Econometrics

ECONOMETRICS
Bruce E. Hansen
c 2000, 20111
University of Wisconsin
www.ssc.wisc.edu/~bhansen
This Revision: January 13, 2011

Comments Welcome
1
This manuscript may be printed and reproduced for individual or instructional use, but may not be
printed for commercial purposes.
Contents
Preface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii
1 Introduction 1
1.1 What is Econometrics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 The Probability Approach to Econometrics . . . . . . . . . . . . . . . . . . . . . . . 1
1.3 Econometric Terms and Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.4 Observational Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.5 Standard Data Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.6 Sources for Economic Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.7 Econometric Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.8 Reading the Manuscript . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2 Moment Estimation 8
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.2 Population and Sample Mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.3 Sample Mean is Unbiased . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.4 Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
2.5 Convergence in Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 Weak Law of Large Numbers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.7 Vector-Valued Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.8 Convergence in Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.9 Functions of Moments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.10 Delta Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
2.11 Stochastic Order Symbols . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.12 Uniform Stochastic Bounds* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.13 Semiparametric E¢ ciency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
2.14 Expectation* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.15 Technical Proofs* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3 Conditional Expectation and Projection 28

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.2 The Distribution of Wages . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.3 Conditional Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.4 Conditional Expectation Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
3.5 Continuous Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.6 Law of Iterated Expectations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.7 Monotonicity of Conditioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.8 CEF Error . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.9 Best Predictor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.10 Conditional Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
3.11 Homoskedasticity and Heteroskedasticity . . . . . . . . . . . . . . . . . . . . . . . . . 42
3.12 Regression Derivative . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
i
CONTENTS ii
3.13 Linear CEF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.14 Linear CEF with Nonlinear E¤ects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
3.15 Linear CEF with Dummy Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
3.16 Best Linear Predictor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
3.17 Linear Predictor Error Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.18 Regression Coe¢ cients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.19 Regression Sub-Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.20 Coe¢ cient Decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.21 Omitted Variable Bias . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.22 Best Linear Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.23 Normal Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.24 Regression to the Mean . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
3.25 Reverse Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.26 Limitations of the Best Linear Predictor . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.27 Random Coe¢ cient Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.28 Causal E¤ects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.29 Existence and Uniqueness of the Conditional Expectation* . . . . . . . . . . . . . . 65
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4 The Algebra of Least Squares 71

4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.2 Least Squares Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
4.3 Solving for Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
4.4 Illustration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.5 Least Squares Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
4.6 Model in Matrix Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
4.7 Projection Matrix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
4.8 Orthogonal Projection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
4.9 Regression Components . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
4.10 Residual Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
4.11 Prediction Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82
4.12 In‡uential Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
4.13 Measures of Fit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
4.14 Normal Regression Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5 Least Squares Regression 91

5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.2 Mean of Least-Squares Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.3 Variance of Least Squares Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
5.4 Gauss-Markov Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
5.5 Residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
5.6 Estimation of Error Variance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
5.7 Covariance Matrix Estimation Under Homoskedasticity . . . . . . . . . . . . . . . . 98
5.8 Covariance Matrix Estimation Under Heteroskedasticity . . . . . . . . . . . . . . . . 99
5.9 Standard Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
5.10 Multicollinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
CONTENTS iii
6 Asymptotic Theory for Least Squares 109

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.2 Consistency of Least-Squares Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 109
6.3 Consistency of Sample Variance Estimators . . . . . . . . . . . . . . . . . . . . . . . 112
6.4 Asymptotic Normality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
6.5 Joint Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.6 Uniformly Consistent Residuals* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
6.7 Asymptotic Leverage* . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119
6.8 Consistent Covariance Matrix Estimation . . . . . . . . . . . . . . . . . . . . . . . . 120
6.9 Functions of Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
6.10 Asymptotic Standard Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
6.11 t statistic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.12 Con…dence Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.13 Regression Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124
6.14 Quadratic Forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
6.15 Con…dence Regions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
6.16 Semiparametric E¢ ciency in the Projection Model . . . . . . . . . . . . . . . . . . . 128
6.17 Semiparametric E¢ ciency in the Homoskedastic Regression Model* . . . . . . . . . . 130
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
7 Restricted Estimation 136

7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
7.2 Constrained Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
7.3 Exclusion Restriction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.4 Minimum Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
7.5 Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
7.6 Asymptotic Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
7.7 E¢ cient Minimum Distance Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . 141
7.8 Exclusion Restriction Revisited . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 142
7.9 Variance and Standard Error Estimation . . . . . . . . . . . . . . . . . . . . . . . . . 143
7.10 Nonlinear Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146
8 Testing 147
8.1 t tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
8.2 t-ratios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
8.3 Wald Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.4 Minimum Distance Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
8.5 F Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150
8.7 Problems with Tests of NonLinear Hypotheses . . . . . . . . . . . . . . . . . . . . . 152
8.8 Monte Carlo Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156
8.9 Estimating a Wage Equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 161
9 Additional Regression Topics 163

9.1 Generalized Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
9.2 Testing for Heteroskedasticity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.3 Forecast Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166
9.4 NonLinear Least Squares . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
CONTENTS iv
9.5 Least Absolute Deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

9.6 Quantile Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
9.7 Testing for Omitted NonLinearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
9.8 Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177
10 The Bootstrap 179

10.1 De…nition of the Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
10.2 The Empirical Distribution Function . . . . . . . . . . . . . . . . . . . . . . . . . . . 179
10.3 Nonparametric Bootstrap . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
10.4 Bootstrap Estimation of Bias and Variance . . . . . . . . . . . . . . . . . . . . . . . 181
10.5 Percentile Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
10.6 Percentile-t Equal-Tailed Interval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.7 Symmetric Percentile-t Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
10.8 Asymptotic Expansions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
10.9 One-Sided Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 187
10.10Symmetric Two-Sided Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
10.11Percentile Con…dence Intervals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189
10.12Bootstrap Methods for Regression Models . . . . . . . . . . . . . . . . . . . . . . . . 190
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
11 Generalized Method of Moments 193

11.1 Overidenti…ed Linear Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193
11.2 GMM Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 194
11.3 Distribution of GMM Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
11.4 Estimation of the E¢ cient Weight Matrix . . . . . . . . . . . . . . . . . . . . . . . . 196
11.5 GMM: The General Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
11.6 Over-Identi…cation Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197
11.7 Hypothesis Testing: The Distance Statistic . . . . . . . . . . . . . . . . . . . . . . . 198
11.8 Conditional Moment Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
11.9 Bootstrap GMM Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 200
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
12 Empirical Likelihood 204

12.1 Non-Parametric Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
12.2 Asymptotic Distribution of EL Estimator . . . . . . . . . . . . . . . . . . . . . . . . 206
12.3 Overidentifying Restrictions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
12.4 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208
12.5 Numerical Computation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209
13 Endogeneity 211
13.1 Instrumental Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
13.2 Reduced Form . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
13.3 Identi…cation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
13.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
13.5 Special Cases: IV and 2SLS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214
13.6 Bekker Asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
13.7 Identi…cation Failure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 217
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
CONTENTS v
14 Univariate Time Series 221

