estimation method) for the skedastic regression. One typical problem with implementation of FGLS estimation is that in the linear specication 2 > 0 for all i. If 2 < 0 for some i, then the FGLS estimator i (9.4), there is no guarantee that i 2 is not well dened. Furthermore, if i 0 for some i then the FGLS estimator will force the regression equation through the point (yi , xi ), which is undesirable. This suggests that there is a need to bound the estimated variances away from zero. A trimming rule takes the form
2 2 i , c 2] i = max[
1 e = X 0D e 1 X e 1 y . X 0D
for some c > 0. For example, setting c = 1/4 means that the conditional variance function is constrained to exceed one-fourth of the unconditional variance. As there is no clear method to select c, this introduces a degree of arbitrariness. In this context it is useful to re-estimate the model with several choices for the trimming parameter. If the estimates turn out to be sensitive to its choice, the estimation method should probably be reconsidered. It is possible to show that if the skedastic regression is correctly specied, then FGLS is asymptotically equivalent to GLS. As the proof is tricky, we just state the result without proof.
CHAPTER 9. REGRESSION EXTENSIONS Theorem 9.2.1 If the skedastic regression is correctly specied, p e e n GLS F GLS 0, d e n N (0, V ) , F GLS 1 2 V = E i xi x0 . i
220
and thus where
e 0 is appropriate when the skedastic which is consistent for V as n . This estimator V regression (9.4) is correctly specied. 2 = It may be the case that 0 z i is only an approximation to the true conditional variance i 2 0 2 should perhaps be E(ei | xi ). In this case we interpret z i as a linear projection of ei on z i . called a quasi-FGLS estimator of . Its asymptotic variance is not that given in Theorem 9.2.1. Instead, 1 1 1 2 2 1 0 0 0 0 E E xi x0 z x x z x x . V = E 0 z i i i i i i i i i
2 = 0 z , V takes a sandwich form similar to the covariance matrix of the OLS estimator. Unless i i 0 e V is inconsistent for V . e0 An appropriate solution is to use a White-type estimator in place of V . This may be written as n !1 n ! n !1 X X X 1 1 1 2 4 2 2 e = i xi x0 i e i xi x0 i xi x0 V i i i n n n i=1 i=1 i=1 1 1 1 0 e 1 1 0 e 1 1 0 e 1 b e 1 XD X X D DD X XD X = n n n
Examining the asymptotic distribution of Theorem 9.2.1, the natural estimator of the asympe is totic variance of n !1 1 X 0 1 1 0 1 2 0 e XD X i xi xi = . V = n n
i=1
b = diag{e where D 2 2 n }. This is estimator is robust to misspecication of the conditional vari1 , ..., e ance, and was proposed by Cragg (1992). In the linear regression model, FGLS is asymptotically superior to OLS. Why then do we not exclusively estimate regression models by FGLS? This is a good question. There are three reasons. First, FGLS estimation depends on specication and estimation of the skedastic regression. Since the form of the skedastic regression is unknown, and it may be estimated with considerable error, the estimated conditional variances may contain more noise than information about the true conditional variances. In this case, FGLS can do worse than OLS in practice. Second, individual estimated conditional variances may be negative, and this requires trimming to solve. This introduces an element of arbitrariness which is unsettling to empirical researchers. Third, and probably most importantly, OLS is a robust estimator of the parameter vector. It is consistent not only in the regression model, but also under the assumptions of linear projection. The GLS and FGLS estimators, on the other hand, require the assumption of a correct conditional mean. If the equation of interest is a linear projection and not a conditional mean, then the OLS
CHAPTER 9. REGRESSION EXTENSIONS
221
and FGLS estimators will converge in probability to dierent limits as they will be estimating two dierent projections. The FGLS probability limit will depend on the particular function selected for the skedastic regression. The point is that the eciency gains from FGLS are built on the stronger assumption of a correct conditional mean, and the cost is a loss of robustness to misspecication.
9.3
Testing for Heteroskedasticity
2 The hypothesis of homoskedasticity is that E e2 i | xi = , or equivalently that H0 : 1 = 0 in the regression (9.4). We may therefore test this hypothesis by the estimation (9.6) and constructing a Wald statistic. In the classic literature it is typical to impose the stronger assumption that ei is independent of xi , in which case i is independent of xi and the asymptotic variance (9.5) for simplies to 1 2 E i . (9.9) V = E z i z 0 i
Hence the standard test of H0 is a classic F (or Wald) test for exclusion of all regressors from the skedastic regression (9.6). The asymptotic distribution (9.7) and the asymptotic variance (9.9) under independence show that this test has an asymptotic chi-square distribution. Theorem 9.3.1 Under H0 and ei independent of xi , the Wald test of H0 is asymptotically 2 q.
Most tests for heteroskedasticity take this basic form. The main dierences between popular tests are which transformations of xi enter z i . Motivated by the form of the asymptotic variance b , White (1980) proposed that the test for heteroskedasticity be based on of the OLS estimator setting z i to equal all non-redundant elements of xi , its squares, and all cross-products. BreuschPagan (1979) proposed what might appear to bea distinct test, but the only dierence isthat they 2 with 2 4 which holds when e is N 0, 2 . If allowed for general choice of z i , and replaced E i i this simplication is replaced by the standard formula (under independence of the error), the two tests coincide. It is important not to misuse tests for heteroskedasticity. It should not be used to determine whether to estimate a regression equation by OLS or FGLS, nor to determine whether classic or White standard errors should be reported. Hypothesis tests are not designed for these purposes. Rather, tests for heteroskedasticity should be used to answer the scientic question of whether or not the conditional variance is a function of the regressors. If this question is not of economic interest, then there is no value in conducting a test for heteorskedasticity
9.4
Testing for Omitted NonLinearity
If the goal is to estimate the conditional expectation E (yi | xi ) , it is useful to have a general test of the adequacy of the specication. One simple test for neglected nonlinearity is to add nonlinear functions of the regressors to the b i has been regression, and test their signicance using a Wald test. Thus, if the model yi = x0 i + e t by OLS, let z i = h(xi ) denote functions of xi which are not linear functions of xi (perhaps 0 +e e squares of non-binary regressors) and then t yi = x0 i by OLS, and form a Wald statistic i + zi for = 0. Another popular approach is the RESET test proposed by Ramsey (1969). The null model is yi = x0 i + ei
CHAPTER 9. REGRESSION EXTENSIONS which is estimated by OLS, yielding predicted values 2 y i . zi = . .

m y i
222 b y i = x0 i . Now let
be a (m 1)-vector of powers of y i . Then run the auxiliary regression

0 e e+e i yi = x0 i + zi d
(9.10)
by OLS, and form the Wald statistic Wn for = 0. It is easy (although somewhat tedious) to show that under the null hypothesis, Wn 2 m1 . Thus the null is rejected at the % level if Wn exceeds the upper % tail critical value of the 2 m1 distribution. To implement the test, m must be selected in advance. Typically, small values such as m = 2, 3, or 4 seem to work best. The RESET test appears to work well as a test of functional form against a wide range of smooth alternatives. It is particularly powerful at detecting single-index models of the form yi = G(x0 i ) + ei where G() is a smooth link function. To see why this is the case, note that (9.10) may be written as 2 3 m 0b e b b 1 + x0 2 + x0 m1 + e i yi = x0 i + xi i i which has essentially approximated G() by a mth order polynomial
9.5
Least Absolute Deviations
We stated that a conventional goal in econometrics is estimation of impact of variation in xi on the central tendency of yi . We have discussed projections and conditional means, but these are not the only measures of central tendency. An alternative good measure is the conditional median. To recall the denition and properties of the median, let y be a continuous random variable. The median = med(y) is the value such that Pr(y ) = Pr (y ) = 0.5. Two useful facts about the median are that (9.11) = argmin E |y |
and where E sgn (y ) = 0 sgn (u) = 1 1 if u 0 if u < 0
is the sign function. These facts and denitions motivate three estimators of . The rst denition is the 50th 1 Pn empirical quantile. The second is the value which minimizes i=1 |yi | , and the third denition n 1 Pn is the solution to the moment equation n i=1 sgn (yi ) . These distinctions are illusory, however, as these estimators are indeed identical. Now lets consider the conditional median of y given a random vector x. Let m(x) = med (y | x) denote the conditional median of y given x. The linear median regression model takes the form yi = x0 i + ei In this model, the linear function med (yi | xi = x) = x0 is the conditional median function, and the substantive assumption is that the median function is linear in x. Conditional analogs of the facts about the median are med (ei | xi ) = 0
CHAPTER 9. REGRESSION EXTENSIONS Pr(yi x0 | xi = x) = Pr(yi > x0 | xi = x) = .5 E (sgn (ei ) | xi ) = 0 E (xi sgn (ei )) = 0 = min E |yi x0 i | These facts motivate the following estimator. Let LADn () =
n 1 X yi x0 i n i=1
223
be the average of absolute deviations. The least absolute deviations (LAD) estimator of minimizes this function b = argmin LADn ()
Equivalently, it is a solution to the moment condition

n
The LAD estimator has an asymptotic normal distribution.
1X b = 0. xi sgn yi x0 i n
i=1
(9.12)
Theorem 9.5.1 Asymptotic Distribution of LAD Estimator When the conditional median is linear in x d b n N (0, V ) where V = 1 1 1 Exi x0 E xi x0 E xi x0 i f (0 | xi ) i i f (0 | xi ) 4
and f (e | x) is the conditional density of ei given xi = x.
The variance of the asymptotic distribution inversely depends on f (0 | x) , the conditional density of the error at its median. When f (0 | x) is large, then there are many innovations near to the median, and this improves estimation of the median. In the special case where the error is independent of xi , then f (0 | x) = f (0) and the asymptotic variance simplies V =
1 (Exi x0 i) 4f (0)2
(9.13)
This simplication is similar to the simplication of the asymptotic covariance of the OLS estimator under homoskedasticity. Computation of standard error for LAD estimates typically is based on equation (9.13). The main diculty is the estimation of f (0), the height of the error density at its median. This can be done with kernel estimation techniques. See Chapter 20. While a complete proof of Theorem 9.5.1 is advanced, we provide a sketch here for completeness. Proof of Theorem 9.5.1: Similar to NLLS, LAD is an optimization estimator. Let 0 denote the true value of 0 .
224
p The rst step is to show that 0 . The general nature of the proof is similar to that for the p NLLS estimator, and is sketched here. For any xed , by the WLLN, LADn () E |yi x0 i | . Furthermore, it can be shown that this convergence is uniform in . (Proving uniform convergence is more challenging than for the NLLS criterion since the LAD criterion is not dierentiable in , the minimizer of LADn ( ), converges in probability to 0 , the minimizer of .) It follows that 0 E |yi xi |. b ) = 0, where gn () = n1 Pn g i () Since sgn (a) = 1 2 1 (a 0) , (9.12) is equivalent to g n ( i=1 and g i () = xi (1 2 1 (yi x0 i )) . Let g ( ) = Eg i ( ). We need three preliminary results. First, by the central limit theorem (Theorem 5.7.1) n X d 1/2 n (g n (0 ) g (0 )) = n g i (0 ) N 0, Exi x0 i i=1
since Egi (0 )g i (0 )0 = Exi x0 i . Second using the law of iterated expectations and the chain rule of dierentiation, 0 0 g ( ) = 0 Exi 1 2 1 yi xi 0 = 2 0 E xi E 1 ei x0 i xi 0 | xi " Z 0 # xi x0 i 0 = 2 0 E xi f (e | xi ) de 0 0 = 2E xi x0 i f xi xi 0 | xi 0 0 g ( ) = 2E xi xi f (0 | xi ) .
so
Together
Third, by a Taylor series expansion and the fact g() = 0 b ) ' g () b . g ( 0 b ' n 0 '
1 ) ng( 0 g ( 0 ) 1 f (0 | x ) n g ( ) g ( ) = 2E xi x0 i n i
1 1 n (gn (0 ) g (0 )) E xi x0 i f (0 | xi ) 2 1 d 1 N 0, Exi x0 E xi x0 i f (0 | xi ) i 2 = N (0, V ) .
p b The third line follows from an asymptotic empirical process argument and the fact that 0 .
9.6
Quantile Regression
Quantile regression has become quite popular in recent econometric practice. For [0, 1] the th quantile Q of a random variable with distribution function F (u) is dened as Q = inf {u : F (u) }
225
When F (u) is continuous and strictly monotonic, then F (Q ) = , so you can think of the quantile as the inverse of the distribution function. The quantile Q is the value such that (percent) of the mass of the distribution is less than Q . The median is the special case = .5. The following alternative representation is useful. If the random variable U has th quantile Q , then (9.14) Q = argmin E (U ) .
where (q ) is the piecewise linear function q (1 ) (q) = q
= q ( 1 (q < 0)) .
q<0 q0
(9.15)
This generalizes representation (9.11) for the median to all quantiles. For the random variables (yi , xi ) with conditional distribution function F (y | x) the conditional quantile function q (x) is Q (x) = inf {y : F (y | x) } . Again, when F (y | x) is continuous and strictly monotonic in y , then F (Q (x) | x) = . For xed , the quantile regression function q (x) describes how the th quantile of the conditional distribution varies with the regressors. As functions of x, the quantile regression functions can take any shape. However for computational convenience it is typical to assume that they are (approximately) linear in x (after suitable transformations). This linear specication assumes that Q (x) = 0 x where the coecients vary across the quantiles . We then have the linear quantile regression model yi = x0 i + ei where ei is the error dened to be the dierence between yi and its th conditional quantile x0 i . By construction, the th conditional quantile of ei is zero, otherwise its properties are unspecied without further restrictions. b for solves the miniGiven the representation (9.14), the quantile regression estimator mization problem b = argmin S () n
where
() Sn
and (q) is dened in (9.15). ( ) does not have an algebraic solution, nuSince the quantile regression criterion function Sn merical methods are necessary for its minimization. Furthermore, since it has discontinuous derivatives, conventional Newton-type optimization methods are inappropriate. Fortunately, fast linear programming methods have been developed for this problem, and are widely available. An asymptotic distribution theory for the quantile regression estimator can be derived using similar arguments as those for the LAD estimator in Theorem 9.5.1.
1X = yi x0 i n
i=1
CHAPTER 9. REGRESSION EXTENSIONS Theorem 9.6.1 Asymptotic Distribution of the Quantile Regression Estimator When the th conditional quantile is linear in x d b n N (0, V ) , 1 1 Exi x0 E xi x0 V = (1 ) E xi x0 i f (0 | xi ) i i f (0 | xi )
226
where
and f (e | x) is the conditional density of ei given xi = x.
In general, the asymptotic variance depends on the conditional density of the quantile regression error. When the error ei is independent of xi , then f (0 | xi ) = f (0) , the unconditional density of ei at 0, and we have the simplication V = 1 (1 ) E xi x0 . i 2 f (0)
A recent monograph on the details of quantile regression is Koenker (2005).
227
Exercises
Exercise 9.1 Suppose that yi = g (xi , )+ei with E (ei | xi ) = 0, is the NLLS estimator, and V is . You are interested in the conditional mean function E (yi | xi = x) = g (x) the estimate of var at some x. Find an asymptotic 95% condence interval for g (x). Exercise 9.2 In Exercise 8.4, you estimated a cost function on a cross-section of electric companies. The equation you estimated was log T Ci = 1 + 2 log Qi + 3 log P Li + 4 log P Ki + 5 log P Fi + ei . (9.16)
(a) Following Nerlove, add the variable (log Qi )2 to the regression. Do so. Assess the merits of this new specication using a hypothesis test. Do you agree with this modication? (b) Now try a non-linear specication. Consider model (9.16) plus the extra term 6 zi , where zi = log Qi (1 + exp ( (log Qi 7 )))1 . In addition, impose the restriction 3 + 4 + 5 = 1. This model is called a smooth threshold model. For values of log Qi much below 7 , the variable log Qi has a regression slope of 2 . For values much above 7 , the regression slope is 2 + 6 , and the model imposes a smooth transition between these regimes. The model is non-linear because of the parameter 7 . The model works best when 7 is selected so that several values (in this example, at least 10 to 15) of log Qi are both below and above 7 . Examine the data and pick an appropriate range for 7 . (c) Estimate the model by non-linear least squares. I recommend the concentration method: Pick 10 (or more if you like) values of 7 in this range. For each value of 7 , calculate zi and estimate the model by OLS. Record the sum of squared errors, and nd the value of 7 for which the sum of squared errors is minimized. (d) Calculate standard errors for all the parameters (1 , ..., 7 ). Exercise 9.3 The data le cps78.dat contains 550 observations on 20 variables taken from the May 1978 current population survey. Variables are listed in the le cps78.pdf. The goal of the exercise is to estimate a model for the log of earnings (variable LNWAGE) as a function of the conditioning variables. (a) Start by an OLS regression of LNWAGE on the other variables. Report coecient estimates and standard errors. (b) Consider augmenting the model by squares and/or cross-products of the conditioning variables. Estimate your selected model and report the results. (c) Are there any variables which seem to be unimportant as a determinant of wages? You may re-estimate the model without these variables, if desired. (d) Test whether the error variance is dierent for men and women. Interpret. (e) Test whether the error variance is dierent for whites and nonwhites. Interpret. (f) Construct a model for the conditional variance. Estimate such a model, test for general heteroskedasticity and report the results. (g) Using this model for the conditional variance, re-estimate the model from part (c) using FGLS. Report the results.
CHAPTER 9. REGRESSION EXTENSIONS (h) Do the OLS and FGLS estimates dier greatly? Note any interesting dierences. (i) Compare the estimated standard errors. Note any interesting dierences. Exercise 9.4 For any predictor g (xi ) for yi , the mean absolute error (MAE) is E |yi g (xi )| .
228
Show that the function g (x) which minimizes the MAE is the conditional median m (x) = med(yi | xi ). Exercise 9.5 Dene g (u) = 1 (u < 0) where 1 () is the indicator function (takes the value 1 if the argument is true, else equals zero). Let satisfy Eg (yi ) = 0. Is a quantile of the distribution of yi ? Exercise 9.6 Verify equation (9.14).
Chapter 10
The Bootstrap
10.1 Denition of the Bootstrap
Let F denote a distribution function for the population of observations (yi , xi ) . Let Tn = Tn ((y1 , x1 ) , ..., (yn , xn ) , F ) or a t-statistic /s( ). Note that we be a statistic of interest, for example an estimator write Tn as possibly a function of F . For example, the t-statistic is a function of the parameter which itself is a function of F. The exact CDF of Tn when the data are sampled from the distribution F is Gn (u, F ) = Pr(Tn u | F ) In general, Gn (u, F ) depends on F, meaning that G changes as F changes. Ideally, inference would be based on Gn (u, F ). This is generally impossible since F is unknown. Asymptotic inference is based on approximating Gn (u, F ) with G(u, F ) = limn Gn (u, F ). When G(u, F ) = G(u) does not depend on F, we say that Tn is asymptotically pivotal and use the distribution function G(u) for inferential purposes. In a seminal contribution, Efron (1979) proposed the bootstrap, which makes a dierent approximation. The unknown F is replaced by a consistent estimate Fn (one choice is discussed in the next section). Plugged into Gn (u, F ) we obtain G n (u) = Gn (u, Fn ). (10.1)
We call G n the bootstrap distribution. Bootstrap inference is based on Gn (u). Let (yi , xi ) denote random variables with the distribution Fn . A random sample from this dis = T ((y , x ) , ..., (y , x ) , F ) constructed tribution is called the bootstrap data. The statistic Tn n n n n 1 1 on this sample is a random variable with distribution G n . That is, Pr(Tn u) = Gn (u). We call Tn the bootstrap statistic. The distribution of Tn is identical to that of Tn when the true CDF is Fn rather than F. The bootstrap distribution is itself random, as it depends on the sample through the estimator Fn . In the next sections we describe computation of the bootstrap distribution.
10.2
The Empirical Distribution Function
Recall that F (y, x) = Pr (yi y, xi x) = E (1 (yi y ) 1 (xi x)) , where 1() is the indicator function. This is a population moment. The method of moments estimator is the corresponding
229
CHAPTER 10. THE BOOTSTRAP
230
Figure 10.1: Empirical Distribution Functions sample moment:
Fn (y, x) is called the empirical distribution function (EDF). Fn is a nonparametric estimate of F. Note that while F may be either discrete or continuous, Fn is by construction a step function. The EDF is a consistent estimator of the CDF. To see this, note that for any (y, x), 1 (yi y) 1 (xi x) p is an iid random variable with expectation F (y, x). Thus by the WLLN (Theorem 5.4.2), Fn (y, x) F (y, x) . Furthermore, by the CLT (Theorem 5.7.1), d n (Fn (y, x) F (y, x)) N (0, F (y, x) (1 F (y, x))) . To see the eect of sample size on the EDF, in the Figure below, I have plotted the EDF and true CDF for three random samples of size n = 25, 50, 100, and 500. The random draws are from the N (0, 1) distribution. For n = 25, the EDF is only a crude approximation to the CDF, but the approximation appears to improve for the large n. In general, as the sample size gets larger, the EDF step function gets uniformly close to the true CDF. The EDF is a valid discrete probability distribution which puts probability mass 1/n at each , x ) with the pair (yi , xi ), i = 1, ..., n. Notationally, it is helpful to think of a random pair (yi i distribution Fn . That is, Pr(yi y, x i x) = Fn (y, x).
, x ) : We can easily calculate the moments of functions of (yi i Z Eh (yi , x h(y, x)dFn (y, x) i) =
1X Fn (y, x) = 1 (yi y ) 1 (xi x) . n

i=1
(10.2)
= = the empirical sample average.
n X i=1
h (yi , xi ) Pr (yi = yi , x i = xi ) n
1X h (yi , xi ) , n
i=1
231
10.3
Nonparametric Bootstrap
The nonparametric bootstrap is obtained when the bootstrap distribution (10.1) is dened using the EDF (10.2) as the estimate Fn of F. G Since the EDF Fn is a multinomial (with n support points), n could 2n1 in principle the distribution , x )}, be calculated by direct methods. However, as there are n possible samples {(y1 , x1 ) , ..., (yn n such a calculation is computationally infeasible. The popular alternative is to use simulation to approximate the distribution. The algorithm is identical to our discussion of Monte Carlo simulation, with the following points of clarication: The sample size n used for the simulation is the same as the sample size.
, x ) are drawn randomly from the empirical distribution. This is The random vectors (yi i equivalent to sampling a pair (yi , xi ) randomly from the sample. = T ((y , x ) , ..., (y , x ) , F ) is calculated for each bootstrap samThe bootstrap statistic Tn n n n n 1 1 ple. This is repeated B times. B is known as the number of bootstrap replications. A theory for the determination of the number of bootstrap replications B has been developed by Andrews and Buchinsky (2000). It is desirable for B to be large, so long as the computational costs are reasonable. B = 1000 typically suces. of F, it is typically through dependence on a parameter. When the statistic Tnis a function For example, the t-ratio /s() depends on . As the bootstrap statistic replaces F with the parameter Fn , it similarly replaces with n , the value of implied by Fn . Typically n = , estimate. (When in doubt use .) Sampling from the EDF is particularly easy. Since Fn is a discrete probability distribution putting probability mass 1/n at each sample point, sampling from the EDF is equivalent to random sampling a pair (yi , xi ) from the observed data with replacement. In consequence, a bootstrap , x ) , ..., (y , x )} will necessarily have some ties and multiple values, which is generally sample {(y1 n n 1 not a problem.
10.4
Bootstrap Estimation of Bias and Variance
is n = E( 0 ). Let Tn () = . Then n = E(Tn (0 )). The bootstrap The bias of n = . The bootstrap estimate counterparts are = ((y1 , x1 ) , ..., (yn , xn )) and Tn = of n is = E(Tn ). n
is If this is calculated by the simulation described in the previous section, the estimate of n B 1 X = Tnb B
1 B
. = is biased, it might be desirable to construct a biased-corrected estimator (one with reduced If bias). Ideally, this would be = n ,
b=1 B X b=1
CHAPTER 10. THE BOOTSTRAP but n is unknown. The (estimated) bootstrap biased-corrected estimator is
= n
232
. = 2
( ) =
. Intuitively, the bootstrap makes Note, in particular, that the biased-corrected estimator is not the following experiment. Suppose that is the truth. Then what is the average value of this suggests that the . If this is lower than , calculated from such samples? The answer is and the estimator is downward-biased, so a biased-corrected estimator of should be larger than , then the estimator is . Similarly if is higher than , and best guess is the dierence between upward-biased and the biased-corrected estimator should be lower than . The variance of is Let Tn = . Vn = E(Tn ETn )2 .
= . It has variance Let Tn 2 Vn = E(Tn ETn ) . B 1 X 2 n b = . V B b=1
The simulation estimate is
is the square root of the bootstrap estimate of variance, A bootstrap standard error for q . ) = V n s ( While this standard error may be calculated and reported, it is not clear if it is useful. The primary use of asymptotic standard errors is to construct asymptotic condence intervals, which are based on the asymptotic normal approximation to the t-ratio. However, the use of the bootstrap presumes that such asymptotic approximations might be poor, in which case the normal approximation is suspected. It appears superior to calculate bootstrap condence intervals, and we turn to this next.
10.5
Percentile Intervals
For a distribution function Gn (u, F ), let qn (, F ) denote its quantile function. This is the function which solves Gn (qn (, F ), F ) = . [When Gn (u, F ) is discrete, qn (, F ) may be non-unique, but we will ignore such complications.] () = q (, F ) Let qn () denote the quantile function of the true sampling distribution, and qn n n denote the quantile function of the bootstrap distribution. Note that this function will change depending on the underlying statistic Tn whose distribution is Gn . an estimate of a parameter of interest. In (1 )% of samples, lies in the region Let Tn = , [qn (/2), qn (1 /2)]. This motivates a condence interval proposed by Efron:
(/2), C1 = [qn qn (1 /2)].
This is often called the percentile condence interval. () is estimated by q (), the th sample quantile of the n Computationally, the quantile qn simulated statistics {Tn1 , ..., TnB }, as discussed in the section on Monte Carlo simulation. The (/2), (1 /2)]. q n (1 )% Efron percentile interval is then [ qn
233
The interval C1 is a popular bootstrap condence interval often used in empirical practice. This is because it is easy to compute, simple to motivate, was popularized by Efron early in the history of the bootstrap, and also has the feature that it is translation invariant. That is, if we dene = f () as the parameter of interest for a monotonically increasing function f, then percentile (/2)), (1 /2))], f (qn method applied to this problem will produce the condence interval [f (qn which is a naturally good property. However, as we show now, C1 is in a deep sense very poorly motivated. and let qn () It will be useful if we introduce an alternative denition of C1 . Let Tn () = be the quantile function of its distribution. (These are the original quantiles, with subtracted.) Then C1 can alternatively be written as + q (/2), C1 = [ n + q (1 /2)]. n + qn (1 /2)].
This is a bootstrap estimate of the ideal condence interval

0 + qn (/2), = [ C1
The latter has coverage probability 0 + qn (/2) 0 + qn (1 /2) = Pr Pr 0 C1 0 qn (/2) = Pr qn (1 /2)
= Gn (qn (/2), F0 ) Gn (qn (1 /2), F0 )
0 has a symmetric distribution which generally is not 1 ! There is one important exception. If about 0, then Gn (u, F0 ) = 1 Gn (u, F0 ), so 0 Pr 0 C1 = Gn (qn (/2), F0 ) Gn (qn (1 /2), F0 ) = (1 Gn (qn (/2), F0 )) (1 Gn (qn (1 /2), F0 )) 1 1 = 1 2 2 =1
0 and C are designed for the case and this idealized condence interval is accurate. Therefore, C1 1 that has a symmetric distribution about 0 . does not have a symmetric distribution, C1 may perform quite poorly. When However, by the translation invariance argument presented above, it also follows that if there ) is symmetrically distributed exists some monotonically increasing transformation f () such that f ( about f (0 ), then the idealized percentile bootstrap method will be accurate. Based on these arguments, many argue that the percentile interval should not be used unless the sampling distribution is close to unbiased and symmetric. The problems with the percentile method can be circumvented, at least in principle, by an alternative method. . Then Let Tn () =
so an exact (1 )% condence interval for 0 would be

0 qn (1 /2), C2 = [
1 = Pr (qn (/2) Tn (0 ) qn (1 /2)) qn (1 /2) 0 qn (/2) , = Pr qn (/2)]. q (/2)]. n
This motivates a bootstrap analog q (1 /2), C2 = [ n
234
Notice that generally this is very dierent from the Efron interval C1 ! They coincide in the special case that G n (u) is symmetric about , but otherwise they dier. Computationally, thisinterval can be estimated from a bootstrap simulation by sorting the These are sorted bootstrap statistics Tn = , which are centered at the sample estimate . (.025) and q (.975). The 95% condence interval is then [ to yield the quantile estimates q n n n (.025)]. q n (.975), q This condence interval is discussed in most theoretical treatments of the bootstrap, but is not widely used in practice.
10.6
Percentile-t Equal-Tailed Interval
we want to test H0 : = 0 against H1 : < 0 at size . We would set Tn () = Suppose /s( ) and reject H0 in favor of H1 if Tn (0 ) < c, where c would be selected so that Pr (Tn (0 ) < c) = . Thus c = qn (). Since this is unknown, a bootstrap test replaces qn () with the bootstrap estimate (), and the test rejects if T ( ) < q (). qn n 0 n (1 ). Similarly, if the alternative is H1 : > 0 , the bootstrap test rejects if Tn (0 ) > qn Computationally, these critical values can be estimated from a bootstrap simulation by sorting /s( ). Note, and this is important, that the bootstrap test the bootstrap t-statistics Tn = and the standard error s( ) is calculated on the bootstrap statistic is centered at the estimate , () and/or q (1 ). n sample. These t-statistics n are sorted to nd the estimated quantiles q /s( ). Then taking the intersection of two one-sided intervals, Let Tn () = 1 = Pr (qn (/2) Tn (0 ) qn (1 /2)) 0 /s( ) qn (1 /2) = Pr qn (/2) s( )qn (1 /2) 0 s( )qn (/2) , = Pr
0 s( )qn (1 /2), C3 = [
so an exact (1 )% condence interval for 0 would be
s( )qn (/2)].
This motivates a bootstrap analog s( )q (1 /2), C3 = [ n s( )q (/2)]. n
This is often called a percentile-t condence interval. It is equal-tailed or central since the probability that 0 is below the left endpoint approximately equals the probability that 0 is above the right endpoint, each /2. Computationally, this is based on the critical values from the one-sided hypothesis tests, discussed above.
10.7
Symmetric Percentile-t Intervals
we want to test H0 : = 0 against H1 : 6= 0 at size . We would set Tn () = Suppose /s() and reject H0 in favor of H1 if |Tn (0 )| > c, where c would be selected so that Pr (|Tn (0 )| > c) = .
CHAPTER 10. THE BOOTSTRAP Note that Pr (|Tn (0 )| < c) = Pr (c < Tn (0 ) < c) = Gn (c) Gn (c) Gn (c),
235
which is a symmetric distribution function. The ideal critical value c = qn () solves the equation Gn (qn ()) = 1 . Equivalently, qn () is the 1 quantile of the distribution of |Tn (0 )| . (), the 1 quantile of the distribution of |T | , or the number The bootstrap estimate is qn n which solves the equation
Gn (qn ()) = G n (qn ()) Gn (qn ()) = 1 . () is estimated from a bootstrap simulation by sorting the bootstrap tComputationally, qn | = ), and taking the upper % quantile. The bootstrap test rejects if statistics |Tn /s( (). |Tn (0 )| > qn Let s( )q (), + s( )q ()], C4 = [ n n
() is the bootstrap critical value for a two-sided hypothesis test. C is called the symmetric where qn 4 percentile-t interval. It is designed to work well since s( )q () 0 + s( )q () Pr (0 C4 ) = Pr n n
' Pr (|Tn (0 )| < qn ()) = 1 . If is a vector, then to test H0 : = 0 against H1 : 6= 0 at size , we would use a Wald statistic 0 1 V Wn () = n
()) = Pr (|Tn (0 )| < qn
() is found as the quantile from simulated values of W . Computationally, the critical value qn n , not 0 . Note in the simulation that the Wald statistic is a quadratic form in [This is a typical mistake made by practitioners.]
or some other asymptotically chi-square statistic. Thus here Tn () = Wn (). The ideal test rejects if Wn qn (), where qn () is the (1 )% quantile of the distribution of Wn . The bootstrap test (), where q () is the (1 )% quantile of the distribution of rejects if Wn qn n 0 1 V . Wn =n
10.8
Asymptotic Expansions
d
Let Tn R be a statistic such that Tn N(0, 2 ). (10.3)
236
u Gn (u, F ) = + o (1) . (10.4) u as n , it says nothing, however, about the rate While (10.4) says that Gn converges to of convergence, or the size of the divergence for any particular sample size n. A better asymptotic approximation may be obtained through an asymptotic expansion. The following notation will be helpful. Let an be a sequence.
In some cases, such as when Tn is a t-ratio, then 2 = 1. In other cases 2 is unknown. Equivalently, writing Tn Gn (u, F ) then for each u and F u , lim Gn (u, F ) = n or
Denition 10.8.1 an = o(1) if an 0 as n Denition 10.8.2 an = O(1) if |an | is uniformly bounded. Denition 10.8.3 an = o(nr ) if nr |an | 0 as n .
Basically, an = O(nr ) if it declines to zero like nr . We say that a function g (u) is even if g (u) = g (u), and a function h(u) is odd if h(u) = h(u). The derivative of an even function is odd, and vice-versa.
Theorem 10.8.1 Under regularity conditions and (10.3), Gn (u, F ) = u + g1 (u, F ) + n1/2 1 1 g2 (u, F ) + O(n3/2 ) n
uniformly over u, where g1 is an even function of u, and g2 is an odd function of u. Moreover, g1 and g2 are dierentiable functions of u and continuous in F relative to the supremum norm on the space of distribution functions.
The expansion in Theorem 10.8.1 is often called an Edgeworth expansion. We can interpret Theorem 10.8.1 as follows. First, Gn (u, F ) converges to the normal limit at rate n1/2 . To a second order of approximation, u + n1/2 g1 (u, F ). Gn (u, F )
which adds a symmetric non-normal component to the approximate density (for example, adding leptokurtosis).
Since the derivative of g1 is odd, the density function is skewed. To a third order of approximation, u Gn (u, F ) + n1/2 g1 (u, F ) + n1 g2 (u, F )
CHAPTER 10. THE BOOTSTRAP [Side Note: When Tn = n Xn /, a standardized sample mean, then
237
where (u) is the standard normal pdf, and
1 g1 (u) = 3 u2 1 (u) 6 5 1 1 3 4 u3 3u + 2 10 u + 15 u (u) u g2 (u) = 24 72 3 3 = E (X )3 / 3
4 = E (X )4 / 4 3
the standardized skewness and excess kurtosis of the distribution of X. Note that when 3 = 0 and 4 = 0, then g1 = 0 and g2 = 0, so the second-order Edgeworth expansion corresponds to the normal distribution.]
Francis Edgeworth
Francis Ysidro Edgeworth (1845-1926) of Ireland, founding editor of the Economic Journal, was a profound economic and statistical theorist, developing the theories of indierence curves and asymptotic expansions. He also could be viewed as the rst econometrician due to his early use of mathematical statistics in the study of economic data.
10.9
One-Sided Tests
Using the expansion of Theorem 10.8.1, we can assess the accuracy of one-sided hypothesis tests and condence regions based on an asymptotically normal t-ratio Tn . An asymptotic test is based on (u). To the second order, the exact distribution is Pr (Tn < u) = Gn (u, F0 ) = (u) + since = 1. The dierence is (u) Gn (u, F0 ) = 1 g1 (u, F0 ) + O(n1 ) 1 n /2 1 g1 (u, F0 ) + O(n1 ) 1 n /2
= O(n1/2 ), so the order of the error is O(n1/2 ). A bootstrap test is based on G n (u), which from Theorem 10.8.1 has the expansion G n (u) = Gn (u, Fn ) = (u) + 1 g1 (u, Fn ) + O(n1 ). 1 n /2
Because (u) appears in both expansions, the dierence between the bootstrap distribution and the true distribution is G n (u) Gn (u, F0 ) = 1 n1/2 (g1 (u, Fn ) g1 (u, F0 )) + O(n1 ).
238
Since Fn converges to F at rate n, and g1 is continuous with respect to F, the dierence (g1 (u, Fn ) g1 (u, F0 )) converges to 0 at rate n. Heuristically, g1 (u, Fn ) g1 (u, F0 )
F g1 (u, F )
g1 (u, F0 ) (Fn F0 ) F = O(n1/2 ),
The derivative
is only heuristic, as F is a function. We conclude that