14.1 Stationarity and Ergodicity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
14.2 Autoregressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
14.3 Stationarity of AR(1) Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
14.4 Lag Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
14.5 Stationarity of AR(k) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
14.6 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225
14.7 Asymptotic Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 226
14.8 Bootstrap for Autoregressions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
14.9 Trend Stationarity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 227
14.10Testing for Omitted Serial Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 228
14.11Model Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
14.12Autoregressive Unit Roots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 229
15 Multivariate Time Series 231

15.1 Vector Autoregressions (VARs) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
15.2 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
15.3 Restricted VARs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
15.4 Single Equation from a VAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 232
15.5 Testing for Omitted Serial Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . 233
15.6 Selection of Lag Length in an VAR . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
15.7 Granger Causality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
15.8 Cointegration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 234
15.9 Cointegrated VARs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235
16 Limited Dependent Variables 237

16.1 Binary Choice . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
16.2 Count Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
16.3 Censored Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
16.4 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
17 Panel Data 242

17.1 Individual-E¤ects Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
17.2 Fixed E¤ects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 242
17.3 Dynamic Panel Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244
18 Nonparametrics 245
18.1 Kernel Density Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
18.2 Asymptotic MSE for Kernel Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . 247
A Matrix Algebra 250

A.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
A.2 Matrix Addition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
A.3 Matrix Multiplication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
A.4 Trace . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
A.5 Rank and Inverse . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253
A.6 Determinant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
A.7 Eigenvalues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
A.8 Positive De…niteness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255
A.9 Matrix Calculus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
A.10 Kronecker Products and the Vec Operator . . . . . . . . . . . . . . . . . . . . . . . . 257
A.11 Vector and Matrix Norms and Inequalities . . . . . . . . . . . . . . . . . . . . . . . . 257
CONTENTS vi
B Probability 260
B.1 Foundations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
B.2 Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
B.3 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 262
B.4 Gamma Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263
B.5 Common Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
B.6 Multivariate Random Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
B.7 Conditional Distributions and Expectation . . . . . . . . . . . . . . . . . . . . . . . . 268
B.8 Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 270
B.9 Normal and Related Distributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 271
B.10 Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
B.11 Maximum Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
C Numerical Optimization 282

C.1 Grid Search . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
C.2 Gradient Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282
C.3 Derivative-Free Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 284
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.
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 asymp-
totic 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 contem-
porary econometric methods and theory.
As this is a manuscript in progress, some parts are quite incomplete, in particular the later
sections of the manuscript. Hopefully one day these sections will be ‡eshed out and completed in
more detail.
I would like to thank Ying-Ying Lee for providing research assistance in preparing some of the
empirical examples presented in the text.
vii
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 Frisch’s own words in the introduction to the …rst issue of
Econometrica for an explanation of the discipline.
A word of explanation regarding the term econometrics may be in order. Its de…ni-
tion 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 uni…cation of the theoretical-
quantitative 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 de…ninitely 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 su¢ cient, condition for a real understanding of the
quantitative relations in modern economic life. It is the uni…cation of all three that is
powerful. And it is this uni…cation that constitutes econometrics.
Ragnar Frisch, Econometrica, (1933), 1, pp. 1-2.
This de…nition remains valid today, although some terms have evolved somewhat in their usage.
Today, we would say that econometrics is the uni…ed 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 (1911-
1999) of Norway, winner of the 1989 Nobel Memorial Prize in Economic Sciences, in his seminal
1
CHAPTER 1. INTRODUCTION 2
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 quan-
tities, and it is incohorent 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 eco-
nomic model is a probability model, it follows naturally that the best way to quantify, estimate,
and conduct inferences about the economy is through the powerful theory of mathematical statis-
tics. The appropriate method for a quantitative economic analysis follows from the probabilistic
construction of the economic model.
Haavelmo’s 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 Haavelmo’s original idea. A probabilistic economic
model is speci…ed, and the quantitative analysis performed under the assumption that the economic
model is correctly speci…ed. 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 speci…ed. 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 quasi-
structural 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
de…ned 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
speci…ed but some features are left unspeci…ed. 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 approx-
imations and hence inherently false. The di¤erence is that the calibrationist literature rejects
mathematical statistics 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 vari-
ables. 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 speci…c 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.
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
1
Ad hoc means “for this purpose” – a method designed for a speci…c problem – and not based on a generalizable
principle.
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.
0 1
x1
B x2 C
B C
x = B . C:
.
@ . A
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 i’th observation is the set (yi ; xi ; z i ):
It is proper mathematical practice to use upper case X for random variables and lower case x
for realizations or speci…c values. This practice is not commonly followed in econometrics because
instead we use upper case to denote matrices. Thus the notation yi will in some places refer to a
random variable, and in other places a speci…c realization. Hopefully there will be no confusion as
the use should be evident from the context.
As we mentioned before, ideally each observation consists of a set of measurements on the
list of variables. In practice it is common to …nd that some variables are not measured for some
observations, and in these cases we describe these variables or observations as unobserved or
missing.
We typically use Greek letters such as ; and 2 to denote unknown parameters of an econo-
metric 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. ^
and ~ are estimates of :
The covariance matrix of an econometric estimator will typically be written using the capital
p b
boldface V ; often with a subscript to denote the estimator, e.g. V b = var n as the
p
covariance matrix for n b : Hopefully without causing confusion, we will use the notation
p
V = avar( b ) to denote the asymptotic covariance matrix of n b (the variance of the
asymptotic distribution). Estimates will be denoted by appending hats or tildes, e.g. Vb is an
estimate of 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 worker’s 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 di¤erent levels of
education to the di¤erent groups, and then follow the children’s wage path after they mature and
enter the labor force. The di¤erences between the groups would be direct measurements of the ef-
fects of di¤erent 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 person’s 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 di¢ cult to infer causality, as we are not able to manipulate one variable to
see the direct e¤ect on the other. For example, a person’s level of education is (at least partially)
determined by that person’s choices. These factors are likely to be a¤ected 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 di¢ cult to infer causality from observational data alone. Causal
inference requires identi…cation, and this is based on strong assumptions. We will return to a
discussion of some of these issues in Chapter 13.
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 tick-by-tick) 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 individual’s observations are mutually dependent. This is a modi…ed 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 i’th
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.
De…nition 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 re-

alization from a joint probability distribution F (y; x; z) which can call the population. This
“population” is in…nitely large. This abstraction can be a source of confusion as it does not cor-
respond to a physical population in the real world. 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 cen-
tury, allowing the application of mathematical statistics to the social sciences. Before this concep-
tual 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 eco-
nomic 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
speci…c 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

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 journal’s website,
as many journals archive data and replication programs online. Second, check the website(s) of
the paper’s 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.
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 signi…cant time,
and programming errors are hard to prevent and di¢ cult to detect and eliminate.
R (www.r-project.org) is an integrated suite of statistical and graphical software that is ‡exible,
open source, and best of all, 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 di¤erent 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

Chapter 2 is a review of moment estimation and asymptotic distribution theory. This material
should be familiar from an earlier course in statistics, but I have included this at the beginning be-
cause of its central importance in econometric distribution theory. Chapters 3 through 9 deal with
the core linear regression and projection models. Chapter 10 introduces the bootstrap. Chapters
11 through 13 deal with the Generalized Method of Moments, empirical likelihood and endogeneity.
Chapters 14 and 15 cover time series, and Chapters 16, 17 and 18 cover limited dependent vari-
ables, panel data, and nonparametrics. Reviews of matrix algebra, probability theory, maximum
likelihood, and numerical optimization can be found in the appendix.
Technical sections which may not be of interest to all readers are marked with an asterisk (*).
Chapter 2
Moment Estimation
2.1 Introduction
Most econometric estimators can be written as functions of sample moments. To understand
econometric estimation we need a thorough understanding of moment estimation. This chapter
provides a concise summary. It will useful for most students to review this material, even if most
is familiar.
2.2 Population and Sample Mean

A random variable y with density f has the expectation or mean1
= E (y)
Z 1
= uf (u)du:
1
This is the average value of y in the population.

We would like to estimate from a random sample. Recall that a random sample fy1 ; :::; yn g
consists of n observations of independent and identically draws from the distribution of y:
Assumption 2.2.1 The observations fy1 ; :::; yn g are a

random sample.
As is the average value of y in the population, it seems reasonable to estimate from the
average value of y in the sample. This is the sample mean, written as
n
1 1X
y = (y1 + yn ) = yi
n n
i=1
It is important to understand the distinction between and y. The population mean is a non-
random feature of the population while the sample mean y is a random feature of a random sample.
is …xed, while y varies with the sample. We use the term “mean” to refer to both, but they are
really quite distinct. Here, as is common in econometrics, we put a bar “ ”over y to indicate that
the quantity is a sample mean. This convention is useful as it helps readers recognize a sample
mean. It is also common to see the notation y n ; where the subscript “n”indicates that the sample
mean depends on the sample size n:
1
For a rigorous treatment of expectation see Section 2.14.
8
CHAPTER 2. MOMENT ESTIMATION 9
Moment estimation uses sample moments as estimates of population moments. In the case
of the mean, the moment estimate of the population mean = Ey is the sample mean b = y: Here,
as is common in econometrics, we put a hat “^”over the parameter to indicate that b is a sample
estimate of : This is a helpful convention, as just by seeing the symbol b we understand that it is
a sample estimate of a population parameter :
2.3 Sample Mean is Unbiased

Since the sample mean is a linear function of the observations, it is simple to calculate its
expectation.
n
! n
1X 1X
Ey = E yi = Eyi = :
n n
i=1 i=1
This shows that the expected value of the sample mean equals the population mean. An estimator
with this property is called unbiased.
De…nition 2.3.1 An estimator b for is unbiased if Eb = .
Theorem 2.3.1 If E jyj < 1 then Ey = and b = y is unbiased for the popula-
tion mean :
You may notice that we slipped in the additional condition “If E jyj < 1”. This assumption
ensures that is …nite and the mean of y is well de…ned.
2.4 Variance
The variance of the random variable y is de…ned as
2
= var (y)
= E (y Ey)2
= Ey 2 (Ey)2 :
Notice that the variance is the function of two moments, Ey 2 and Ey:
We can calculate the variance of the sample mean b: It is convenient to de…ne the centered
observations ui = yi which have mean zero and variance 2 . Then
n
1X
^ = ui
n
i=1
and
n
!0 n
1
n n n
1X @1
X 1 XX 1 X 1
var (^ ) = E (^ )2 = E ui uj A = 2 E (ui uj ) = 2 2
= 2
n n n n n
i=1 j=1 i=1 j=1 i=1
where the second-to-last inequality is because E (ui uj ) = 2 for i = j yet E (ui uj ) = 0 for i 6= j due
to independence.
2 1 2
Theorem 2.4.1 If < 1 then var (b) = n
This result links the variance of the estimator b with the variance of the individual observation
yi and with the sample size n: In particular, var (b) is proportional to 2 ; and inversely proportional
to n and thus decreases as n increases.
2.5 Convergence in Probability

In Theorem 2.4.1 we showed that the variance of b decreases with the sample size n: This
implies that the sampling distribution of b concentrates as the sample size increases. We now give
a formal de…nition.
De…nition 2.5.1 A random variable zn 2 R converges in probability

p
to z as n ! 1; denoted zn ! z; if for all > 0;
lim Pr (jzn zj ) = 1: (2.1)

n!1
The de…nition looks quite abstract, but it formalizes the concept of a distribution concen-
trating about a point. The event fjzn zj g is the event that zn is within of the point z:
Pr (jzn zj ) is the probability of this event –that zn is within of the point z. The statement
(2.1) is that this probability approaches 1 as the sample size n increases. The de…nition of conver-
gence in probability requires that this holds for any : So even for very small intervals about z; the
distribution of zn concentrates within this interval for large n:
p
When zn ! z we call z the probability limit (or plim) of zn .
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 also important to include the phrase “as n ! 1”to be speci…c about how the limit is obtained.
Students often confuse convergence in probability with convergence in expectation:
Ezn ! Ez (2.2)
but these are distinct concepts. Neither (2.1) nor (2.2) 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 n 1 and the value an 6= 0 with
probability 1 n 1 , or
1
Pr (zn = an ) = (2.3)
n
n 1
Pr (zn = 0) = :
n
In this example the probability distribution of zn concentrates at zero as n increases. You can
p
check that zn ! 0 as n ! 1; regardless of the sequence an :
In this example we can also calculate that the expectation of zn is
an
Ezn = :
n
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 in…nity 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 arti…cial, 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 must not be used.
2.6 Weak Law of Large Numbers

As we mentioned in the two previous sections, the variance of the sample mean decreases to
zero as the sample size increases. We now show 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 Chebyshev’s
inequality (B.26). The latter states that for any random variable zn and constant > 0
var(zn )
Pr (jzn Ezn j > ) 2 :
1 2
Set zn = b; for which Ezn = and var(zn ) = n (by Theorems 2.3.1 and 2.4.1). Then
2
Pr (jb j> ) 2:
n
For …xed 2 and ; the bound on the right-hand-side shrinks to zero as n ! 1: Thus the probability
that b is within of approaches 1 as n gets large, so b converges in probability to .
We have shown that the sample mean b converges in probability to the population mean .
This result is called the weak law of large numbers. Our derivation assumed that y has a …nite
variance, but this is not necessary. It is only necessary for y to have a …nite mean.
Theorem 2.6.1 Weak Law of Large Numbers (WLLN)

If E jyj < 1 then as n ! 1,
n
1X p
y= yi ! E(yi ):
n
i=1
The proof of Theorem 2.6.1 is presented in Section 2.15.