1 G n (u) Gn (u, F0 ) = O (n ),
or
u) = Pr (Tn u) + O(n1 ), Pr (Tn
which is an improved rate of convergence over the asymptotic test (which converged at rate O(n1/2 )). This rate can be used to show that one-tailed bootstrap inference based on the tratio achieves a so-called asymptotic renement the Type I error of the test converges at a faster rate than an analogous asymptotic test.
10.10
Symmetric Two-Sided Tests
If a random variable y has distribution function H (u) = Pr(y u), then the random variable |y| has distribution function H (u) = H (u) H (u) since Pr (|y| u) = Pr (u y u) = H (u) H (u).
= Pr (y u) Pr (y u)
For example, if Z N(0, 1), then |Z | has distribution function (u) = (u) (u) = 2(u) 1. Similarly, if Tn has exact distribution Gn (u, F ), then |Tn | has the distribution function Gn (u, F ) = Gn (u, F ) Gn (u, F ). A two-sided hypothesis test rejects H0 for large values of |Tn | . Since Tn Z, then |Tn | |Z | . Thus asymptotic critical values are taken from the distribution, and exact critical values are taken from the Gn (u, F0 ) distribution. From Theorem 10.8.1, we can calculate that Gn (u, F ) = Gn (u, F ) Gn (u, F ) 1 1 = (u) + 1/2 g1 (u, F ) + g2 (u, F ) n n 1 1 (u) + 1/2 g1 (u, F ) + g2 (u, F ) + O(n3/2 ) n n 2 = (u) + g2 (u, F ) + O(n3/2 ), n
d d
(10.5)
where the simplications are because g1 is even and g2 is odd. Hence the dierence between the asymptotic distribution and the exact distribution is (u) Gn (u, F0 ) = 2 g2 (u, F0 ) + O(n3/2 ) = O(n1 ). n
239
The order of the error is O(n1 ). Interestingly, the asymptotic two-sided test has a better coverage rate than the asymptotic one-sided test. This is because the rst term in the asymptotic expansion, g1 , is an even function, meaning that the errors in the two directions exactly cancel out. Applying (10.5) to the bootstrap distribution, we nd Gn (u) = Gn (u, Fn ) = (u) +
2 g2 (u, Fn ) + O(n3/2 ). n
Thus the dierence between the bootstrap and exact distributions is 2 (g2 (u, Fn ) g2 (u, F0 )) + O(n3/2 ) n = O(n3/2 ), the last equality because Fn converges to F0 at rate n, and g2 is continuous in F. Another way of writing this is | < u) = Pr (|Tn | < u) + O(n3/2 ) Pr (|Tn Gn (u) Gn (u, F0 ) =
so the error from using the bootstrap distribution (relative to the true unknown distribution) is O(n3/2 ). This is in contrast to the use of the asymptotic distribution, whose error is O(n1 ). Thus a two-sided bootstrap test also achieves an asymptotic renement, similar to a one-sided test. A reader might get confused between the two simultaneous eects. Two-sided tests have better rates of convergence than the one-sided tests, and bootstrap tests have better rates of convergence than asymptotic tests. The analysis shows that there may be a trade-o between one-sided and two-sided tests. Twosided tests will have more accurate size (Reported Type I error), but one-sided tests might have more power against alternatives of interest. Condence intervals based on the bootstrap can be asymmetric if based on one-sided tests (equal-tailed intervals) and can therefore be more informative and have smaller length than symmetric intervals. Therefore, the choice between symmetric and equal-tailed condence intervals is unclear, and needs to be determined on a case-by-case basis.
10.11
d
Percentile Condence Intervals

n 0 . We know that
To evaluate the coverage rate of the percentile interval, set Tn =
Tn N(0, V ), which is not pivotal, as it depends on the unknown V. Theorem 10.8.1 shows that a rst-order approximation u + O(n1/2 ), Gn (u, F ) = where = V , and for the bootstrap u + O(n1/2 ), ( u ) = G ( u, F ) = G n n n where = V (Fn ) is the bootstrap estimate of . The dierence is u u + O(n1/2 ) G ( u ) G ( u, F ) = n 0 n u u ( ) + O(n1/2 ) = = O(n1/2 )
Hence the order of the error is O(n1/2 ). The good news is that the percentile-type methods (if appropriately used) can yield nconvergent asymptotic inference. Yet these methods do not require the calculation of standard
240
errors! This means that in contexts where standard errors are not available or are dicult to calculate, the percentile bootstrap methods provide an attractive inference method. The bad news is that the rate of convergence is disappointing. It is no better than the rate obtained from an asymptotic one-sided condence region. Therefore if standard errors are available, it is unclear if there are any benets from using the percentile bootstrap over simple asymptotic methods. Based on these arguments, the theoretical literature (e.g. Hall, 1992, Horowitz, 2001) tends to advocate the use of the percentile-t bootstrap methods rather than percentile methods.
10.12
Bootstrap Methods for Regression Models
The bootstrap methods we have discussed have set G n (u) = Gn (u, Fn ), where Fn is the EDF. Any other consistent estimate of F may be used to dene a feasible bootstrap estimator. The advantage of the EDF is that it is fully nonparametric, it imposes no conditions, and works in nearly any context. But since it is fully nonparametric, it may be inecient in contexts where more is known about F. We discuss bootstrap methods appropriate for the linear regression model yi = x0 i + ei
, x ) from the EDF, which implies The non-parametric bootstrap resamples the observations (yi i 0 yi = xi + ei E (x i ei ) = 0
E (ei | xi ) = 0.
but generally The bootstrap distribution does not impose the regression assumption, and is thus an inecient estimator of the true distribution (when in fact the regression assumption is true.) One approach to this problem is to impose the very strong assumption that the error i is independent of the regressor xi . The advantage is that in this case it is straightforward to construct bootstrap distributions. The disadvantage is that the bootstrap distribution may be a poor approximation when the error is not independent of the regressors. To impose independence, it is sucient to sample the x i and ei independently, and then create 0 yi = xi + ei . There are dierent ways to impose independence. A non-parametric method 1 , ..., e n }. A parametric is to sample the bootstrap errors e i randomly from the OLS residuals {e method is to generate the bootstrap errors ei from a parametric distribution, such as the normal 2 ). e i N(0, For the regressors x i , a nonparametric method is to sample the xi randomly from the EDF or sample values {x1 , ..., xn }. A parametric method is to sample xi from an estimated parametric distribution. A third approach sets x i = xi . This is equivalent to treating the regressors as xed in repeated samples. If this is done, then all inferential statements are made conditionally on the observed values of the regressors, which is a valid statistical approach. It does not really matter, however, whether or not the xi are really xed or random. The methods discussed above are unattractive for most applications in econometrics because they impose the stringent assumption that xi and ei are independent. Typically what is desirable is to impose only the regression condition E (ei | xi ) = 0. Unfortunately this is a harder problem. One proposal which imposes the regression condition without independence is the Wild Bootstrap. The idea is to construct a conditional distribution for e i so that E (e i | xi ) = 0 2 2 E ei | xi = e i 3 E ei | xi = e 3 i.
E (e i | xi ) 6= 0.
241
A conditional distribution with these features will preserve the main important features of the data. This can be achieved using a two-point distribution of the form ! ! 1 + 5 51 Pr e e i = i = 2 2 5 ! ! 1 5 5+1 Pr e e i = i = 2 2 5 For each xi , you sample e i using this two-point distribution.
242
Exercises
Exercise 10.1 Let Fn (x) denote the EDF of a random sample. Show that d n (Fn (x) F0 (x)) N (0, F0 (x) (1 F0 (x))) . Exercise 10.2 Take a random sample {y1 , ..., yn } with = Eyi and 2 = var (yi ) . Let the statistic of interest be the sample mean Tn = y n . Find the population moments ETn and var (Tn ) . Let , ..., y } be a random sample from the empirical distribution function and let T = y be its {y1 n n n and var (T ) . sample mean. Find the bootstrap moments ETn n Exercise 10.3 Consider the following bootstrap procedure for a regression of yi on xi . Let the OLS residuals. denote the OLS estimator from the regression of y on X , and e = y X i ) : i = 1, ..., n} . That is, draw a random (a) Draw a random vector (x , e ) from the pair {(xi , e + e . Draw (with i0 . Set y = x0 integer i0 from [1, 2, ..., n], and set x = xi0 and e = e replacement) n such vectors, creating a random bootstrap data set (y , X ). and any other statistic of interest. (b) Regress y on X , yielding OLS estimates Show that this bootstrap procedure is (numerically) identical to the non-parametric bootstrap. Exercise 10.4 Consider the following bootstrap procedure. Using the non-parametric bootstrap, on these samples and then calculate generate bootstrap samples, calculate the estimate
)/s( ), Tn = (
) is the standard error in the original data. Let q (.05) and q (.95) denote the 5% and where s( n n , and dene the bootstrap condence interval 95% quantiles of Tn h i s( )q (.95), s( )q (.05) . C= n n Show that C exactly equals the Alternative percentile interval (not the percentile-t interval). Exercise 10.5 You want to test H0 : = 0 against H1 : > 0. The test for H0 is to reject if ( ) > c where c is picked so that Type I error is . You do this as follows. Using the nonTn = /s on these samples parametric bootstrap, you generate bootstrap samples, calculate the estimates and then calculate /s( ). = Tn
(.95) denote the 95% quantile of T . You replace c with q (.95), and thus reject H if Let qn 0 n n ( ) > q (.95). What is wrong with this procedure? Tn = /s n
= 1.2 and s( ) = .2. Using the non-parametric Exercise 10.6 Suppose that in an application, is calculated on each bootstrap, 1000 samples are generated from the bootstrap distribution, and are .75 and 1.3, respectively. are sorted, and the 2.5% and 97.5% quantiles of the sample. The (a) Report the 95% Efron Percentile interval for . (b) Report the 95% Alternative Percentile interval for . (c) With the given information, can you report the 95% Percentile-t interval for ? Exercise 10.7 The datale hprice1.dat contains data on house prices (sales), with variables listed in the le hprice1.pdf. Estimate a linear regression of price on the number of bedrooms, lot size, size of house, and the colonial dummy. Calculate 95% condence intervals for the regression coecients using both the asymptotic normal approximation and the percentile-t bootstrap.
Chapter 11
NonParametric Regression
11.1 Introduction
When components of x are continuously distributed then the conditional expectation function E (yi | xi = x) = m(x) can take any nonlinear shape. Unless an economic model restricts the form of m(x) to a parametric function, the CEF is inherently nonparametric, meaning that the function m(x) is an element of an innite dimensional class. In this situation, how can we estimate m(x)? What is a suitable method, if we acknowledge that m(x) is nonparametric? There are two main classes of nonparametric regression estimators: kernel estimators, and series estimators. In this chapter we introduce kernel methods. To get started, suppose that there is a single real-valued regressor xi . We consider the case of vector-valued regressors later.
11.2
Binned Estimator
For clarity, x the point x and consider estimation of the single point m(x). This is the mean of yi for random pairs (yi , xi ) such that xi = x. If the distribution of xi were discrete then we could estimate m(x) by taking the average of the sub-sample of observations yi for which xi = x. But when xi is continuous then the probability is zero that xi exactly equals any specic x. So there is no sub-sample of observations with xi = x and we cannot simply take the average of the corresponding yi values. However, if the CEF m(x) is continuous, then it should be possible to get a good approximation by taking the average of the observations for which xi is close to x, perhaps for the observations for which |xi x| h for some small h > 0. Later we will call h a bandwidth. This estimator can be written as Pn i=1 1 (|xi x| h) yi (11.1) m b (x) = P n i=1 1 (|xi x| h) where 1() is the indicator function. Alternatively, (11.1) can be written as m b (x) =
n X i=1
wi (x)yi
(11.2)
where Pn
Notice that
i=1 wi (x)
= 1, so (11.2) is a weighted average of the yi . 243
1 (|xi x| h) wi (x) = Pn . j =1 1 (|xj x| h)
CHAPTER 11. NONPARAMETRIC REGRESSION
244
Figure 11.1: Scatter of (yi , xi ) and Nadaraya-Watson regression It is possible P that for some values of x there are no values of xi such that |xi x| h, which implies that n i=1 1 (|xi x| h) = 0. In this case the estimator (11.1) is undened for those values of x. To visualize, Figure 11.1 displays a scatter plot of 100 observations on a random pair (yi , xi ) generated by simulation1 . (The observations are displayed as the open circles.) The estimator (11.1) of the CEF m(x) at x = 2 with h = 1/2 is the average of the yi for the observations such that xi falls in the interval [1.5 xi 2.5]. (Our choice of h = 1/2 is somewhat arbitrary. Selection of h will be discussed later.) The estimate is m b (2) = 5.16 and is shown on Figure 11.1 by the rst solid square. We repeat the calculation (11.1) for x = 3, 4, 5, and 6, which is equivalent to partitioning the support of xi into the regions [1.5, 2.5], [2.5, 3.5], [3.5, 4.5], [4.5, 5.5], and [5.5, 6.5]. These partitions are shown in Figure 11.1 by the verticle dotted lines, and the estimates (11.1) by the solid squares. These estimates m b (x) can be viewed as estimates of the CEF m(x). Sometimes called a binned estimator, this is a step-function approximation to m(x) and is displayed in Figure 11.1 by the horizontal lines passing through the solid squares. This estimate roughly tracks the central tendency of the scatter of the observations (yi , xi ). However, the huge jumps in the estimated step function at the edges of the partitions are disconcerting, counter-intuitive, and clearly an artifact of the discrete binning. If we take another look at the estimation formula (11.1) there is no reason why we need to evaluate (11.1) only on a course grid. We can evaluate m b (x) for any set of values of x. In particular, we can evaluate (11.1) on a ne grid of values of x and thereby obtain a smoother estimate of the CEF. This estimator with h = 1/2 is displayed in Figure 11.1 with the solid line. This is a generalization of the binned estimator and by construction passes through the solid squares. The bandwidth h determines the degree of smoothing. Larger values of h increase the width of the bins in Figure 11.1, thereby increasing the smoothness of the estimate m b (x) as a function of x. Smaller values of h decrease the width of the bins, resulting in less smooth conditional mean estimates.
1
The distribution is xi N (4, 1) and yi | xi N (m(xi ), 16) with m(x) = 10 log(x).
245
11.3
Kernel Regression
One deciency with the estimator (11.1) is that it is a step function in x, as it is discontinuous at each observation x = xi . That is why its plot in Figure 11.1 is jagged. The source of the discontinuity is that the weights wi (x) are constructed from indicator functions, which are themselves discontinuous. If instead the weights are constructed from continuous functions then the CEF estimator will also be continuous in x. To generalize (11.1) it is useful to write the weights 1 (|xi x| h) in terms of the uniform density function on [1, 1] 1 k0 (u) = 1 (|u| 1) . 2 Then xi x xi x 1 = 2k0 . 1 (|xi x| h) = 1 h h and (11.1) can be written as xi x yi i=1 k0 h . m b (x) = Pn xi x i=1 k0 h Pn (11.3)
The uniform density k0 (u) is a special case of what is known as a kernel function.
Denition 11.3.1 A second-order kernel function R k (u) satises 0 R 2 = u2 k (u)du < . k (u) < , k(u) = k (u), k(u)du = 1 and k Essentially, a kernel function is a probability density function which is bounded and symmetric about zero. A generalization of (11.1) is obtained by replacing the uniform kernel with any other kernel function: Pn xi x yi i=1 k h . (11.4) m b (x) = Pn xi x k i=1 h
The estimator (11.4) is known as the Nadaraya-Watson estimator, the kernel regression estimator, or the local constant estimator. The bandwidth h plays the same role in (11.4) as it does in (11.1). Namely, larger values of h will result in estimates m b (x) which are smoother in x, and smaller values of h will result in estimates which are more erratic. It might be helpful to consider the two extreme cases h 0 and b (x) is h . As h 0 we can see that m b (xi ) yi (if the values of xi are unique), so that m b (x) y, the sample mean, simply the scatter of yi on xi . In contrast, as h then for all x, m so that the nonparametric CEF estimate is a constant function. For intermediate values of h, m b (x) will lie between these two extreme cases.
The estimator (11.4) also takes the form (11.2) with xi x k h . wi (x) = Pn xj x j =1 k h
246
The uniform density is not a good kernel choice as it produces discontinuous CEF estimates. To obtain a continuous CEF estimate m b (x) it is necessary for the kernel k(u) to be continuous. The two most commonly used choices are the Epanechnikov kernel k1 (u) = and the normal or Gaussian kernel 3 1 u2 1 (|u| 1) 4
2 u 1 . k (u) = exp 2 2 For computation of the CEF estimate (11.4) the scale of the kernel is not important so long as u the bandwidth is selected appropriately. That is, for any b > 0, kb (u) = b1 k is a valid kernel b function with the identical shape as k(u). Kernel regression with the kernel k(u) and bandwidth h is identical to kernel regression with the kernel kb (u) and bandwidth h/b. The estimate (11.4) using the Epanechnikov kernel and h = 1/2 is also displayed in Figure 11.1 with the dashed line. As you can see, this estimator appears to be much smoother than that using the uniform kernel. 2 and roughTwo important constants associated with a kernel function k (u) are its variance k ness Rk , which are dened as Z 2 k = u2 k (u)du (11.5) Z k (u)2 du. (11.6) Rk =
Some common kernels and their roughness and variance values are reported in Table 9.1. Table 9.1: Common Second-Order Kernels Kernel Uniform Epanechnikov Biweight Triweight Gaussian Equation k0 (u) = 1 1) 21 (|u| 2 1 (|u| 1) 1 u k1 (u) = 3 4 2 2 k2 (u) = 15 16 1 u 1 (|u| 1) 2 3 k3 (u) = 35 |u| 1) 32 1 u 1( k (u) =
1 2
exp u 2
Rk 1/2 3/5 5/7 350/429 1/ (2 )
2 k 1/3 1/5 1/7 1/9
11.4
Local Linear Estimator
The Nadaraya-Watson (NW) estimator is often called a local constant estimator as it locally (about x) approximates the CEF m(x) as a constant function. One way to see this is to observe that m b (x) solves the minimization problem This is a weighted regression of yi on an intercept only. Without the weights, this estimation problem reduces to the sample mean. The NW estimator generalizes this to a local mean. This interpretation suggests that we can construct alternative nonparametric estimators of the CEF by alternative local approximations. Many such local approximations are possible. A popular choice is the local linear (LL) approximation. Instead of approximating m(x) locally as a constant, n X xi x (yi )2 . k m b (x) = argmin h
i=1
247
The local linear estimator of m(x) is the estimated intercept m b (x) = b(x)
the local linear approximation approximates the CEF locally by a linear function, and estimates this local approximation by locally weighted least squares. Specically, for each x we solve the following minimization problem n n o X xi x b (yi (xi x))2 . b(x), (x) = argmin k h ,
i=1
and the local linear estimator of the regression derivative m(x) is the estimated slope coecient b(x). d m(x) = z i (x) = and ki (x) = k Then b(x) b(x) = n X
i=1
Computationally, for each x set
1 xi x xi x h
ki (x)z i (x)z i (x)
where K = diag{k1 (x), ..., kn (x)}. To visualize, Figure 11.2 displays the scatter plot of the same 100 observations from Figure 11.1, divided into three regions depending on the regressor xi : [1, 3], [3, 5], [5, 7]. A linear regression is t to the observations in each region, with the observations weighted by the Epanechnikov kernel with h = 1. The three tted regression lines are displayed by the three straight solid lines. The values of these regression lines at x = 2, x = 4 and x = 6, respectively, are the local linear estimates m b (x) at x = 2, 4, and 6. This estimation is repeated for all x in the support of the regressors, and plotted as the continuous solid line in Figure 11.2. One interesting feature is that as h , the LL estimator approaches the full-sample linear . That is because as h all observations receive equal least-squares estimator m b (x) + x weight regardless of x. In this sense we can see that the LL estimator is a exible generalization of the linear OLS estimator. Which nonparametric estimator should you use in practice: NW or LL? The theoretical literature shows that neither strictly dominates the other, but we can describe contexts where one or the other does better. Roughly speaking, the NW estimator performs better than the LL estimator when m(x) is close to a at line, but the LL estimator performs better when m(x) is meaningfully non-constant. The LL estimator also performs better for values of x near the boundary of the support of xi .
1 0 = Z 0 KZ Z Ky
!1
n X i=1
ki (x)z i (x)yi
11.5
Nonparametric Residuals and Regression Fit
The tted regression at x = xi is m b (xi ) and the tted residual is e i = yi m b (xi ).
248
Figure 11.2: Scatter of (yi , xi ) and Local Linear tted regression As a general rule, but especially when the bandwidth h is small, it is hard to view e i as a good i 0. This clearly measure of the t of the regression. As h 0 then m b (xi ) yi and therefore e indicates overtting as the true error is not zero. In general, since m b (xi ) is a local average which i small, and the degree includes yi , the tted value will be necessarily close to yi and the residual e of this overtting increases as h decreases. A standard solution is to measure the t of the regression at x = xi by re-estimating the model excluding the ith observation. For Nadaraya-Watson regression, the leave-one-out estimator of m(x) excluding observation i is P xj x yj j 6=i k h . m e i (x) = P xj x k j 6=i h Notationally, the i subscript is used to indicate that the ith observation is omitted. The leave-one-out predicted value for yi at x = xi equals P xj xi yj j 6=i k h . y i = m e i (xi ) = P xj xi j 6=i k h
The leave-one-out residuals (or prediction errors) are the dierence between the leave-one-out predicted values and the actual observation i . e i = yi y Since y i is not a function of yi , there is no tendency for y i to overt for small h. Consequently, e i is a good measure of the t of the estimated nonparametric regression. Similarly, the leave-one-out local-linear residual is e i = yi ei with 1 X X ei 0 k z z kij z ij yj , = ij ij ij ei
j 6=i j 6=i
CHAPTER 11. NONPARAMETRIC REGRESSION z ij = and 1 xj xi xj xi h .
249
kij = k
11.6
Cross-Validation Bandwidth Selection
As we mentioned before, the choice of bandwidth h is crucial. As h increases, the kernel regression estimators (both NW and LL) become more smooth, ironing out the bumps and wiggles. This reduces estimation variance but at the cost of increased bias and oversmoothing. As h decreases the estimators become more wiggly, erratic, and noisy. It is desirable to select h to trade-o these features. How can this be done systematically? To be explicit about the dependence of the estimator on the bandwidth, let us write the estimator of m(x) with a given bandwidth h as m b (x, h), and our discussion will apply equally to the NW and LL estimators. Ideally, we would like to select h to minimize the mean-squared error (MSE) of m b (x, h) as a estimate of m(x). For a given value of x the MSE is We are typically interested in estimating m(x) for all values in the support of x. A common measure for the average t is the integrated MSE Z IM SEn (h) = M SEn (x, h)fx (x)dx Z = E (m b (x, h) m(x))2 fx (x)dx b (x, h) m(x))2 . M SEn (x, h) = E (m
where fx (x) is the marginal density of xi . Notice that we have dened the IMSE as an integral with respect to the density fx (x). Other weight functions could be used, but it turns out that this is a convenient choice. The IMSE is closely related with the MSFE of Section 4.9. Let (yn+1 , xn+1 ) be out-of-sample observations (and thus independent of the sample) and consider predicting yn+1 given xn+1 and b (xn+1 , h) which has the nonparametric estimate m b (x, h). The natural point estimate for yn+1 is m mean-squared forecast error b (xn+1 , h))2 M SF En (h) = E (yn+1 m = E (en+1 + m(xn+1 ) m b (xn+1 , h))2
b (x, h). We thus see that where the nal equality uses the fact that xn+1 is independent of m M SF En (h) = 2 + IM SEn (h). Since 2 is a constant independent of the bandwidth h, M SF En (h) and IM SEn (h) are equivalent measures of the t of the nonparameric regression. The optimal bandwidth h is the value which minimizes IM SEn (h) (or equivalently M SF En (h)). While these functions are unknown, we learned in Theorem 4.9.1 that (at least in the case of linear regression) M SF En can be estimated by the sample mean-squared prediction errors. It turns out that this fact extends to nonparametric regression. The nonparametric leave-one-out residuals are e i (xi , h) e i (h) = yi m
= 2 + E (m(xn+1 ) m b (xn+1 , h))2 Z = 2 + E (m b (x, h) m(x))2 fx (x)dx
250
This function of h is known as the cross-validation criterion. The cross-validation bandwidth b h is the value which minimizes CV (h) b h = argmin CV (h)
hh
where we are being explicit about the dependence on the bandwidth h. The mean squared leaveone-out residuals is n 1X CV (h) = e i (h)2 . n
i=1
(11.7)
for some h > 0. The restriction h h is imposed so that CV (h) is not evaluated over unreasonably small bandwidths. There is not an explicit solution to the minimization problem (11.7), so it must be solved numerically. A typical practical method is to create a grid of values for h, e.g. [h1 , h2 , ..., hJ ], evaluate CV (hj ) for j = 1, ..., J, and set b h= argmin
h[h1 ,h2 ,...,hJ ]
CV (h).
Evaluation using a coarse grid is typically sucient for practical application. Plots of CV (h) against h are a useful diagnostic tool to verify that the minimum of CV (h) has been obtained. We said above that the cross-validation criterion is an estimator of the MSFE. This claim is based on the following result.
Theorem 11.6.1 E (CV (h)) = M SF En1 (h) = IM SEn1 (h) + 2 (11.8)
Theorem 11.6.1 shows that CV (h) is an unbiased estimator of IM SEn1 (h) + 2 . The rst term, IM SEn1 (h), is the integrated MSE of the nonparametric estimator using a sample of size n 1. If n is large, IM SEn1 (h) and IM SEn (h) will be nearly identical, so CV (h) is essentially unbiased as an estimator of IM SEn (h) + 2 . Since the second term ( 2 ) is unaected by the bandwidth h, it is irrelevant for the problem of selection of h. In this sense we can view CV (h) as an estimator of the IMSE, and more importantly we can view the minimizer of CV (h) as an estimate of the minimizer of IM SEn (h). To illustrate, Figure 11.3 displays the cross-validation criteria CV (h) for the Nadaraya-Watson and Local Linear estimators using the data from Figure 11.1, both using the Epanechnikov kernel. The CV functions are computed on a grid with intervals 0.01. The CV-minimizing bandwidths are h = 1.09 for the Nadaraya-Watson estimator and h = 1.59 for the local linear estimator. Figure 11.3 shows the minimizing bandwidths by the arrows. It is not surprising that the CV criteria recommend a larger bandwidth for the LL estimator than for the NW estimator, as the LL employs more smoothing for a given bandwidth. The CV criterion can also be used to select between dierent nonparametric estimators. The CV-selected estimator is the one with the lowest minimized CV criterion. For example, in Figure 11.3, the NW estimator has a minimized CV criterion of 16.88, while the LL estimator has a minimized CV criterion of 16.81. Since the LL estimator achieves a lower value of the CV criterion, LL is the CV-selected estimator. The dierence (0.07) is small, suggesting that the two estimators are near equivalent in IMSE. Figure 11.4 displays the tted CEF estimates (NW and LL) using the bandwidths selected by cross-validation. Also displayed is the true CEF m(x) = 10 ln(x). Notice that the nonparametric
251
Figure 11.3: Cross-Validation Criteria, Nadaraya-Watson Regression and Local Linear Regression
Figure 11.4: Nonparametric Estimates using data-dependent (CV) bandwidths
252
estimators with the CV-selected bandwidths (and especially the LL estimator) track the true CEF quite well. e i (xi , h) is a function only of (x1 , ..., xn ) and Proof of Theorem 11.6.1. Observe that m(xi ) m (e1 , ..., en ) excluding ei , and is thus uncorrelated with ei . Since e i (h) = m(xi ) m e i (xi , h) + ei , then E (CV (h)) = E e i (h)2 = E e2 e i (xi , h) m(xi ))2 i + E (m + 2E ((m e i (xi , h) m(xi )) ei ) e i (x, h), which are indepenThe second term is an expectation over the random variables xi and m dent as the second is not a function of the ith observation. Thus taking the conditional expectation given the sample excluding the ith observation, this is the expectation over xi only, which is the integral with respect to its density Z e i (xi , h) m(xi ))2 = (m e i (x, h) m(x))2 fx (x)dx. Ei (m E (m e i (xi , h) m(xi )) = E
2
= 2 + E (m e i (xi , h) m(xi ))2 .
(11.9)
Taking the unconditional expecation yields
= IM SEn1 (h)
(m e i (x, h) m(x))2 fx (x)dx
where this is the IMSE of a sample of size n 1 as the estimator m e i uses n 1 observations. Combined with (11.9) we obtain (11.8), as desired.
11.7
There is no nite sample distribution theory for kernel estimators, but there is a well developed asymptotic distribution theory. The theory is based on the approximation that the bandwidth h decreases to zero as the sample size n increases. This means that the smoothing is increasingly localized as the sample size increases. So long as the bandwidth does not decrease to zero too quickly, the estimator can be shown to be asymptotically normal, 2 but with a non-trivial bias. 2 Let fx (x) denote the marginal density of xi and (x) = E ei | xi = x denote the conditional variance of ei = yi m(xi ).
CHAPTER 11. NONPARAMETRIC REGRESSION Theorem 11.7.1 Let m b (x) denote either the Nadarya-Watson or Local Linear estimator of m(x). If x is interior to the support of xi and fx (x) > 0, then as n and h 0 such that nh , d Rk 2 (x) 2 (11.10) nh m b (x) m(x) h2 k B (x) N 0, fx (x)
2 and R are dened in (11.5) and (11.6). For the Nadarayawhere k k Watson estimator
253
1 0 (x)m0 (x) B (x) = m00 (x) + fx (x)1 fx 2 and for the local linear estimator 1 B (x) = fx (x)m00 (x) 2
There are several interesting features about the asymptotic distribution which are noticeably dierent than for the estimator converges at the rate nh, not n. parametric estimators. First, Since h 0, nh diverges slower than n, thus the nonparametric estimator converges more slowly than a parametric estimator. Second, the asymptotic distribution contains a non-neglible 2 B (x). This term asymptotically disappears since h 0. Third, the assumptions that bias term h2 k nh and h 0 mean that the estimator is consistent for the CEF m(x). The fact that the estimator converges at the rate nh has led to the interpretation of nh as the eective sample size. This is because the number of observations being used to construct m b (x) is proportional to nh, not n as for a parametric estimator. It is helpful to understand that the nonparametric estimator has a reduced convergence rate because the object being estimated m(x) is nonparametric. This is harder than estimating a nite dimensional parameter, and thus comes at a cost. Unlike parametric estimation, the asymptotic distribution of the nonparametric estimator includes a term representing the bias of the estimator. The asymptotic distribution (11.10) shows the form of this bias. Not only is it proportional to the squared bandwidth h2 (the degree of smoothing), it is proportional to the function B (x) which depends on the slope and curvature of the CEF m(x). Interestingly, when m(x) is constant then B (x) = 0 and the kernel estimator has no asymptotic bias. The bias is essentially increasing in the curvature of the CEF function m(x). This is because the local averaging smooths m(x), and the smoothing induces more bias when m(x) is curved. Theorem 11.7.1 shows that the asymptotic distributions of the NW and LL estimators are similar, with the only dierence arising in the bias function B (x). The bias term for the NW estimator has an extra component which depends on the rst derivative of the CEF m(x) while the bias term of the LL estimator is invariant to the rst derivative. The fact that the bias formula for the LL estimator is simpler and is free of dependence on the rst derivative of m(x) suggests that the LL estimator will generally have smaller bias than the NW estimator (but this is not a precise ranking). Since the asymptotic variances in the two distributions are the same, this means that the LL estimator achieves a reduced bias without an eect on asymptotic variance. This analysis has led to the general preference for the LL estimator over the NW estimator in the nonparametrics literature. One implication of Theorem 11.7.1 is that we can dene the asymptotic MSE (AMSE) of m b (x) as the squared bias plus the asymptotic variance 2 Rk 2 (x) 2 . B (x) + AM SE (m b (x)) = h2 k nhfx (x) (11.11)
CHAPTER 11. NONPARAMETRIC REGRESSION Focusing on rates, this says AM SE (m b (x)) h4 + 1 nh
254
(11.12)
which means that the AMSE is dominated by the larger of h4 and (nh)1 . Notice that the bias is increasing in h and the variance is decreasing in h. (More smoothing means more observations are used for local estimation: this increases the bias but decreases estimation variance.) To select h to minimize the AMSE, these two components should balance each other. Setting h4 (nh)1 means setting h n1/5 . Another way to see this is to pick h to minimize the right-hand-side of (11.12). The rst-order condition for h is 1 1 h4 + = 4h3 2 = 0 h nh nh which when solved for h yields h = n1/5 . What this means is that for AMSE-ecient estimation of m(x), the optimal rate for the bandwidth is h n1/5 . Theorem 11.7.2 The bandwidth which minimizes the AMSE is (11.12) 1 / 5 1 / 5 4 / 5 and . With h n then AM SE (m b (x)) = O n of order h n m b (x) = m(x) + Op n2/5 . This result means that the bandwidth should take the form h = cn1/5 . The optimal constant c depends on the kernel k, the bias function B (x) and the marginal density fx (x). A common misinterpretation is to set h = n1/5 , which is equivalent to setting c = 1 and is completely arbitrary. Instead, an empirical bandwidth selection rule such as cross-validation should be used in practice. When h = cn1/5 we can rewrite the asymptotic distribution (11.10) as Rk 2 (x) d 2 2 n2/5 (m b (x) m(x)) N c k B (x), 1/2 c fx (x)
In this representation, we see that m b (x) is asymptotically normal, but with a n2/5 rate of convergence and non-zero mean. The asymptotic distribution depends on the constant c through the bias (positively) and the variance (inversely). The asymptotic distribution in Theorem 11.7.1 allows for the optimal rate h = cn1/5 but this rate is not required. particular, consider an undersmoothing (smaller than optimal) bandwith In with rate h = o n1/5 . For example, we could specify that h = cn for some c > 0 and 1/5 < < 1. Then nhh2 = O(n(15)/2 ) = o(1) so the bias term in (11.10) is asymptotically negligible so Theorem 11.7.1 implies Rk 2 (x) d . nh (m b (x) m(x)) N 0, fx (x) That is, the estimator is asymptotically normal without a bias component. Not having an asymptotic bias component is convenient theoretical manipuations, so many authors impose the for some undersmoothing condition h = o n1/5 to ensure this situation. This convenience comes at a cost. First, the resulting estimator is inecient as its convergence rate is is Op n(1)/2 > Op n2/5 since > 1/5. Second, the distribution theory is an inherently misleading approximation as it misses a critically key ingredient of nonparametric estimation the trade-o between bias and variance. The approximation (11.10) is superior precisely because it contains the asymptotic bias component which is a realistic implication of nonparametric estimation. Undersmoothing assumptions should be avoided when possible.
255
11.8
Conditional Variance Estimation

2 (x) = var (yi | xi = x) = E e2 i | xi = x .
Lets consider the problem of estimation of the conditional variance
Even if the conditional mean m(x) is parametrically specied, it is natural to view 2 (x) as inherently nonparametric as economic models rarely specify the form of the conditional variance. Thus it is quite appropriate to estimate 2 (x) nonparametrically. 2 2 We know that 2 (x) is the CEF of e2 i given xi . Therefore if ei were observed, (x) could be nonparametrically estimated using NW or LL regression. For example, the ideal NW estimator is Pn 2 2 i=1 ki (x)ei . (x) = P n i=1 ki (x)
Since the errors ei are not observed, we need to replace them with an empirical residual, such as b (xi ) where m b (x) is the estimated CEF. (The latter could be a nonparametric estimator e i = yi m such as NW or LL, or even a parametric estimator.) Even better, use the leave-one-out prediction b i (xi ), as these are not subject to overtting. errors e i = yi m With this substitution the NW estimator of the conditional variance is Pn e2 2 i i=1 ki (x) . (11.13) (x) = P n k ( x ) i=1 i
This estimator depends on a set of bandwidths h1 , ..., hq , but there is no reason for the bandwidths to be the same as those used to estimate the conditional mean. Cross-validation can be used to select the bandwidths for estimation of 2 (x) separately from cross-validation for estimation of m b (x). There is one subtle dierence between CEF and conditional variance estimation. The conditional variance is inherently non-negative 2 (x) 0 and it is desirable for our estimator to satisfy this property. Interestingly, the NW estimator (11.13) is necessarily non-negative, since it is a smoothed average of the non-negative squared residuals, but the LL estimator is not guarenteed to be nonnegative for all x. For this reason, the NW estimator is preferred for conditional variance estimation. Fan and Yao (1998, Biometrika) derive the asymptotic distribution of the estimator (11.13). They obtain the surprising result that the asymptotic distribution of this two-step estimator is identical to that of the one-step idealized estimator 2 (x).
11.9
Standard Errors
Theorem 11.7.1 shows the asymptotic variances of both the NW and LL nonparametric regression estimators equal Rk 2 (x) . V (x) = fx (x) For standard errors we need an estimate of V (x) . A plug-in estimate replaces the unknowns by estimates. The roughness Rk can be found from Table 9.1. The conditional variance can be estimated using (11.13). The density of xi can be estimated using the methods from Section 20.1. Replacing these estimates into the formula for V (x) we obtain the asymptotic variance estimate 2 (x) (x) = Rk . V f x (x)
CHAPTER 11. NONPARAMETRIC REGRESSION Then an asymptotic standard error for the kernel estimate m b (x) is r 1 s (x) = V (x). nh
256
Plots of the estimated CEF m b (x) can be accompanied by condence intervals m b (x) 2 s(x). These are known as pointwise condence intervals, as they are designed to have correct coverage at each x, not uniformly in x. One important caveat about the interpretation of nonparametric condence intervals is that they are not centered at the true CEF m(x), but rather are centered at the biased or pseudo-true value 2 B (x). m (x) = m(x) + h2 k Consequently, a correct statement about the condence interval m b (x) 2 s(x) is that it asymptoti cally contains m (x) with probability 95%, not that it asymptotically contains m(x) with probability 2 B (x). 95%. The discrepancy is that the condence interval does not take into account the bias h2 k Unfortunately, nothing constructive can be done about this. The bias is dicult and noisy to estimate, so making a bias-correction only inates estimation and decreases overall precision. variance A technical trick is to assume undersmoothing h = o n1/5 but this does not really eliminate the bias, it only assumes it away. The plain fact is that once we honestly acknowledge that the true CEF is nonparametric, it then follows that any nite sample estimate will have nite sample bias, and this bias will be inherently unknown and thus impossible to incorporate into condence intervals.
11.10
Multiple Regressors
Our analysis has focus on the case of real-valued xi for simplicity of exposition, but the methods of kernel regression extend easily to the multiple regressor case, at the cost of a reduced rate of convergence. In this section we consider the case of estimation of the conditional expectation function E (yi | xi = x) = m(x) when x1i . xi = . . xdi
is a d-vector. For any evaluation point x and observation i, dene the kernel weights x1i x1 x2i x2 xdi xd k k , ki (x) = k h1 h2 hd a d-fold product kernel. The kernel weights ki (x) assess if the regressor vector xi is close to the evaluation point x in the Euclidean space Rd . These weights depend on a set of d bandwidths, hj , one for each regressor. We can group them together into a single vector for notational convenience: h1 . h= . . . hd Given these weights, the Nadaraya-Watson estimator takes the form Pn i=1 ki (x)yi . m b (x) = P n i=1 ki (x)
CHAPTER 11. NONPARAMETRIC REGRESSION For the local-linear estimator, dene z i (x) = 1 xi x
257
and then the local-linear estimator can be written as m b (x) = b(x) where b(x) b(x) = = Z KZ n X
i=1 0
ki (x)z i (x)z i (x) 1 Z Ky