The WLLN shows that the estimator b = y converges in probability to the true population
mean . An estimator which converges in probability to the population value is called consistent.
p
De…nition 2.6.1 An estimator ^ of a parameter is consistent if ^ ! as n ! 1:
Consistency is a good property for an estimator to possess. It means that for any given data
distribution; there is a sample size n su¢ ciently large such that the estimator ^ will be arbitrarily
close to the true value with high probability. Unfortunately it does not mean that ^ will actually
be close to in a given …nite sample, but it is minimal property for an estimator to be considered
a “good” estimator.
Theorem 2.6.2 Under Assumption 2.2.1 and E jyj < 1; b = y is consis-

tent for the population mean :
Almost Sure Convergence and the Strong Law*
Convergence in probability is sometimes called weak convergence. A related con-

cept 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.)
De…nition 2.6.2 A random variable zn 2 R converges almost surely

a:s:
to z as n ! 1; denoted zn ! z; if for every > 0
Pr lim jzn zj = 1: (2.4)

n!1
The convergence (2.4) is stronger than (2.1) because it computes the probability of
a limit rather than the limit of a probability. Almost sure convergence is stronger than
a:s: p
convergence in probability in the sense that zn ! z implies zn ! z.
In the example (2.3) of Section 2.5, the sequence zn converges in probability to zero
for any sequence an ; but this is not su¢ cient 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 2.6.3 Strong Law of Large Numbers (SLLN)

If E jyj < 1; then as n ! 1,
n
1 X a:s:
y= yi ! E(yi ):
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 su¢ cient for most purposes in econometrics, so we will not use the
SLLN in this text.
2.7 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 2 Rm is a vector. To …x notation, the
elements of y are 0 1
y1
B y2 C
B C
y=B .. C:
@ . A
ym
The population mean of y is just the vector of marginal means
0 1
E (y1 )
B E (y2 ) C
B C
= E(y) = B .. C:
@ . A
E (ym )
When working with random vectors y it is convenient to measure their magnitude with the
Euclidean norm
2 1=2
kyk = y12 + + ym :
This is the classic Euclidean length of the vector y. Notice that
kyk2 = y 0 y:
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 2.7.1 For y 2 Rm ; E kyk < 1 if and only if E jyj j < 1 for
j = 1; :::; m:
Theorem 2.7.1 implies that the components of are …nite if and only if E kyk < 1.
The m m variance matrix of y is
V = var (y) = E (y ) (y )0 :
V is often called a variance-covariance matrix. You can show that the elements of V are …nite if
E kyk2 < 1:
A random sample fy 1 ; :::; y n g consists of n observations of independent and identically draws
from the distribution of y: (Each draw is an m-vector.) The vector sample mean
0 1
y1
1X
n B y C
B 2 C
y= yi = B . C
n @ .. A
i=1
ym
is the vector of means of the individual variables.
Convergence in probability of a vector is de…ned as convergence in probability of all elements
p p
in the vector. Thus y ! if and only if y j ! j for j = 1; :::; m: Since the latter holds if
E jyj j < 1 for j = 1; :::; m; or equivalently E kyk < 1; we can state this formally as follows.
Theorem 2.7.2 Weak Law of Large Numbers (WLLN) for random vectors
If E kyk < 1 then as n ! 1,
n
1X p
y= y i ! E(y i ):
n
i=1
2.8 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.
De…nition 2.8.1 Let z n be a random vector with distribution Fn (u) = Pr (z n u) : We

d
say that z n converges in distribution to z as n ! 1, denoted z n ! z; if for all u at
which F (u) = Pr (z u) is continuous, Fn (u) ! F (u) as n ! 1:
d
When z n ! z, it is common to refer to z as the asymptotic distribution or limit distri-
bution of z n .
When the limit distribution z is degenerate (that is, Pr (z = c) = 1 for some c) we can write
d p
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.
Theorem 2.8.1 Central Limit Theorem (CLT). If E kyk2 < 1 then

as n ! 1
n
p 1 X d
n (y n )= p (y i ) ! N (0; V )
n
i=1
where = Ey and V = E (y ) (y )0 :
p
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 1733. This was extended to cover an approximation to the binomial dis-
tribution in 1812 by Pierre-Simon Laplace, and the general statement is credited to the Russian
mathematician Aleksandr Lyapunov in 1901.
2.9 Functions of Moments

We now expand our investigation and consider estimation of parameters which can be written
as a continuous function of . That is, the parameter of interest is the vector of functions
= g( ) (2.5)
where g : Rm ! Rk : As one example, the geometric mean of wages w is
= exp (E (log (w))) (2.6)

which is (2.5) with

g(u) = exp (u)
and = E (log (w)) : As another example, the skewness of the wage distribution is
E (w Ew)3
sk = 3=2
E (w Ew)2
= g Ew; Ew2 ; Ew3
where w = wage and

3 3
3 2 1+2 1
g( 1; 2; 3) = : (2.7)
2 3=2
2 1
In this case we can set 0
1
w
y = @ w2 A
w3
so that 0 1
Ew
= @ Ew2 A : (2.8)
Ew3
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
with its point estimate b so that
b = g (b ) : (2.9)
Again, the hat “^” indicates that b is a sample estimate of :
For example, the plug-in estimate of the geometric mean of the wage distribution from (2.6)
is
b = exp(b)
with
n
1X
b= log (wagei ) :
n
i=1
The plug-in estimate of the skewness of the wage distribution is

1 Pn
c n i=1 (wi w)3
sk =
1 Pn
3=2
n i=1 (wi w)2
b3 3b2 b1 + 2b31
= 3=2
b2 b21
where
n
1X j
bj = wi :
n
i=1
A useful property is that continuous functions are limit-preserving.
p
Theorem 2.9.1 Continuous Mapping Theorem (CMT). If z n ! c
p
as n ! 1 and g ( ) is continuous at c; then g(z n ) ! g(c) as n ! 1.
The proof of Theorem 2.9.1 is given in Section 2.15.

p
For example, if zn ! c as n ! 1 then
p
zn + a ! c + a
p
azn ! ac
p
zn2 ! c2
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:
We need the following assumption in order for b to be consistent for .
Theorem 2.9.2 If E kyk < 1 and g (u) is continuous at u = then
b = g (b ) p! g ( ) =
as n ! 1; and thus b is consistent for :
To apply Theorem 2.9.2 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 2.7.2 and
Theorem 2.9.2 that if E jlog (wage)j < 1 then as n ! 1
p
b ! :
In our second example g de…ned in (2.7) is continuous for all such that var(w) = 2 2 > 0;
1
3
which holds unless w has a degenerate distribution. Thus if E jwj < 1 and var(w) > 0 then as
n!1
c p! sk:
sk
2.10 Delta Method