0
!1
n X i=1
ki (x)z i (x)yi
where K = diag{k1 (x), ..., kn (x)}. In multiple regressor kernel regression, cross-validation remains a recommended method for bandwidth selection. The leave-one-out residuals e i and cross-validation criterion CV (h) are dened identically as in the single regressor case. The only dierence is that now the CV criterion is a function over the d-dimensional bandwidth h. This is a critical practical dierence since nding b which minimizes CV (h) can be computationally dicult when h is high the bandwidth vector h dimensional. Grid search is cumbersome and costly, since G gridpoints per dimension imply evaulation of CV (h) at Gd distinct points, which can be a large number. Furthermore, plots of CV (h) against h are challenging when d > 2. The asymptotic distribution of the estimators in the multiple regressor case is anextension of the single regressor case. Let fx (x) denote the marginal density of xi and 2 (x) = E e2 i | xi = x the conditional variance of ei = yi m(xi ). Let |h| = h1 h2 hd . Theorem 11.10.1 Let m b (x) denote either the Nadarya-Watson or Local Linear estimator of m(x). If x is interior to the support of xi and fx (x) > 0, then as n and hj 0 such that n |h| , d d 2 (x) X p Rk d 2 2 n |h| m hj B,j (x) N 0, b (x) m(x) k fx (x)
j =1
where for the Nadaraya-Watson estimator Bj (x) =
1 2 m(x) + fx (x)1 fx (x) m(x) 2 x2 x x j j j
and for the Local Linear estimator Bj (x) = 1 2 m(x) 2 x2 j
That is, the squared bias is of order h4 , the same as in the single regressor case, but the variance is of larger order (nhd )1 . Setting h to balance these two components requires setting h n1/(4+d) .
For notational simplicity consider the case that there is a single common bandwidth h. In this case the AMSE takes the form 1 AM SE (m b (x)) h4 + d nh
CHAPTER 11. NONPARAMETRIC REGRESSION Theorem 11.10.2 The bandwidth which minimizes the AMSEis of order b (x)) = O n4/(4+d) h n1/(4+d) . With h n1/(4+d) then AM SE (m and m b (x) = m(x) + Op n2/(4+d)
258
In all estimation problems an increase in the dimension decreases estimation precision. For example, in parametric estimation an increase in dimension typically increases the asymptotic variance. In nonparametric estimation an increase in the dimension typically decreases the convergence rate, which is a more decrease in precision. For example, in kernel regression the con fundamental b (x) is a local vergence rate Op n2/(4+d) decreases as d increases. The reason is the estimator m average of the yi for observations such that xi is close to x, and when there are multiple regressors the number of such observations is inherently smaller. This phenomenon that the rate of convergence of nonparametric estimation decreases as the dimension increases is called the curse of dimensionality.
Chapter 12
Series Estimation
12.1 Approximation by Series
As we mentioned at the beginning of Chapter 11, there are two main methods of nonparametric regression: kernel estimation and series estimation. In this chapter we study series methods. Series methods approximate an unknown function (e.g. the CEF m(x)) with a exible parametric function, with the number of parameters treated similarly to the bandwidth in kernel regression. A series approximation to m(x) takes the form mK (x) = mK (x, K ) where mK (x, K ) is a known parametric family and K is an unknown coecient. The integer K is the dimension of K and indexes the complexity of the approximation. A linear series approximation takes the form mK (x) =
K X j =1
zjK (x)jK (12.1)
= z K (x)0 K
where zjK (x) are (nonlinear) functions of x, and are known as basis functions or basis function transformations of x. For real-valued x, a well-known linear series approximation is the pth-order polynomial mK (x) =
p X j =0
xj jK
where K = p + 1. When x Rd is vector-valued, a pth-order polynomial is mK (x) =

p X
j1 =0
jd =0
p X
jd 1 xj 1 xd j1 ,...,jd K .
This includes all powers and cross-products, and the coecient vector has dimension K = (p + 1)d . In general, a common method to create a series approximation for vector-valued x is to include all non-redundant cross-products of the basis function transformations of the components of x.
12.2
Splines
Another common series approximation is a continuous piecewise polynomial function known as a spline. While splines can be of any polynomial order (e.g. linear, quadratic, cubic, etc.), a common choice is cubic. To impose smoothness it is common to constrain the spline function to have continuous derivatives up to the order of the spline. Thus a quadratic spline is typically 259
CHAPTER 12. SERIES ESTIMATION
260
constrained to have a continuous rst derivative, and a cubic spline is typically constrained to have a continuous rst and second derivative. There is more than one way to dene a spline series expansion. All are based on the number of knots the join points between the polynomial segments. To illustrate, a piecewise linear function with two segments and a knot at t is x<t m1 (x) = 00 + 01 (x t) mK (x) = xt m2 (x) = 10 + 11 (x t) (For convenience we have written the segments functions as polyomials in x t.) The function mK (x) equals the linear function m1 (x) for x < t and equals m2 (t) for x > t. Its left limit at x = t is 00 and its right limit is 10 , so is continuous if (and only if) 00 = 10 . Enforcing this constraint is equivalent to writing the function as mK (x) = 0 + 1 (x t) + 2 (x t) 1 (x t) or after transforming coecients, as mK (x) = 0 + 1 x + 2 (x t) 1 (x t) Notice that this function has K = 3 coecients, the same as a quadratic polynomial. A piecewise quadratic function with one knot at t is 2 x<t m1 (x) = 00 + 01 (x t) + 02 (x t) mK (x) = xt m2 (x) = 10 + 11 (x t) + 12 (x t)2 mK (x) = 0 + 1 x + 2 x2 + 3 (x t)2 1 (x t) . Here, K = 4. Furthermore, a piecewise cubic function with one knot and a continuous second derivative is mK (x) = 0 + 1 x + 2 x2 + 3 x3 + 4 (x t)3 1 (x t) which has K = 5. The polynomial order p is selected to control the smoothness of the spline, as mK (x) has continuous derivatives up to p 1. In general, a pth-order spline with N knots at t1 , t2 , ..., tN with t1 < t2 < < tN is mK (x) =
p X j =0
This function is continuous at x = t if 00 = 10 , and has a continuous rst derivative if 01 = 11 . Imposing these contraints and rewriting, we obtain the function
j xj +
N X k=1
k (x tk )p 1 (x tk )
which has K = N + p + 1 coecients. In spline approximation, the typical approach is to treat the polynomial order p as xed, and select the number of knots N to determine the complexity of the approximation. The knots tk are typically treated as xed. A common choice is to set the knots to evenly partition the support X of xi .
261
12.3
Partially Linear Model
A common use of a series expansion is to allow the CEF to be nonparametric with respect to one variable, yet linear in the other variables. This allows exibility in a particular variable of interest. A partially linear CEF with vector-valued regressor x1 and real-valued continuous x2 takes the form m (x1 , x2 ) = x0 1 1 + m2 (x2 ). This model is commonly used when x1 are discrete (e.g. binary variables) and x2 is continuously distributed. Series methods are particularly convenient for estimation of partially linear models, as we can replace the unknown function m2 (x2 ) with a series expansion to obtain m (x) ' mK (x) = x0 K K where z K = z K (x2 ) are the basis transformations of x2 (typically polynomials or splines) and 2K 0 are coecients. After transformation the regressors are xK = (x0 1 , z K ). and the coecients are 0 0 0 K = (1 , 2K ) .
0 = x0 1 1 + z K 2K
12.4
Additively Separable Models
When x is multivariate a common simplication is to treat the regression function m (x) as additively separable in the individual regressors, which means that m (x) = m1 (x1 ) + m2 (x2 ) + + md (xd ) . Series methods are quite convenient for estimation of additively separable models, as we simply apply series expansions (polynomials or splines) separately for each component mj (xj ) . The advantage of additive separability is the reduction in dimensionality. While an unconstrained pth order polynomial has (p + 1)d coecients, an additively separable polynomial model has only (p + 1)d coecients. This can be a major reduction in the number of coecients. The disadvantage of this simplication is that the interaction eects have been eliminated. The decision to impose additive separability can be based on an economic model which suggests the absence of interaction eects, or can be a model selection decision similar to the selection of the number of series terms. We will discuss model selection methods below.
12.5
Uniform Approximations
A good series approximation mK (x) will have the property that it gets close to the true CEF m(x) as the complexity K increases. Formal statements can be derived from the theory of functional analysis. An elegant and famous theorem is the Stone-Weierstrass theorem, (Weierstrass, 1885, Stone 1937, 1948) which states that any continuous function can be arbitrarily uniformly well approximated by a polynomial of suciently high order. Specically, the theorem states that for x Rd , if m(x) is continuous on a compact set X , then for any > 0 there exists a polynomial mK (x) of some order K which is uniformly within of m(x):
xX
sup |mK (x) m(x)| .
(12.2)
Thus the true unknown m(x) can be arbitrarily well approximately by selecting a suitable polynomial.
262
Figure 12.1: True CEF and Best Approximations The result (12.2) can be stengthened. In particular, if the sth derivative of m(x) is continuous then the uniform approximation error satises sup |mK (x) m(x)| = O K (12.3)
xX
as K where = s/d. This result is more useful than (12.2) because it gives a rate at which the approximation mK (x) approaches m(x) as K increases. Both (12.2) and (12.3) hold for spline approximations as well. Intuitively, the number of derivatives s indexes the smoothness of the function m(x). (12.3) says that the best rate at which a polynomial or spline approximates the CEF m(x) depends on the underlying smoothness of m(x). The more smooth is m(x), the fewer series terms (polynomial order or spline knots) are needed to obtain a good approximation. To illustrate polynomial approximation, Figure 12.1 displays the CEF m(x) = x1/4 (1 x)1/2 on x [0, 1]. In addition, the best approximations using polynomials of order K = 3, K = 4, and K = 6 are displayed. You can see how the approximation with K = 3 is fairly crude, but improves with K = 4 and especially K = 6. Approximations obtained with cubic splines are quite similar so not displayed. As a series approximation can be written as mK (x) = z K (x)0 K as in (12.1), then the coecient of the best uniform approximation (12.3) is then 0 (12.4) K = argmin sup z K (x) K m(x) .
K xX
The approximation error is We can write this as
rK (x) = m(x) z K (x)0 K. m(x) = z K (x)0 K + rK (x)
(12.5)
to emphasize that the true conditional mean can be written as the linear approximation plus error. A useful consequence of equation (12.3) is sup |rK (x)| O K . (12.6)
xX
263
Figure 12.2: True CEF, polynomial interpolation, and spline interpolation
12.6
Runges Phenomenon
Despite the excellent approximation implied by the Stone-Weierstrass theorem, polynomials have the troubling disadvantage that they are very poor at simple interpolation. The problem is known as Runges phenomenon, and is illustrated in Figure 12.2. The solid line is the CEF m(x) = (1 + x2 )1 displayed on [5, 5]. The circles display the function at the K = 11 integers in this interval. The long dashes display the 10th order polynomial t through these points. Notice that the polynomial approximation is erratic and far from the smooth CEF. This discrepancy gets worse as the number of evaluation points increases, as Runge (1901) showed that the discrepancy increases to innity with K. In contrast, splines do not exhibit Runges phenomenon. In Figure 12.2 the short dashes display a cubic spline with seven knots t through the same points as the polynomial. While the tted spline displays some oscillation relative to the true CEF, they are relatively moderate. Because of Runges phenomenon, high-order polynomials are not used for interpolation, and are not popular choices for high-order series approximations. Instead, splines are widely used.
12.7
Approximating Regression
For each observation i we observe (yi , xi ) and then construct the regressor vector z Ki = z K (xi ) using the series transformations. Stacking the observations in the matrices y and Z K , the least squares estimate of the coecient K in the series approximation z K (x)0 K is b K = Z 0 Z K 1 Z 0 y , K K bK . m b K (x) = z K (x)0 (12.7)
and the least squares estimate of the regression function is
As we learned in Chapter 2, the least-squares coecient is estimating the best linear predictor of yi given z Ki . This is 1 E (z Ki yi ) . K = E z Ki z 0 Ki
CHAPTER 12. SERIES ESTIMATION Given this coecient, the series approximation is z K (x)0 K with approximation error rK (x) = m(x) z K (x)0 K . The true CEF equation for yi is yi = m(xi ) + ei with ei the CEF error. Dening rKi = rK (xi ), we nd yi = z 0 Ki K + eKi where the equation error is eKi = rKi + ei .
264
(12.8)
(12.9)
Observe that the error eKi includes the approximation error and thus does not have the properties of a CEF error. In matrix notation we can write these equations as y = Z K K + rK + e = Z K K + eK . (12.10)
K K will increase with K . For example, if the support of the variables z K (xi ) is the unit cube [0, 1] , then you can compute that K = K . As discussed in Newey (1997) and Li and Racine (2007, Corollary 15.1) if the support of xi is compact then K = O(K ) for polynomials and K = O(K 1/2 ) for splines.
let X denote the support of xi , and dene the largest normalized length of the regressor vector in the support of xi 1/2 1 K = sup z K (x)0 Q . (12.11) K z K (x)
xX
We now impose some regularity conditions on the regression model to facilitate the theory. Dene the K K expected design matrix QK = E z Ki z 0 Ki ,
Assumption 12.7.1 1. For some > 0 the series approximation satises (12.3). 2 < . 2. E e2 i | xi 3. min (QK ) > 0.
2 K/n 4. K = K (n) is a function of n which satises K/n 0 and K 0 as n .
Assumptions 12.7.1.1 through 12.7.1.3 concern properties of the regression model. Assumption 12.7.1.1 holds with = s/d if X is compact and the sth derivative of m(x) is continuous. Assumption 12.7.1.2 allows for conditional heteroskedasticity, but requires the conditional variance to be bounded. Assumption 12.7.1.3 excludes near-singular designs. Since estimates of the conditional mean are unchanged if we replace z Ki with z Ki = B K z Ki for any non-singular B K , Assumption 12.7.1.3 can be viewed as holding after transformation by an appropriate non-singular B K .
265
Assumption 12.7.1.4 concerns the choice of the number of series terms, which is under the control of the user. It species that K can increase with sample size, but at a controlled rate of growth. Since K = O(K ) for polynomials and K = O(K 1/2 ) for splines, Assumption 12.7.1.4 is satised if K 3 /n 0 for polynomials and K 2 /n 0 for splines. This means that while the number of series terms K can increase with the sample size, K must increase at a much slower rate. In Section 12.5 we introduced the best uniform approximation, and in this section we introduced the best linear predictor. What is the relationship? They may be similar in practice, but they are not the same and we should be careful to maintain the distinction. Note that from (12.5) we can ) from (12.6). Then write m(xi ) = z 0 Ki K + rKi where rKi = rK (xi ) satises supi |rKi | = O (K the best linear predictor equals 1 E (z Ki yi ) K = E z Ki z 0 Ki 1 E (z Ki m(xi )) = E z Ki z 0 Ki 1 E z Ki (z 0 = E z Ki z 0 Ki Ki K + rKi ) 1 0 = E (z Ki rKi ). K + E z Ki z Ki Thus the dierence between the two approximations is
(x) = z K (x)0 ( rK (x) rK K K ) 1 0 E (z Ki rKi ). = z K (x) E z Ki z 0 Ki
(12.12)
and by (12.6)
Observe that by the properties of projection 2 1 0 0 E r E (z Ki rKi )0 Ki E (r Ki z Ki ) E z Ki z Ki 2 = E rKi Z (x)2 fx (x)dx O K 2 . rK
(12.13)
(12.14)
Then applying the Schwarz inequality to (12.12), Denition (12.11), (12.13) and (12.14), we nd 1/2 1 (x)| z K (x)0 E z Ki z 0 z K (x) |rK (x) rK Ki 1/2 1 E (rKi z Ki )0 E z Ki z 0 E ( z r ) Ki Ki Ki O K K . (12.15) It follows that the best linear predictor approximation error satises sup |rK (x)| O K K .
xX
(12.16)
the nal inequality by (12.14).
The bound (12.16) is probably not the best possible, but it shows that the best linear predictor satises a uniform approximation bound. Relative to (12.6), the rate is slower by the factor K . The bound (12.16) term is o(1) as K if K K 0. A sucient condition is that > 1 (s > d) for polynomials and > 1/2 (s > d/2) for splines, where d = dim(x) and s is the number of continuous derivatives of m(x). It is also useful to observe that since K is the best linear approximation to m(xi ) in meansquare (see Section 2.24), then 2 2 = E m(xi ) z 0 ErKi Ki K 2 E m(xi ) z 0 Ki K O K 2 (12.17)
266
12.8
Residuals and Regression Fit
b The tted regression at x = xi is m b K (xi ) = z 0 Ki K and the tted residual is The leave-one-out prediction errors are b K (xi ). e iK = yi m
0 where hKii = z 0 z Ki . Ki (Z K Z K ) As for kernel regression, the prediction errors e iK are better estimates of the errors than the tted residuals e iK , as they do not have the tendency to over-t when the number of series terms is large. To assess the t of the nonparametric regression, the estimate of the mean-square prediction error is n n 1X 2 1X 2 2 = e iK = e iK (1 hKii )2 K n n i=1 i=1
b where K,i is the least-squares coecient with the ith observation omitted. Using (3.38) we can also write iK (1 hKii )1 e iK = e
1
b K,i (xi ) e iK = yi m b = yi z 0
Ki K,i
and the prediction
R2
is
12.9
Cross-Validation Model Selection

n
Pn 2 e 2 e RK = 1 Pn i=1 iK 2 . ) i=1 (yi y
The cross-validation criterion for selection of the number of series terms is the MSPE CV (K ) =
2 K
1X 2 = e iK (1 hKii )2 . n
i=1
e2 , we have a dataBy selecting the series terms to minimize CV (K ), or equivalently maximize R K dependent rule which is designed to produce estimates with low integrated mean-squared error (IMSE) and mean-squared forecast error (MSFE). As shown in Theorem 11.6.1, CV (K ) is an approximately unbiased estimated of the MSFE and IMSE, so nding the model which produces the smallest value of CV (K ) is a good indicator that the estimated model has small MSFE and IMSE. The proof of the result is the same for all nonparametric estimators (series as well as kernels) so does not need to be repeated here. As a practical matter, an estimator corresponds to a set of regressors z Ki , that is, a set of transformations of the original variables xi . For each set of regressions, the regression is estimated and CV (K ) calculated, and the estimator is selected which has the smallest value of CV (K ). If there are p ordered regressors, then there are p possible estimators. Typically, this calculation is simple even if p is large. However, if the p regressors are unordered (and this is typical) then there are 2p possible subsets of conceivable models. If p is even moderately large, 2p can be immensely large so brute-force computation of all models may be computationally demanding.
267
12.10
Convergence in Mean-Square
b K are indexed by K . The point of nonparametric estimation is to let The series estimate K be exible so as to incorporate greater complexity when the data are suciently informative. This means that K will typically be increasing with sample size n. This invalidates conventional asymptotic distribution theory. However, we can develop extensions which use appropriate matrix norms, and by focusing on real-valued functions of the parameters including the estimated regression function itself. The asymptotic theory we present in this and the next several sections is largely taken from Newey (1997). Our rst main result shows that the least-squares estimate converges to K in mean-square distance. Theorem 12.10.1 Under Assumption 12.7.1, as n , 0 K b b + op K 2 K K QK K K = Op n
(12.18)
The proof of Theorem 12.10.1 is rather technical and deferred to Section 12.16. The rate of convergence in (12.18) has two terms. The Op (K/n) term is due to estimation variance. Note in contrast that the corresponding rate would be Op (1/n) in the parametric case. The dierence is that in the parametric case we assume that the number of regressors K is xed as n increases, while in the nonparametric case we allow the number of regressors K to be exible. As 2 term in (12.18) is due to the series K increases, the estimation variance increases. The op K approximation error. Using Theorem 12.10.1 we can establish the following convergence rate for the estimated regression function. Theorem 12.10.2 Under Assumption 12.7.1, as n , Z K + Op K 2 (m b K (x) m(x))2 fx (x)dx = Op n
(12.19)
Theorem 12.10.2 shows that the integrated squared dierence between the tted regression and the true CEF converges in probability to zero if K as n . The convergence results of Theorem 12.10.2 show that the number of series terms K involves a trade-o similar to the role of the bandwidth h in kernel regression. Larger K implies smaller approximation error but increased estimation variance. The optimal rate which minimizes the average squared error in (12.19) is K = O n1/(1+2) , yielding an optimal rate of convergence in (12.19) of Op n2/(1+2) . This rate depends on the unknown smoothness of the true CEF (the number of derivatives s) and so does not directly syggest a practical rule for determining K. Still, the implication is that when the function being estimated is less smooth ( is small) then it is necessary to use a larger number of series terms K to reduce the bias. In contrast, when the function is more smooth then it is better to use a smaller number of series terms K to reduce the variance. To establish (12.19), using (12.7) and (12.8) we can write b (12.20) m b K (x) m(x) = z K (x)0 K K rK (x).
268
Since eR Ki are projection errors, they satisfy E (z Ki eKi ) = R0 and thus E (z Ki rKi ) = 0. This 2 = means z K (x)rK (x)fx (x)dx = 0. Also observe that QK = z K (x)z K (x)0 fx (x)dx and ErKi R rK (x)2 fx (x)dx. Then Z (m b K (x) m(x))2 fx (x)dx 0 b K K + Er2 b K K QK = Ki K + Op K 2 Op n by (12.18) and (12.17), establishing (12.19).
12.11
Uniform Convergence
Theorem 12.10.2 established conditions under which m b K (x) is consistent in a squared error norm. It is also of interest to know the rate at which the largest deviation converges to zero. We have the following rate. Theorem 12.11.1 Under Assumption 12.7.1, then as n , ! r 2K K + Op K K . b K (x) m(x)| = Op sup |m n xX
(12.21)
Relative to Theorem 12.10.2, the error has been increased multiplicatively by K . This slower convergence rate is a penalty for the stronger uniform convergence, though it is probably not the best possible rate. Examining the bound in (12.21) notice that the rst term is op (1) under Assumption 12.7.1.4. The second term is op (1) if K K 0, which requires that K and that be suciently large. A sucient condition is that s > d for polynomials and s > d/2 for splines, where d = dim(x) and s is the number of continuous derivatives of m(x). Thus higher dimensional x require a smoother CEF m(x) to ensure that the series estimate m b K (x) is uniformly consistent. The convergence (12.21) is straightforward to show using (12.18). Using (12.20), the Triangle Inequality, the Schwarz inequality (A.10), Denition (12.11), (12.18) and (12.16),
xX
0 1/2 1/2 1 b b z ( x ) Q sup z K (x)0 Q K K K K K K K xX + O K K 2 1/2 K + Op K + O K K , K Op n r ! 2K K = Op + Op K K . n This is (12.21).
b K (x) m(x)| sup |m b sup z K (x)0 K K + sup |rK (x)|

xX xX
(12.22)
269
12.12
Asymptotic Normality
One advantage of series methods is that the estimators are (in nite samples) equivalent to parametric estimators, so it is easy to calculate covariance matrix estimates. We now show that we can also justify normal asymptotic approximations. The theory we present in this section will apply to any linear function of the regression function. That is, we allow the parameter of interest to be aany non-trivial real-valued linear function of the entire regression function m() = a (m) . This includes the regression function m(x) at a given point x, derivatives of m(x), and integrals b as an estimator for m(x), the estimator for is over m(x). Given m b K (x) = z K (x)0 K b for some K 1 vector of constants aK 6= 0. (The relationship a (m b K ) = a0 K K follows since a is b K .) linear in m and m b K is linear in K to be If K were xed as n , then by standard asymptotic theory we would expect asymptotically normal with variance
1 1 vK = a0 K QK K QK aK
K = a (m b b K ) = a0 K K
where
The standard justication, however, is not valid in the nonparametric case, in part because vK may diverge as K , and in part due to the nite sample bias due to the approximation error. K is Therefore a new theory is required. Interestingly, it turns out that in the nonparametric case K . The proof is dierent still asymptotically normal, and vK is still the appropriate variance for than the parametric case as the dimensions of the matrices are increasing with K, and we need to be attentive to the estimators bias due to the series approximation. Theorem 12.12.1 Under Assumption 12.7.1, if in addition E e4 i |xi 2 = O (1), then as n , 4 < , E e2 i |xi > 0, and K K n K + a (rK ) d N (0, 1) (12.23) 1/2 vK
2 K = E z Ki z 0 Ki eKi .
The proof of Theorem 12.12.1 can be found in Section 12.16. K is approximately normal with bias a (rK ) and Theorem 12.12.1 shows that the estimator variance vK /n. The variance is the same as in the parametric case, but the asymptotic distribution contains an asymptotic bias, similar as is found in kernel regression. We discuss the bias in more detail below. Notice that Theorem 12.12.1 requires K K = O(1), which is similar to that found in Theorem 12.11.1 to establish uniform convergence. The the bound K K = O(1) allows K to be constant with n or to increase with n. However, when K is increasing the bound requires that be sucient large so that K grows faster than K . A sucient condition is that s = d for polynomials and s = d/2 for splines. The fact that the condition allows for K to be constant means that Theorem 12.12.1 includes parametric least-squares as a special case with explicit attention to estimation bias.
270
K still One useful message from Theorem 12.12.1 is that the classic variance formula vK for applies for series regression. Indeed, we can estimate the asymptotic variance using the standard White formula b 1 b b 1 v K = a0 K QK K QK aK n X bK = 1 z Ki z 0 2 Ki e iK n
i=1 n X i=1
K is Hence a standard error for
bK = 1 Q n r
z Ki z 0 Ki .
s (K ) =
It can be shown (Newey, 1997) that v K /vK 1 as n and thus the distribution in (12.23) is K . unchanged if vK is replaced with v K has a bias term a (rK ) . What is this? It is the Theorem 12.12.1 shows that the estimator same transformation of the function rK (x) as = a (m) is of the regression function m(x). For example, if = m(x) is the regression at a xed point x , then a (rK ) = rK (x), the approximation d d m(x) is the regression derivative, then a (rK ) = rK (x) is the error at the same point. If = dx dx derivative of the approximation error. K for shown in Theorem 12.12.1 is simply the This means that the bias in the estimator approximation error, transformed by the functional of interest. If we are estimating the regression function then the bias is the error in approximating the regression function; if we are estimating the regression derivative then the bias is the error in the derivative in the approximation error for the regression function.
1 0 b 1 b b 1 a Q K QK aK . n K K
p
12.13
Asymptotic Normality with Undersmoothing
An unpleasant aspect about Theorem 12.12.1 is the bias term. An interesting trick is that this bias term can be made asymptotically negligible if we assume that K increases with n at a suciently fast rate. Theorem 12.13.1 Assumption 12.7.1, if in addition E e4 i |xi 2 Under ) O (K ) , nK 2 0, and 4 < , E ei |xi 2 > 0, a (rK 1 a0 K QK aK is bounded away from zero, then n K d N (0, 1) . (12.24) 1/2 vK
) O (K ) states that the function of interest (for example, the regression The condition a (rK function, its derivative, or its integral) applied to the uniform approximation error converges to zero as the number of terms K in the series approximation increases. If a (m) = m(x) then this condition holds by (12.6). 1 The condition that a0 K QK aK is bounded away from zero is simply a technical requirement to exclude degeneracy.
271
The critical condition is the assumption that nK 2 0. This requires that K at a 1/2 . This is a troubling condition. The optimal rate for estimation of m(x) is rate faster n 1/than K = O n (1+2) . If we set K = n1/(1+2) by this rule then nK 2 = n1/(1+2) , not zero. Thus this assumption is equivalent to assuming that K is much larger than optimal. The reason why this trick works (that is, why the bias is negligible) is that by increasing K, the asymptotic bias decreases and the asymptotic variance increases and thus the variance dominates. Because K is larger than optimal, we typically say that m b K (x) is undersmoothed relative to the optimal series estimator. Many authors like to focus their asymptotic theory on the assumptions in Theorem 12.13.1, as the distribution (12.24) appears cleaner. However, it is a poor use of asymptotic theory. There are three problems with the assumption nK 2 0 and the approximation (12.24). First, it says that if we intentionally pick K to be larger than optimal, we can increase the estimation variance relative to the bias so the variance will dominate the bias. But why would we want to intentionally use an estimator which is sub-optimal? Second, the assumption nK 2 0 does not eliminate the asymptotic bias, it only makes it of lower order than the variance. So the approximation (12.24) is technically valid, but the missing asymptotic bias term is just slightly smaller in asymptotic order, and thus still relevant in nite samples. Third, the condition nK 2 0 is just an assumption, it has nothing to do with actual empirical practice. Thus the dierence between (12.23) and (12.24) is in the assumptions, not in the actual reality or in the actual empirical practice. Eliminating a nuisance (the asymptotic bias) through an assumption is a trick, not a substantive use of theory. My strong view is that the result (12.23) is more informative than (12.24). It shows that the asymptotic distribution is normal but has a non-trivial nite sample bias.
12.14
Regression Estimation
A special yet important example of a linear estimator of the regression function is the regression function at a xed point x. In the notation of the previous section, a (m) = m(x) and aK = z K (x). b . As this is a key problem of interest, we K = m b K (x) = z K (x)0 The series estimator of m(x) is K restate the asymptotic results of Theorems 12.12.1 and 12.13.1 for this estimator. 4 Theorem 12.14.1 Under Assumption 12.7.1, if in addition E ei |xi 2 2 4 < , E ei |xi > 0, and K K = O(1), then as n , n (m b K (x) m(x) + rK (x))
1/2 vK (x)
N (0, 1)
(12.25)
where
1 1 vK (x) = z K (x)0 Q K K QK z K (x). 1 If K K = O(1) is replaced by nK 2 0, and z K (x)0 Q K z K (x) is bounded away from zero, then n (m b K (x) m(x)) d N (0, 1) (12.26) 1/2 vK (x)
There are two important features about the asymptotic distribution (12.25). First, as mentioned in the previous section, it shows how to construct asymptotic standard errors for the CEF m(x). These are r 1 b 1 b KQ b 1 z K (x). z K (x)0 Q s (x) = K K n
272
Second, (12.25) shows that the estimator has the asymptotic bias component rK (x). This is due to the fact that the nite order series is an approximation to the unknown CEF m(x), and this results in nite sample bias. The asymptotic distribution (12.26) shows that the bias term is negligable if K diverges fast enough so that nK 2 0. As discussed in the previous section, this means that K is larger than optimal. 1 The assumption that z K (x)0 Q K z K (x) is bounded away from zero is a technical condition to exclude degenerate cases, and is automatically satised if z K (x) includes an intercept. b K (x) Plots of the CEF estimate m b K (x) can be accompanied by 95% condence intervals m 2 s(x). As we discussed in the chapter on kernel regression, this can be viewed as a condence interval for the pseudo-true CEF m K (x) = m(x) r K (x), not for the true m(x). As for kernel regression, the dierence is the unavoidable consequence of nonparametric estimation.
12.15
Kernel Versus Series Regression
In this and the previous chapter we have presented two distinct methods of nonparametric regression based on kernel methods and series methods. Which should be used in practice? Both methods have advantages and disadvantages and there is no clear overall winner. First, while the asymptotic theory of the two estimators appear quite dierent, they are actually rather closely related. When the regression function m(x) is twice dierentiable (s = 2) then the rate of convergence of both the MSE of the kernel regression estimator with optimal bandwidth h and the series estimator with optimal K is n2/(d+4) . There is no dierence. If the regression function is smoother than twice dierentiable (s > 2) then the rate of the convergence of the series estimator improves. This may appear to be an advantage for series methods, but kernel regression can also take advantage of the higher smoothness by using so-called higher-order kernels or local polynomial regression, so perhaps this advantage is not too large. Both estimators are asymptotically normal and have straightforward asymptotic standard error formulae. The series estimators are a bit more convenient for this purpose, as classic parametric standard error formula work without amendment. An advantage of kernel methods is that their distributional theory is easier to derive. The theory is all based on local averages which is relatively straightforward. In contrast, series theory is more challenging, dealing with increasing parameter spaces. An important dierence in the theory is that for kernel estimators we have explicit representations for the bias while we only have rates for series methods. This means that plug-in methods can be used for bandwidth selection in kernel regression. However, typically we rely on cross-validation, which is equally applicable in both kernel and series regression. Kernel methods are also relatively easy to implement when the dimension d is large. There is not a major change in the methodology as d increases. In contrast, series methods become quite cumbersome as d increases as the number of cross-terms increases exponentially. A major advantage of series methods is that it has inherently a high degree of exibility, and the user is able to implement shape restrictions quite easily. For example, in series estimation it is relatively simple to implement a partial linear CEF, an additively separable CEF, monotonicity, concavity or convexity. These restrictions are harder to implement in kernel regression.
12.16
Technical Proofs
1/2
Dene z Ki = z K (xi ) and let QK denote the positive denite square root of QK . As mentioned before Theorem 12.10.1, the regression problem is unchanged if we replace z Ki with a rotated 1/2 0 regressor such as z z Ki . This is a convenient choice for then E (z Ki = QK Ki z Ki ) = I K . For notational convenience we will simply write the transformed regressors as z Ki and set QK = I K .
CHAPTER 12. SERIES ESTIMATION We start with some convergence results for the sample design matrix X b K = 1 Z0 ZK = 1 Q z Ki z 0 K Ki . n n
i=1 n
273
Theorem 12.16.1 Under Assumption 12.7.1 and QK = I K , as n , b (12.27) QK I K = op (1) and b K ) 1. min (Q
n p
(12.28)
Proof. Since
then
K K X 2 X b QK I K = j =1 =1
1X (z jKi z n
i=1
Ki
Ez jKi z
Ki )
!2
K K X 2 X b var E QK I K = j =1 =1 1
=n
K K X X j =1 =1 K X j =1
1X z jKi z n
i=1
Ki
var (z jKi z
K X
Ki )
n =n
1 1
E E
z2 jKi
z 2Ki (12.29)
2 Since z 0 Ki z Ki K by denition (12.11) and using (A.1) we nd 0 E z0 Ki z Ki = tr Ez Ki z Ki = tr I K = K,
=1 2 0 z Ki z Ki .
(12.30) (12.31)
so that
and hence (12.29) is o(1) under Assumption 12.7.1.4. Theorem 5.11.1 shows that this implies (12.27). b K I K which are real as Q b K I K is symmetric. Then Let 1 , 2 , ..., K be the eigenvalues of Q K !1/2 X b b K ) 1 = min (Q b K I K ) 2 = Q I min (Q K K
=1
2 2 K K E z0 Ki z Ki
where the second equality is (A.8). This is op (1) by (12.27), establishing (12.28).
Proof of Theorem 12.10.1. As above, assume that the regressors have been transformed so that QK = I K . From expression (12.10) we can substitute to nd b = Z 0 Z K 1 Z 0 eK . K K K K 1 1 0 bK Z eK (12.32) =Q n K
CHAPTER 12. SERIES ESTIMATION Using (12.32) and the Quadratic Inequality (A.14), 0 b K K b K K 1 1 0 b b = n2 e0 K Z K QK QK Z K eK 1 2 0 bK max Q n2 e0 K Z K Z K eK . Observe that (12.28) implies 1 1 b bK = Q = Op (1). max Q max K
i
274
(12.33)
(12.34)
As eKi are projection errors, they satisfy E (z Ki eKi ) = 0. Since the observations are independent, using (12.30) and (12.35), then n n X X 0 2 eKi z 0 z Kj eKj n2 E e0 K Z K Z K eK = n E Ki
i=1 ij =1
Since eKi = ei + rKi , and using Assumption 12.7.1.2 and (12.16), then 2 2 2 2 2 . sup E e2 Ki |xi = + sup rKi + O K K
i
(12.35)
= n2
1
Together, (12.33), (12.34) and (12.37) imply (12.18).
2 K/n = o(1) by Assumption 12.7.1.4. Theorem 5.11.1 shows that this implies since K 2 0 + op K 2 . n2 e0 K Z K Z K eK = Op n
2 E z0 Ki z Ki sup E eKi |xi i 2 12 K K 2 + O K n n K = 2 + o K 2 n n
n X i=1
2 E z0 Ki z Ki eKi
(12.36)
(12.37)
Proof of Theorem 12.12.1. As above, assume that the regressors have been transformed so that QK = I K . Using m(x) = z K (x)0 K + rK (x) and linearity = a (m) = a z K (x)0 K + a (rK ) = a0 K K + a (rK ) Combined with (12.32) we nd b K K K + a (rK ) = a0 K =
1 0 b 1 0 a Q Z eK n K K K
CHAPTER 12. SERIES ESTIMATION and thus r rn n b K K + a (rK ) = a0 K K vk vk K r 1 0 b 1 0 = a Q Z eK nvk K K K 1 = a0 Z 0 eK nvK K K 1 1 0 b + a0 I Q K ZK e K nvK K 1 1 0 b + a0 K QK I K Z K r K . nvK
275
(12.38) (12.39) (12.40)
where we have used eK = e + rK . We now take the terms in (12.38)-(12.40) separately. First, take (12.38). We can write
n 1 1 X 0 0 a0 Z e = aK z Ki eKi . K nvK K K nvK i=1
(12.41)
Observe that a0 K z Ki eKi are independent across i, mean zero, and have variance 2 0 2 = a0 E a0 K z Ki eKi K E z Ki z Ki eKi aK = vK .
Using (12.43), the Schwarz Inequality, and (12.31) 4 4 4 4 a0 E a0 K z Ki eKi = E K z Ki E eKi |xi 4 8 ( + 1 ) E a0 K z Ki 2 0 2 8 ( + 1 ) a0 E z Ki z Ki K aK 2 2 = 8 ( + 1 ) a0 K K. K aK 2 2 2 Since E e2 Ki |xi = E ei |xi + rKi , 0 2 vK = a0 K E z Ki z Ki eKi aK 0 2 a0 K E z Ki z Ki aK = 2 a0 K aK . Equation (12.44) and (12.45) combine to show that 2K 4 4 8 ( + 1 ) K 1 0 = o(1) E a z e Ki K Ki 2 4 n nvK
= O (1) means K for some < . Then by the c The assumption that K 1 1 r K K 4 inequality and E ei |xi 4 4 (12.43) sup E e4 Ki |xi 8 sup E ei |xi + rKi 8 ( + 1 ) . i i
We will apply the Lindeberg CLT 5.7.2, for which it is sucient to verify Lyapunovs condition (5.6): n 4 4 4 1 X 0 1 0 (12.42) E a z e = E a z K Ki Ki K Ki eKi 0. 2 2 n2 vK nvK i=1
(12.44)
(12.45)
276
under Assumption 12.7.1.4. This establishes Lyapunovs condition (12.42). Hence the Lindeberg CLT applies to (12.41) and we conclude 1 d 0 a0 (12.46) K Z K eK N (0, 1) . nvK Second, take (12.39). Since E (e | X ) = 0, then applying E e2 2 , the Schwarz and Norm i |xi Inequalities, (12.45), (12.34) and (12.27), ! 2 1 1 0 0 b IK Z e | X Q a E K K nvK K 1 0 1 0 b 1 b = Q aK QK I K Z 0 E ee | X Z I K K aK K K nvK 1 2 0 b 1 bK Q b K I K aK aK QK I K Q vK 1 2 0 b bK Q b K I K aK = aK QK I K Q vK 2 1 2 a0 K aK bK b K IK max Q Q vK 2 2 op (1). This establishes 1 1 p 0 b a0 K QK I K Z K e 0. nvK (12.47)
2 Observe that since the observations are independent and Ez Ki rKi = 0, z 0 Ki z Ki K , and (12.17) n n X X 1 0 1 r ZK Z0 E rKi z 0 z Kj rKj K rK = E Ki n K n i=1 ij =1 ! n 1X 0 2 z Ki z Ki rKi =E n i=1 2 2 K E rKi 2 2 = O K K
Third, take (12.40). By the Cauchy-Schwarz inequality, (12.45), and the Quadratic Inequality, 2 1 1 0 0 b I K Z rK Q a K K nvK K 1 1 a0 aK 0 b K IK Q b K IK Z0 K rK Z K Q K rK nvK 1 2 1 b K I K 1 r0 2 max Q ZK Z0 (12.48) K rK . n K
= O(1)
Equivalently,
1 ZK Z0 since K K 2 = O(1). Thus r0 K r K = Op (1). This means that (12.48) is op (1) since (12.28) n K implies 1 1 b K I K = max Q b K 1 = op (1). (12.49) max Q 1 1 p 0 b a0 K QK I K Z K r K 0. nvK (12.50)
CHAPTER 12. SERIES ESTIMATION Equations (12.46), (12.47) and (12.50) applied to (12.38)-(12.40) show that r n d K K + a (rK ) N (0, 1) vk completing the proof.
277
Proof of Theorem 12.13.1. The assumption that nK 2 = o(1) implies K = o n1/2 . Thus K K
2 K n
1/2 !
2K K n
1/2 !
= o(1)
so the conditions of Theorem 12.12.1 are satised. It is thus sucient to show that r n a (rK ) = o(1). vk From (12.12)
(x) + z K (x)0 K rK (x) = rK 1 K = E z Ki z 0 E (z Ki rKi ). Ki
Thus by linearity, applying (12.45), and the Schwarz inequality r r n n a (rK ) = ) + a0 a (rK K K vk vk + 2
0 )1/2 (nK K .
0 1/2 a (rK ) aK aK
n1/2
(12.51) (12.52)
) = O n1/2 K = o(1). By (12.14) and nK 2 = o(1) By assumption, n1/2 a (rK 0 1 0 K = nE rKi z Ki E z Ki z 0 E (z Ki rKi ) nK Ki 2 nO K = o(1).
Together, both (12.51) and (12.52) are o(1), as required.
Chapter 13
Generalized Method of Moments

13.1 Overidentied Linear Model
yi = x0 i + ei
0 = x0 1i 1 + x2i 2 + ei
Consider the linear model
E (xi ei ) = 0 where x1i is k 1 and x2i is r 1 with = k + r. We know that without further restrictions, an asymptotically ecient estimator of is the OLS estimator. Now suppose that we are given the information that 2 = 0. Now we can write the model as yi = x0 1i 1 + ei E (xi ei ) = 0. In this case, how should 1 be estimated? One method is OLS regression of yi on x1i alone. This method, however, is not necessarily ecient, as there are restrictions in E (xi ei ) = 0, while 1 is of dimension k < . This situation is called overidentied. There are k = r more moment restrictions than free parameters. We call r the number of overidentifying restrictions. This is a special case of a more general class of moment condition models. Let g (y, x, z , ) be an 1 function of a k 1 parameter with k such that Eg (yi , xi , z i , 0 ) = 0 (13.1)
where 0 is the true value of . In our previous example, g(y, z , ) = z (y x0 1 1 ). In econometrics, this class of models are called moment condition models. In the statistics literature, these are known as estimating equations. As an important special case we will devote special attention to linear moment condition models, which can be written as yi = x0 i + ei E (z i ei ) = 0. where the dimensions of xi and z i are k 1 and 1 , with k. If k = the model is just identied, otherwise it is overidentied. The variables xi may be components and functions of z i , but this is not required. This model falls in the class (13.1) by setting g (y, x, z , 0 ) = z (y x0 ) 278 (13.2)
CHAPTER 13. GENERALIZED METHOD OF MOMENTS
279
13.2
GMM Estimator
n n 1 0 1X 1X g i () = z i yi x0 Z y Z 0X . i = n n n i=1 i=1
Dene the sample analog of (13.2) g n () = (13.3)
The method of moments estimator for is dened as the parameter value which sets gn () = 0. This is generally not possible when > k, as there are more equations than free parameters. The idea of the generalized method of moments (GMM) is to dene an estimator which sets g n () close to zero. For some weight matrix W n > 0, let Jn () = n g n ()0 W n gn (). This is a non-negative measure of the length of the vector g n (). For example, if W n = I , then, Jn ( ) = n g n ()0 g n () = n kg n ()k2 , the square of the Euclidean length. The GMM estimator minimizes Jn ().
b GMM = argmin Jn () . Denition 13.2.1
b ) = 0, and the GMM estimator is the method of moments Note that if k = , then g n ( estimator. The rst order conditions for the GMM estimator are 0= b) Jn ( b )0 W n g ( b) = 2 g n ( n 1 0 1 0 b X Z Wn Z y X = 2 n n
so
which establishes the following.
b = 2 X 0Z W n Z 0y 2 X 0Z W n Z 0X
Proposition 13.2.1 b GMM = X 0 Z W n Z 0 X 1 X 0 Z W n Z 0 y . While the estimator depends on W n , the dependence is only up to scale, for if W n is replaced b GMM does not change. by cW n for some c > 0,
280
13.3
Distribution of GMM Estimator

p
Assume that W n W > 0. Let and where gi = z i ei . Then
and
We conclude:
1 0 1 0 d X Z W n Z e Q0 W N (0, ) . n n
1 0 1 0 p X Z Wn Z X Q0 W Q n n
0 2 = E zi z0 i ei = E g i g i ,
Q = E z i x0 i
where
Theorem 13.3.1 Asymptotic Distribution of GMM Estimator d b n N (0, V ) , 1 0 1 Q W W Q Q0 W Q . V = Q0 W Q
Thus we have
In general, GMM estimators are asymptotically normal with sandwich form asymptotic variances. The optimal weight matrix W 0 is one which minimizes V . This turns out to be W 0 = 1 . The proof is left as an exercise. This yields the ecient GMM estimator: b = X 0 Z 1 Z 0 X 1 X 0 Z 1 Z 0 y . Theorem 13.3.2 Asymptotic Distribution of Ecient GMM Estimator 1 d b . n N 0, Q0 1 Q W 0 = 1 is not known in practice, but it can be estimated consistently. For any W n W 0 , b the ecient GMM estimator, as it has the same asymptotic distribution. we still call By ecient, we mean that this estimator has the smallest asymptotic variance in the class of GMM estimators with this set of moment conditions. This is a weak concept of optimality, as we are only considering alternative weight matrices W n . However, it turns out that the GMM estimator is semiparametrically ecient, as shown by Gary Chamberlain (1987). If it is known that E (g i ()) = 0, and this is all that is known, this is a semi-parametric problem, as the distribution of the data is unknown. Chamberlain showed that in this context, no semiparametric estimator (one which is consistent globally for the class of models considered) 1 where G = E g (). Since the GMM can have a smaller asymptotic variance than G0 1 G 0 i estimator has this asymptotic variance, it is semiparametrically ecient. This result shows that in the linear model, no estimator has greater asymptotic eciency than the ecient linear GMM estimator. No estimator can do better (in this rst-order asymptotic sense), without imposing additional assumptions.
p
281
13.4
Estimation of the Ecient Weight Matrix
b is consistent yet inecient. For Given any weight matrix W n > 0, the GMM estimator 1 example, we can set W n = I . In the linear model, a better choice is W n = (Z 0 Z ) . Given b any such rst-step estimator, we can dene the residuals e i = yi x0 i and moment equations b ). Construct g i = z i e i = g (yi , xi , z i , X b) = 1 gn = g n ( g i , n
i=1 n
g i
and dene Wn =
p
=g i gn , =
Then W n 1 = W 0 , and GMM using W n as the weight matrix is asymptotically ecient. A common alternative choice is to set n !1 1X 0 g i g i Wn = n
i=1
1 X 0 g i g i n
i=1
!1
1X 0 g i g 0 i gngn n
i=1
!1
(13.4)
which uses the uncentered moment conditions. Since Eg i = 0, these two estimators are asymptotically equivalent under the hypothesis of correct specication. However, Alastair Hall (2000) has shown that the uncentered estimator is a poor choice. When constructing hypothesis tests, under the alternative hypothesis the moment conditions are violated, i.e. Egi 6= 0, so the uncentered estimator will contain an undesirable bias term and the power of the test will be adversely aected. A simple solution is to use the centered moment conditions to construct the weight matrix, as in (13.4) above. Here is a simple way to compute the ecient GMM estimator for the linear model. First, set b using this weight matrix, and construct the residual e b i = yi x0 W n = (Z 0 Z )1 , estimate i . i , and let g be the associated n matrix. Then the ecient GMM estimator is Then set g i = z i e 1 0 1 0 0 1 0 0 b = X 0Z g g ng n g0 ZX XZ g g ng n g 0 Z y. n n In most cases, when we say GMM, we actually mean ecient GMM. There is little point in using an inecient GMM estimator when the ecient estimator is easy to compute. can be seen from the above formula. Set An estimator of the asymptotic variance of 0 1 0 1 b = n X 0Z g g ng n g0 ZX . V n
1b . Asymptotic standard errors are given by the square roots of the diagonal elements of V n There is an important alternative to the two-step GMM estimator just described. Instead, we can let the weight matrix be considered as a function of . The criterion function is then n !1 1X 0 0 J () = n g n () gi ()g i () gn (). n
i=1
where b which minimizes this function is called the continuously-updated GMM estimator, and The was introduced by L. Hansen, Heaton and Yaron (1996). The estimator appears to have some better properties than traditional GMM, but can be numerically tricky to obtain in some cases. This is a current area of research in econometrics. g i ( ) = g i ( ) g n ( )
282
13.5
GMM: The General Case
In its most general form, GMM applies whenever an economic or statistical model implies the 1 moment condition E (g i ()) = 0. Often, this is all that is known. Identication requires l k = dim(). The GMM estimator minimizes J () = n g n ()0 W n gn () where g n () = and Wn = 1X g i () n
i=1 n
e ) constructed using a preliminary consistent estimator e , perhaps obtained by rst with g i = gi ( setting W n = I . Since the GMM estimator depends upon the rst-stage estimator, often the weight b recomputed. This estimator can be iterated if needed. matrix W n is updated, and then Theorem 13.5.1 Distribution of Nonlinear GMM Estimator Under general regularity conditions, 1 d 0 1 b , n N 0, G G where and = E gig0 i
1X 0 g i g 0 i gngn n
i=1
!1
G=E
g (). 0 i
b may be estimated by The variance of where
0 1 1 b = G G V = n1 X
i 0 g ig i
and = n1 G
X 0 g i ( ). i
The general theory of GMM estimation and testing was exposited by L. Hansen (1982).
13.6
Over-Identication Test
Overidentied models ( > k) are special in the sense that there may not be a parameter value such that the moment condition
283
Eg (yi , xi , z i , ) = 0 holds. Thus the model the overidentifying restrictions are testable. 0 For example, take the linear model yi = 0 1 x1i + 2 x2i + ei with E (x1i ei ) = 0 and E (x2i ei ) = 0. It is possible that 2 = 0, so that the linear equation may be written as yi = 0 1 x1i + ei . However, it is possible that 2 6= 0, and in this case it would be impossible to nd a value of 1 so that 0 both E (x1i (yi x0 1i 1 )) = 0 and E (x2i (yi x1i 1 )) = 0 hold simultaneously. In this sense an exclusion restriction can be seen as an overidentifying restriction. p Note that g n Egi , and thus g n can be used to assess whether or not the hypothesis that Eg i = 0 is true or not. The criterion function at the parameter estimates is Jn = n g 0 nW ngn 0 1 = n2 g 0 ng n g 0 gn. g n g n
is a quadratic form in g n , and is thus a natural test statistic for H0 : Eg i = 0.
Theorem 13.6.1 (Sargan-Hansen). Under the hypothesis of correct specication, and if the weight matrix is asymptotically ecient,
d b ) 2k . Jn = Jn (
The proof of the theorem is left as an exercise. This result was established by Sargan (1958) for a specialized case, and by L. Hansen (1982) for the general case. The degrees of freedom of the asymptotic distribution are the number of overidentifying restrictions. If the statistic J exceeds the chi-square critical value, we can reject the model. Based on this information alone, it is unclear what is wrong, but it is typically cause for concern. The GMM overidentication test is a very useful by-product of the GMM methodology, and it is advisable to report the statistic J whenever GMM is the estimation method. When over-identied models are estimated by GMM, it is customary to report the J statistic as a general test of model adequacy.
13.7
Hypothesis Testing: The Distance Statistic
We described before how to construct estimates of the asymptotic covariance matrix of the GMM estimates. These may be used to construct Wald tests of statistical hypotheses. If the hypothesis is non-linear, a better approach is to directly use the GMM criterion function. This is sometimes called the GMM Distance statistic, and sometimes called a LR-like statistic (the LR is for likelihood-ratio). The idea was rst put forward by Newey and West (1987). For a given weight matrix W n , the GMM criterion function is Jn () = n g n ( )0 W n g n () For h : Rk Rr , the hypothesis is H0 : h() = 0. The estimates under H1 are b = argmin Jn ()
CHAPTER 13. GENERALIZED METHOD OF MOMENTS and those under H0 are
284
b ) and Jn ( e ). The GMM distance statistic is the The two minimizing criterion functions are Jn ( dierence e ) Jn ( b ). Dn = Jn ( Proposition 13.7.1 If the same weight matrix W n is used for both null and alternative, 1. D 0 2. D 2 r 3. If h is linear in , then D equals the Wald statistic.
d
e = argmin J ().
h()=0
If h is non-linear, the Wald statistic can work quite poorly. In contrast, current evidence suggests that the Dn statistic appears to have quite good sampling properties, and is the preferred test statistic. Newey and West (1987) suggested to use the same weight matrix W n for both null and alternative, as this ensures that Dn 0. This reasoning is not compelling, however, and some current research suggests that this restriction is not necessary for good performance of the test. This test shares the useful feature of LR tests in that it is a natural by-product of the computation of alternative models.
13.8
Conditional Moment Restrictions
In many contexts, the model implies more than an unconditional moment restriction of the form Eg i () = 0. It implies a conditional moment restriction of the form E (ei ( ) | z i ) = 0 where ei () is some s 1 function of the observation and the parameters. In many cases, s = 1. The variable z i is often called an instrument. It turns out that this conditional moment restriction is much more powerful, and restrictive, than the unconditional moment restriction discussed above. As discussed later in Chapter 15, the linear model yi = x0 i + ei with instruments z i falls into this class under the assumption E (ei | z i ) = 0. In this case, ei () = yi x0 i . It is also helpful to realize that conventional regression models also fall into this class, except that in this case xi = z i . For example, in linear regression, ei ( ) = yi x0 i , while in a nonlinear regression model ei () = yi g (xi , ). In a joint model of the conditional mean and variance yi x0 i ei (, ) = . 2 0 0 (yi xi ) f (xi )
Here s = 2. Given a conditional moment restriction, an unconditional moment restriction can always be constructed. That is for any 1 function (xi , ) , we can set g i () = (xi , ) ei () which
285
satises Eg i () = 0 and hence denes a GMM estimator. The obvious problem is that the class of functions is innite. Which should be selected? This is equivalent to the problem of selection of the best instruments. If xi R is a valid 3 instrument satisfying E (ei | xi ) = 0, then xi , x2 i , xi , ..., etc., are all valid instruments. Which should be used? One solution is to construct an innite list of potent instruments, and then use the rst k instruments. How is k to be determined? This is an area of theory still under development. A recent study of this problem is Donald and Newey (2001). Another approach is to construct the optimal instrument. The form was uncovered by Chamberlain (1987). Take the case s = 1. Let ei () | z i Ri = E and Then the optimal instrument is so the optimal moment is g i () = Ai ei ( ). Setting g i ( ) to be this choice (which is k 1, so is just-identied) yields the best GMM estimator possible. In practice, Ai is unknown, but its form does help us think about construction of optimal instruments. In the linear model ei () = yi x0 i , note that Ri = E (xi | z i ) and so 2 = E e2 i i | zi , 2 = E ei ()2 | z i . i
2 Ri Ai = i
2 z i . Hence ecient GMM is GLS, as we In the case of linear regression, xi = z i , so Ai = i discussed earlier in the course. In the case of endogenous variables, note that the ecient instrument Ai involves the estimation of the conditional mean of xi given z i . In other words, to get the best instrument for xi , we need the best conditional mean model for xi given z i , not just an arbitrary linear projection. The ecient instrument is also inversely proportional to the conditional variance of ei . This is the same as the GLS estimator; namely that improved eciency can be obtained if the observations are weighted inversely to the conditional variance of the errors.
2 E (xi | z i ) . Ai = i
13.9
Bootstrap GMM Inference
b be the 2SLS or GMM estimator of . Using the EDF of (yi , z i , xi ), we can apply the Let b, bootstrap methods discussed in Chapter 10 to compute estimates of the bias and variance of and construct condence intervals for , identically as in the regression model. However, caution should be applied when interpreting such results. A straightforward application of the nonparametric bootstrap works in the sense of consistently achieving the rst-order asymptotic distribution. This has been shown by Hahn (1996). However, it fails to achieve an asymptotic renement when the model is over-identied, jeopardizing the
286
, z , x ) do not satisfy the same moment conditions as the population distribThis means that (yi i i ution. A correction suggested by Hall and Horowitz (1996) can solve the problem. Given the bootstrap sample (y , Z , X ), dene the bootstrap GMM criterion 0 Jn () = n g ( ) g ( ) W g ( ) g ( ) n n n n n
theoretical justication for percentile-t methods. Furthermore, the bootstrap applied J test will yield the wrong answer. b is the true value and yet g ( ) 6= 0. Thus according to The problem is that in the sample, n random variables (yi , z i , xi ) drawn from the EDF Fn , b ) 6= 0. E gi = g n (
b are the in-sample residuals. The bootstrap J statistic is J ( b ). where e = y X n Brown and Newey (2002) have an alternative solution. They note that we can sample from the observations i described in Chapter 14. Since with the empirical likelihood probabilities p Pn b = 0, this sampling scheme preserves the moment conditions of the model, so no i g i i=1 p recentering or adjustments is needed. Brown and Newey argue that this bootstrap procedure will be more ecient than the Hall-Horowitz GMM bootstrap.
b ) is from the in-sample data, not from the bootstrap data. where gn ( b minimize J (), and dene all statistics and tests accordingly. In the linear model, this Let n implies that the bootstrap estimator is b = X 0 Z W Z 0 X 1 X 0 Z W Z 0 y Z 0 e . n n n
287
Exercises
Exercise 13.1 Take the model yi = x0 i + ei E (xi ei ) = 0
0 e2 i = z i + i
E (z i i ) = 0. , Find the method of moments estimators for ( , ) . Exercise 13.2 Take the single equation y = X + e 1 0 2 , Assume E e2 i | z i = . Show that if is estimated by GMM with weight matrix W n = (Z Z ) then 1 d b n N 0, 2 Q0 M 1 Q
0 where Q = E (z i x0 i ) and M = E (z i z i ) .
E (e | Z ) = 0
b b Exercise 13.3 Take the model yi = x0 i = yi x0 i + ei with E (z i ei ) = 0. Let e i where is consistent for (e.g. a GMM estimator with arbitrary weight matrix). Dene the estimate of the optimal GMM weight matrix n !1 1X 0 2 zi zi e i . Wn = n i=1 0 2 p 1 Show that W n where = E z i z i ei . Exercise 13.4 In the linear model estimated by GMM with general weight matrix W , the asympGMM is totic variance of 1 0 1 V = Q0 W Q Q W W Q Q0 W Q
(b) We want to show that for any W , V V 0 is positive semi-denite (for then V 0 is the smaller possible covariance matrix and W = 1 is the ecient weight matrix). To do this, start by nding matrices A and B such that V = A0 A and V 0 = B 0 B . (c) Show that B 0 A = B 0 B and therefore that B 0 (A B ) = 0. (d) Use the expressions V = A0 A, A = B + (A B ) , and B 0 (A B ) = 0 to show that V V 0. Exercise 13.5 The equation of interest is yi = m(xi , ) + ei E (z i ei ) = 0. The observed data is (yi , z i , xi ). z i is 1 and is k 1, k. Show how to construct an ecient GMM estimator for .
1 (a) Let V 0 be this matrix when W = 1 . Show that V 0 = Q0 1 Q .
288
Exercise 13.6 In the linear model y = X + e with E(xi ei ) = 0, a Generalized Method of Moments (GMM) criterion function for is dened as Jn () = 1 b 1 X 0 (y X ) (y X )0 X n e = argmin Jn ().
h()=0 h()=0
(13.5)
1 0 2 = y x0 b b = 1 Pn xi x0 e where X 0 y is LS. The i i i i , e i are the OLS residuals, and = (X X ) i=1 n GMM estimator of , subject to the restriction h() = 0, is dened as
The GMM test statistic (the distance statistic) of the hypothesis h() = 0 is e ) = min Jn (). D = Jn ( (13.6)
(a) Show that you can rewrite Jn () in (13.5) as 0 1 b b V Jn () = n (b) Show that in this setting, the distance statistic D in (13.6) equals the Wald statistic. Exercise 13.7 Take the linear model yi = x0 i + ei E (z i ei ) = 0. of . Let and consider the GMM estimator b )0 b) b 1 g n ( Jn = ng n ( e is the same as the minimum distance estimator. thus
denote the test of overidentifying restrictions. Show that Jn 2k as n by demonstrating each of the following: (a) Since > 0, we can write 1 = CC 0 and = C 01 C 1 1 0 b) b) C C 0 (b) Jn = n C 0 gn ( C 0 g n (
b ) = Dn C 0 g n (0 ) where (c) C 0 g n ( 1 1 1 0 1 0 1 0 1 0 0 b b 1 C 01 ZX XZ ZX XZ Dn = I C n n n n g n ( 0 ) = 1 0 Z e. n
p 1
(d) Dn I R (R0 R)
d
R0 where R = C 0 E (z i x0 i)
1 (g) u0 I R (R0 R) R0 u 2k . Hint: I R (R0 R)

1
(e) n1/2 C 0 g n (0 ) u N (0, I ) d 1 (f) Jn u0 I R (R0 R) R0 u
R0 is a projection matrix.
Chapter 14
Empirical Likelihood
14.1 Non-Parametric Likelihood
An alternative to GMM is empirical likelihood. The idea is due to Art Owen (1988, 2001) and has been extended to moment condition models by Qin and Lawless (1994). It is a non-parametric analog of likelihood estimation. The idea is to construct a multinomial distribution F (p1 , ..., pn ) which places probability pi at each observation. To be a valid multinomial distribution, these probabilities must satisfy the requirements that pi 0 and n X pi = 1. (14.1)
i=1
First let us consider a just-identied model. In this case the moment condition places no additional restrictions on the multinomial distribution. The maximum likelihood estimators of the probabilities (p1 , ..., pn ) are those which maximize the log-likelihood subject to the constraint (14.1). This is equivalent to maximizing n ! n X X log(pi ) pi 1
i=1 i=1 1 where is a Lagrange multiplier. The n rst order conditions are 0 = p i . Combined with the constraint (14.1) we nd that the MLE is pi = n1 yielding the log-likelihood n log(n). Now consider the case of an overidentied model with moment condition
Since each observation is observed once in the sample, the log-likelihood function for this multinomial distribution is n X log(pi ). (14.2) log L (p1 , ..., pn ) =
i=1
Eg i (0 ) = 0 where g is 1 and is k 1 and for simplicity we write gi ( ) = g (yi , z i , xi , ). The multinomial distribution which places probability pi at each observation (yi , xi , z i ) will satisfy this condition if and only if n X pi g i () = 0 (14.3)
i=1
The empirical likelihood estimator is the value of which maximizes the multinomial loglikelihood (14.2) subject to the restrictions (14.1) and (14.3).
289
CHAPTER 14. EMPIRICAL LIKELIHOOD The Lagrangian for this maximization problem is n ! n n X X X L (, p1 , ..., pn , , ) = log(pi ) pi 1 n0 pi g i ()
i=1 i=1 i=1
290
where and are Lagrange multipliers. The rst-order-conditions of L with respect to pi , , and are 1 = + n0 g i ( ) pi
n X i=1
pi = 1
n X i=1
pi g i () = 0.
Substituting into L we nd
Multiplying the rst equation by pi , summing over i, and using the second and third equations, we nd = n and 1 . pi = n 1 + 0 g i () R (, ) = n log (n)
n X i=1
log 1 + 0 g i () .
(14.4)
For given , the Lagrange multiplier () minimizes R (, ) : ( ) = argmin R(, ).
(14.5)
This minimization problem is the dual of the constrained maximization problem. The solution (when it exists) is well dened since R( , ) is a convex function of . The solution cannot be obtained explicitly, but must be obtained numerically (see section 6.5). This yields the (prole) empirical log-likelihood function for . R() = R( , ()) = n log (n)
n X i=1
log 1 + ()0 g i ()
is the value which maximizes R(), or equivalently minimizes its negative The EL estimate = argmin [R()]
(14.6)
(see Section 14.5). Numerical methods are required for calculation of = ( ), probabilities As a by-product of estimation, we also obtain the Lagrange multiplier p i = and maximized empirical likelihood 1 . 0 gi n 1+
n X i=1
) = R(
log ( pi ) .
(14.7)
CHAPTER 14. EMPIRICAL LIKELIHOOD
291
14.2
Asymptotic Distribution of EL Estimator

g () 0 i
Let 0 denote the true value of and dene Gi () = (14.8)
and
G = EGi (0 ) = E g i (0 ) g i (0 )0 V
0 2 0 For example, in the linear model, Gi () = z i x0 i , G = E (z i xi ), and = E z i z i ei . Theorem 14.2.1 Under regularity conditions, d 0 n N (0, V ) d n 1 N (0, V ) where V and V are dened in (14.9) and (14.10), and n are asymptotically independent.
1 V = G0 1 G 1 0 = G G0 1 G G
(14.9) (14.10)
n 0 and
is the same as for ecient GMM. The theorem shows that the asymptotic variance V for Thus the EL estimator is asymptotically ecient. Chamberlain (1987) showed that V is the semiparametric eciency bound for in the overidentied moment condition model. This means that no consistent estimator for this class of models can have a lower asymptotic variance than V . Since the EL estimator achieves this bound, it is an asymptotically ecient estimator for . , ) jointly solve Proof of Theorem 14.2.1. ( g i R(, ) = 0= 0 i=1 1 + g i 0 n G X i . R( , ) = 0= 0 i=1 1 + g i
n X
(14.11)
(14.12)
0 1 Pn 1 Pn 1 Pn Let Gn = n i=1 Gi ( 0 ) , g n = n i=1 g i ( 0 ) and n = n i=1 g i ( 0 ) g i ( 0 ) . Expanding (14.12) around = 0 and = 0 = 0 yields 0 ' G0 . 0 n
(14.13)
Expanding (14.11) around = 0 and = 0 = 0 yields 0 + n 0 ' gn Gn
(14.14)