In this section we introduce two tools –an extended version of the CMT and the Delta Method
–which allow us to calculate the asymptotic distribution of the parameter estimate b .
We …rst present an extended version of the continuous mapping theorem which allows conver-
gence in distribution.
Theorem 2.10.1 Continuous Mapping Theorem

d
If z n ! z as n ! 1 and g : Rm ! Rk has the set of discontinuity points
d
Dg such that Pr (z 2 Dg ) = 0; then g(z n ) ! g(z) as n ! 1.
For a proof of Theorem 2.10.1 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 2.10.1 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) = u 1 is discontinuous at u = 0; but if
d d
zn ! z N (0; 1) then Pr (z = 0) = 0 so zn 1 ! z 1 :
A special case of the Continuous Mapping Theorem is known as Slutsky’s Theorem.
Theorem 2.10.2 Slutsky’s Theorem

d p
If zn ! z and cn ! c as n ! 1 then
d
1. zn + cn ! z + c
d
2. zn cn ! zc
zn d z
3. ! if c 6= 0
cn c
Even though Slutsky’s 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.
Despite the fact that the plug-in estimator b is a function of b for which we have an asymptotic
distribution, Theorem 2.10.1 does not directly give us an asymptotic distribution for b : This is
p
because b = g (b ) is written as a function of b , not of the standardized sequence n (b ):
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 2.10.3 Delta Method:

p d
If n ( n 0 ) ! ; where is m 1; and g( ) : Rm ! Rk ; k m; is
continuously di¤ erentiable in a neighborhood of then as n ! 1
p d
a n (g ( n) g( 0 )) ! G0 (2.10)
@
where G( ) = @ g( )0 and G = G( 0 ): In particular, if
p d
n( n 0) ! N (0; V )
where V is m m; then as n ! 1
p d
n (g ( n) g( 0 )) ! N 0; G0 V G : (2.11)
The Delta Method allows us to complete our derivation of the asymptotic distribution of the
estimator b of . Relative to consistency, it requires the stronger smoothness condition that g( )
is continuously di¤erentiable.
Now by combining Theorems 2.8.1 and 2.10.3 we can …nd the asymptotic distribution of the
plug-in estimator b .
@
Theorem 2.10.4 If E kyk2 < 1 and G (u) = g (u)0 is continuous in
@u
a neighborhood of u = then as n ! 1
p d
n b ! N 0; G0 V G
where G = G ( ) :
2.11 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. The notation zn = op (1) (pronounced “small
p
oh-P-one”) means that zn ! 0 as n ! 1: We also say that
zn = op (an )
if an is a sequence such that an 1 zn = op (1): For example, for any consistent estimator b for we
then can write
b = + op (1)
Similarly, the notation zn = Op (1) (pronounced “big on-P-one”) means that zn is bounded in
probability. Precisely, for any " > 0 there is a constant M" < 1 such that
lim Pr (jzn j > M" ) ":

n!1
We say that
zn = Op (an )
1z
if an is a sequence such that an n = 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 zn ! z (for example, if z N (0; V )) then
b
zn = Op (1): It follows that for estimators which satisfy the convergence of Theorem 2.10.4 then
we can write
b = + Op (n 1=2 ):
There are many simple rules for manipulating op (1) and Op (1) sequences which can be deduced
from the continuous mapping theorem or Slutsky’s 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)
2.12 Uniform Stochastic Bounds*

For some applications it can be useful to obtain the stochastic order of the random variable
max jyi j :
1 i n
This is the magnitude of the largest observation in the sample fy1 ; :::; yn g: 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 2.12.1 If E jyjr < 1 then as n ! 1

1=r p
n max jyi j ! 0
1 i n
Equivalently,
max jyi j = op (n1=r ): (2.12)
1 i n
Theorem 2.12.1 says that the largest observation will diverge at a rate slower than n1=r . As r
increases this rate decreases. Thus the higher the moment, the slower the rate of divergence of the
largest observation.
To simplify the notation, we write (2.12) as
yi = op (n1=r )
uniformly in 1 i n. It is important to understand when the Op or op symbols are applied to

subscript i random variables we typically mean uniform convergence in the sense of (2.12).
Theorem 2.12.1 applies to random vectors. If E kykr < 1 then
max ky i k = op (n1=r ):
1 i n
We now prove Theorem 2.12.1. Take any : The event max1 i n jyi j > n1=r means that at
S
least one of the jyi j exceeds n1=r ; which is the same as the event ni=1 jyi j > n1=r or equivalently
S n r r
i=1 fjyi j > ng : Since the probability of the union of events is smaller than the sum of the
probabilities,
n
!
[
Pr n 1=r max jyi j > = Pr fjyi jr > r ng
1 i n
i=1
n
X
Pr (jyi jr > n r )
i=1
n
1 X
E (jyi jr 1 (jyi jr > n r ))
n r
i=1
1 r
= r E (jyi j 1 (jyi jr > n r ))
where the second inequality is the strong form of Markov’s inequality (Theorem B.25) and the …nal
equality is since the yi are iid. Since E jyjr < 1 this …nal expectation converges to zero as n ! 1:
This is because Z
E jyi j = jyjr dF (y) < 1
r
implies Z
r r
E (jyi j 1 (jyi j > c)) = jyjr dF (y) ! 0
r
jyj >c
1=r p
as c ! 1: We have established that n max1 i n jyi j ! 0; as required.
2.13 Semiparametric E¢ ciency

In this section we argue that the sample mean b and plug-in estimator b = g (b ) are e¢ cient
estimators of the parameters and . Our demonstration is based on the rich but technically
challenging theory of semiparametric e¢ ciency 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).
We start by examining the sample mean b ; for the asymptotic e¢ ciency of b will follow from
that of b :
Recall, we know that if E kyk2 < 1 then the sample mean has the asymptotic distribution
p d
n (b ) ! N (0; V ) : We want to know if b is the best feasible estimator, or if there is another
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?
When we ask if b is the best estimator, we need to be clear about the class of models –the class
of permissible distributions. For estimation of the mean of the distribution of y the broadest
conceivable class is L1 = fF : E kyk < 1g : This class is too broad n for our currentn purposes, as
o
b is not asymptotically N (0; V ) for all F 2 L1 : A more realistic choice is L2 = F : E kyk2 < 1
– the class of …nite-variance distributions. When we seek an e¢ cient 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 2 L2 :
To show that the answer is not immediately obvious, it might be helpful to review a set-
ting where the sample mean is ine¢ p cient. Suppose that y 2 R has the double exponential den-
sity f (y j ) = 2 1=2 exp jy j 2 : Since var (y) = 1 we see that the sample mean sat-
p d
is…es n (^ ) ! N (0; 1). In this model the maximum likelihood estimator (MLE) ~ for
is the sample median. Recall from the theory of maximum likelhood that the MLE satis…es
p d 1 p
n (~ ) ! N 0; ES 2 where S = @@ log f (y j ) = 2 sgn (y ) is the score. We can
p 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 ine¢ cient.
But the estimator which achieves this improved e¢ ciency –the sample median –is not generi-
cally 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 p is that the sample
median is e¢ cient in the class of densities f (y j ) = 2 1=2 exp jy j 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 e¢ cient 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 unspeci…ed. In the
semiparametric context an estimator is called semiparametrically e¢ cient 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 “submod-
els”. 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 de…ned 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 =
yf (y)dy: A parametric submodel for f (y) is a density f (y j ) which is a smooth function of
a parameter , and there is a true value 0 such that f (y j 0 ) = f (y): The index indicates the
submodels. The equality f (y j 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 j ) ; the mean is ( ) = yf (y j ) dy which varies
with the parameter . Let 2 @ be the class of all submodels for f:
Since each submodel is parametric we can calculate the e¢ ciency bound for estimation of
within this submodel. Speci…cally, given the density f (y j ) its likelihood score is
@
S = log f (y j 0) ;
@
1
so the Cramer-Rao lower bound for estimation of is ES S 0 : De…ning M = @@ ( 0)
0;
by Theorem B.11.5 the Cramer-Rao lower bound for estimation of within the submodel is
1
0
V =M ES S 0 M .
As V is the e¢ ciency bound for the submodel class f (y j ) ; no estimator can have an
asymptotic variance smaller than V for any density f (y j ) 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 de…ne2
V = sup V :
2@
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 e¢ ciency bound for estimation of , as it is a lower bound on the asymptotic
variance for any semiparametric estimator. If the asymptotic variance of a speci…c semiparametric
estimator equals the bound V we say that the estimator is semiparametrically e¢ cient.
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 satis…es 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
satis…es V 0 < V 1 but this would imply V < V 1 which contradicts the Cramer-Rao Theorem.
We now …nd this submodel for the sample mean b : Our goal is to …nd a parametric submodel
whose Cramer-Rao bound for is V : This can be done by creating a tilted version of the true
density. Consider the parametric submodel
0 1
f (y j ) = f (y) 1 + V (y ) (2.13)
where f (y) is the true density and = Ey: Note that