1 Premultiplying by G0 n n and using (14.13) yields
292
and using the WLLN and CLT yields Solving for 1 0 1 1 0 ' G0 n Gn n ng n n n Gn 1 0 1 d G0 1 G G N (0, ) = N (0, V )
1 0 1 0 1 g G G 0 ' G0 n n 0 + Gn n n n n n n 1 0 1 = G0 g G G n n 0 n n n n (14.15)
and using (14.15) yields Solving (14.14) for 1 0 1 1 1 ' I Gn G0 n Gn n ng n n n n Gn 1 0 1 d N (0, ) 1 I G G0 1 G G = 1 N (0, V ) Furthermore, since 1 0 G0 I 1 G G0 1 G G =0
(14.16)
n 0 and n are asymptotically uncorrelated and hence independent.
14.3
Overidentifying Restrictions
In a parametric likelihood context, tests are based on the dierence in the log likelihood functions. The same statistic can be constructed for empirical likelihood. Twice the dierence between the unrestricted empirical log-likelihood n log (n) and the maximized empirical log-likelihood for the model (14.7) is n X 0gi . 2 log 1 + LRn = (14.17)
i=1
Theorem 14.3.1 If Eg i (0 ) = 0 then LRn 2k .
The EL overidentication test is similar to the GMM overidentication test. They are asymptotically rst-order equivalent, and have the same interpretation. The overidentication test is a very useful by-product of EL estimation, and it is advisable to report the statistic LRn whenever EL is the estimation method. Proof of Theorem 14.3.1. First, by a Taylor expansion, (14.15), and (14.16),
n 1 X 0 g i ' n g n + Gn n i=1 1 0 1 1 ' I Gn G0 Gn n ng n n n Gn . ' n n
CHAPTER 14. EMPIRICAL LIKELIHOOD Second, since log(1 + u) ' u u2 /2 for u small, LRn =
n X i=1
293
' 2
= 2k
N (0, V )0 1 N (0, V )
0 n ' n
d
n 0X i=1
0 gi 2 log 1 +
i=1
n 0 X 0 gi gi gi
where the proof of the nal equality is left as an exercise.
14.4
Testing
Eg i () = 0 (14.18)
Let the maintained model be where g is 1 and is k 1. By maintained we mean that the overidentfying restrictions contained in (14.18) are assumed to hold and are not being challenged (at least for the test discussed in this section). The hypothesis of interest is h() = 0. where h : Rk Ra . The restricted EL estimator and likelihood are the values which solve = argmax R()
h()=0
) = max R(). R(
h()=0
is simply an EL estimator with k +a overidentifying Fundamentally, the restricted EL estimator relative to . To test restrictions, so there is no fundamental change in the distribution theory for the hypothesis h( ) while maintaining (14.18), the simple overidentifying restrictions test (14.17) is not appropriate. Instead we use the dierence in log-likelihoods: ) R( ) . LRn = 2 R( This test statistic is a natural analog of the GMM distance statistic.
Theorem 14.4.1 Under (14.18) and H0 : h() = 0, LRn 2 a.
The proof of this result is more challenging and is omitted.
294
14.5
Numerical Computation
Gauss code which implements the methods discussed below can be found at http://www.ssc.wisc.edu/~bhansen/progs/elike.prc Derivatives The numerical calculations depend on derivatives of the dual likelihood function (14.4). Dene g i ( , ) = gi () 1 + 0 g i ()
G i ( , ) = The rst derivatives of (14.4) are
Gi ()0 1 + 0 g i ()
X R (, ) = R = g i ( , )
i=1
R = The second derivatives are R R
R (, ) =
n X i=1
G i ( , ) .
Inner Loop The so-called inner loop solves (14.5) for given . The modied Newton method takes a quadratic approximation to Rn (, ) yielding the iteration rule j +1 = j (R ( , j ))1 R (, j ) . (14.19)
X 2 0 = R ( , ) = g i ( , ) g i ( , ) 0 i=1 n 2 X Gi () 0 g i ( , ) Gi ( , ) = R (, ) = 0 1 + 0 g i () i=1 0 2 n X 0 g i ( ) 2 0 G = R (, ) = i ( , ) Gi ( , ) 0 1 + 0 g i () i=1
where > 0 is a scalar steplength (to be discussed next). The starting value 1 can be set to the zero vector. The iteration (14.19) is continued until the gradient R (, j ) is smaller than some prespecied tolerance. Ecient convergence requires a good choice of steplength . One method uses the following quadratic approximation. Set 0 = 0, 1 = 1 2 and 2 = 1. For p = 0, 1, 2, set p = j p (R (, j ))1 R (, j ))
Rp = R (, p )
A quadratic function can be t exactly through these three points. The value of which minimizes this quadratic is = R2 + 3R0 4R1 . 4R2 + 4R0 8R1 yielding the steplength to be plugged into (14.19).
CHAPTER 14. EMPIRICAL LIKELIHOOD A complication is that must be constrained so that 0 pi 1 which holds if n 1 + 0 g i () 1
295
(14.20)
for all i. If (14.20) fails, the stepsize needs to be decreased. Outer Loop The outer loop is the minimization (14.6). This can be done by the modied Newton method described in the previous section. The gradient for (14.6) is R = R() = R(, ) = R + 0 R = R
since R (, ) = 0 at = (), where = 1 () = R R , 0
the second equality following from the implicit function theorem applied to R (, ()) = 0. The Hessian for (14.6) is R = R() 0 = 0 R (, ()) + 0 R ( , ( )) 0 0 = R (, ()) + R0 + R + R
1 = R0 R R R .
It is not guaranteed that R > 0. If not, the eigenvalues of R should be adjusted so that all are positive. The Newton iteration rule is
1 j +1 = j R R
where is a scalar stepsize, and the rule is iterated until convergence.
Chapter 15
Endogeneity
We say that there is endogeneity in the linear model y = x0 i + ei if is the parameter of interest and E(xi ei ) 6= 0. This cannot happen if is dened by linear projection, so requires a structural interpretation. The coecient must have meaning separately from the denition of a conditional mean or linear projection. Example: Measurement error in the regressor. Suppose that (yi , x i ) are joint random 0 is linear, is the parameter of interest, and x is not observed. Instead ) = x variables, E(yi | x i i i + u where u is an k 1 measurement error, independent of yi and x we observe xi = x i i i i . Then
0 yi = x i + ei
= (xi ui )0 + ei = x0 i + vi where vi = ei u0 i . The problem is that 0 E (xi vi ) = E (x = E ui u0 i + ui ) ei ui i 6= 0 1 p = E xi x0 E ui u0 i i 6= .
if 6= 0 and E (ui u0 i ) 6= 0. It follows that if is the OLS estimator, then
This is called measurement error bias. Example: Supply and Demand. The variables qi and pi (quantity and price) are determined jointly by the demand equation qi = 1 pi + e1i and the supply equation e1i is iid, Eei = 0, 1 + 2 = 1 and Eei e0 i = I 2 (the latter for simplicity). e2i The question is, if we regress qi on pi , what happens? It is helpful to solve for qi and pi in terms of the errors. In matrix notation, qi e1i 1 1 = 1 2 pi e2i qi = 2 pi + e2i .
Assume that ei =
296
CHAPTER 15. ENDOGENEITY so
297
qi pi
= = =
The projection of qi on pi yields
1 1 1 2
2 1 1 1
2 e1i + 1 e2i (e1i e2i )
e1i e2i
e1i e2i .
qi = pi + i E (pi i ) = 0 where = 1 E (pi qi ) = 2 2 E p2 i
p , which does not equal either 1 or 2 . This is called Hence if it is estimated by OLS, simultaneous equations bias.
15.1
Instrumental Variables
yi = x0 i + ei (15.1)
Let the equation of interest be where xi is k 1, and assume that E(xi ei ) 6= 0 so there is endogeneity. We call (15.1) the structural equation. In matrix notation, this can be written as y = X + e. (15.2) Any solution to the problem of endogeneity requires additional information which we call instruments. Denition 15.1.1 The 1 random vector z i is an instrumental variable for (15.1) if E (z i ei ) = 0.
In a typical set-up, some regressors in xi will be uncorrelated with ei (for example, at least the intercept). Thus we make the partition k1 x1i (15.3) xi = x2i k2 where E(x1i ei ) = 0 yet E(x2i ei ) 6= 0. We call x1i exogenous and x2i endogenous. By the above denition, x1i is an instrumental variable for (15.1), so should be included in z i . So we have the partition k1 x1i zi = (15.4) z 2i 2 where x1i = z 1i are the included exogenous variables, and z 2i are the excluded exogenous variables. That is z 2i are variables which could be included in the equation for yi (in the sense that they are uncorrelated with ei ) yet can be excluded, as they would have true zero coecients in the equation. The model is just-identied if = k (i.e., if 2 = k2 ) and over-identied if > k (i.e., if > k2 ). 2 We have noted that any solution to the problem of endogeneity requires instruments. This does not mean that valid instruments actually exist.
CHAPTER 15. ENDOGENEITY
298
15.2
Reduced Form
be the k matrix of coecients from a projection of xi on z i , and dene ui = xi 0 z i
The reduced form relationship between the variables or regressors xi and the instruments z i is found by linear projection. Let 1 E z i x0 = E zi z0 i i
as the projection error. Then the reduced form linear relationship between xi and z i is xi = 0 z i + ui . In matrix notation, we can write (15.5) as X = Z + U where U is n k. By construction, E(z i u0 i ) = 0, so (15.5) is a projection and can be estimated by OLS: b 0 zi + u xi = i . b = Z 0 Z 1 Z 0 X . y = (Z + U ) + e = Z + v, where = and v = U + e. Observe that E (z i vi ) = E z i u0 i + E (z i ei ) = 0. +v y = Z , 0 1 0 = ZZ Zy (15.8) (15.7) b +U b X = Z (15.6) (15.5)
or where
Substituting (15.6) into (15.2), we nd
Thus (15.7) is a projection equation and may be estimated by OLS. This is
The equation (15.7) is the reduced form for y . (15.6) and (15.7) together are the reduced form equations for the system y = Z + v X = Z + U . , As we showed above, OLS yields the reduced-form estimates
299
15.3
Identication
The structural parameter relates to (, ) through (15.8). The parameter is identied, meaning that it can be recovered from the reduced form, if rank () = k. (15.9)
Assume that (15.9) holds. If = k, then = 1 . If > k, then for any W > 0, = 1 (0 W ) 0 W . If (15.9) is not satised, then cannot be recovered from (, ) . Note that a necessary (although not sucient) condition for (15.9) is k. Since Z and X have the common variables X 1 , we can rewrite some of the expressions. Using (15.3) and (15.4) to make the matrix partitions Z = [Z 1 , Z 2 ] and X = [Z 1 , X 2 ] , we can partition as 11 12 = 21 22 I 12 = 0 22 (15.6) can be rewritten as X1 = Z1 X 2 = Z 1 12 + Z 2 22 + U 2 . (15.10) is identied if rank() = k, which is true if and only if rank(22 ) = k2 (by the upper-diagonal structure of ). Thus the key to identication of the model rests on the 2 k2 matrix 22 in (15.10).
15.4
Estimation
yi = x0 i + ei E (z i ei ) = 0
The model can be written as
or Eg i () = 0 g i () = z i yi x0 i .
This is a moment condition model. Appropriate estimators include GMM and EL. The estimators and distribution theory developed in those Chapter 8 and 9 directly apply. Recall that the GMM estimator, for given weight matrix W n , is = X 0 ZW n Z 0 X 1 X 0 ZW n Z 0 y .
15.5
Special Cases: IV and 2SLS
If the model is just-identied, so that k = , then the formula for GMM simplies. We nd that b = X 0 ZW n Z 0 X 1 X 0 ZW n Z 0 y 1 1 0 1 0 Wn X Z X ZW n Z 0 y = Z 0X 1 0 = Z 0X Zy
300
This estimator is often called the instrumental variables estimator (IV) of , where Z is used as an instrument for X . Observe that the weight matrix W n has disappeared. In the just-identied case, the weight matrix places no role. This is also the method of moments estimator of , and the EL estimator. Another interpretation stems from the fact that since = 1 , we can construct the Indirect Least Squares (ILS) estimator: b b= b 1 1 0 1 0 1 0 0 = ZZ ZX Zy ZZ 0 1 0 0 1 0 = ZX ZZ ZZ Zy 0 1 0 = ZX Zy .
which again is the IV estimator.
This is called the two-stage-least squares (2SLS) estimator. It was originally proposed by Theil (1953) and Basmann (1957), and is the classic estimator for linear equations with instruments. Under the homoskedasticity assumption, the 2SLS estimator is ecient GMM, but otherwise it is inecient. It is useful to observe that writing 1 0 Z P = Z Z 0Z c = P X = Z b X then the 2SLS estimator is b = X 0 P X 1 X 0 P y 0 1 0 c c y. cX = X X
2 Recall that the optimal weight matrix is an estimate of the inverse of = E z i z 0 i ei . In the 2 0 2 0 special case that E e2 i | z i = (homoskedasticity), then = E (z i z i ) E (z i z i ) suggesting 1 0 the weight matrix W n = (Z Z ) . Using this choice, the GMM estimator equals 0 1 0 1 0 0 1 0 0 b = X Z ZZ ZX XZ ZZ Zy 2SLS
The source of the two-stage name is since it can be computed as follows c = Z b = P X. b = (Z 0 Z )1 (Z 0 X ) and X First regress X on Z , vis., 0 1 0 b= X c c y. c, vis., cX X Second, regress y on X
c. Recall, X = [X 1 , X 2 ] and Z = [X 1 , Z 2 ]. Then It is useful to scrutinize the projection X i h c2 c= X c1 , X X = [P X 1 , P X 2 ] = [X 1 , P X 2 ] i h c2 , = X 1, X
c2 . So only the since X 1 lies in the span of Z . Thus in the second stage, we regress y on X 1 and X endogenous variables X 2 are replaced by their tted values: b 12 + Z 2 b 22 . c2 = Z 1 X
301
15.6
Bekker Asymptotics
Bekker (1994) used an alternative asymptotic framework to analyze the nite-sample bias in the 2SLS estimator. Here we present a simplied version of one of his results. In our notation, the model is y = X + e X = Z + U = (e, U ) E ( | Z ) = 0 0 E |Z =S (15.11) (15.12)
As before, Z is n l so there are l instruments. First, lets analyze the approximate bias of OLS applied to (15.11). Using (15.12), 1 0 X e = E (xi ei ) = 0 E (z i ei ) + E (ui ei ) = s21 E n and 1 0 XX E n = E xi x0 i 0 0 0 0 = 0 E z i z 0 i + E ui z i + E z i ui + E ui ui = 0 Q + S 22
Let P = Z (Z 0 Z ) Z 0 . By the spectral decomposition of an idempotent matrix, P = H H 0 0 where = diag (I l , 0) . Let Q = H 0 S 1/2 which satises EQ0 Q = I n and partition Q = (q 0 1 Q2 ) where q 1 is l 1. Hence 1 1 0 P | Z = S 1/20 E Q0 Q | Z S 1/2 E n n 1 1/20 1 0 q q S 1/2 = S E n n 1 1 l = S 1/20 S 1/2 n = S where = Using (15.12) and this result, 1 0 1 0 0 1 0 E X P e = E Z e + E U P e = s21 , n n n l . n
which is zero only when s21 = 0 (when X is exogenous). We now derive a similar result for the 2SLS estimator. 2SLS = X 0 P X 1 X 0 P y .
1
where Q = E (z i z 0 i ) . Hence by a rst-order approximation 1 1 1 0 0 OLS E E XX Xe E n n 1 = 0 Q + S 22 s21
(15.13)
CHAPTER 15. ENDOGENEITY and 1 0 1 0 0 0 E X P X = 0 E z i z 0 i + E (z i ui ) + E ui z i + E U P U n n = 0 Q + S 22 . Together 1 1 1 0 0 X PX X Pe E 2SLS E E n n 1 = 0 Q + S 22 s21 .
302
(15.14)
In general this is non-zero, except when s21 = 0 (when X is exogenous). It is also close to zero when = 0. Bekker (1994) pointed out that it also has the reverse implication that when = l/n is large, the bias in the 2SLS estimator will be large. Indeed as 1, the expression in (15.14) approaches that in (15.13), indicating that the bias in 2SLS approaches that of OLS as the number of instruments increases. Bekker (1994) showed further that under the alternative asymptotic approximation that is xed as n (so that the number of instruments goes to innity proportionately with sample 2SLS size) then the expression in (15.14) is the probability limit of
15.7
Identication Failure
Recall the reduced form equation X 2 = Z 1 12 + Z 2 22 + U 2 . The parameter fails to be identied if 22 has decient rank. The consequences of identication failure for inference are quite severe. Take the simplest case where k = l = 1 (so there is no Z 1 ). Then the model may be written as yi = xi + ei xi = zi + ui
2 . We see that is identied if and only if 6= 0, which occurs and 22 = = E (zi xi ) /Ezi when E (xi zi ) 6= 0. Thus identication hinges on the existence of correlation between the excluded exogenous variable and the included endogenous variable. Suppose this condition fails, so E (xi zi ) = 0. Then by the CLT
2 2 1 X d zi ei N1 N 0, E zi ei n
i=1
(15.15)
therefore
2 2 1 X 1 X d zi xi = zi ui N2 N 0, E zi ui n n
i=1 i=1
(15.16)
since the ratio of two normals is Cauchy. This is particularly nasty, as the Cauchy distribution does not have a nite mean. This result carries over to more general settings, and was examined by Phillips (1989) and Choi and Phillips (1992).
1 Pn i=1 zi ei n Pn 1 i=1 zi xi n
N1 Cauchy, N2
303
Suppose that identication does not completely fail, but is weak. This occurs when 22 is full rank, but small. This can be handled in an asymptotic analysis by modeling it as local-to-zero, viz 22 = n1/2 C , where C is a full rank matrix. The n1/2 is picked because it provides just the right balancing to allow a rich distribution theory. To see the consequences, once again take the simple case k = l = 1. Here, the instrument xi is weak for zi if = n1/2 c. Then (15.15) is unaected, but (15.16) instead takes the form 1 X 1 X 2 1 X zi xi = zi + zi ui n n n
i=1 i=1 i=1 n n 1X 2 1 X = zi c + zi ui n n i=1 i=1 n n n
Qc + N2 therefore
d
N1 . Qc + N2
is inconsistent for and the As in the case of complete identication failure, we nd that asymptotic distribution of is non-normal. In addition, standard test statistics have non-standard distributions, meaning that inferences about parameters of interest can be misleading. The distribution theory for this model was developed by Staiger and Stock (1997) and extended to nonlinear GMM estimation by Stock and Wright (2000). Further results on testing were obtained by Wang and Zivot (1998). The bottom line is that it is highly desirable to avoid identication failure. Once again, the equation to focus on is the reduced form X 2 = Z 1 12 + Z 2 22 + U 2 and identication requires rank(22 ) = k2 . If k2 = 1, this requires 22 6= 0, which is straightforward to assess using a hypothesis test on the reduced form. Therefore in the case of k2 = 1 (one RHS endogenous variable), one constructive recommendation is to explicitly estimate the reduced form equation for X 2 , construct the test of 22 = 0, and at a minimum check that the test rejects H0 : 22 = 0. When k2 > 1, 22 6= 0 is not sucient for identication. It is not even sucient that each column of 22 is non-zero (each column corresponds to a distinct endogenous variable in Z 2 ). So while a minimal check is to test that each columns of 22 is non-zero, this cannot be interpreted as denitive proof that 22 has full rank. Unfortunately, tests of decient rank are dicult to implement. In any event, it appears reasonable to explicitly estimate and report the reduced form equations for Z 2 , and attempt to assess the likelihood that 22 has decient rank.
304
Exercises
1. Consider the single equation model yi = zi + ei , denote the IV estimator of using as an where yi and zi are both real-valued (1 1). Let instrument a dummy variable di (takes only the values 0 and 1). Find a simple expression for the IV estimator in this context. 2. In the linear model yi = x0 i + ei 2 = E e2 | x suppose i i is known. Show that the GLS estimator of can be written as an i IV estimator using some instrument z i . (Find an expression for z i .) 3. Take the linear model y = X + e. and the OLS residual be e . Let the OLS estimator for be = y X and the IV residual be e . Let the IV estimator for using some instrument Z be = y X 0 0 If X is indeed endogeneous, will IV t better than OLS, in the sense that e e <e e , at least in large samples? 4. The reduced form between the regressors xi and instruments z i takes the form xi = 0 z i + ui or X = Z + U where xi is k 1, z i is l 1, X is n k, Z is n l, U is n k, and is l k. The parameter is dened by the population moment condition E z i u0 i =0 = (Z 0 Z )1 (Z 0 X ) . Show that the method of moments estimator for is 5. In the structural model y = X + e X = Z + U with l k, l k, we claim that is identied (can be recovered from the reduced form) if rank() = k. Explain why this is true. That is, show that if rank() < k then cannot be identied. 6. Take the linear model yi = xi + ei E (ei | xi ) = 0. where xi and are 1 1. E (ei | xi ) = 0
CHAPTER 15. ENDOGENEITY (a) Show that E (xi ei ) = 0 and E x2 i ei = 0. Is z i = (xi for estimation of ?
305
0 x2 i ) a valid instrumental variable
(b) Dene the 2SLS estimator of , using z i as an instrument for xi . How does this dier from OLS? (c) Find the ecient GMM estimator of based on the moment condition E (z i (yi xi )) = 0. Does this dier from 2SLS and/or OLS? 7. Suppose that price and quantity are determined by the intersection of the linear demand and supply curves Demand : Supply : Q = a0 + a1 P + a2 Y + e1 Q = b0 + b1 P + b2 W + e2
where income (Y ) and wage (W ) are determined outside the market. In this model, are the parameters identied? 8. The data le card.dat is taken from Card (1995). There are 2215 observations with 29 variables, listed in card.pdf. We want to estimate a wage equation log(W age) = 0 + 1 Educ + 2 Exper + 3 Exper2 + 4 South + 5 Black + e where Educ = Eduation (Years) Exper = Experience (Years), and South and Black are regional and racial dummy variables. (a) Estimate the model by OLS. Report estimates and standard errors. (b) Now treat Education as endogenous, and the remaining variables as exogenous. Estimate the model by 2SLS, using the instrument near4, a dummy indicating that the observation lives near a 4-year college. Report estimates and standard errors. (c) Re-estimate by 2SLS (report estimates and standard errors) adding three additional instruments: near2 (a dummy indicating that the observation lives near a 2-year college), fatheduc (the education, in years, of the father) and motheduc (the education, in years, of the mother). (d) Re-estimate the model by ecient GMM. I suggest that you use the 2SLS estimates as the rst-step to get the weight matrix, and then calculate the GMM estimator from this weight matrix without further iteration. Report the estimates and standard errors. (e) Calculate and report the J statistic for overidentication. (f) Discuss your ndings.
Chapter 16
Univariate Time Series

A time series yt is a process observed in sequence over time, t = 1, ..., T . To indicate the dependence on time, we adopt new notation, and use the subscript t to denote the individual observation, and T to denote the number of observations. Because of the sequential nature of time series, we expect that yt and yt1 are not independent, so classical assumptions are not valid. We can separate time series into two categories: univariate (yt R is scalar); and multivariate (yt Rm is vector-valued). The primary model for univariate time series is autoregressions (ARs). The primary model for multivariate time series is vector autoregressions (VARs).
16.1
Stationarity and Ergodicity

Denition 16.1.1 {yt } is covariance (weakly) stationary if E(yt ) = is independent of t, and cov (yt , ytk ) = (k) is independent of t for all k. (k) is called the autocovariance function. (k) = (k)/ (0) = corr(yt , ytk ) is the autocorrelation function.
Denition 16.1.2 {yt } is strictly stationary if the joint distribution of (yt , ..., ytk ) is independent of t for all k.
Denition 16.1.3 A stationary time series is ergodic if (k ) 0 as k . 306
CHAPTER 16. UNIVARIATE TIME SERIES
307
The following two theorems are essential to the analysis of stationary time series. The proofs are rather dicult, however. Theorem 16.1.1 If yt is strictly stationary and ergodic and xt = f (yt , yt1 , ...) is a random variable, then xt is strictly stationary and ergodic.
Theorem 16.1.2 (Ergodic Theorem). If yt is strictly stationary and ergodic and E |yt | < , then as T ,
T 1X p yt E(yt ). T t=1
This allows us to consistently estimate parameters using time-series moments: The sample mean: T 1X yt = T t=1 The sample autocovariance
T 1X (yt ) (ytk ) . (k) = T t=1
The sample autocorrelation
(k ) =
(k) . (0)
2 < , Theorem 16.1.3 If yt is strictly stationary and ergodic and Eyt then as T ,
1. E(yt ); 2. (k) (k); 3. (k) (k ).

p p
Proof of Theorem 16.1.3. Part (1) is a direct consequence of the Ergodic theorem. For Part (2), note that
T 1X (k) = (yt ) (ytk ) T t=1
T T T 1X 1X 1X yt ytk yt ytk + 2 . T t=1 T t=1 T t=1
308
By Theorem 16.1.1 above, the sequence yt ytk is strictly stationary and ergodic, and it has a nite 2 < . Thus an application of the Ergodic Theorem yields mean by the assumption that Eyt
T 1X p yt ytk E(yt ytk ). T t=1
Thus
(k) E(yt ytk ) 2 2 + 2 = E(yt ytk ) 2 = (k). Part (3) follows by the continuous mapping theorem: (k ) = (k)/ (0) (k)/ (0) = (k).
p
16.2
Autoregressions
In time-series, the series {..., y1 , y2 , ..., yT , ...} are jointly random. We consider the conditional expectation E (yt | Ft1 ) where Ft1 = {yt1 , yt2 , ...} is the past history of the series. An autoregressive (AR) model species that only a nite number of past lags matter: E (yt | Ft1 ) = E (yt | yt1 , ..., ytk ) . A linear AR model (the most common type used in practice) species linearity: E (yt | Ft1 ) = 0 + 1 yt1 + 2 yt1 + + k ytk . Letting et = yt E (yt | Ft1 ) , then we have the autoregressive model yt = 0 + 1 yt1 + 2 yt1 + + k ytk + et
E (et | Ft1 ) = 0.
The last property denes a special time-series process.
Denition 16.2.1 et is a martingale dierence sequence (MDS) if E (et | Ft1 ) = 0.
Regression errors are naturally a MDS. Some time-series processes may be a MDS as a consequence of optimizing behavior. For example, some versions of the life-cycle hypothesis imply that either changes in consumption, or consumption growth rates, should be a MDS. Most asset pricing models imply that asset returns should be the sum of a constant plus a MDS. The MDS property for the regression error plays the same role in a time-series regression as does the conditional mean-zero property for the regression error in a cross-section regression. In fact, it is even more important in the time-series context, as it is dicult to derive distribution theories without this property. A useful property of a MDS is that et is uncorrelated with any function of the lagged information Ft1 . Thus for k > 0, E (ytk et ) = 0.
309
16.3
Stationarity of AR(1) Process

yt = yt1 + et .
A mean-zero AR(1) is Assume that et is iid, E(et ) = 0 and By back-substitution, we nd Ee2 t = 2 < .
Loosely speaking, this series converges if the sequence k etk gets small as k . This occurs when || < 1. Theorem 16.3.1 If and only if || < 1 then yt is strictly stationary and ergodic.
yt = et + et1 + 2 et2 + ... X = k etk .

k=0
We can compute the moments of yt using the innite sum: Eyt =

X k=0
k E (etk ) = 0 2 . 1 2
var(yt ) =
If the equation for yt has an intercept, the above results are unchanged, except that the mean of yt can be computed from the relationship Eyt = 0 + 1 Eyt1 , and solving for Eyt = Eyt1 we nd Eyt = 0 /(1 1 ).
X k=0
2k var (etk ) =
16.4
Lag Operator
An algebraic construct which is useful for the analysis of autoregressive models is the lag operator. Denition 16.4.1 The lag operator L satises Lyt = yt1 . Dening L2 = LL, we see that L2 yt = Lyt1 = yt2 . In general, Lk yt = ytk . The AR(1) model can be written in the format yt yt1 = et or The operator (L) = (1 L) is a polynomial in the operator L. We say that the root of the polynomial is 1/, since (z ) = 0 when z = 1/. We call (L) the autoregressive polynomial of yt . From Theorem 16.3.1, an AR(1) is stationary i || < 1. Note that an equivalent way to say this is that an AR(1) is stationary i the root of the autoregressive polynomial is larger than one (in absolute value). (1 L) yt = et .
310
16.5
Stationarity of AR(k)
yt = 1 yt1 + 2 yt2 + + k ytk + et .
The AR(k) model is Using the lag operator, yt 1 Lyt 2 L2 yt k Lk yt = et , or (L)yt = et where (L) = 1 1 L 2 L2 k Lk . We call (L) the autoregressive polynomial of yt . The Fundamental Theorem of Algebra says that any polynomial can be factored as 1 1 1 1 (z ) = 1 1 z 2 z 1 k z
where the 1 , ..., k are the complex roots of (z ), which satisfy (j ) = 0. We know that an AR(1) is stationary i the absolute value of the root of its autoregressive polynomial is larger than one. For an AR(k), the requirement is that all roots are larger than one. Let || denote the modulus of a complex number . Theorem 16.5.1 The AR(k) is strictly stationary and ergodic if and only if |j | > 1 for all j.
One way of stating this is that All roots lie outside the unit circle. If one of the roots equals 1, we say that (L), and hence yt , has a unit root. This is a special case of non-stationarity, and is of great interest in applied time series.
16.6
Let
Estimation
0
xt = = Then the model can be written as The OLS estimator is
1 yt1 yt2 0 1 2 yt = x0 t + et .
ytk 0 k .
b , it is helpful to dene the process ut = xt et . Note that ut is a MDS, since To study E (ut | Ft1 ) = E (xt et | Ft1 ) = xt E (et | Ft1 ) = 0. By Theorem 16.1.1, it is also strictly stationary and ergodic. Thus
T T 1X 1X p xt et = ut E (ut ) = 0. T t=1 T t=1
b = X 0 X 1 X 0 y .
(16.1)
311
Combined with (16.1) and the continuous mapping theorem, we see that b =
T 1X xt x0 t T t=1
The vector xt is strictly stationary and ergodic, and by Theorem 16.1.1, so is xt x0 t . Thus by the Ergodic Theorem, T 1X p 0 xt x0 t E xt xt = Q. T
t=1
!1
T 1X xt et T t=1
Q1 0 = 0.
We have shown the following:
Theorem 16.6.1 If the AR(k) process yt is strictly stationary and ergodic p 2 < , then b and Eyt as T .
16.7
Theorem 16.7.1 MDS CLT. If ut is a strictly stationary and ergodic MDS and E (ut u0 t ) = < , then as T ,
T 1 X d ut N (0, ) . T t=1
Since xt et is a MDS, we can apply Theorem 16.7.1 to see that

T 1 X d xt et N (0, ) , T t=1
where
2 = E(xt x0 t et ).
Theorem 16.7.2 If the AR(k) process yt is strictly stationary and ergodic 4 < , then as T , and Eyt d b T N 0, Q1 Q1 . This is identical in form to the asymptotic distribution of OLS in cross-section regression. The implication is that asymptotic inference is the same. In particular, the asymptotic covariance matrix is estimated just as in the cross-section case.
312
16.8
Bootstrap for Autoregressions
In the non-parametric bootstrap, we constructed the bootstrap sample by randomly resampling from the data values {yt , xt }. This creates an iid bootstrap sample. Clearly, this cannot work in a time-series application, as this imposes inappropriate independence. Briey, there are two popular methods to implement bootstrap resampling for time-series data. Method 1: Model-Based (Parametric) Bootstrap. and residuals e 1. Estimate t . 2. Fix an initial condition (yk+1 , yk+2 , ..., y0 ). 3. Simulate iid draws e 1 , ..., e T }. i from the empirical distribution of the residuals {e
by the recursive formula 4. Create the bootstrap series yt yt = 0 + 1 yt 2 yt k yt 1 + 2 + + k + et .
This construction imposes homoskedasticity on the errors e i , which may be dierent than the properties of the actual ei . It also presumes that the AR(k) structure is the truth. Method 2: Block Resampling 1. Divide the sample into T /m blocks of length m. 2. Resample complete blocks. For each simulated sample, draw T /m blocks.
. 3. Paste the blocks together to create the bootstrap time-series yt
4. This allows for arbitrary stationary serial correlation, heteroskedasticity, and for modelmisspecication. 5. The results may be sensitive to the block length, and the way that the data are partitioned into blocks. 6. May not work well in small samples.
16.9
Trend Stationarity
yt = 0 + 1 t + St St = 1 St1 + 2 St2 + + k Stk + et , or yt = 0 + 1 t + 1 yt1 + 2 yt1 + + k ytk + et . You can estimate (16.4) by OLS.
(16.2) (16.3)
(16.4)
There are two essentially equivalent ways to estimate the autoregressive parameters (1 , ..., k ).
You can estimate (16.2)-(16.3) sequentially by OLS. That is, rst estimate (16.2), get the t , and then perform regression (16.3) replacing St with S t . This procedure is someresidual S times called Detrending.
313
The reason why these two procedures are (essentially) the same is the Frisch-Waugh-Lovell theorem. Seasonal Eects There are three popular methods to deal with seasonal data. Include dummy variables for each season. This presumes that seasonality does not change over the sample. Use seasonally adjusted data. The seasonal factor is typically estimated by a two-sided weighted average of the data for that season in neighboring years. Thus the seasonally adjusted data is a ltered series. This is a exible approach which can extract a wide range of seasonal factors. The seasonal adjustment, however, also alters the time-series correlations of the data. First apply a seasonal dierencing operator. If s is the number of seasons (typically s = 4 or s = 12), s yt = yt yts , or the season-to-season change. The series s yt is clearly free of seasonality. But the long-run trend is also eliminated, and perhaps this was of relevance.
16.10
Testing for Omitted Serial Correlation

yt = 0 + 1 yt1 + ut . ut = ut1 + et (16.5)
For simplicity, let the null hypothesis be an AR(1):
We are interested in the question if the error ut is serially correlated. We model this as an AR(1): (16.6)
with et a MDS. The hypothesis of no omitted serial correlation is H0 : = 0 H1 : 6= 0. We want to test H0 against H1 . To combine (16.5) and (16.6), we take (16.5) and lag the equation once: yt1 = 0 + 1 yt2 + ut1 . We then multiply this by and subtract from (16.5), to nd yt yt1 = 0 0 + 1 yt1 1 yt1 + ut ut1 , or yt = 0 (1 ) + (1 + ) yt1 1 yt2 + et = AR(2). Thus under H0 , yt is an AR(1), and under H1 it is an AR(2). H0 may be expressed as the restriction that the coecient on yt2 is zero. An appropriate test of H0 against H1 is therefore a Wald test that the coecient on yt2 is zero. (A simple exclusion test). In general, if the null hypothesis is that yt is an AR(k), and the alternative is that the error is an AR(m), this is the same as saying that under the alternative yt is an AR(k+m), and this is equivalent to the restriction that the coecients on ytk1 , ..., ytkm are jointly zero. An appropriate test is the Wald test of this restriction.
314
16.11
Model Selection
What is the appropriate choice of k in practice? This is a problem of model selection. A good choice is to minimize the AIC information criterion AIC (k ) = log 2 (k) + 2k , T
where 2 (k) is the estimated residual variance from an AR(k) One ambiguity in dening the AIC criterion is that the sample available for estimation changes as k changes. (If you increase k, you need more initial conditions.) This can induce strange behavior in the AIC. The appropriate remedy is to x a upper value k, and then reserve the rst k as initial conditions, and then estimate the models AR(1), AR(2), ..., AR(k) on this (unied) sample.
16.12
Autoregressive Unit Roots
The AR(k) model is (L)yt = 0 + et (L) = 1 1 L k Lk . As we discussed before, yt has a unit root when (1) = 0, or 1 + 2 + + k = 1. In this case, yt is non-stationary. The ergodic theorem and MDS CLT do not apply, and test statistics are asymptotically non-normal. A helpful way to write the equation is the so-called Dickey-Fuller reparameterization: yt = 0 yt1 + 1 yt1 + + k1 yt(k1) + et . (16.7)
These models are equivalent linear transformations of one another. The DF parameterization is convenient because the parameter 0 summarizes the information about the unit root, since (1) = 0 . To see this, observe that the lag polynomial for the yt computed from (16.7) is (1 L) 0 L 1 (L L2 ) k1 (Lk1 Lk ) But this must equal (L), as the models are equivalent. Thus (1) = (1 1) 0 (1 1) (1 1) = 0 . Hence, the hypothesis of a unit root in yt can be stated as H0 : 0 = 0. Note that the model is stationary if 0 < 0. So the natural alternative is H1 : 0 < 0. Under H0 , the model for yt is yt = + 1 yt1 + + k1 yt(k1) + et , which is an AR(k-1) in the rst-dierence yt . Thus if yt has a (single) unit root, then yt is a stationary AR process. Because of this property, we say that if yt is non-stationary but d yt is stationary, then yt is integrated of order d, or I (d). Thus a time series with unit root is I (1).
315
Since 0 is the parameter of a linear regression, the natural test statistic is the t-statistic for H0 from OLS estimation of (16.7). Indeed, this is the most popular unit root test, and is called the Augmented Dickey-Fuller (ADF) test for a unit root. It would seem natural to assess the signicance of the ADF statistic using the normal table. However, under H0 , yt is non-stationary, so conventional normal asymptotics are invalid. An alternative asymptotic framework has been developed to deal with non-stationary data. We do not have the time to develop this theory in detail, but simply assert the main results.
Theorem 16.12.1 Dickey-Fuller Theorem. If 0 = 0 then as T , T 0 (1 1 2 k1 ) DF ADF = 0 DFt . s( 0 )

d
The limit distributions DF and DFt are non-normal. They are skewed to the left, and have negative means. The rst result states that 0 converges to its true value (of zero) at rate T, rather than the 1 / 2 conventional rate of T . This is called a super-consistent rate of convergence. The second result states that the t-statistic for 0 converges to a limit distribution which is non-normal, but does not depend on the parameters . This distribution has been extensively 0 ) is the tabulated, and may be used for testing the hypothesis H0 . Note: The standard error s( conventional (homoskedastic) standard error. But the theorem does not require an assumption of homoskedasticity. Thus the Dickey-Fuller test is robust to heteroskedasticity. Since the alternative hypothesis is one-sided, the ADF test rejects H0 in favor of H1 when ADF < c, where c is the critical value from the ADF table. If the test rejects H0 , this means that the evidence points to yt being stationary. If the test does not reject H0 , a common conclusion is that the data suggests that yt is non-stationary. This is not really a correct conclusion, however. All we can say is that there is insucient evidence to conclude whether the data are stationary or not. We have described the test for the setting of with an intercept. Another popular setting includes as well a linear time trend. This model is yt = 1 + 2 t + 0 yt1 + 1 yt1 + + k1 yt(k1) + et . (16.8)
This is natural when the alternative hypothesis is that the series is stationary about a linear time trend. If the series has a linear trend (e.g. GDP, Stock Prices), then the series itself is nonstationary, but it may be stationary around the linear time trend. In this context, it is a silly waste of time to t an AR model to the level of the series without a time trend, as the AR model cannot conceivably describe this data. The natural solution is to include a time trend in the tted OLS equation. When conducting the ADF test, this means that it is computed as the t-ratio for 0 from OLS estimation of (16.8). If a time trend is included, the test procedure is the same, but dierent critical values are required. The ADF test has a dierent distribution when the time trend has been included, and a dierent table should be consulted. Most texts include as well the critical values for the extreme polar case where the intercept has been omitted from the model. These are included for completeness (from a pedagogical perspective) but have no relevance for empirical practice where intercepts are always included.
Chapter 17
Multivariate Time Series

A multivariate time series y t is a vector process m 1. Let Ft1 = (y t1 , y t2 , ...) be all lagged information at time t. The typical goal is to nd the conditional expectation E (y t | Ft1 ) . Note that since y t is a vector, this conditional expectation is also a vector.
17.1
Vector Autoregressions (VARs)
Observe that a0 is m 1,and each of A1 through Ak are m m matrices. Dening the m 1 regression error et = y t E (y t | Ft1 ) , we have the VAR model E (et | Ft1 ) = 0.
A linear VAR species that this conditional mean is linear in the arguments: E y t | y t1 , ..., y tk = a0 + A1 y t1 + A2 y t2 + Ak y tk .
A VAR model species that the conditional mean is a function of only a nite number of lags: E (y t | Ft1 ) = E y t | y t1 , ..., y tk .
y t = a0 + A1 y t1 + A2 y t2 + Ak y tk + et
Alternatively, dening the mk + 1 vector y t1 xt = y t2 . . . y tk A= then 1 Ak ,
and the m (mk + 1) matrix
a0 A1 A2 y t = Axt + et .
The VAR model is a system of m equations. One way to write this is to let a0 j be the j th row of A. Then the VAR system can be written as the equations Yjt = a0 j xt + ejt . Unrestricted VARs were introduced to econometrics by Sims (1980). 316
CHAPTER 17. MULTIVARIATE TIME SERIES
317
17.2
Estimation
E (xt ejt ) = 0,
Consider the moment conditions j = 1, ..., m. These are implied by the VAR model, either as a regression, or as a linear projection. The GMM estimator corresponding to these moment conditions is equation-by-equation OLS a j = (X 0 X )1 X 0 y j . An alternative way to compute this is as follows. Note that
0 0 1 a 0 j = y j X (X X ) .
we nd And if we stack these to create the estimate A, y0 1 y0 2 A = . X (X 0 X )1 . . y0 m+1 = Y 0 X (X 0 X )1 , y1 y2 where Y =
ym
the T m matrix of the stacked y 0 t. This (system) estimator is known as the SUR (Seemingly Unrelated Regressions) estimator, and was originally derived by Zellner (1962)
17.3
Restricted VARs
The unrestricted VAR is a system of m equations, each with the same set of regressors. A restricted VAR imposes restrictions on the system. For example, some regressors may be excluded from some of the equations. Restrictions may be imposed on individual equations, or across equations. The GMM framework gives a convenient method to impose such restrictions on estimation.
17.4
Single Equation from a VAR

yjt = a0 j xt + et ,
Often, we are only interested in a single equation out of a VAR system. This takes the form
0 s. In this case, it is convenient to re-dene and xt consists of lagged values of yjt and the other ylt the variables. Let yt = yjt , and z t be the other variables. Let et = ejt and = aj . Then the single equation takes the form (17.1) yt = x0 t + et ,
and xt =
This is just a conventional regression with time series data.
1 y t1
y tk z 0 t1
z0 tk
0 i
318
17.5
Testing for Omitted Serial Correlation
Consider the problem of testing for omitted serial correlation in equation (17.1). Suppose that et is an AR(1). Then yt = x0 t + et E (ut | Ft1 ) = 0. Then the null and alternative are H0 : = 0 H1 : 6= 0. et = et1 + ut (17.2)
Take the equation yt = x0 t + et , and subtract o the equation once lagged multiplied by , to get 0 yt yt1 = x0 t + et xt1 + et1 = x0 t xt1 + et et1 , or
0 yt = yt1 + x0 t + xt1 + ut ,
(17.3)
which is a valid regression model. So testing H0 versus H1 is equivalent to testing for the signicance of adding (yt1 , xt1 ) to the regression. This can be done by a Wald test. We see that an appropriate, general, and simple way to test for omitted serial correlation is to test the signicance of extra lagged values of the dependent variable and regressors. You may have heard of the Durbin-Watson test for omitted serial correlation, which once was very popular, and is still routinely reported by conventional regression packages. The DW test is appropriate only when regression yt = x0 t + et is not dynamic (has no lagged values on the RHS), and et is iid N(0, 2 ). Otherwise it is invalid. Another interesting fact is that (17.2) is a special case of (17.3), under the restriction = . This restriction, which is called a common factor restriction, may be tested if desired. If valid, the model (17.2) may be estimated by iterated GLS. (A simple version of this estimator is called Cochrane-Orcutt.) Since the common factor restriction appears arbitrary, and is typically rejected empirically, direct estimation of (17.2) is uncommon in recent applications.
17.6
Selection of Lag Length in an VAR
If you want a data-dependent rule to pick the lag length k in a VAR, you may either use a testingbased approach (using, for example, the Wald statistic), or an information criterion approach. The formula for the AIC and BIC are (k) + 2 p AIC (k) = log det T p log(T ) (k) + BIC (k) = log det T T X (k) = 1 e t (k)e t (k )0 T t=1 p = m(km + 1) where p is the number of parameters in the model, and e t (k ) is the OLS residual vector from the model with k lags. The log determinant is the criterion from the multivariate normal likelihood.
319
17.7
Granger Causality
The information set F1t is generated only by the history of y t , and the information set F2t is generated by both y t and z t . The latter has more information. We say that z t does not Granger-cause y t if E (y t | F1,t1 ) = E (y t | F2,t1 ) . That is, conditional on information in lagged y t , lagged z t does not help to forecast y t . If this condition does not hold, then we say that z t Granger-causes y t . The reason why we call this Granger Causality rather than causality is because this is not a physical or structure denition of causality. If z t is some sort of forecast of the future, such as a futures price, then z t may help to forecast y t even though it does not cause y t . This denition of causality was developed by Granger (1969) and Sims (1972). In a linear VAR, the equation for y t is
0 y t = + 1 y t1 + + k y tk + z 0 t1 1 + + z tk k + et .
Partition the data vector into (y t , z t ). Dene the two information sets F1t = y t , y t1 , y t2 , ... F2t = y t , z t , y t1 , z t1 , y t2 , z t2 , , ...
In this equation, z t does not Granger-cause y t if and only if H0 : 1 = 2 = = k = 0. This may be tested using an exclusion (Wald) test. This idea can be applied to blocks of variables. That is, y t and/or z t can be vectors. The hypothesis can be tested by using the appropriate multivariate Wald test. If it is found that z t does not Granger-cause y t , then we deduce that our time-series model of E (y t | Ft1 ) does not require the use of z t . Note, however, that z t may still be useful to explain other features of y t , such as the conditional variance.
Clive W. J. Granger
Clive Granger (1934-2009) of England was one of the leading gures in timeseries econometrics, and co-winner in 2003 of the Nobel Memorial Prize in Economic Sciences (along with Robert Engle). In addition to formalizing the denition of causality known as Granger causality, he invented the concept of cointegration, introduced spectral methods into econometrics, and formalized methods for the combination of forecasts.
17.8
Cointegration
The idea of cointegration is due to Granger (1981), and was articulated in detail by Engle and Granger (1987).
CHAPTER 17. MULTIVARIATE TIME SERIES Denition 17.8.1 The m 1 series y t is cointegrated if y t is I (1) yet there exists , m r, of rank r, such that z t = 0 y t is I (0). The r vectors in are called the cointegrating vectors.
320
If the series y t is not cointegrated, then r = 0. If r = m, then y t is I (0). For 0 < r < m, y t is I (1) and cointegrated. In some cases, it may be believed that is known a priori. Often, = (1 1)0 . For example, if y t is a pair of interest rates, then = (1 1)0 species that the spread (the dierence in returns) is stationary. If y = (log(C ) log(I ))0 , then = (1 1)0 species that log(C/I ) is stationary. In other cases, may not be known. If y t is cointegrated with a single cointegrating vector (r = 1), then it turns out that can be consistently estimated by an OLS regression of one component of y t on the others. Thus y t = p 2 = (y 0 y )1 y 0 y 2 . Furthermore (Y1t , Y2t ) and = (1 2 ) and normalize 1 = 1. Then 2 2 2 1 d this estimation is super-consistent: T (2 2 ) Limit, as rst shown by Stock (1987). This is not, in general, a good method to estimate , but it is useful in the construction of alternative estimators and tests. We are often interested in testing the hypothesis of no cointegration: H0 : r = 0 H1 : r > 0. Suppose that is known, so z t = 0 y t is known. Then under H0 z t is I (1), yet under H1 z t is I (0). Thus H0 can be tested using a univariate ADF test on z t . When is unknown, Engle and Granger (1987) suggested using an ADF test on the estimated 0 y t , from OLS of y1t on y2t . Their justication was Stocks result that is superresidual z t = is not consistent, so the ADF critical values are not consistent under H1 . Under H0 , however, appropriate. The asymptotic distribution was worked out by Phillips and Ouliaris (1990). When the data have time trends, it may be necessary to include a time trend in the estimated cointegrating regression. Whether or not the time trend is included, the asymptotic distribution of the test is aected by the presence of the time trend. The asymptotic distribution was worked out in B. Hansen (1992).
17.9
Cointegrated VARs
We can write a VAR as A(L)y t = et A(L) = I A1 L A2 L2 Ak Lk or alternatively as y t = y t1 + D(L)y t1 + et where = A(1)
= I + A1 + A2 + + Ak .
CHAPTER 17. MULTIVARIATE TIME SERIES Theorem 17.9.1 Granger Representation Theorem y t is cointegrated with m r if and only if rank() = r and = 0 where is m r, rank () = r.
321
Thus cointegration imposes a restriction upon the parameters of a VAR. The restricted model can be written as y t = 0 y t1 + D(L)y t1 + et y t = z t1 + D(L)y t1 + et . If is known, this can be estimated by OLS of y t on z t1 and the lags of y t . If is unknown, then estimation is done by reduced rank regression, which is least-squares subject to the stated restriction. Equivalently, this is the MLE of the restricted parameters under the assumption that et is iid N(0, ). One diculty is that is not identied without normalization. When r = 1, we typically just normalize one element to equal unity. When r > 1, this does not work, and dierent authors have adopted dierent identication schemes. In the context of a cointegrated VAR estimated by reduced rank regression, it is simple to test for cointegration by testing the rank of . These tests are constructed as likelihood ratio (LR) tests. As they were discovered by Johansen (1988, 1991, 1995), they are typically called the Johansen Max and Trace tests. Their asymptotic distributions are non-standard, and are similar to the Dickey-Fuller distributions.
Chapter 18
Limited Dependent Variables

A limited dependent variable y is one which takes a limited set of values. The most common cases are Binary: y {0, 1} Multinomial: y {0, 1, 2, ..., k} Integer: y {0, 1, 2, ...} Censored: y R+ The traditional approach to the estimation of limited dependent variable (LDV) models is parametric maximum likelihood. A parametric model is constructed, allowing the construction of the likelihood function. A more modern approach is semi-parametric, eliminating the dependence on a parametric distributional assumption. We will discuss only the rst (parametric) approach, due to time constraints. They still constitute the majority of LDV applications. If, however, you were to write a thesis involving LDV estimation, you would be advised to consider employing a semi-parametric estimation approach. For the parametric approach, estimation is by MLE. A major practical issue is construction of the likelihood function.
18.1
Binary Choice
The dependent variable yi {0, 1}. This represents a Yes/No outcome. Given some regressors xi , the goal is to describe Pr (yi = 1 | xi ) , as this is the full conditional distribution. The linear probability model species that Pr (yi = 1 | xi ) = x0 i . As Pr (yi = 1 | xi ) = E (yi | xi ) , this yields the regression: yi = x0 i + ei which can be estimated by OLS. However, the linear probability model does not impose the restriction that 0 Pr (yi | xi ) 1. Even so estimation of a linear probability model is a useful starting point for subsequent analysis. The standard alternative is to use a function of the form Pr (yi = 1 | xi ) = F x0 i
where F () is a known CDF, typically assumed to be symmetric about zero, so that F (u) = 1 F (u). The two standard choices for F are Logistic: F (u) = (1 + eu )
1
. 322
CHAPTER 18. LIMITED DEPENDENT VARIABLES Normal: F (u) = (u).
323
If F is logistic, we call this the logit model, and if F is normal, we call this the probit model. This model is identical to the latent variable model
yi = x0 i + ei
ei F () 1 yi = 0 For then
>0 if yi . otherwise
Estimation is by maximum likelihood. To construct the likelihood, we need the conditional distribution of an individual observation. Recall that if y is Bernoulli, such that Pr(y = 1) = p and Pr(y = 0) = 1 p, then we can write the density of y as f (y) = py (1 p)1y , y = 0, 1. In the Binary choice model, yi is conditionally Bernoulli with Pr (yi = 1 | xi ) = pi = F (x0 i ) . Thus the conditional density is
1yi i f (yi | xi ) = py i (1 pi ) yi 1y i = F x0 (1 F x0 . i i ) n X i=1 n X i=1
> 0 | xi ) Pr (yi = 1 | xi ) = Pr (yi 0 = Pr xi + ei > 0 | xi = Pr ei > x0 i | xi = 1 F x0 i = F x0 i .
Hence the log-likelihood function is log L() = =
log f (yi | xi ) yi 1y i log F x0 (1 F x0 i i )
n X 0 yi log F x0 = i + (1 yi ) log(1 F xi ) i=1 yi =1
X log F x0 log(1 F x0 i + i ).
yi =0
is the value of which maximizes log L(). Standard errors and test statistics are The MLE computed by asymptotic approximations. Details of such calculations are left to more advanced courses.
18.2
Count Data
exp (i ) k i , k! i = exp(x0 i ).
If y {0, 1, 2, ...}, a typical approach is to employ Poisson regression. This model species that Pr (yi = k | xi ) = k = 0, 1, 2, ...
CHAPTER 18. LIMITED DEPENDENT VARIABLES
324
The conditional density is the Poisson with parameter i . The functional form for i has been picked to ensure that i > 0. The log-likelihood function is log L() =
n X i=1 n X 0 exp(x0 log f (yi | xi ) = i ) + yi xi log(yi !) . i=1
which maximizes log L(). The MLE is the value Since E (yi | xi ) = i = exp(x0 i ) is the conditional mean, this motivates the label Poisson regression. Also observe that the model implies that var (yi | xi ) = i = exp(x0 i ), so the model imposes the restriction that the conditional mean and variance of yi are the same. This may be considered restrictive. A generalization is the negative binomial.
18.3
Censored Data
The idea of censoring is that some data above or below a threshold are mis-reported at the with unbounded support, but we threshold. Thus the model is that there is some latent process yi observe only 0 yi if yi (18.1) yi = <0 . 0 if yi (This is written for the case of the threshold being zero, any known value can substitute.) The observed data yi therefore come from a mixed continuous/discrete distribution. Censored models are typically applied when the data set has a meaningful proportion (say 5% or higher) of data at the boundary of the sample support. The censoring process may be explicit in data collection, or it may be a by-product of economic constraints. An example of a data collection censoring is top-coding of income. In surveys, incomes above a threshold are typically reported at the threshold. The rst censored regression model was developed by Tobin (1958) to explain consumption of durable goods. Tobin observed that for many households, the consumption level (purchases) in a particular period was zero. He proposed the latent variable model
= x0 yi i + ei
ei N(0, 2 ) with the observed variable yi generated by the censoring equation (18.1). This model (now called the Tobit) species that the latent (or ideal) value of consumption may be negative (the household would prefer to sell than buy). All that is reported is that the household purchased zero units of the good. The naive approach to estimate is to regress yi on xi . This does not work because regression | x ) = x0 , and the latter is of interest. Thus OLS will be biased estimates E (yi | xi ) , not E (yi i i for the parameter of interest . [Note: it is still possible to estimate E (yi | xi ) by LS techniques. The Tobit framework postulates that this is not inherently interesting, that the parameter of is dened by an alternative statistical structure.]
iid
CHAPTER 18. LIMITED DEPENDENT VARIABLES
325
Consistent estimation will be achieved by the MLE. To construct the likelihood, observe that the probability of being censored is
< 0 | xi ) Pr (yi = 0 | xi ) = Pr (yi 0 = Pr xi + ei < 0 | xi ei x0 i = Pr < | xi x0 = i .
The conditional density function above zero is normal: y x0 1 i ,
y > 0.
Therefore, the density function for y 0 can be written as
1(y=0) 1(y>0) z x0 x0 1 i i f (y | xi ) = ,
where 1 () is the indicator function. Hence the log-likelihood is a mixture of the probit and the normal: log L( ) =
n X i=1
X x0 yi x0 1 i i + . log log = y =0 y >0 X

i i
log f (yi | xi )
which maximizes log L(). The MLE is the value
18.4
Sample Selection
The problem of sample selection arises when the sample is a non-random selection of potential observations. This occurs when the observed data is systematically dierent from the population of interest. For example, if you ask for volunteers for an experiment, and they wish to extrapolate the eects of the experiment on a general population, you should worry that the people who volunteer may be systematically dierent from the general population. This has great relevance for the evaluation of anti-poverty and job-training programs, where the goal is to assess the eect of training on the general population, not just on the volunteers. A simple sample selection model can be written as the latent model yi = x0 i + e1i Ti = 1 z 0 i + e0i > 0
where 1 () is the indicator function. The dependent variable yi is observed if (and only if) Ti = 1. Else it is unobserved. For example, yi could be a wage, which can be observed only if a person is employed. The equation for Ti is an equation specifying the probability that the person is employed. The model is often completed by specifying that the errors are jointly normal 1 e0i N 0, . e1i 2
CHAPTER 18. LIMITED DEPENDENT VARIABLES It is presumed that we observe {xi , z i , Ti } for all observations. Under the normality assumption, e1i = e0i + vi ,
326
where vi is independent of e0i N(0, 1). A useful fact about the standard normal distribution is that (x) , E (e0i | e0i > x) = (x) = (x) and the function (x) is called the inverse Mills ratio. The naive estimator of is OLS regression of yi on xi for those observations for which yi is available. The problem is that this is equivalent to conditioning on the event {Ti = 1}. However, E (e1i | Ti = 1, z i ) = E e1i | {e0i > z 0 i }, z i 0 = E e0i | {e0i > z 0 i }, z i + E vi | {e0i > z i }, z i = z 0 i , which is non-zero. Thus where Hence e1i = z 0 i + ui ,
E (ui | Ti = 1, z i ) = 0. 0 yi = x0 i + z i + ui (18.2)
is a valid regression equation for the observations for which Ti = 1. Heckman (1979) observed that we could consistently estimate and from this equation, if were known. It is unknown, but also can be consistently estimated by a Probit model for selection. The Heckit estimator is thus calculated as follows Estimate from a Probit, using regressors z i . The binary dependent variable is Ti . , Estimate from OLS of yi on xi and (z 0 ). i
The OLS standard errors will be incorrect, as this is a two-step estimator. They can be corrected using a more complicated formula. Or, alternatively, by viewing the Probit/OLS estimation equations as a large joint GMM problem.
The Heckit estimator is frequently used to deal with problems of sample selection. However, the estimator is built on the assumption of normality, and the estimator can be quite sensitive to this assumption. Some modern econometric research is exploring how to relax the normality assumption. ) does not have much in-sample variation. The estimator can also work quite poorly if (z 0 i This can happen if the Probit equation does not explain much about the selection choice. Another ) can be highly collinear with xi , so the second potential problem is that if z i = xi , then (z 0 i step OLS estimator will not be able to precisely estimate . Based this observation, it is typically recommended to nd a valid exclusion restriction: a variable should be in z i which is not in xi . If ) is not collinear with xi , and hence improve the second stage this is valid, it will ensure that (z 0 i estimators precision.
Chapter 19
Panel Data
A panel is a set of observations on individuals, collected over time. An observation is the pair {yit , xit }, where the i subscript denotes the individual, and the t subscript denotes time. A panel may be balanced : {yit , xit } : t = 1, ..., T ; i = 1, ..., n, or unbalanced : {yit , xit } : For i = 1, ..., n, t = ti , ..., ti .
19.1
Individual-Eects Model
The standard panel data specication is that there is an individual-specic eect which enters linearly in the regression yit = x0 it + ui + eit . The typical maintained assumptions are that the individuals i are mutually independent, that ui and eit are independent, that eit is iid across individuals and time, and that eit is uncorrelated with xit . OLS of yit on xit is called pooled estimation. It is consistent if E (xit ui ) = 0 (19.1)
If this condition fails, then OLS is inconsistent. (19.1) fails if the individual-specic unobserved eect ui is correlated with the observed explanatory variables xit . This is often believed to be plausible if ui is an omitted variable. If (19.1) is true, however, OLS can be improved upon via a GLS technique. In either event, OLS appears a poor estimation choice. Condition (19.1) is called the random eects hypothesis. It is a strong assumption, and most applied researchers try to avoid its use.
19.2
Fixed Eects
This is the most common technique for estimation of non-dynamic linear panel regressions. The motivation is to allow ui to be arbitrary, and have arbitrary correlated with xi . The goal is to eliminate ui from the estimator, and thus achieve invariance. There are several derivations of the estimator. First, let if i = j 1 dij = , 0 else 327
CHAPTER 19. PANEL DATA and di1 . di = . . , din Let .
328
an n 1 dummy vector with a 1 in the i0 th place. u1 . u= . . un Then note that and
ui = d0 i u,
0 yit = x0 it + di u + eit .
(19.2)
Observe that E (eit | xit , di ) = 0, a regressor along with xi . so (19.2) is a valid regression, with di as OLS on (19.2) yields estimator , u . Conventional inference applies. Observe that This is generally consistent. If xit contains an intercept, it will be collinear with di , so the intercept is typically omitted from xit . Any regressor in xit which is constant over time for all individuals (e.g., their gender) will be collinear with di , so will have to be omitted. There are n + k regression parameters, which is quite large as typically n is very large. Computationally, you do not want to actually implement conventional OLS estimation, as the parameter space is too large. OLS estimation of proceeds by the FWL theorem. Stacking the observations together: y = X + Du + e, then by the FWL theorem, = X 0 (I P D ) X 1 X 0 (I P D ) y 1 0 = X 0 X X y , y = y D(D0 D)1 D0 y
where
X = X D(D0 D)1 D0 X . Since the regression of yit on di is a regression onto individual-specic dummies, the predicted value from these regressions is the individual specic mean y i , and the residual is the demean value
yit = yit yi .
is OLS of y on x , the dependent variable and regressors in deviationThe xed eects estimator it it from-mean form.
CHAPTER 19. PANEL DATA Another derivation of the estimator is to take the equation yit = x0 it + ui + eit , and then take individual-specic means by taking the average for the i0 th individual:
ti ti ti 1 X 1 X 1 X yit = x0 + u + eit i Ti t=t Ti t=t it Ti t=t
i i i
329
or y i = x0 i + ui + ei . Subtracting, we nd
0 yit = x it + eit ,
which is free of the individual-eect ui .
19.3
Dynamic Panel Regression

yit = yit1 + x0 it + ui + eit .
A dynamic panel regression has a lagged dependent variable (19.3)
This is a model suitable for studying dynamic behavior of individual agents. Unfortunately, the xed eects estimator is inconsistent, at least if T is held nite as n . This is because the sample mean of yit1 is correlated with that of eit . The standard approach to estimate a dynamic panel is to combine rst-dierencing with IV or GMM. Taking rst-dierences of (19.3) eliminates the individual-specic eect: yit = yit1 + x0 it + eit . However, if eit is iid, then it will be correlated with yit1 :
2 . E (yit1 eit ) = E ((yit1 yit2 ) (eit eit1 )) = E (yit1 eit1 ) = e
(19.4)
So OLS on (19.4) will be inconsistent. But if there are valid instruments, then IV or GMM can be used to estimate the equation. Typically, we use lags of the dependent variable, two periods back, as yt2 is uncorrelated with eit . Thus values of yitk , k 2, are valid instruments. Hence a valid estimator of and is to estimate (19.4) by IV using yt2 as an instrument for yt1 (which is just identied). Alternatively, GMM using yt2 and yt3 as instruments (which is overidentied, but loses a time-series observation). A more sophisticated GMM estimator recognizes that for time-periods later in the sample, there are more instruments available, so the instrument list should be dierent for each equation. This is conveniently organized by the GMM principle, as this enables the moments from the dierent timeperiods to be stacked together to create a list of all the moment conditions. A simple application of GMM yields the parameter estimates and standard errors.
Chapter 20
Nonparametric Density Estimation

20.1 Kernel Density Estimation
d Let X be a random variable with continuous distribution F (x) and density f (x) = dx F (x). , ..., X } While F ( x ) can be estimated by The goal is to estimate f ( x ) from a random sample ( X 1 n (x) = n1 Pn 1 (Xi x) , we cannot dene d F (x) since F (x) is a step function. The the EDF F i=1 dx standard nonparametric method to estimate f (x) is based on smoothing using a kernel. While we are typically interested in estimating the entire function f (x), we can simply focus on the problem where x is a specic xed number, and then see how the method generalizes to estimating the entire function.
Denition 20.1.1 K (u) is a second-order kernel function if it is a symmetric zero-mean density function.
Three common choices for kernels include the Normal 2 u 1 K (u) = exp 2 2 the Epanechnikov K (u) = and the Biweight or Quartic K (u) =
15 16
3 4
1 u2 , |u| 1 0 |u| > 1
2 1 u2 , |u| 1 0 |u| > 1
In practice, the choice between these three rarely makes a meaningful dierence in the estimates. The kernel functions are used to smooth the data. The amount of smoothing is controlled by the bandwidth h > 0. Let 1 u . Kh (u) = K h h be the kernel K rescaled by the bandwidth h. The kernel density estimator of f (x) is X (x) = 1 f Kh (Xi x) . n
i=1 n
330
CHAPTER 20. NONPARAMETRIC DENSITY ESTIMATION
331
This estimator is the average of a set of weights. If a large number of the observations Xi are near (x) is larger. Conversely, if only a few Xi are near x, x, then the weights are relatively large and f then the weights are small and f (x) is small. The bandwidth h controls the meaning of near. (x) is a valid density. That is, f (x) 0 for all x, and Interestingly, f Z X Z n 1 Kh (Xi x) dx f (x)dx = n i=1 n Z 1X = Kh (Xi x) dx n i=1 n Z 1X = K (u) du = 1 n
i=1
where the second-to-last equality makes the change-of-variables u = (Xi x)/h. (x). The mean is We can also calculate the moments of the density f Z n Z 1X xf (x)dx = xKh (Xi x) dx n i=1 n Z 1X = (Xi + uh) K (u) du n i=1 Z Z n n 1X 1X Xi K (u) du + h uK (u) du = n n = 1 n
i=1 n X i=1 i=1
Xi
the sample mean of the Xi , where the second-to-last equality used the change-of-variables u = (Xi x)/h which has Jacobian h. The second moment of the estimated density is Z n Z 1X 2 2 x f (x)dx = x Kh (Xi x) dx n i=1 n Z 1X (Xi + uh)2 K (u) du = n i=1 Z Z n n n X 1 2X 1X 2 2 2 = Xi + Xi h K (u)du + h u K (u) du n n n
i=1 i=1 i=1 n 1X 2 2 Xi + h2 K = n i=1 2 K
where =
(x) is is the variance of the kernel. It follows that the variance of the density f n !2 2 Z Z n X X 1 1 2 2 2 2 (x)dx = x f (x)dx xf Xi + h K Xi n n
i=1 i=1 2 = 2 + h2 K 2 relative to the sample Thus the variance of the estimated density is inated by the factor h2 K moment.
u2 K (u) du
332
20.2
Asymptotic MSE for Kernel Estimates
For xed x and bandwidth h observe that Z Kh (z x) f (z )dz EKh (X x) = Z = Kh (uh) f (x + hu)hdu Z K (u) f (x + hu)du =
The second equality uses the change-of variables u = (z x)/h. The last expression shows that the expected value is an average of f (z ) locally about x. This integral (typically) is not analytically solvable, so we approximate it using a second order Taylor expansion of f (x + hu) in the argument hu about hu = 0, which is valid as h 0. Thus 1 f (x + hu) ' f (x) + f 0 (x)hu + f 00 (x)h2 u2 2 and therefore 1 00 0 2 2 K (u) f (x) + f (x)hu + f (x)h u du EKh (X x) ' 2 Z Z = f (x) K (u) du + f 0 (x)h K (u) udu Z 1 + f 00 (x)h2 K (u) u2 du 2 1 00 2 . = f (x) + f (x)h2 K 2 Z
(x) is then The bias of f (x) f (x) = Bias(x) = Ef 1 1X 2 EKh (Xi x) f (x) = f 00 (x)h2 K . n 2
i=1 n
(x) at x depends on the second derivative f 00 (x). The sharper the derivative, We see that the bias of f (x) smooths data local to Xi = x, so is estimating the greater the bias. Intuitively, the estimator f a smoothed version of f (x). The bias results from this smoothing, and is larger the greater the curvature in f (x). (x). Since it is an average of iid random variables, using We now examine the variance of f rst-order Taylor approximations and the fact that n1 is of smaller order than (nh)1 var (x) = = ' = ' = 1 var (Kh (Xi x)) n 1 1 EKh (Xi x)2 (EKh (Xi x))2 n n Z 1 zx 2 1 K f (z )dz f (x)2 nh2 h n Z 1 K (u)2 f (x + hu) du nh Z f (x) K (u)2 du nh f (x) R(K ) . nh
333
R where R(K ) = K (u)2 du is called the roughness of K. Together, the asymptotic mean-squared error (AMSE) for xed x is the sum of the approximate squared bias and approximate variance 1 f (x) R(K ) 4 . + AM SEh (x) = f 00 (x)2 h4 K 4 nh A global measure of precision is the asymptotic mean integrated squared error (AMISE) Z 4 R(f 00 ) h4 K R(K ) AM ISEh = AM SEh (x)dx = + . (20.1) 4 nh R where R(f 00 ) = (f 00 (x))2 dx is the roughness of f 00 . Notice that the rst term (the squared bias) is increasing in h and the second term (the variance) is decreasing in nh. Thus for the AMISE to decline with n, we need h 0 but nh . That is, h must tend to zero, but at a slower rate than n1 . Equation (20.1) is an asymptotic approximation to the MSE. We dene the asymptotically optimal bandwidth h0 as the value which minimizes this approximate MSE. That is, h0 = argmin AM ISEh
h
It can be found by solving the rst order condition d R(K ) 4 AM ISEh = h3 K R(f 00 ) =0 dh nh2 yielding h0 = R(K ) 4 K R(f 00 ) 1/5
n1/2 .
(20.2)
This solution takes the form h0 = cn1/5 where c is a function of K and f, but not of n. We thus say that the optimal bandwidth is of order O(n1/5 ). Note that this h declines to zero, but at a very slow rate. In practice, how should the bandwidth be selected? This is a dicult problem, and there is a large and continuing literature on the subject. The asymptotically optimal choice given in (20.2) 2 , and R(f 00 ). The rst two are determined by the kernel function. Their depends on R(K ), K values for the three functions introduced in the previous section are given here. R R 2 = u2 K (u) du R(K ) = K (u)2 du K K Gaussian 1 1/(2 ) Epanechnikov 1/5 1/5 Biweight 1/7 5/7 An obvious diculty is that R(f 00 ) is unknown. A classic simple solution proposed by Silverman (1986) has come to be known as the reference bandwidth or Silvermans Rule-of-Thumb. It 5 R(00 ), where is the N(0, 1) distribution and 2 is uses formula (20.2) but replaces R(f 00 ) with an estimate of 2 = var(X ). This choice for h gives an optimal rule when f (x) is normal, and gives a nearly optimal rule when f (x) is close to normal. The downside is that if the density is very far from normal, the rule-of-thumb h can be quite inecient. We can calculate that R(00 ) = 3/ (8 ) . Together with the above table, we nd the reference rules for the three kernel functions introduced earlier. n1/5 Gaussian Kernel: hrule = 1.06 Epanechnikov Kernel: hrule = 2.34 n1/5 Biweight (Quartic) Kernel: hrule = 2.78 n1/5
334
Unless you delve more deeply into kernel estimation methods the rule-of-thumb bandwidth is a good practical bandwidth choice, perhaps adjusted by visual inspection of the resulting estimate (x). There are other approaches, but implementation can be delicate. I now discuss some of these f choices. The plug-in approach is to estimate R(f 00 ) in a rst step, and then plug this estimate into the formula (20.2). This is more treacherous than may rst appear, as the optimal h for estimation of the roughness R(f 00 ) is quite dierent than the optimal h for estimation of f (x). However, there are modern versions of this estimator work well, in particular the iterative method of Sheather and Jones (1991). Another popular choice for selection of h is cross-validation. This works by constructing an estimate of the MISE using leave-one-out estimators. There are some desirable properties of cross-validation bandwidths, but they are also known to converge very slowly to the optimal values. They are also quite ill-behaved when the data has some discretization (as is common in economics), in which case the cross-validation rule can sometimes select very small bandwidths leading to dramatically undersmoothed estimates. Fortunately there are remedies, which are known as smoothed cross-validation which is a close cousin of the bootstrap.
Appendix A
Matrix Algebra
A.1 Notation
Equivalently, a vector a is an element of Euclidean k space, written as a Rk . If k = 1 then a is a scalar. A matrix A is a k r rectangular array of numbers, written as a11 a12 a1r a21 a22 a2r A= . . . . . . . . . ak 1 ak 2 akr
A scalar a is a single number. A vector a is a k 1 list of numbers, typically arranged in a column. We write this as a1 a2 a= . . . ak
By convention aij refers to the element in the i0 th row and j 0 th column of A. If r = 1 then A is a column vector. If k = 1 then A is a row vector. If r = k = 1, then A is a scalar. A standard convention (which we will follow in this text whenever possible) is to denote scalars by lower-case italics (a), vectors by lower-case bold italics (a), and matrices by upper-case bold italics (A). Sometimes a matrix A is denoted by the symbol (aij ). A matrix can be written as a set of column vectors or as a set of row vectors. That is, 1 2 A = a1 a2 ar = . . . k where ai = a1i a2i . . . aki ajr
are column vectors and j =
aj 1 aj 2 335
APPENDIX A. MATRIX ALGEBRA are row vectors. The transpose of a matrix, denoted A0 , is obtained by Thus a11 a21 ak1 a12 a22 ak2 A0 = . . . . . . . . . a1r a2r akr
336
ipping the matrix on its diagonal.
Alternatively, letting B = A0 , then bij = aji . Note that if A is k r, then A0 is r k. If a is a k 1 vector, then a0 is a 1 k row vector. An alternative notation for the transpose of A is A> . A matrix is square if k = r. A square matrix is symmetric if A = A0 , which requires aij = aji . A square matrix is diagonal if the o-diagonal elements are all zero, so that aij = 0 if i 6= j. A square matrix is upper (lower) diagonal if all elements below (above) the diagonal equal zero. An important diagonal matrix is the identity matrix, which has ones on the diagonal. The k k identity matrix is denoted as 1 0 0 0 1 0 Ik = . . . . . . . . . . 0 0 1 A partitioned matrix takes the form A11 A12 A21 A22 A= . . . . . . Ak1 Ak2 A1r A2r . . . Akr
where the Aij denote matrices, vectors and/or scalars.
A.2
Matrix Addition
If the matrices A = (aij ) and B = (bij ) are of the same order, we dene the sum A + B = (aij + bij ) . Matrix addition follows the commutative and associative laws: A+B =B+A A + (B + C ) = (A + B ) + C .
A.3
Matrix Multiplication
If A is k r and c is real, we dene their product as Ac = cA = (aij c) . If a and b are both k 1, then their inner product is a0 b = a1 b1 + a2 b2 + + ak bk =
k X j =1
aj bj .
Note that a0 b = b0 a. We say that two vectors a and b are orthogonal if a0 b = 0.
APPENDIX A. MATRIX ALGEBRA
337
If A is k r and B is r s, so that the number of columns of A equals the number of rows of B , we say that A and B are conformable. In this event the matrix product AB is dened. Writing A as a set of row vectors and B as a set of column vectors (each of length r), then the matrix product is dened as 0 a1 a0 2 AB = . b1 b2 bs . . a0 k =
0 a0 1 b1 a1 b2 0 a0 2 b1 a2 b2 . . . . . . 0 0 ak b1 ak b2
a0 1 bs a0 2 bs . . . a0 k bs
Matrix multiplication is not commutative: in general AB 6= BA. However, it is associative and distributive: A (BC ) = (AB ) C A (B + C ) = AB + AC An alternative way to write the matrix product is to use matrix partitions. For example, A11 A12 B 11 B 12 AB = A21 A22 B 21 B 22 = As another example, B1 B2 . . . Br A11 B 11 + A12 B 21 A11 B 12 + A12 B 22 A21 B 11 + A22 B 21 A21 B 12 + A22 B 22 .
AB =
A1 A2
Ar
= A1 B 1 + A2 B 2 + + Ar B r r X = Aj B j
j =1
An important property of the identity matrix is that if A is k r, then AI r = A and I k A = A. The k r matrix A, r k , is called orthogonal if A0 A = I r .
A.4
Trace
The trace of a k k square matrix A is the sum of its diagonal elements tr (A) =
k X i=1
aii .
APPENDIX A. MATRIX ALGEBRA Some straightforward properties for square matrices A and B and real c are tr (cA) = c tr (A) tr A0 = tr (A) tr (I k ) = k. Also, for k r A and r k B we have tr (AB ) = tr (BA) . Indeed, tr (AB ) = tr =
k X i=1 k X i=1
338
tr (A + B ) = tr (A) + tr (B )
(A.1)
0 a0 1 b1 a1 b2 0 a2 b1 a0 2 b2 . . . . . . 0 0 ak b1 ak b2
a0 1 bk a0 2 bk . . . a0 k bk
a0 i bi b0 i ai
= tr (BA) .
A.5
Rank and Inverse

The rank of the k r matrix (r k ) A=
a1 a2
ar
is the number of linearly independent columns aj , and is written as rank (A) . We say that A has full rank if rank (A) = r. A square k k matrix A is said to be nonsingular if it is has full rank, e.g. rank (A) = k. This means that there is no k 1 c 6= 0 such that Ac = 0. If a square k k matrix A is nonsingular then there exists a unique matrix k k matrix A1 called the inverse of A which satises AA1 = A1 A = I k . For non-singular A and C , some important properties include AA1 = A1 A = I k 1 0 0 1 = A A
Also, if A is an orthogonal matrix, then A1 = A0 .
(AC )1 = C 1 A1 1 1 (A + C )1 = A1 A1 + C 1 C 1 A1 (A + C )1 = A1 A1 + C 1 A1
APPENDIX A. MATRIX ALGEBRA Another useful result for non-singular A is known as the Woodbury matrix identity 1 CDA1 . (A + BCD)1 = A1 A1 BC C + CDA1 BC
339
(A.2)
In particular, for C = 1, B = b and D = b0 for vector b we nd what is known as the Sherman Morrison formula 1 1 1 0 1 A bb0 = A1 + 1 b0 A1 b A bb A . (A.3) The following fact about inverting partitioned matrices is quite useful. A11 A12 A21 A22 1 = A11 A12 A21 A22 =
1 A 112 1 1 A 221 A21 A11 1 1 A 112 A12 A22 1 A 221
(A.4)
1 1 where A112 = A11 A12 A 22 A21 and A221 = A22 A21 A11 A12 . There are alternative algebraic representations for the components. For example, using the Woodbury matrix identity you can show the following alternative expressions 1 1 1 1 A11 = A 11 + A11 A12 A221 A21 A11
1 1 1 1 A22 = A 22 + A22 A21 A112 A12 A22 1 1 A12 = A 11 A12 A221 1 1 A21 = A 22 A21 A112
Even if a matrix A does not possess an inverse, we can still dene the Moore-Penrose generalized inverse A as the matrix which satises AA A = A A AA = A AA is symmetric A A is symmetric For any matrix A, the Moore-Penrose generalized inverse A exists and is unique. For example, if A11 0 A= 0 0
1 and when A 11 exists then
A =
1 0 A 11 0 0
A.6
Determinant
The determinant is a measure of the volume of a square matrix. While the determinant is widely used, its precise denition is rarely needed. However, we present the denition here for completeness. Let A = (aij ) be a general k k matrix . Let = (j1 , ..., jk ) denote a permutation of (1, ..., k) . There are k! such permutations. There is a unique count of the number of inversions of the indices of such permutations (relative to the natural order (1, ..., k) , and let = +1 if this count is even and = 1 if the count is odd. Then the determinant of A is dened as X a1j1 a2j2 akjk . det A =
340
For example, if A is 2 2, then the two permutations of (1, 2) are (1, 2) and (2, 1) , for which (1,2) = 1 and (2,1) = 1. Thus det A = (1,2) a11 a22 + (2,1) a21 a12 = a11 a22 a12 a21 . Some properties include det (A) = det (A0 ) det (cA) = ck det A det (AB ) = (det A) (det B ) det A1 = (det A)1 A B det = (det D) det A BD1 C if det D 6= 0 C D det A 6= 0 if and only if A is nonsingular If A is orthogonal, then det A = 1 If A is triangular (upper or lower), then det A = Qk
i=1 aii
A.7
Eigenvalues
The characteristic equation of a k k square matrix A is det (A I k ) = 0. The left side is a polynomial of degree k in so it has exactly k roots, which are not necessarily distinct and may be real or complex. They are called the latent roots or characteristic roots or eigenvalues of A. If i is an eigenvalue of A, then A i I k is singular so there exists a non-zero vector hi such that (A i I k ) hi = 0. The vector hi is called a latent vector or characteristic vector or eigenvector of A corresponding to i . We now state some useful properties. Let i and hi , i = 1, ..., k denote the k eigenvalues and eigenvectors of a square matrix A. Let be a diagonal matrix with the characteristic roots in the diagonal, and let H = [h1 hk ]. Q det(A) = k i=1 i Pk tr(A) = i=1 i
A is non-singular if and only if all its characteristic roots are non-zero. If A has distinct characteristic roots, there exists a nonsingular matrix P such that A = P 1 P and P AP 1 = . If A is symmetric, then A = H H 0 and H 0 AH = , and the characteristic roots are all real. A = H H 0 is called the spectral decomposition of a matrix. When the eigenvalues of k k A are real they are written in decending order 1 2 k . We also write min (A) = k = min{ } and max (A) = 1 = max{ }.