Z Z Z
f (y j ) dy = f (y)dy + 0 V 1
f (y) (y ) dy = 1
and for all close to zero f (y j ) 0: Thus f (y j ) is a valid density function. It is a parametric
submodel since f (y j 0 ) = f (y) when 0 = 0: This parametric submodel has the mean
Z
( ) = yf (y j ) dy
Z Z
= yf (y)dy + f (y)y (y )0 V 1 dy
= +
which is a smooth function of :

Since
@ @ 0 1 V 1 (y )
log f (y j ) = log 1 + V (y ) =
@ @ 1 + 0 V 1 (y )
it follows that the score function for is
@ 1
S = log f (y j 0) =V (y ): (2.14)
@
By Theorem B.11.3 the Cramer-Rao lower bound for is
1 1
E S S0 = V 1
E (y ) (y )0 V 1
=V: (2.15)
2
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.
The Cramer-Rao lower bound for ( ) = + is also V , and this equals the asymptotic variance
of the moment estimator b : This was what we set out to show.
In summary, we have shown that in the submodel (2.13) 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 2.13.1 In the class of distributions F 2 L2 ; the semipara-

metric variance bound for estimation of is V = var(yi ); and the sample
mean b is a semiparametrically e¢ cient estimator of the population mean
.
We call this result a proposition rather than a theorem as we have not attended to the regularity
conditions.
It is a simple matter to extend this result to the plug-in estimator b = g (b ). We know from
Theorem 2.10.4 that if E kyk2 < 1 and g (u) is continuously di¤erentiable at u = then the plug-in
p d
estimator has the asymptotic distribution n b ! N (0; G0 V G) : We therefore consider the
n o
class of distributions L2 (g) = F : E kyk2 < 1; g (u) is continuously di¤erentiable at u = Ey :
For example, if = 1 = 2 where 1 = Ey1 and 2 = Ey2 then L2 (g) = F : Ey12 < 1; Ey22 < 1; and Ey2 6= 0 :
For any submodel the Cramer-Rao lower bound for estimation of = g ( ) is G0 V G by
Theorem B.11.5. For the submodel (2.13) this bound is G0 V G which equals the asymptotic variance
of b from Theorem 2.10.4. Thus b is semiparametrically e¢ cient.
Proposition 2.13.2 In the class of distributions F 2 L2 (g) the semipara-

metric variance bound for estimation of = g ( ) is G0 V G; and the plug-in
estimator b = g (b ) is a semiparametrically e¢ cient estimator of .
The result in Proposition 2.13.2 is quite general. Smooth functions of sample moments are
e¢ cient 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.
2.14 Expectation*
For any random variable y we de…ne the mean or expectation Ey as follows. If y is discrete,
1
X
Ey = j Pr (y = j) ;
j=1
and if y is continuous with density f

Z 1
Ey = yf (y)dy:
1
We can unify these de…nitions by writing the expectation as the Lebesgue integral with respect to
the distribution function F Z 1
Ey = ydF (y):
1
The mean is well de…ned and …nite if

Z 1
E jyj = jyj dF (y) < 1:
1
If this does not hold, we evaluate

Z 1
I1 = ydF (y)
0
Z 0
I2 = ydF (y)
1
If I1 = 1 and I2 < 1 then we de…ne Ey = 1: If I1 < 1 and I2 = 1 then we de…ne Ey = 1: If

both I1 = 1 and I2 = 1 then Ey is unde…ned.
If = Ey is well de…ned we say that is identi…ed, meaning that the parameter is uniquely
determined by the distribution of the observed variables. The demonstration that the parameters
of an econometric model are identi…ed is an important precondition for estimation. Typically,
identi…cation holds under a set of restrictions, and an identi…cation theorem carefully describes a
set of such conditions which are su¢ cient for identi…cation. In the case of the mean ; a su¢ cient
condition for identi…cation is E jyj < 1:
The mean of y is …nite if E jyj < 1: More generally, y has a …nite r’th moment if
E jyjr < 1:
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
a 1
f (y) = ay ; y > 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 the …gure 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
8 R a
1 r a 1
>
< a 1 y dy = if r < a
a r
E jyjr =
>
:
1 if r a
Thus to allow for stricter …nite moments (larger r) we need to restrict the class of permissible
densities (require larger a).
2.0
f(y)
a=2
1.5
1.0
0.5 a=1
0.0
1 2 3 4
y
Pareto Densities, a = 1 and a = 2
Thus, broadly speaking, the restriction that y has a …nite r’th moment means that the tail of y’s
density declines to zero faster than y r 1 : The faster decline of the tail means that the probability
of observing an extreme value of y is a more rare event.
It is helpful to know that the existence of …nite moments is monotonic. Liapunov’s Inequality
(B.23) implies that if E jyjp < 1 for some p > 0, then E jyjr < 1 for all 0 r p: For example,
Ey 2 < 1 implies E jyj < 1 and thus both the mean and variance of y are …nite.
2.15 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 2.6.1: 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 < 1 so that for all n N;
Pr (jyj > ) : Fix and : Set " = =3: Pick C < 1 large enough so that
E (jyi j 1 (jyi j > C)) " (2.16)
(where 1 ( ) is the indicator function) which is possible since E jyi j < 1: De…ne the random variables
wi = yi 1 (jyi j C) E (yi 1 (jyi j C))
zi = yi 1 (jyi j > C) E (yi 1 (jyi j > C))
so that
y =w+z
and
E jyj E jwj + E jzj : (2.17)
We now show that sum of the expectations on the right-hand-side can be bounded below 3":
First, by the Triangle Inequality (A.9) and the Expectation Inequality (B.18),
E jzi j = E jyi 1 (jyi j > C) E (yi 1 (jyi j > C))j
E jyi 1 (jyi j > C)j + jE (yi 1 (jyi j > C))j
2E jyi 1 (jyi j > C)j
2"; (2.18)
and thus by the Triangle Inequality (A.9) and (2.18)
n n
1X 1X
E jzj = E zi E jzi j 2": (2.19)
n n
i=1 i=1
Second, by a similar argument