1 1 1 The characteristic roots of A1 are 1 , 2 , ..., k .
341
The matrix H has the orthonormal properties H 0 H = I and HH 0 = I . H 1 = H 0 and (H 0 )

1
=H
A.8
c0 Ac
Positive Deniteness
We say that a k k symmetric square matrix A is positive semi-denite if for all c 6= 0, 0. This is written as A 0. We say that A is positive denite if for all c 6= 0, c0 Ac > 0. This is written as A > 0. Some properties include: If A = G0 G for some matrix G, then A is positive semi-denite. (For any c 6= 0, c0 Ac = 0 0 where = Gc.) If G has full rank, then A is positive denite. If A is positive denite, then A is non-singular and A1 exists. Furthermore, A1 > 0. A > 0 if and only if it is symmetric and all its characteristic roots are positive. By the spectral decomposition, A = H H 0 where H 0 H = I and is diagonal with nonnegative diagonal elements. All diagonal elements of are strictly positive if (and only if) A > 0. If A > 0 then A1 = H 1 H 0 . If A 0 and rank (A) = r < k then A = H H 0 where A is the Moore-Penrose 1 1 1 generalized inverse, and = diag 1 , 2 , ..., k , 0, ..., 0
If A 0 we can nd a matrix B such that A = BB 0 . We call B a matrix square root of A. The matrix B need not be unique. One way to construct B is to use the spectral decomposition A = H H 0 where is diagonal, and then set B = H 1/2 . There is a unique root root B which is also positive semi-denite B 0.
A square matrix A is idempotent if AA = A. If A is idempotent and symmetric then all its characteristic roots equal either zero or one and is thus positive semi-denite. To see this, note that we can write A = H H 0 where H is orthogonal and contains the r (real) characteristic roots. Then A = AA = H H 0 H H 0 = H 2 H 0 . By the uniqueness of the characteristic roots, we deduce that 2 = and 2 i = i for i = 1, ..., r. Hence they must equal either 0 or 1. It follows that the spectral decomposition of idempotent A takes the form I kr 0 (A.5) H0 A=H 0 0 with H 0 H = I k . Additionally, tr(A) = rank(A). If A is idempotent then I A is also idempotent. One useful fact is that A is idempotent then for any conformable vector c,
0
c (I A) c c c To see this, note that c0 c = c0 Ac + c0 (I A) c.
c0 Ac c0 c
0
(A.6) (A.7)
Since A and I A are idempotent, they are both positive semi-denite, so both c0 Ac and c0 (I A) c are negative. Thus they must satisfy (A.6)-(A.7),.
342
A.9
Matrix Calculus
and
g (x) = x0 Some properties are now summarized.

x x0 x
Let x = (x1 , ..., xk ) be k 1 and g (x) = g (x1 , ..., xk ) : Rk R. The vector derivative is x1 g (x) . . g (x) = . x xk g (x)
x1 g (x)
xk g (x)
(a0 x) =
(x0 a) = a
(Ax) = A (x0 Ax) = (A + A0 ) x (x0 Ax) = A + A0
2 x x0
A.10
Kronecker Products and the Vec Operator
Some important properties are now summarized. These results hold for matrices for which all matrix multiplications are conformable. (A + B ) C = A C + B C (A B ) (C D) = AC BD A (B C ) = (A B ) C (A B )0 = A0 B 0 tr (A B ) = tr (A) tr (B ) (A B )1 = A1 B 1
Let A = (aij ) be an m n matrix and let B be any matrix. The Kronecker product of A and B , denoted A B , is the matrix a11 B a12 B a1n B a21 B a22 B a2n B AB = . . . . . . . . . . am1 B am2 B amn B
Let A = [a1 a2 an ] be m n. The vec of A, denoted by vec (A) , is the mn 1 vector a1 a2 vec (A) = . . . . an
If A is m m and B is n n, det(A B ) = (det (A))n (det (B ))m If A > 0 and B > 0 then A B > 0 vec (ABC ) = (C 0 A) vec (B ) tr (ABCD) = vec (D0 ) (C 0 A) vec (B )
0
343
A.11
Vector and Matrix Norms
The Euclidean norm of an m 1 vector a is 1/2 kak = a0 a m !1/2 X = a2 . i

i=1
The Frobenius norm of an m n matrix A is kAk = kvec (A)k 1/2 = tr A0 A 1/2 n m X X . = a2 ij

i=1 j =1
If an m m matrix A is symmetric with eigenvalues , kAk = m X

=1
= 1, ..., m, then
!1/2
To see this, by the spectral decomposition A = H H 0 with H 0 H = I and = diag{1 , ..., m }, so m !1/2 X 1/2 0 0 1/2 2 = (tr ()) = . (A.8) kAk = tr H H H H
=1
A useful calculation is for any m 1 vectors a and b, using (A.1),
and in particular
1/2 0 1/2 ab = tr ba0 ab0 = b0 ba0 a = kak kbk 0 aa = kak2 . (A.9)
The are other matrix norms. Another norm of frequent use is the spectral norm 1/2 kAkS = max A0 A
where max (B ) denotes the largest eigenvalue of the matrix B .
A.12
Matrix Inequalities
Schwarz Inequality: For any m 1 vectors a and b, 0 a b kak kbk .
(A.10)
Schwarz Matrix Inequality: For any m n matrices A and B, 0 A B kAk kB k .
(A.11)
APPENDIX A. MATRIX ALGEBRA Triangle Inequality: For any m n matrices A and B, kA + B k kAk + kB k .
344
(A.12)
Trace Inequality. For any m m matrices A and B such that A is symmetric and B 0 tr (AB ) max (A) tr (B ) where max (A) is the largest eigenvalue of A. Quadratic Inequality. For any m 1 b and m m symmetric matrix A b0 Ab max (A) b0 b Norm Inequality. For any m m matrices A and B such that A 0 and B 0 kAB k max (A) kB k (A.15) (A.14) (A.13)
where max (A) is the largest eigenvalue of A. Eigenvalue Product Inequaly. For any k k matrices A 0 and B 0, the eigenvalues (AB ) are real and satisfy min (A) min (B ) (AB ) = A1/2 BA1/2 max (A) max (B ) (A.16) (Zhang and Zhang, 2006, Corollary 11) Jensens Inequality. If g () : R R is convex, then for any non-negative weights aj such that Pm j =1 aj = 1, and any real numbers xj m m X X g aj xj aj g (xj ) .
j =1 j =1
(A.17)
In particular, setting aj = 1/m, then
m m X 1 1 X g xj g (xj ) . m m
j =1 j =1
(A.18)
Loves cr Inequality. For r > 0, r X m X m aj cr |aj |r j =1 j =1 (A.19)
where cr = 1 when r 1 and cr = mr1 when r 1. c2 Inequality. For any m 1 vectors a and b,
(a + b)0 (a + b) 2a0 a + 2b0 b
(A.20)
Proof of Schwarz Inequality: First, suppose that kbk = 0. Then b = 0 and both |a0 b| = 0 and 1 0 kak kbk = 0 so the inequality is true. Second, suppose that kbk > 0 and dene c = a b b0 b b a. 0 Since c is a vector, c c 0. Thus 2 0 c0 c = a0 a a0 b / b0 b .
APPENDIX A. MATRIX ALGEBRA Rearranging, this implies that 2 a0 b a0 a b0 b .
345
Taking the square root of each side yields the result.
Proof of Schwarz Matrix Inequality: Partition A = [a1 , ..., an ] and B = [b1 , ..., bn ]. Then by partitioned matrix multiplication, the denition of the matrix Euclidean norm and the Schwarz inequality 0 a b1 a0 b2 1 1 0 0 b A B = a2 b1 a0 2 2 . . .. . . . . . ka1 k kb1 k ka1 k kb2 k ka2 k kb1 k ka2 k kb2 k . . . . . .. . . 1/2 n n X X = kai k2 kbj k2
i=1 j =1
1/2 1/2 m m n X n X X X = a2 kbji k2 ji

i=1 j =1 i=1 j =1
n X
i=1
kai k
!1/2 n X
i=1
kbi k
!1/2
= kAk kB k
Proof of Triangle Inequality: Let a = vec (A) and b = vec (B ) . Then by the denition of the matrix norm and the Schwarz Inequality kA + B k2 = ka + bk2
= a0 a + 2a0 b + b0 b a0 a + 2 a0 b + b0 b = (kak + kbk)2 = (kAk + kB k)2
kak2 + 2 kak kbk + kbk2
Proof of Trace Inequality. By the spectral decomposition for symmetric matices, A = H H 0 where has the eigenvalues j of A on the diagonal and H is orthonormal. Dene C = H 0 BH which has non-negative diagonal elements Cjj since B is positive semi-denite. Then tr (AB ) = tr (C ) =
m X j =1
j Cjj max j
j
m X j =1
Cjj = max (A) tr (C )
where the inequality uses the fact that Cjj 0. But note that tr (C ) = tr H 0 BH = tr HH 0 B = tr (B )
APPENDIX A. MATRIX ALGEBRA since H is orthonormal. Thus tr (AB ) max (A) tr (B ) as stated.
346
Proof of Quadratic Inequality: In the Trace Inequality set B = bb0 and note tr (AB ) = b0 Ab and tr (B ) = b0 b. Proof of Norm Inequality. Using the Trace Inequality kAB k = (tr (BAAB ))1/2 = (tr (AABB ))1/2 (max (AA) tr (BB ))1/2 = max (A) kB k . The nal equality holds since A 0 implies that max (AA) = max (A)2 . Proof of Jensens Inequality (A.17). By the denition of convexity, for any [0, 1] g (x1 + (1 ) x2 ) g (x1 ) + (1 ) g (x2 ) . This implies m m X X g aj xj = g a1 g (x1 ) + (1 a1 )
j =1 j =2 j =2
(A.21)
m m X X a1 g (x1 )+(1 a1 ) b2 g (x2 ) + (1 b2 )g cj xj = a1 g (x1 )+a2 g (x2 )+(1 a1 ) (1b2 )g cj xj

j =2 j =2
where bj = aj /(1 a1 ) and
Pm
j =2 bj
m X a1 g (x1 ) + (1 a1 ) g bj xj .
aj xj 1 a1
= 1. By another application of (A.21) this is bounded by
where cj = bj /(1 b2 ). By repeated application of (A.21) we obtain (A.17).
a rewriting Proof of Loves cr Inequality. For r 1 this is simply of the nite form Jensens Pm r inequality (A.18) with g (u) = u . For r < 1, dene bj = |aj | / j =1 |aj | . The facts that 0 bj 1 r and r < 1 imply bj bj and thus m m X X 1= bj br j
j =1 j =1
which implies
The proof is completed by observing that r r m m X X aj |aj | .

j =1 j =1
r m m X X |aj | |aj |r .
j =1 j =1

2 Proof of c2 Inequality. By the cr inequality, (aj + bj )2 2a2 j + 2bj . Thus
347
(a + b) (a + b) =
m X
(aj + bj )2 a2 j +2
0 m X j =1
= 2a a + 2b b
j =1 m X j =1 0
b2 j
Appendix B
Probability
B.1 Foundations
The set S of all possible outcomes of an experiment is called the sample space for the experiment. Take the simple example of tossing a coin. There are two outcomes, heads and tails, so we can write S = {H, T }. If two coins are tossed in sequence, we can write the four outcomes as S = {HH, HT, T H, T T }. An event A is any collection of possible outcomes of an experiment. An event is a subset of S, including S itself and the null set . Continuing the two coin example, one event is A = {HH, HT }, the event that the rst coin is heads. We say that A and B are disjoint or mutually exclusive if A B = . For example, the sets {HH, HT } and {T H } are disjoint. Furthermore, if the sets A1 , A2 , ... are pairwise disjoint and i=1 Ai = S, then the collection A1 , A2 , ... is called a partition of S. The following are elementary set operations: Union: A B = {x : x A or x B }. Intersection: A B = {x : x A and x B }. / A}. Complement: Ac = {x : x The following are useful properties of set operations. Commutatitivity: A B = B A; A B = B A. Associativity: A (B C ) = (A B ) C ; A (B C ) = (A B ) C. Distributive Laws: A (B C ) = (A B ) (A C ) ; A (B C ) = (A B ) (A C ) . c c c c (A B ) = Ac B c . DeMorgans Laws: (A B ) = A B ; A probability function assigns probabilities (numbers between 0 and 1) to events A in S. This is straightforward when S is countable; when S is uncountable we must be somewhat more careful. A set B is called a sigma algebra (or Borel eld) if B , A B implies Ac B, and A1 , A2 , ... B implies i=1 Ai B . A simple example is {, S } which is known as the trivial sigma algebra. For any sample space S, let B be the smallest sigma algebra which contains all of the open sets in S. When S is countable, B is simply the collection of all subsets of S, including and S. When S is the real line, then B is the collection of all open and closed intervals. We call B the sigma algebra associated with S. We only dene probabilities for events contained in B. We now can give the axiomatic denition of probability. Given S and B, a probability function Pr satises Pr( S ) = 1, Pr(A) 0 for all A B, and if A1 , A2 , ... B are pairwise disjoint, then P A ) = Pr ( i i=1 i=1 Pr(Ai ). Some important properties of the probability function include the following Pr () = 0 Pr(A) 1 Pr (Ac ) = 1 Pr(A) 348
APPENDIX B. PROBABILITY Pr (B Ac ) = Pr(B ) Pr(A B ) Pr (A B ) = Pr(A) + Pr(B ) Pr(A B ) If A B then Pr(A) Pr(B ) Bonferronis Inequality: Pr(A B ) Pr(A) + Pr(B ) 1 Booles Inequality: Pr (A B ) Pr(A) + Pr(B )
349
For some elementary probability models, it is useful to have simple rules to count the number of objects in a set. These counting rules are facilitated by using the binomial coecients which are dened for nonnegative integers n and r, n r, as n n! . = r! (n r)! r When counting the number of objects in a set, there are two important distinctions. Counting may be with replacement or without replacement. Counting may be ordered or unordered. For example, consider a lottery where you pick six numbers from the set 1, 2, ..., 49. This selection is without replacement if you are not allowed to select the same number twice, and is with replacement if this is allowed. Counting is ordered or not depending on whether the sequential order of the numbers is relevant to winning the lottery. Depending on these two distinctions, we have four expressions for the number of objects (possible arrangements) of size r from n objects. Without Replacement Ordered Unordered
n! (n r)! n r
In the lottery example, 49 if counting is unordered and without replacement, the number of potential combinations is 6 = 13, 983, 816. If Pr(B ) > 0 the conditional probability of the event A given the event B is Pr (A | B ) = Pr (A B ) . Pr(B )
With Replacement nr n+r1

r
For any B, the conditional probability function is a valid probability function where S has been replaced by B. Rearranging the denition, we can write Pr(A B ) = Pr (A | B ) Pr(B ) which is often quite useful. We can say that the occurrence of B has no information about the likelihood of event A when Pr (A | B ) = Pr(A), in which case we nd Pr(A B ) = Pr (A) Pr(B ) (B.1) We say that the events A and B are statistically independent when (B.1) holds. Furthermore, we say that the collection of events A1 , ..., Ak are mutually independent when for any subset {Ai : i I }, ! \ Y Pr Ai = Pr (Ai ) .
iI iI
Theorem 3 (Bayes Rule). For any set B and any partition A1 , A2 , ... of the sample space, then for each i = 1, 2, ... Pr (B | Ai ) Pr(Ai ) Pr (Ai | B ) = P j =1 Pr (B | Aj ) Pr(Aj )
APPENDIX B. PROBABILITY
350
B.2
Random Variables
A random variable X is a function from a sample space S into the real line. This induces a new sample space the real line and a new probability function on the real line. Typically, we denote random variables by uppercase letters such as X, and use lower case letters such as x for potential values and realized values. (This is in contrast to the notation adopted for most of the textbook.) For a random variable X we dene its cumulative distribution function (CDF) as F (x) = Pr (X x) . (B.2)
Sometimes we write this as FX (x) to denote that it is the CDF of X. A function F (x) is a CDF if and only if the following three properties hold: 1. limx F (x) = 0 and limx F (x) = 1 2. F (x) is nondecreasing in x 3. F (x) is right-continuous We say that the random variable X is discrete if F (x) is a step function. In the latter case, the range of X consists of a countable set of real numbers 1 , ..., r . The probability function for X takes the form j = 1, ..., r (B.3) Pr (X = j ) = j , Pr where 0 j 1 and j =1 j = 1. We say that the random variable X is continuous if F (x) is continuous in x. In this case Pr(X = ) = 0 for all R so the representation (B.3) is unavailable. Instead, we represent the relative probabilities by the probability density function (PDF) f (x) = so that F (x) = and Z d F (x) dx
x
f (u)du Z
b
Pr (a X b) =
f (u)du.
These expressions only make sense if F (x) is dierentiable. While there are examples of continuous random variables which do not possess a PDF, these cases are unusualRand are typically ignored. A function f (x) is a PDF if and only if f (x) 0 for all x R and f (x)dx = 1.
B.3
Expectation
For any measurable real function g, we dene the mean or expectation Eg (X ) as follows. If X is discrete, r X g (j )j , Eg (X ) =
j =1
and if X is continuous
Eg (X ) = The latter is well dened and nite if Z
g (x)f (x)dx.
|g (x)| f (x)dx < .
(B.4)
APPENDIX B. PROBABILITY If (B.4) does not hold, evaluate I1 =
351
g (x)f (x)dx Z g (x)f (x)dx
g (x)>0
I2 =
g (x)<0
If I1 = and I2 < then we dene Eg (X ) = . If I1 < and I2 = then we dene Eg (X ) = . If both I1 = and I2 = then Eg (X ) is undened. Since E (a + bX ) = a + bEX, we say that expectation is a linear operator. For m > 0, we dene the m0 th moment of X as EX m and the m0 th central moment as E (X EX )m . Two special moments are the mean = EX and variance 2 = E (X )2 = EX 2 2 . We 2 call = the standard deviation of X. We can also write 2 = var(X ). For example, this allows the convenient expression var(a + bX ) = b2 var(X ). The moment generating function (MGF) of X is M () = E exp (X ) . The MGF does not necessarily exist. However, when it does and E |X |m < then dm M ( ) = E (X m ) dm =0 which is why it is called the moment generating function. More generally, the characteristic function (CF) of X is C () = E exp (iX ) where i = 1 is the imaginary unit. The CF always exists, and when E |X |m < dm C () = im E (X m ) . m d =0 The Lp norm, p 1, of the random variable X is kX kp = (E |X |p )1/p .
B.4
Gamma Function
It satises the property
The gamma function is dened for > 0 as Z () = x1 exp (x) .

0
(1 + ) = () so for positive integers n, Special values include and (n) = (n 1)! (1) = 1 1 = 1/2 . 2 Sterlings formula is an expansion for the its logarithm 1 1 1 1 1 log z + + + log () = log(2 ) + 3 2 2 12 360 12605
352
B.5
Common Distributions
For reference, we now list some important discrete distribution function. Bernoulli Pr (X = x) = px (1 p)1x , EX = p var(X ) = p(1 p) Binomial n x Pr (X = x) = p (1 p)nx , x EX = np var(X ) = np(1 p) Geometric Pr (X = x) = p(1 p)x1 , 1 EX = p 1p var(X ) = p2 Multinomial Pr (X1 = x1 , X2 = x2 , ..., Xm = xm ) = n! m px1 px2 px m , x1 !x2 ! xm ! 1 2 x1 + + xm = n; p1 + + pm = 1 EXi = pi x = 1, 2, ...; 0p1 x = 0, 1, ..., n; 0p1 x = 0, 1; 0p1
cov (Xi , Xj ) = npi pj Negative Binomial Pr (X = x) = (r + x) r p (1 p)x1 , x! (r) r (1 p) EX = p r (1 p) var(X ) = p2
var(Xi ) = npi (1 pi )
x = 0, 1, 2, ...;
0p1
Poisson Pr (X = x) = exp () x , x! EX = x = 0, 1, 2, ..., >0
var(X ) = We now list some important continuous distributions.
APPENDIX B. PROBABILITY Beta ( + ) 1 x (1 x) 1 , ()( ) = + var(X ) = ( + + 1) ( + )2 f (x) = Cauchy f (x) = 1 , (1 + x2 ) EX = < x < 0 x 1; > 0, > 0
353
var(X ) = Exponential f (x) =
x 1 exp , EX =
0 x < ;
>0
var(X ) = 2 Logistic f (x) = exp (x) , (1 + exp (x))2 EX = 0 2 3 < x < ;
var(X ) = Lognormal
Pareto
! (log x )2 1 exp , f (x) = 2 2 2x EX = exp + 2 /2 var(X ) = exp 2 + 22 exp 2 + 2 f (x) =
0 x < ;
>0
, x < , > 0, x +1 , >1 EX = 1 2 var(X ) = , >2 ( 1)2 ( 2) Uniform 1 , ba a+b EX = 2 (b a)2 var(X ) = 12 f (x) = axb
>0
APPENDIX B. PROBABILITY Weibull x 1 , 0 x < ; exp f (x) = x 1 1/ EX = 1 + 2 1 2/ 2 var(X ) = 1+ 1+ Gamma f (x) = x 1 1 , x exp () EX = 0 x < ; > 0, > 0 > 0, > 0
354
var(X ) = 2 Chi-Square f (x) = x 1 r/21 , x exp 2 (r/2)2r/2 EX = r 0 x < ; r>0
var(X ) = 2r Normal ! 1 (x )2 f (x) = , exp 2 2 2 EX = var(X ) = 2 Student t r+1 r+1 x2 ( 2 ) 2 f (x) = , 1+ r r r 2 < x < ; r>0 < x < ; < < , 2 > 0
EX = 0 if r > 1 r if r > 2 var(X ) = r2
B.6
Multivariate Random Variables
A pair of bivariate random variables (X, Y ) is a function from the sample space into R2 . The joint CDF of (X, Y ) is F (x, y) = Pr (X x, Y y) . If F is continuous, the joint probability density function is f (x, y ) = For a Borel measurable set A R2 , Pr ((X, Y ) A) = 2 F (x, y ). xy Z Z
f (x, y )dxdy
APPENDIX B. PROBABILITY For any measurable function g (x, y ), Eg (X, Y ) = The marginal distribution of X is FX (x) = Pr(X x) Z
355
g (x, y)f (x, y )dxdy.
= lim F (x, y ) y Z x Z f (x, y)dydx =

so the marginal density of X is fX (x) = Similarly, the marginal density of Y is fY (y ) = Z
d FX (x) = dx
f (x, y )dy.
f (x, y)dx.
The random variables X and Y are dened to be independent if f (x, y) = fX (x)fY (y ). Furthermore, X and Y are independent if and only if there exist functions g (x) and h(y ) such that f (x, y ) = g (x)h(y ). If X and Y are independent, then Z Z E (g (X )h(Y )) = g (x)h(y)f (y, x)dydx Z Z = g (x)h(y)fY (y )fX (x)dydx Z Z = g (x)fX (x)dx h(y)fY (y )dy = Eg (X ) Eh (Y ) . if the expectations exist. For example, if X and Y are independent then E(XY ) = EX EY. Another implication of (B.5) is that if X and Y are independent and Z = X + Y, then MZ () = E exp ( (X + Y )) = E (exp (X ) exp (Y )) = E exp 0 X E exp 0 Y
(B.5)
= MX ()MY (). The covariance between X and Y is
(B.6)
cov(X, Y ) = XY = E ((X EX ) (Y EY )) = EXY EX EY. The correlation between X and Y is corr (X, Y ) = XY = XY . X Y
APPENDIX B. PROBABILITY The Cauchy-Schwarz Inequality implies that |XY | 1.
356
(B.7)
The correlation is a measure of linear dependence, free of units of measurement. If X and Y are independent, then XY = 0 and XY = 0. The reverse, however, is not true. For example, if EX = 0 and EX 3 = 0, then cov(X, X 2 ) = 0. A useful fact is that var (X + Y ) = var(X ) + var(Y ) + 2 cov(X, Y ). An implication is that if X and Y are independent, then var (X + Y ) = var(X ) + var(Y ), the variance of the sum is the sum of the variances. A k 1 random vector X = (X1 , ..., Xk )0 is a function from S to Rk . Let x = (x1 , ..., xk )0 denote a vector in Rk . (In this Appendix, we use bold to denote vectors. Bold capitals X are random vectors and bold lower case x are nonrandom vectors. Again, this is in distinction to the notation used in the bulk of the text) The vector X has the distribution and density functions F (x) = Pr(X x) f (x) =
k F (x). x1 xk
For a measurable function g : Rk Rs , we dene the expectation Z g (x)f (x)dx Eg (X ) =

Rk
where the symbol dx denotes dx1 dxk . In particular, we have the k 1 multivariate mean = EX and k k covariance matrix = E (X ) (X )0 = EXX 0 0
If the elements of X are mutually independent, then is a diagonal matrix and k ! k X X Xi = var (X i ) var
i=1 i=1
B.7
Conditional Distributions and Expectation

f (x, y) fX (x)
The conditional density of Y given X = x is dened as fY |X (y | x) =
357
if fX (x) > 0. One way to derive this expression from the denition of conditional probability is fY |X (y | x) = lim Pr (Y y | x X x + ) y 0 Pr ({Y y} {x X x + }) lim = 0 y Pr(x X x + ) F (x + , y ) F (x, y ) lim = y 0 FX (x + ) FX (x) lim y 0
x F (x + , y )
= =
fX (x 2 xy F (x, y )
+ )
fX (x) f (x, y) . = fX (x) The conditional mean or conditional expectation is the function Z m(x) = E (Y | X = x) = yfY |X (y | x) dy.
The conditional mean m(x) is a function, meaning that when X equals x, then the expected value of Y is m(x). Similarly, we dene the conditional variance of Y given X = x as 2 (x) = var (Y | X = x) = E (Y m(x))2 | X = x = E Y 2 | X = x m(x)2 .
Evaluated at x = X , the conditional mean m(X ) and conditional variance 2 (X ) are random variables, functions of X . We write this as E(Y | X ) = m(X ) and var (Y | X ) = 2 (X ). For example, if E (Y | X = x) = + 0 x, then E (Y | X ) = + 0 X , a transformation of X . The following are important facts about conditional expectations. Simple Law of Iterated Expectations: E (E (Y | X )) = E (Y ) Proof : E (E (Y | X )) = E (m(X )) Z m(x)fX (x)dx = Z Z = yfY |X (y | x) fX (x)dydx Z Z = yf (y, x) dydx = E(Y ).

(B.8)
Law of Iterated Expectations: E (E (Y | X , Z ) | X ) = E (Y | X ) (B.9)
APPENDIX B. PROBABILITY Conditioning Theorem. For any function g (x), E (g (X )Y | X ) = g (X ) E (Y | X ) Proof : Let h(x) = E (g (X )Y | X = x) Z g (x)yfY |X (y | x) dy = Z = g (x) yfY |X (y | x) dy = g (x)m(x)
358
(B.10)
where m(x) = E (Y | X = x) . Thus h(X ) = g (X )m(X ), which is the same as E (g (X )Y | X ) = g (X ) E (Y | X ) .
B.8
Transformations
Suppose that X Rk with continuous distribution function FX (x) and density fX (x). Let Y = g(X ) where g(x) : Rk Rk is one-to-one, dierentiable, and invertible. Let h(y ) denote the inverse of g (x). The Jacobian is h(y ) . J (y ) = det y0 Consider the univariate case k = 1. If g (x) is an increasing function, then g (X ) Y if and only if X h(Y ), so the distribution function of Y is FY (y ) = Pr (g (X ) y) = FX (h(Y )) . = Pr (X h(Y ))
Taking the derivative, the density of Y is fY (y ) = d d FY (y ) = fX (h(Y )) h(y ). dy dy
If g (x) is a decreasing function, then g (X ) Y if and only if X h(Y ), so FY (y) = Pr (g (X ) y ) = 1 Pr (X h(Y )) = 1 FX (h(Y )) and the density of Y is fY (y ) = fX (h(Y )) We can write these two cases jointly as fY (y) = fX (h(Y )) |J (y )| . (B.11) d h(y ). dy
This is known as the change-of-variables formula. This same formula (B.11) holds for k > 1, but its justication requires deeper results from analysis. As one example, take the case X U [0, 1] and Y = log(X ). Here, g (x) = log(x) and h(y ) = exp(y ) so the Jacobian is J (y ) = exp(y ). As the range of X is [0, 1], that for Y is [0,). Since fX (x) = 1 for 0 x 1 (B.11) shows that fY (y ) = exp(y), an exponential density. 0 y ,
359
B.9
Normal and Related Distributions