jwi j = jyi 1 (jyi j C) E (yi 1 (jyi j C))j
jyi 1 (jyi j C)j + jE (yi 1 (jyi j C))j
2 jyi 1 (jyi j C)j
2C (2.20)
where the …nal inequality is (2.16). Then by Jensen’s Inequality (B.15), the fact that the wi are
iid and mean zero, and (2.20),
Ewi2 4C 2
(E jwj)2 E jwj2 = = "2 (2.21)
n n
2 2
the …nal inequality holding for n 4C 2 ="2 = 36C 2 = . Equations (2.17), (2.19) and (2.21)
together show that
E jyj 3"2 (2.22)
as desired.
Finally, by Markov’s Inequality (B.24) and (2.22),
E jyj 3"
Pr (jyj > ) = ;
the …nal equality by the de…nition of ": We have shown that for any > 0 and > 0 then for all
n 36C 2 = 2 2 ; Pr (jyj > ) ; as needed.
Proof of Theorem 2.7.1: By Loève’s cr Inequality (B.14)

0 11=2
Xm Xm
kyk = @ 2A
yj jyj j :
j=1 j=1
Thus if E jyj j < 1 for j = 1; :::; m; then

m
X
E kyk E jyj j < 1:
j=1
For the reverse inequality, the Euclidean norm of a vector is larger than the length of any individual
component, so for any j; jyj j kyk : Thus, if E kyk < 1; then E jyj j < 1 for j = 1; :::; m:
Proof of Theorem 2.8.1: The moment bound Ey 0i y i < 1 is su¢ cient to guarantee that the
elements of and V are well de…ned and …nite. Without loss of generality, it is su¢ cient to
consider the case = 0:
p
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 Lévy’s Continuity Theorem (see Van der
p
Vaart (2008) Theorem 2.13) this is su¢ cient to established that ny n converges in distribution to
N (0; V ) :
For 2 Rm ; let C ( ) = E exp i 0 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 i 0 y i
@
@2
0 C( ) = i2 E y i y 0i exp i 0 y i :
@ @
When evaluated at =0
C(0) = 1
@
C(0) = iE (y i ) = 0
@
@2
0 C(0) = E y i y 0i = V:
@ @
Furthermore,
@ @
c ( ) = c( ) = C( ) 1 C( )
@ @
@2 1 @
2
2 @ @
c ( ) = 0 c( ) = C( ) C( ) C( ) C( ) C( )
@ @ @ @ 0 @ @ 0
so when evaluated at =0
c(0) = 0
c (0) = 0
c (0) = V:
By a second-order Taylor series expansion of c( ) about = 0;

1 1
c( ) = c(0) + c (0)0 + 0
c ( ) = 0
c ( ) (2.23)
2 2
where lies on the line segment joining 0 and :
p p
We now compute Cn ( ) = E exp i 0 ny n ; the characteristic function of ny n : By the prop-
erties of the exponential function, the independence of the y i ; the de…nition of c( ) and (2.23)
n
!
1 X 0
log Cn ( ) = log E exp i p yi
n
i=1
n
Y 1 0
= log E exp i p yi
n
i=1
n
Y 1 0
= log E exp i p yi
n
i=1
n
X 1 0
= log E exp i p yi
n
i=1
= nc p
n
1 0
= c ( n)
2
p
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;
1 0
log Cn ( ) ! 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 2.9.1: 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 ! 1 by the assumption that z n ! c:
p
Hence g(z n ) ! g(c) as n ! 1.
Proof of Theorem 2.10.3: 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 :
p
It follows that ajn = gj ( jn ) gj ! 0: Stacking across elements of g; we …nd
p p d
n (g ( n) g( )) = (G + an )0 n( n ) ! G0 : (2.24)
d p d
The convergence is by Theorem 2.10.1, as G + an ! G; n ( n ) ! ; and their product is
continuous. This establishes (2.10)
When N (0; V ) ; the right-hand-side of (2.24) equals
G0 = G0 N (0; V ) = N 0; G0 V G
establishing (2.11).
Chapter 3
Conditional Expectation and

Projection
3.1 Introduction
The most commonly applied econometric tool is least-squares estimation, also known as regres-
sion. 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.
3.2 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 person’s 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 un-
observed 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 di¤erentiable we de…ne the probability density function
d
f (u) = F (u):
du
The density contains the same information as the distribution function, but the density is typically
easier to visually interpret.
28
CHAPTER 3. CONDITIONAL EXPECTATION AND PROJECTION 29
1.0
0.9
0.8
0.7
Wage Distribution
0.6
Wage Density
0.5
0.4
0.3
0.2
0.1
0.0
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70 80 90 100
Dollars per Hour Dollars per Hour
Figure 3.1: Wage Distribution and Density. All full-time U.S. workers
In Figure 3.1 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 3.1 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. For this reason the median is not the dominant measure of central tendency in
econometrics.
As discussed in Sections 2.2 and 2.14, the expectation or mean of a random variable y with
density f is
= E (y)
Z 1
= uf (u)du:
1
The mean U.S. wage ($23.90) is indicated in the right panel of Figure 3.1 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.
The mean is 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 distribution
1
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 individual
wage and salary earnings divided by hours worked.
2 1
If F is not continuous the de…nition is m = inffu : F (u) g
3
2
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.
as can be seen easily in the right panel of Figure 3.1. 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 $23.90
as a “typical” wage rate.
Log Wage Density
1 2 3 4 5 6
Log Dollars per Hour
Figure 3.2: Log Wage Density
In this context it is useful to transform the data by taking the natural logarithm5 . Figure 3.2
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.
3.3 Conditional Expectation

We saw in Figure 3.2 the density of log wages. Is this wage distribution the same for all workers,
or does the wage distribution vary across subpopulations? To answer this question, we can compare
wage distributions for di¤erent groups –for example, men and women. The plot on the left in Figure
3.3 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.
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 speci…c values as
E (log(wage) j gender = man) = 3:05 (3.1)
E (log(wage) j gender = woman) = 2:81: (3.2)

We call these means conditional as they are conditioning on a …xed value of the variable gender.
While you might not think of gender as a random variable, it is random from the viewpoint of
5
Throughout the text, we will use log(y) to denote the natural logarithm of y:
6
More precisely, the geometric mean exp (E (log w)) = $19:11 is a robust measure of central tendency.
white men
white women
black men
black women
Log Wage Density
Log Wage Density