2 x 1 , (x) = exp 2 2
The standard normal density is < x < .
It is conventional to write X N (0, 1) , and to denote the standard normal density function by (x) and its distribution function by (x). The latter has no closed-form solution. The normal density has all moments nite. Since it is symmetric about zero all odd moments are zero. By iterated integration by parts, we can also show that EX 2 = 1 and EX 4 = 3. In fact, for any positive integer m, EX 2m = (2m 1)!! = (2m 1) (2m 3) 1. Thus EX 4 = 3, EX 6 = 15, EX 8 = 105, and EX 10 = 945. If Z is standard normal and X = + Z, then using the change-of-variables formula, X has density ! (x )2 1 exp , < x < . f (x) = 2 2 2 which is the univariate normal density mean and variance of the distribution are and . The 2 , and it is conventional to write X N , 2 . For x Rk , the multivariate normal density is 1 (x )0 1 (x ) , x Rk . f (x) = exp k/2 1/2 2 (2 ) det () The mean and covariance matrix of the distribution are and , and it is conventional to write X N (, ). The MGF and CF of the multivariate normal are exp 0 + 0 /2 and exp i0 0 /2 , respectively. If X Rk is multivariate normal and the elements of X are mutually uncorrelated, then 2 } is a diagonal matrix. In this case the density function can be written as = diag{j
2 + + (x )2 / 2 (x1 1 )2 /1 k k k exp f (x) = k/2 2 (2 ) 1 k ! k Y (xj j )2 1 exp = 2 1/2 2j j j =1 (2 )
!!
which is the product of marginal univariate normal densities. This shows that if X is multivariate normal with uncorrelated elements, then they are mutually independent. Theorem B.9.1 If X N (, ) and Y = a + BX with B an invertible matrix, then Y N (a + B , B B 0 ) . Theorem B.9.2 Let X N (0, I r ) . Then Q = X 0 X is distributed chi-square with r degrees of freedom, written 2 r. Theorem B.9.3 If Z N (0, A) with A > 0, q q, then Z 0 A1 Z 2 q.
p Theorem B.9.4 Let Z N (0, 1) and Q 2 r be independent. Then Tr = Z/ Q/r is distributed as students t with r degrees of freedom.
360
Proof of Theorem B.9.1. By the change-of-variables formula, the density of Y = a + BX is ! 1 (y Y )0 1 Y (y Y ) , y Rk . exp f (y ) = k/2 1/2 2 (2 ) det (Y ) where Y = a+B and Y = B B 0 , where we used the fact that det (B B 0 )
1/2
= det ()1/2 det (B ) .
Proof of Theorem B.9.2. First, suppose a random variable Q is distributed chi-square with r degrees of freedom. It has the MGF Z 1 r xr/21 exp (tx) exp (x/2) dy = (1 2t)r/2 E exp (tQ) = r/2 2 0 2 R where the second equality uses the fact that 0 y a1 exp (by) dy = ba (a), which can be found by applying change-of-variables to the gamma function. Our goal is to calculate the MGF of r/2 will establish that Q 2 Q = X 0 X and show that it equals (1 2t) r. Pr , which 0 2 Note that we can write Q = X X = j =1 Zj where the Zj are independent N (0, 1) . The 2 is distribution of each of the Zj 2 Pr Zj y = 2 Pr (0 Zj y) 2 Z y 1 x exp dx =2 2 2 0 Z y s 1 1/2 = s exp ds 1 1/2 2 0 2 2 2 is using the changeof-variables s = x2 and the fact 1 . Thus the density of Zj 2 = f1 (x) = 1 1
2
Proof of Theorem B.9.3. The fact that A > 0 means that we can write A = CC 0 where C is non-singular. Then A1 = C 10 C 1 and C 1 Z N 0, C 1 AC 10 = N 0, C 1 CC 0 C 10 = N (0, I q ) . Thus 0 Z 0 A1 Z = Z 0 C 10 C 1 Z = C 1 Z C 1 Z 2 q.
1/2 2 which is the 2 . 1 and by our above calculation has the MGF of E exp tZj = (1 2t) Pr 2 2 Since the Zj are mutually independent, (B.6) implies that the MGF of Q = j =1 Zj is ir h (1 2t)1/2 = (1 2t)r/2 , which is the MGF of the 2 r density as desired.
x 1/2 x exp 2 21/2
Proof of Theorem B.9.4. Using the simple law of iterated expectations, Tr has distribution
APPENDIX B. PROBABILITY function
361
! Z x F (x) = Pr p Q/r ( r ) Q =E Zx r !# " r Q |Q = E Pr Z x r r ! Q = E x r
Thus its density is r ! Q d f (x) = E x dx r r !r ! Q Q =E x r r ! r Z qx2 1 q 1 r exp = q r/21 exp (q/2) dq 2r r 2 2r/2 2 0 r+1 r+1 x2 ( 2 ) 2 1+ = r r r 2
which is that of the student t with r degrees of freedom.
B.10
Inequalities
Jensens Inequality. If g () : Rm R is convex, then for any random vector x for which E kxk < and E |g (x)| < , g (E(x)) E (g (x)) . (B.12)
Conditional Jensens Inequality. If g () : Rm R is convex, then for any random vectors (y , x) for which E ky k < and E kg (y )k < , g (E(y | x)) E (g (y ) | x) . Conditional Expectation Inequality. For any r such that E |y|r < , then E |E(y | x)|r E |y |r < . Expectation Inequality. For any random matrix Y for which E kY k < , kE(Y )k E kY k . (B.15) (B.14) (B.13)
1 Hlders Inequality. If p > 1 and q > 1 and 1 p + q = 1, then for any random m n matrices X and Y, (B.16) E X 0 Y (E kX kp )1/p (E kY kq )1/q .
APPENDIX B. PROBABILITY Cauchy-Schwarz Inequality. For any random m n matrices X and Y, 1/2 1/2 E kY k2 . E X 0 Y E kX k2
362
(B.17)
Matrix Cauchy-Schwarz Inequality. Tripathi (1999). For any random x Rm and y R , Eyx0 Exx0 Exy 0 Eyy 0 (B.18) Minkowskis Inequality. For any random m n matrices X and Y, (E kX + Y kp )1/p (E kX kp )1/p + (E kY kp )1/p Liapunovs Inequality. For any random m n matrix X and 1 r p, (E kX kr )1/r (E kX kp )1/p (B.20) (B.19)
Markovs Inequality (standard form). For any random vector x and non-negative function g (x) 0, (B.21) Pr(g (x) > ) 1 Eg (x). Markovs Inequality (strong form). For any random vector x and non-negative function g (x) 0, (B.22) Pr(g (x) > ) 1 E (g (x) 1 (g (x) > )) . Chebyshevs Inequality. For any random variable x, Pr(|x Ex| > ) var (x) . 2 (B.23)
Proof of Jensens Inequality (B.12). Since g (u) is convex, at any point u there is a nonempty set of subderivatives (linear surfaces touching g (u) at u but lying below g (u) for all u). Let a + b0 u be a subderivative of g (u) at u = Ex. Then for all u, g (u) a + b0 u yet g (Ex) = a + b0 Ex. Applying expectations, Eg (x) a + b0 Ex = g (Ex), as stated. Proof of Conditional Jensens Inequality. The same as the proof of (B.12), but using conditional expectations. The conditional expectations exist since E ky k < and E kg (y )k < . Proof of Conditional Expectation Inequality. As the function |u|r is convex for r 1, the Conditional Jensens inequality implies |E(y | x)|r E (|y |r | x) . Taking unconditional expectations and the law of iterated expectations, we obtain E |E(y | x)|r EE (|y|r | x) = E |y |r < as required.
Proof of Expectation Inequality. By the Triangle inequality, for [0, 1], kU 1 + (1 )U 2 k kU 1 k + (1 ) kU 2 k
363
which shows that the matrix norm g (U ) = kU k is convex. Applying Jensens Inequality (B.12) we nd (B.15).
1 Proof of Hlders Inequality. Since 1 p + q = 1 an application of Jensens Inequality (A.17) shows that for any real a and b 1 1 1 1 exp a + b exp (a) + exp (b) . p q p q
Setting u = exp (a) and v = exp (b) this implies u1/p v1/q u v + p q
and this inequality holds for any u > 0 and v > 0. Set u = kX kp /E kX kp and v = kY kq /E kY kq . Note that Eu = Ev = 1. By the matrix Schwarz Inequality (A.11), kX 0 Y k kX k kY k. Thus (E kX k )
p 1/p
E kX 0 Y k
(E kY k )
q 1/q
(E kX kp )1/p (E kY kq )1/q = E u1/p v 1/q u v + E p q 1 1 = + p q = 1,
E (kX k kY k)
which is (B.16).
Proof of Cauchy-Schwarz Inequality. Special case of Hlders with p = q = 2. Proof of Matrix Cauchy-Schwarz Inequality. Dene e = y (Eyx0 ) (Exx0 ) x. Note that Eee0 0 is positive semi-denite. We can calculate that Eee0 = Eyy 0 Eyx0 Exx0 Exy 0 . Proof of Liapunovs Inequality. The function g (u) = up/r is convex for u > 0 since p r. Set u = kX kr . By Jensens inequality, g (Eu) Eg (u) or Raising both sides to the power 1/p yields (E kX kr )1/r (E kX kp )1/p as claimed. (E kX kr )p/r E (kX kr )p/r = E kX kp .
Since the left-hand-side is positive semi-denite, so is the right-hand-side, which means Eyy 0 (Eyx0 ) (Exx0 ) Exy 0 as stated.
Proof of Minkowskis Inequality. Note that by rewriting, using the triangle inequality (A.12), and then Hlders Inequality to the two expectations E kX + Y kp = E kX + Y k kX + Y kp1 E kX k kX + Y kp1 + E kY k kX + Y kp1 1/q (E kX kp )1/p E kX + Y kq(p1) 1/q + (E kY kp )1/p E kX + Y kq(p1) = (E kX kp )1/p + (E kY kp )1/p E (kX + Y kp )(p1)/p
364
where the second equality picks q to satisfy 1/p+1/q = 1, and the nal equality uses this fact to make the substitution q = p/(p 1) and then collects terms. Dividing both sides by E (kX + Y kp )(p1)/p , we obtain (B.19). Proof of Markovs Inequality. Let F denote the distribution function of x. Then Z dF (u) Pr (g (x) ) = Z
{g (u)} {g (u)}
= 1
= 1 E (g (x) 1 (g (x) > )) the inequality using the region of integration {g (u) > }. This establishes the strong form (B.22). Since 1 (g (x) > ) 1, the nal expression is less than 1 E (g (x)) , establishing the standard form (B.21). Proof of Chebyshevs Dene y = (x Ex)2 and note that Ey = var (x) . The events Inequality. {|x Ex| > } and y > 2 are equal, so by an application Markovs inequality we nd Pr(|x Ex| > ) = Pr(y > 2 ) 2 E (y ) = 2 var (x) as stated.
g (u) dF (u)
1 (g (u) > ) g (u)dF (u)
B.11
Maximum Likelihood
In this section we provide a brief review of the asymptotic theory of maximum likelihood estimation. When the density of y i is f (y | ) where F is a known distribution function and is an unknown m 1 vector, we say that the distribution is parametric and that is the parameter of the distribution F. The space is the set of permissible value for . In this setting the method of maximum likelihood is an appropriate technique for estimation and inference on . We let denote a generic value of the parameter and let 0 denote its true value. The joint density of a random sample (y 1 , ..., y n ) is
n Y f (y i | ) . fn (y 1 , ..., y n | ) = i=1
The likelihood of the sample is this joint density evaluated at the observed sample values, viewed as a function of . The log-likelihood function is its natural logarithm log L() =
n X i=1
log f (y i | ) .
The likelihood score is the derivative of the log-likelihood, evaluated at the true parameter value. log f (y i | 0 ) . Si = We also dene the Hessian 2 log f (y i | 0 ) (B.24) H = E 0
APPENDIX B. PROBABILITY and the outer product matrix = E SiS0 i .
365
(B.25)
We now present three important features of the likelihood.
Theorem B.11.1 E log f (y | ) =0 =0 ES i = 0 and H=I (B.28) (B.26) (B.27)
The matrix I is called the information, and the equality (B.28) is called the information matrix equality. is the parameter value which maximizes the The maximum likelihood estimator (MLE) likelihood (equivalently, which maximizes the log-likelihood). We can write this as = argmax log L().
(B.29)
as a function of the data, but these In some simple cases, we can nd an explicit expression for cases are rare. More typically, the MLE must be found by numerical methods. is a natural estimator for the parameter observe that the To understand why the MLE standardized log-likelihood is a sample average and an estimator of E log f (y i | ) : 1 1X p log L() = log f (y i | ) E log f (y i | ) . n n
i=1 n
maximizes the left-hand-side, we can see that it is an estimator of the maximizer of As the MLE the right-hand-side. The rst-order condition for the latter problem is 0= E log f (y i | )
is an estimator of 0 . In. fact, under which holds at = 0 by (B.26). This suggests that p is consistent, 0 as n . Furthermore, we can derive conventional regularity conditions, its asymptotic distribution. n 0
Theorem B.11.2 Under N 0, I 1 .
regularity
conditions,
We omit the regularity conditions for Theorem B.11.2, but the result holds quite broadly for models which are smooth functions of the parameters. Theorem B.11.2 gives the general form for the asymptotic distribution of the MLE. A famous result shows that the asymptotic variance is the smallest possible.
APPENDIX B. PROBABILITY e is an unbiased regTheorem B.11.3 Cramer-Rao Lower Bound. If e ular estimator of , then var() (nI ) .
366
all (regular) estimators is I 1 . An estimator is called asymptotically ecient if its asymptotic variance equals this lower bound. Theorem B.11.2 shows that the MLE has this asymptotic variance, and is thus asymptotically ecient.
The Cramer-Rao Theorem shows that the nite sample variance of an unbiased estimator is bounded below by (nI )1 . This means that the asymptotic variance of the standardized estimator e n 0 is bounded below by I 1 . In other words, the best possible asymptotic variance among
Theorem B.11.4 The MLE is asymptotically ecient in the sense that its asymptotic variance equals the Cramer-Rao Lower Bound.
b is an asymptotically ecient estimator for since it where G = g (0 )0 . By Theorem B.11.4, 0 1 is the MLE. The asymptotic variance G I G is the Cramer-Rao lower bound for estimation of . Theorem B.11.5 The Cramer-Rao lower bound for = g () is G0 I 1 G, b = g ( b) is asymptotically ecient. and the MLE Proof of Theorem B.11.1. To see (B.26), Z E log f (y | ) log f (y | ) f (y | 0 ) dy = =0 =0 Z f (y | 0 ) f (y | ) dy = f (y | ) =0 Z f (y | ) dy = =0 1 = = 0. =0 E log f (y | 0 ) = E log f (y | 0 ) = 0.
Theorem B.11.4 gives a strong endorsement for the MLE in parametric models. b = g ( b). Finally, consider functions of parameters. If = g () then the MLE of is This is because maximization (e.g. (B.29)) is unaected by parameterization and transformation. Applying the Delta Method to Theorem B.11.2 we conclude that d b ' G0 n b n N 0, G0 I 1 G (B.30)
Equation (B.27) follows by exchanging integration and dierentiation
APPENDIX B. PROBABILITY Similarly, we can show that E By direction computation, 2 log f (y | 0 ) = 0 = Taking expectations yields (B.28).
2 f 0 2 f 0
367 !
2 f 0
(y | 0 ) f (y | 0 )
= 0.
(y | 0 ) f (y | 0 )
(y | 0 ) f (y | 0 )0 f (y | 0 )2
(y | 0 ) log f (y | 0 ) log f (y | 0 )0 . f (y | 0 )
Proof of Theorem B.11.2 Taking the rst-order condition for maximization of log L(), and making a rst-order Taylor series expansion, log L() 0= = n X log f y i | = =
i=1 n X i=1
X 2 log f (y i | 0 ) + log f ( y | ) n 0 , i 0
n i=1
and 0 . (Technically, the specic value of n varies by where n lies on a line segment joining row in this expansion.) Rewriting this equation, we nd 0 =
n X i=1
!1 n ! X 2 log f (y i | n ) Si 0 i=1
where S i are the likelihood scores. Since the score S i is mean-zero (B.27) with covariance matrix (equation B.25) an application of the CLT yields 1 X d S i N (0, ) . n
i=1 n
The analysis of the sample Hessian is somewhat more complicated due to the presence of n . p 2 log f (y i , ) . If it is continuous in , then since n 0 it follows that Let H() = 0 H(n ) H and so
p
n n 2 1X 1 X 2 log f (y i , n ) = log f (y i , n ) H(n ) + H(n ) n n 0 0 i=1 i=1 H

p
the nal equality using Theorem B.11.1 .
by an application of a uniform WLLN. (By uniform, we mean that the WLLN holds uniformly over the parameter value. This requires the second derivative to be a smooth function of the parameter.) Together, d 0 n H1 N (0, ) = N 0, H1 H1 = N 0, I 1 ,
APPENDIX B. PROBABILITY Proof of Theorem B.11.3. Let Y = (y 1 , ..., y n ) be the sample, and set S= X log fn (Y , 0 ) = Si
i=1 n
368
e= e (Y ) as a which by Theorem (B.11.1) has mean zero and variance nI . Write the estimator e function of the data. Since is unbiased for any , Z e e (Y ) f (Y , ) dY . = E = Dierentiating with respect to and evaluating at 0 yields Z e (Y ) f (Y , ) dY I= 0 Z e (Y ) log f (Y , ) f (Y , 0 ) dY = 0 eS 0 =E e 0 S 0 =E
the nal equality since E (S ) = 0 e 0 S 0 = I , and var (S ) = E (SS 0 ) = By the matrix Cauchy-Schwarz inequality (B.18), E nI , 0 e =E e 0 e 0 var 0 0 0 e e E 0 S E SS E S 0 = E SS 0 = (nI ) as stated.
Appendix C
Numerical Optimization
Many econometric estimators are dened by an optimization problem of the form = argmin Q()
(C.1)
where the parameter is Rm and the criterion function is Q() : R. For example NLLS, GLS, MLE and GMM estimators take this form. In most cases, Q() can be computed is not available in closed form. In this case, numerical methods are required to for given , but . obtain
C.1
Grid Search
Many optimization problems are either one dimensional (m = 1) or involve one-dimensional optimization as a sub-problem (for example, a line search). In this context grid search may be employed. Grid Search. Let = [a, b] be an interval. Pick some > 0 and set G = (b a)/ to be the number of gridpoints. Construct an equally spaced grid on the region [a, b] with G gridpoints, which is {(j ) = a + j (b a)/G : j = 0, ..., G}. At each point evaluate the criterion function and nd the gridpoint which yields the smallest value of the criterion, which is ( j) where j = j) is the gridpoint estimate of . If the grid is su argmin0j G Q((j )). This value ( ne to ciently j) . capture small oscillations in Q(), the approximation error is bounded by , that is, ( Plots of Q((j )) against (j ) can help diagnose errors in grid selection. This method is quite robust but potentially costly. Two-Step Grid Search. The gridsearch method can be rened by a two-step execution. For an error bound of pick G so that G2 = (b a)/ For the rst step dene an equally spaced grid on the region [a, b] with G gridpoints, which is {(j ) = a + j (b a)/G : j = 0, ..., G}. At each point evaluate the criterion function and let j = argmin0j G Q((j )). For the second step dene an equally spaced grid on [( j 1), ( j + 1)] with G gridpoints, which is {0 (k ) = 0 2 ( j 1) + 2k(b a)/G : k = 0, ..., G}. Let k = argmin0kG Q( (k)). The estimate of is . The advantage of the two-step method over a one-step grid search is that the number of k p function evaluations has been reduced from (b a)/ to 2 (b a)/ which can be substantial. The which disadvantage is that if the function Q() is irregular, the rst-step grid may not bracket thus would be missed.
C.2
Gradient Methods
Gradient Methods are iterative methods which produce a sequence i : i = 1, 2, ... which . All require the choice of a starting value 1 , and all require the are designed to converge to 369
APPENDIX C. NUMERICAL OPTIMIZATION computation of the gradient of Q() g () = and some require the Hessian Q()
370
2 Q(). 0 If the functions g() and H() are not analytically available, they can be calculated numerically. Take the j 0 th element of g (). Let j be the j 0 th unit vector (zeros everywhere except for a one in the j 0 th row). Then for small H() = gj () ' Similarly, Q( + j + k ) Q( + k ) Q( + j ) + Q() 2 In many cases, numerical derivatives can work well but can be computationally costly relative to analytic derivatives. In some cases, however, numerical derivatives can be quite unstable. Most gradient methods are a variant of Newtons method which is based on a quadratic approximation. By a Taylors expansion for close to ) ' g() + H() 0 = g ( gjk () ' which implies = H()1 g (). This suggests the iteration rule i+1 = i H(i )1 g (i ). where One problem with Newtons method is that it will send the iterations in the wrong direction if H(i ) is not positive denite. One modication to prevent this possibility is quadratic hill-climbing which sets i+1 = i (H(i ) + i I m )1 g (i ). where i is set just above the smallest eigenvalue of H (i ) if H () is not positive denite. Another productive modication is to add a scalar steplength i . In this case the iteration rule takes the form (C.2) i+1 = i Di g i i Q( + j ) Q() .
where gi = g (i ) and Di = H(i )1 for Newtons method and Di = (H(i ) + i I m )1 for quadratic hill-climbing. Allowing the steplength to be a free parameter allows for a line search, a one-dimensional optimization. To pick i write the criterion function as a function of Q() = Q(i + Di g i )
a one-dimensional optimization problem. There are two common methods to perform a line search. A quadratic approximation evaluates the rst and second derivatives of Q() with respect to , and picks i as the value minimizing this approximation. The half-step method considers the sequence = 1, 1/2, 1/4, 1/8, ... . Each value in the sequence is considered and the criterion Q(i + Di g i ) evaluated. If the criterion has improved over Q(i ), use this value, otherwise move to the next element in the sequence.
APPENDIX C. NUMERICAL OPTIMIZATION
371
Newtons method does not perform well if Q() is irregular, and it can be quite computationally costly if H () is not analytically available. These problems have motivated alternative choices for the weight matrix Di . These methods are called Quasi-Newton methods. Two popular methods are do to Davidson-Fletcher-Powell (DFP) and Broyden-Fletcher-Goldfarb-Shanno (BFGS). Let g i = gi g i1
i = i i1 and . The DFP method sets Di = Di1 + The BFGS methods sets Di = Di1 + i 0 i 0 i g 0 Di1 gi 0 i i i 0 i D i1 g D g + + . i 1 i i 0 0 0 2 0 i gi i g i i g i i gi i 0 Di1 gi g 0 i i D i1 + . 0 0 g i Di1 g i i g i
For any of the gradient methods, the iterations continue until the sequence has converged in some sense. This can be dened by examining whether |i i1 | , |Q (i ) Q (i1 )| or |g (i )| has become small.
C.3
Derivative-Free Methods
All gradient methods can be quite poor in locating the global minimum when Q() has several local minima. Furthermore, the methods are not well dened when Q() is non-dierentiable. In these cases, alternative optimization methods are required. One example is the simplex method of Nelder-Mead (1965). A more recent innovation is the method of simulated annealing (SA). For a review see Goe, Ferrier, and Rodgers (1994). The SA method is a sophisticated random search. Like the gradient methods, it relies on an iterative sequence. At each iteration, a random variable is drawn and added to the current value of the parameter. If the resulting criterion is decreased, this new value is accepted. If the criterion is increased, it may still be accepted depending on the extent of the increase and another randomization. The latter property is needed to keep the algorithm from selecting a local minimum. As the iterations continue, the variance of the random innovations is shrunk. The SA algorithm stops when a large number of iterations is unable to improve the criterion. The SA method has been found to be successful at locating global minima. The downside is that it can take considerable computer time to execute.
Bibliography
[1] Abadir, Karim M. and Jan R. Magnus (2005): Matrix Algebra, Cambridge University Press. [2] Aitken, A.C. (1935): On least squares and linear combinations of observations, Proceedings of the Royal Statistical Society, 55, 42-48. [3] Akaike, H. (1973): Information theory and an extension of the maximum likelihood principle. In B. Petroc and F. Csake, eds., Second International Symposium on Information Theory. [4] Anderson, T.W. and H. Rubin (1949): Estimation of the parameters of a single equation in a complete system of stochastic equations, The Annals of Mathematical Statistics, 20, 46-63. [5] Andrews, Donald W. K. (1988): Laws of large numbers for dependent non-identically distributed random variables, Econometric Theory, 4, 458-467. [6] Andrews, Donald W. K. (1991), Asymptotic normality of series estimators for nonparameric and semiparametric regression models, Econometrica, 59, 307-345. [7] Andrews, Donald W. K. (1993), Tests for parameter instability and structural change with unknown change point, Econometrica, 61, 821-8516. [8] Andrews, Donald W. K. and Moshe Buchinsky: (2000): A three-step method for choosing the number of bootstrap replications, Econometrica, 68, 23-51. [9] Andrews, Donald W. K. and Werner Ploberger (1994): Optimal tests when a nuisance parameter is present only under the alternative, Econometrica, 62, 1383-1414. [10] Ash, Robert B. (1972): Real Analysis and Probability, Academic Press. [11] Basmann, R. L. (1957): A generalized classical method of linear estimation of coecients in a structural equation, Econometrica, 25, 77-83. [12] Bekker, P.A. (1994): Alternative approximations to the distributions of instrumental variable estimators, Econometrica, 62, 657-681. [13] Billingsley, Patrick (1968): Convergence of Probability Measures. New York: Wiley. [14] Billingsley, Patrick (1995): Probability and Measure, 3rd Edition, New York: Wiley. [15] Bose, A. (1988): Edgeworth correction by bootstrap in autoregressions, Annals of Statistics, 16, 1709-1722. [16] Box, George E. P. and Dennis R. Cox, (1964). An analysis of transformations, Journal of the Royal Statistical Society, Series B, 26, 211-252. [17] Breusch, T.S. and A.R. Pagan (1979): The Lagrange multiplier test and its application to model specication in econometrics, Review of Economic Studies, 47, 239-253. 372
BIBLIOGRAPHY
373
[18] Brown, B. W. and Whitney K. Newey (2002): GMM, ecient bootstrapping, and improved inference , Journal of Business and Economic Statistics. [19] Card, David (1995): Using geographic variation in college proximity to estimate the return to schooling, in Aspects of Labor Market Behavior: Essays in Honour of John Vanderkamp, L.N. Christodes, E.K. Grant, and R. Swidinsky, editors. Toronto: University of Toronto Press. [20] Carlstein, E. (1986): The use of subseries methods for estimating the variance of a general statistic from a stationary time series, Annals of Statistics, 14, 1171-1179. [21] Casella, George and Roger L. Berger (2002): Statistical Inference, 2nd Edition, Duxbury Press. [22] Chamberlain, Gary (1987): Asymptotic eciency in estimation with conditional moment restrictions, Journal of Econometrics, 34, 305-334. [23] Choi, In and Peter C.B. Phillips (1992): Asymptotic and nite sample distribution theory for IV estimators and tests in partially identied structural equations, Journal of Econometrics, 51, 113-150. [24] Chow, G.C. (1960): Tests of equality between sets of coecients in two linear regressions, Econometrica, 28, 591-603. [25] Cragg, John (1992): Quasi-Aitken Estimation for Heterskedasticity of Unknown Form" Journal of Econometrics, 54, 179-201. [26] Davidson, James (1994): Stochastic Limit Theory: An Introduction for Econometricians. Oxford: Oxford University Press. [27] Davison, A.C. and D.V. Hinkley (1997): Bootstrap Methods and their Application. Cambridge University Press. [28] Dickey, D.A. and W.A. Fuller (1979): Distribution of the estimators for autoregressive time series with a unit root, Journal of the American Statistical Association, 74, 427-431. [29] Donald Stephen G. and Whitney K. Newey (2001): Choosing the number of instruments, Econometrica, 69, 1161-1191. [30] Dufour, J.M. (1997): Some impossibility theorems in econometrics with applications to structural and dynamic models, Econometrica, 65, 1365-1387. [31] Efron, Bradley (1979): Bootstrap methods: Another look at the jackknife, Annals of Statistics, 7, 1-26. [32] Efron, Bradley (1982): The Jackknife, the Bootstrap, and Other Resampling Plans. Society for Industrial and Applied Mathematics. [33] Efron, Bradley and R.J. Tibshirani (1993): An Introduction to the Bootstrap, New York: Chapman-Hall. [34] Eicker, F. (1963): Asymptotic normality and consistency of the least squares estimators for families of linear regressions, Annals of Mathematical Statistics, 34, 447-456. [35] Engle, Robert F. and Clive W. J. Granger (1987): Co-integration and error correction: Representation, estimation and testing, Econometrica, 55, 251-276. [36] Frisch, Ragnar (1933): Editorial, Econometrica, 1, 1-4.
BIBLIOGRAPHY
374
[37] Frisch, Ragnar and F. Waugh (1933): Partial time regressions as compared with individual trends, Econometrica, 1, 387-401. [38] Gallant, A. Ronald and D.W. Nychka (1987): Seminonparametric maximum likelihood estimation, Econometrica, 55, 363-390. [39] Gallant, A. Ronald and Halbert White (1988): A Unied Theory of Estimation and Inference for Nonlinear Dynamic Models. New York: Basil Blackwell. [40] Galton, Francis (1886): Regression Towards Mediocrity in Hereditary Stature, The Journal of the Anthropological Institute of Great Britain and Ireland, 15, 246-263. [41] Goldberger, Arthur S. (1964): Econometric Theory, Wiley. [42] Goldberger, Arthur S. (1968): Topics in Regression Analysis, Macmillan [43] Goldberger, Arthur S. (1991): A Course in Econometrics. Cambridge: Harvard University Press. [44] Goe, W.L., G.D. Ferrier and J. Rogers (1994): Global optimization of statistical functions with simulated annealing, Journal of Econometrics, 60, 65-99. [45] Gosset, William S. (a.k.a. Student) (1908): The probable error of a mean, Biometrika, 6, 1-25. [46] Gauss, K.F. (1809): Theoria motus corporum coelestium, in Werke, Vol. VII, 240-254. [47] Granger, Clive W. J. (1969): Investigating causal relations by econometric models and cross-spectral methods, Econometrica, 37, 424-438. [48] Granger, Clive W. J. (1981): Some properties of time series data and their use in econometric specication, Journal of Econometrics, 16, 121-130. [49] Granger, Clive W. J. and Timo Tersvirta (1993): Modelling Nonlinear Economic Relationships, Oxford University Press, Oxford. [50] Gregory, A. and M. Veall (1985): On formulating Wald tests of nonlinear restrictions, Econometrica, 53, 1465-1468, [51] Haavelmo, T. (1944): The probability approach in econometrics, Econometrica, supplement, 12. [52] Hall, A. R. (2000): Covariance matrix estimation and the power of the overidentifying restrictions test, Econometrica, 68, 1517-1527, [53] Hall, P. (1992): The Bootstrap and Edgeworth Expansion, New York: Springer-Verlag. [54] Hall, P. (1994): Methodology and theory for the bootstrap, Handbook of Econometrics, Vol. IV, eds. R.F. Engle and D.L. McFadden. New York: Elsevier Science. [55] Hall, P. and J.L. Horowitz (1996): Bootstrap critical values for tests based on GeneralizedMethod-of-Moments estimation, Econometrica, 64, 891-916. [56] Hahn, J. (1996): A note on bootstrapping generalized method of moments estimators, Econometric Theory, 12, 187-197. [57] Hamilton, James D. (1994) Time Series Analysis.
BIBLIOGRAPHY
375
[58] Hansen, Bruce E. (1992): Ecient estimation and testing of cointegrating vectors in the presence of deterministic trends, Journal of Econometrics, 53, 87-121. [59] Hansen, Bruce E. (1996): Inference when a nuisance parameter is not identied under the null hypothesis, Econometrica, 64, 413-430. [60] Hansen, Bruce E. (2006): Edgeworth expansions for the Wald and GMM statistics for nonlinear restrictions, Econometric Theory and Practice: Frontiers of Analysis and Applied Research, edited by Dean Corbae, Steven N. Durlauf and Bruce E. Hansen. Cambridge University Press. [61] Hansen, Lars Peter (1982): Large sample properties of generalized method of moments estimators, Econometrica, 50, 1029-1054. [62] Hansen, Lars Peter, John Heaton, and A. Yaron (1996): Finite sample properties of some alternative GMM estimators, Journal of Business and Economic Statistics, 14, 262-280. [63] Hausman, J.A. (1978): Specication tests in econometrics, Econometrica, 46, 1251-1271. [64] Heckman, J. (1979): Sample selection bias as a specication error, Econometrica, 47, 153161. [65] Horn, S.D., R.A. Horn, and D.B. Duncan. (1975) Estimating heteroscedastic variances in linear model, Journal of the American Statistical Association, 70, 380-385. [66] Horowitz, Joel (2001): The Bootstrap, Handbook of Econometrics, Vol. 5, J.J. Heckman and E.E. Leamer, eds., Elsevier Science, 3159-3228. [67] Imbens, G.W. (1997): One step estimators for over-identied generalized method of moments models, Review of Economic Studies, 64, 359-383. [68] Imbens, G.W., R.H. Spady and P. Johnson (1998): Information theoretic approaches to inference in moment condition models, Econometrica, 66, 333-357. [69] Jarque, C.M. and A.K. Bera (1980): Ecient tests for normality, homoskedasticity and serial independence of regression residuals, Economic Letters, 6, 255-259. [70] Johansen, S. (1988): Statistical analysis of cointegrating vectors, Journal of Economic Dynamics and Control, 12, 231-254. [71] Johansen, S. (1991): Estimation and hypothesis testing of cointegration vectors in the presence of linear trend, Econometrica, 59, 1551-1580. [72] Johansen, S. (1995): Likelihood-Based Inference in Cointegrated Vector Auto-Regressive Models, Oxford University Press. [73] Johansen, S. and K. Juselius (1992): Testing structural hypotheses in a multivariate cointegration analysis of the PPP and the UIP for the UK, Journal of Econometrics, 53, 211-244. [74] Kitamura, Y. (2001): Asymptotic optimality and empirical likelihood for testing moment restrictions, Econometrica, 69, 1661-1672. [75] Kitamura, Y. and M. Stutzer (1997): An information-theoretic alternative to generalized method of moments, Econometrica, 65, 861-874.. [76] Koenker, Roger (2005): Quantile Regression. Cambridge University Press. [77] Kunsch, H.R. (1989): The jackknife and the bootstrap for general stationary observations, Annals of Statistics, 17, 1217-1241.
BIBLIOGRAPHY
376
[78] Kwiatkowski, D., P.C.B. Phillips, P. Schmidt, and Y. Shin (1992): Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time series have a unit root? Journal of Econometrics, 54, 159-178. [79] Lafontaine, F. and K.J. White (1986): Obtaining any Wald statistic you want, Economics Letters, 21, 35-40. [80] Lehmann, E.L. and George Casella (1998): Theory of Point Estimation, 2nd Edition, Springer. [81] Lehmann, E.L. and Joseph P. Romano (2005): Testing Statistical Hypotheses, 3rd Edition, Springer. [82] Lindeberg, Jarl Waldemar, (1922): Eine neue Herleitung des Exponentialgesetzes in der Wahrscheinlichkeitsrechnung, Mathematische Zeitschrift, 15, 211-225. [83] Li, Qi and Jerey Racine (2007) Nonparametric Econometrics. [84] Lovell, M.C. (1963): Seasonal adjustment of economic time series, Journal of the American Statistical Association, 58, 993-1010. [85] MacKinnon, James G. (1990): Critical values for cointegration, in Engle, R.F. and C.W. Granger (eds.) Long-Run Economic Relationships: Readings in Cointegration, Oxford, Oxford University Press. [86] MacKinnon, James G. and Halbert White (1985): Some heteroskedasticity-consistent covariance matrix estimators with improved nite sample properties, Journal of Econometrics, 29, 305-325. [87] Magnus, J. R., and H. Neudecker (1988): Matrix Dierential Calculus with Applications in Statistics and Econometrics, New York: John Wiley and Sons. [88] Mann, H.B. and A. Wald (1943). On stochastic limit and order relationships, The Annals of Mathematical Statistics 14, 217226. [89] Muirhead, R.J. (1982): Aspects of Multivariate Statistical Theory. New York: Wiley. [90] Nelder, J. and R. Mead (1965): A simplex method for function minimization, Computer Journal, 7, 308-313. [91] Nerlove, Marc (1963): Returns to Scale in Electricity Supply, Chapter 7 of Measurement in Economics (C. Christ, et al, eds.). Stanford: Stanford University Press, 167-198. [92] Newey, Whitney K. (1990): Semiparametric eciency bounds, Journal of Applied Econometrics, 5, 99-135. [93] Newey, Whitney K. (1997): Convergence rates and asymptotic normality for series estimators, Journal of Econometrics, 79, 147-168. [94] Newey, Whitney K. and Daniel L. McFadden (1994): Large Sample Estimation and Hypothesis Testing, in Robert Engle and Daniel McFadden, (eds.) Handbook of Econometrics, vol. IV, 2111-2245, North Holland: Amsterdam. [95] Newey, Whitney K. and Kenneth D. West (1987): Hypothesis testing with ecient method of moments estimation, International Economic Review, 28, 777-787. [96] Owen, Art B. (1988): Empirical likelihood ratio condence intervals for a single functional, Biometrika, 75, 237-249.
BIBLIOGRAPHY [97] Owen, Art B. (2001): Empirical Likelihood. New York: Chapman & Hall.
377
[98] Park, Joon Y. and Peter C. B. Phillips (1988): On the formulation of Wald tests of nonlinear restrictions, Econometrica, 56, 1065-1083, [99] Phillips, Peter C.B. (1989): Partially identied econometric models, Econometric Theory, 5, 181-240. [100] Phillips, Peter C.B. and Sam Ouliaris (1990): Asymptotic properties of residual based tests for cointegration, Econometrica, 58, 165-193. [101] Politis, D.N. and J.P. Romano (1996): The stationary bootstrap, Journal of the American Statistical Association, 89, 1303-1313. [102] Potscher, B.M. (1991): Eects of model selection on inference, Econometric Theory, 7, 163-185. [103] Qin, J. and J. Lawless (1994): Empirical likelihood and general estimating equations, The Annals of Statistics, 22, 300-325. [104] Ramsey, J. B. (1969): Tests for specication errors in classical linear least-squares regression analysis, Journal of the Royal Statistical Society, Series B, 31, 350-371. [105] Rudin, W. (1987): Real and Complex Analysis, 3rd edition. New York: McGraw-Hill. [106] Runge, Carl (1901): ber empirische Funktionen und die Interpolation zwischen quidistanten Ordinaten, Zeitschrift fr Mathematik und Physik, 46, 224-243. [107] Said, S.E. and D.A. Dickey (1984): Testing for unit roots in autoregressive-moving average models of unknown order, Biometrika, 71, 599-608. [108] Secrist, Horace (1933): The Triumph of Mediocrity in Business. Evanston: Northwestern University. [109] Shao, J. and D. Tu (1995): The Jackknife and Bootstrap. NY: Springer. [110] Sargan, J.D. (1958): The estimation of economic relationships using instrumental variables, Econometrica, 2 6, 393-415. [111] Shao, Jun (2003): Mathematical Statistics, 2nd edition, Springer. [112] Sheather, S.J. and M.C. Jones (1991): A reliable data-based bandwidth selection method for kernel density estimation, Journal of the Royal Statistical Society, Series B, 53, 683-690. [113] Shin, Y. (1994): A residual-based test of the null of cointegration against the alternative of no cointegration, Econometric Theory, 10, 91-115. [114] Silverman, B.W. (1986): Density Estimation for Statistics and Data Analysis. London: Chapman and Hall. [115] Sims, C.A. (1972): Money, income and causality, American Economic Review, 62, 540-552. [116] Sims, C.A. (1980): Macroeconomics and reality, Econometrica, 48, 1-48. [117] Staiger, D. and James H. Stock (1997): Instrumental variables regression with weak instruments, Econometrica, 65, 557-586. [118] Stock, James H. (1987): Asymptotic properties of least squares estimators of cointegrating vectors, Econometrica, 55, 1035-1056.
BIBLIOGRAPHY
378
[119] Stock, James H. (1991): Condence intervals for the largest autoregressive root in U.S. macroeconomic time series, Journal of Monetary Economics, 28, 435-460. [120] Stock, James H. and Jonathan H. Wright (2000): GMM with weak identication, Econometrica, 68, 1055-1096. [121] Stock, James H. and Mark W. Watson (2010): Introduction to Econometrics, 3rd edition, Addison-Wesley. [122] Stone, Marshall H. (1937): Applications of the Theory of Boolean Rings to General Topology, Transactions of the American Mathematical Society, 41, 375-481. [123] Stone, Marshall H. (1948): The Generalized Weierstrass Approximation Theorem, Mathematics Magazine, 21, 167-184. [124] Theil, Henri. (1953): Repeated least squares applied to complete equation systems, The Hague, Central Planning Bureau, mimeo. [125] Theil, Henri (1961): Economic Forecasts and Policy. Amsterdam: North Holland. [126] Theil, Henri. (1971): Principles of Econometrics, New York: Wiley. [127] Tobin, James (1958): Estimation of relationships for limited dependent variables, Econometrica, 2 6, 24-36. [128] Tripathi, Gautam (1999): A matrix extension of the Cauchy-Schwarz inequality, Economics Letters, 63, 1-3. [129] van der Vaart, A.W. (1998): Asymptotic Statistics, Cambridge University Press. [130] Wald, A. (1943): Tests of statistical hypotheses concerning several parameters when the number of observations is large, Transactions of the American Mathematical Society, 54, 426-482. [131] Wang, J. and E. Zivot (1998): Inference on structural parameters in instrumental variables regression with weak instruments, Econometrica, 66, 1389-1404. [132] Weierstrass, K. (1885): ber die analytische Darstellbarkeit sogenannter willkrlicher Functionen einer reellen Vernderlichen, Sitzungsberichte der Kniglich Preuischen Akademie der Wissenschaften zu Berlin, 1885. [133] White, Halbert (1980): A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity, Econometrica, 48, 817-838. [134] White, Halbert (1984): Asymptotic Theory for Econometricians, Academic Press. [135] Wooldridge, Jerey M. (2010) Econometric Analysis of Cross Section and Panel Data, 2nd edition, MIT Press. [136] Wooldridge, Jerey M. (2009) Introductory Econometrics: A Modern Approach, 4th edition, Southwestern [137] Zellner, Arnold. (1962): An ecient method of estimating seemingly unrelated regressions, and tests for aggregation bias, Journal of the American Statistical Association, 57, 348-368. [138] Zhang, Fuzhen and Qingling Zhang (2006): Eigenvalue inequalities for matrix product, IEEE Transactions on Automatic Control, 51, 1506-1509.

Econometrics Note

Uploaded by

Copyright:

Available Formats

Econometrics Note

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Econometrics Note

Uploaded by

Copyright:

Available Formats

What topics are covered in the book?

What topics are covered in the book?

What are some common statistical/econometric terms introduced?

What are some common statistical/econometric terms introduced?

ECONOMETRICS

This Revision: January 3, 2014 Comments Welcome

ii 41 43 47 49 50 51 55 57 57 57 58 58 59 60 62 62 63 65 66 67 68 68 70 71 72 74 76 78 82 83 86 86 86 87 88 89 91 92 93 95 96 97 100 101 102 103 105 108 110

vi 296 297 298 299 299 299 301 302 304

369 . 369 . 369 . 371

The Probability Approach to Econometrics

Econometric Terms and Notation

Standard Data Structures

Sources for Economic Data

Reading the Manuscript

plimn N(, 2 ) N(0, 1) 2 k In tr A A0 A1 A > 0, A 0 kak kAk = log

Conditional Expectation and Projection

The Distribution of Wages

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Dollars per Hour

Dollars per Hour

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Log Wage Density

Log Dollars per Hour

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

white men white women black men black women

Log Wage Density

3 Log Dollars per Hour

Log Wage Density

3 Log Dollars per Hour

(a) Women and Men

(b) By Gender and Race

white black other

A useful approximation for the natural logarithm for small x is

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Figure 2.4: log(1 + x)

Conditional Expectation Function

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Log Dollars per Hour

white men white women

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Exp=5 Exp=10 Exp=25 Exp=40

Conditional Mean Linear Projection Quadratic Projection 0 10 20 30 40 50

Labor Market Experience (Years)

(a) Joint density of log(wage) and experience and conditional mean

Log Wage Conditional Density

Log Dollars per Hour

(b) Conditional density

Law of Iterated Expectations

and when x is continuous E (E (y | x)) =

+ E (log(wage) | gender = woman) Pr (gender = woman)

+ E (log(wage) | gender = man, race = black) Pr (race = black|gender = man)

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Log Dollars per Hour

Labor Market Experience (Years)

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Theorem 2.10.2 If Ey 2 < then var (y ) var (y E (y | x1 )) var (y E (y | x1 , x2 )) .

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Clearly, the economist in question ignored variance!

CHAPTER 2. CONDITIONAL EXPECTATION AND PROJECTION

Homoskedasticity and Heteroskedasticity