Women Men
0 1 2 3 4 5 6 1 2 3 4 5
Log Dollars per Hour Log Dollars per Hour
Figure 3.3: Left: Log Wage Density for Women and Men. Right: Log Wage Density by Gender
and Race
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 3.3 appear similar, a hasty inference might be that there is not
a meaningful di¤erence between the wage distributions of men and women. Before jumping to this
conclusion let us examine the di¤erences in the distributions of Figure 3.3 more carefully. As we
mentioned above, the primary di¤erence between the two densities appears to be their means. This
di¤erence equals
E (log(wage) j gender = man) E (log(wage) j gender = woman) = 3:05 2:81
= 0:24 (3.3)
A di¤erence in expected log wages of 0.24 implies an average 24% di¤erence between the wages
of men and women, which is quite substantial. (For an explanation of logarithmic and percentage
di¤erences see the box on Log Di¤erences below.)
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 3.3. Again we see that the primary di¤erence between the four density
functions is their central tendency.
Focusing on the means of these distributions, Table 3.1 reports the mean log wage for each of
the six sub-populations.
men women
white 3.07 2.82
black 2.86 2.73
other 3.03 2.86
Table 3.1: Mean Log Wages by Sex and Race
The entries in Table 3.1 are the conditional means of log(wage) given gender and race. For
example
E (log(wage) j gender = man; race = white) = 3:07
and
E (log(wage) j gender = woman; race = black) = 2:73
One bene…t 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 3.1 allows us to easily calculate average wage di¤erences 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%.
Log Di¤erences
A useful approximation for the natural logarithm for small x is
log (1 + x) x: (3.4)
This can be derived from the in…nite series expansion of log (1 + x) :
x2 x3 x4
log (1 + x) = x + +
2 3 4
= x + O(x2 ):
The symbol O(x2 ) means that the remainder is bounded by Ax2 as x ! 0 for some A < 1:
A plot of log (1 + x) and the linear approximation x is shown in the following …gure. We
can see that log (1 + x) and the linear approximation x are very close for jxj 0:1, and
reasonably close for jxj 0:2, but the di¤erence increases with jxj.
0.4
0.2
-0.4 -0.2 0.2 0.4

x
-0.2
-0.4
log(1 + 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)
or
c
log y log y = log(1 + c=100)
100
where the approximation is (3.4). This shows that 100 multiplied by the di¤erence in
logarithms is approximately the percentage di¤erence between y and y , and this approx-
imation is quite good for jcj 10%:
3.4 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 7 .
7
Here, education is de…ned as years of schooling beyond kindergarten. A high school graduate has education=12,
a college graduate has education=16, a Master’s degree has education=18, and a professional degree (medical, law or
The conditional mean of log wages given gender, race, and education is a single number for each
category. For example
E (log(wage) j gender = man; race = white; education = 12) = 2:84
We display in Figure 3.4 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 di¤erent rate for schooling levels above and below nine years. Another
striking feature of Figure 3.4 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.
4.0
●
3.5
●
●
3.0
● ●
2.5
●
●
● ●
2.0
● white men
white women
4 6 8 10 12 14 16 18 20
Years of Education
Figure 3.4: Mean Log Wage as a Function of Years of Education
In many cases it is convenient to simplify the notation by writing variables using single charac-
ters, 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 j 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 3.1. The CEF is a function of (gender ;
race) as it varies across the entries.
For greater compactness, we will typically write the conditioning variables as a vector in Rk :
0 1
x1
B x2 C
B C
x = B . C: (3.5)
@ .. A
xk
PhD) has education=20.
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 j x) = m (x) :
Sometimes, it is useful to notationally distinguish E (y j x) as the CEF evaluated at the random

vector x from E (y j x = x0 ) as the CEF evaluated at the …xed value x0 : (And it is mathematically
correct to do so.) The …rst expression E (y j x) is a random variable and the second expression
E (y j x = x0 ) is a function. We will not always enforce this distinction as it can become notationally
burdensome. Hopefully, the use of E (y j x) should be apparent from the context.
3.5 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 labor market
experience. The contours of their joint density are plotted on the left side of Figure 3.5 for the
population of white men with 12 years of education.
4.0
Exp=5
Exp=10
Exp=25
Exp=40
3.5
Log Wage Conditional Density

3.0
2.5
2.0
0 10 20 30 40 50 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5
Labor Market Experience (Years) Log Dollars per Hour
Figure 3.5: Left: Joint density of log(wage) and experience and conditional mean of log(wage)
given experience for white men with education=12. Right: Conditional densities of log(wage) for
white men with education=12.
Given the joint density f (y; x) the variable x has the marginal density
Z
fx (x) = f (y; x)dy:
R
For any x such that fx (x) > 0 the conditional density of y given x is de…ned as
f (y; x)
fyjx (y j x) = : (3.6)
fx (x)
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 speci…c value of x parallel with the y-axis. For example,
take the density contours on the left side of Figure 3.5 and slice through the contour plot at a
speci…c 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 3.5. We can see that the
distribution of wages shifts to the right and becomes more di¤use 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 (3.6)
Z
m (x) = E (y j x) = yfyjx (y j x) dy: (3.7)
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 in…nitely
small.
In Figure 3.5 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.
3.6 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.
Theorem 3.6.1 Simple Law of Iterated Expectations

If E jyj < 1 then for any random vector x,
E (E (y j x)) = E (y)
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
1
X
E (E (y j x)) = E (y j x j ) Pr (x = x j )
j=1
and when x is continuous Z

E (E (y j x)) = E (y j x ) fx (x)dx:
Rk
Going back to our investigation of average log wages for men and women, the simple law states
that
E (log(wage) j gender = man) Pr (gender = man)

+E (log(wage) j gender = woman) Pr (gender = woman)
= E (log(wage)) :
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 3.6.2 Law of Iterated Expectations

If E jyj < 1 then for any random vectors x1 and x2 ,
E (E (y j x1 ; x2 ) j x1 ) = E (y j 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 j x1 ) ; the expectation conditional on x1 alone. Sometimes we phrase this as: “The smaller
information set wins.”
As an example
E (log(wage) j gender = man; race = white) Pr (race = whitejgender = man)

+E (log(wage) j gender = man; race = black) Pr (race = blackjgender = man)
+E (log(wage) j gender = man; race = other) Pr (race = otherjgender = man)
= E (log(wage) j 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 e¤ectively treat it as if it is constant. For example, E (x j x) = x and E (g (x) j x) = g (x) for
any function g( ): The general property is known as the conditioning theorem.
Theorem 3.6.3 Conditioning Theorem

If
E jg (x) yj < 1 (3.8)
then
E (g (x) y j x) = g (x) E (y j x) (3.9)
and
E (g (x) y) = E (g (x) E (y j x)) (3.10)
The proofs of Theorems 3.6.1, 3.6.2 and 3.6.3 are given in Section 3.30.
3.7 Monotonicity of Conditioning

What is the e¤ect of increasing the amount of information when constructing a conditional
expectation? That is, how do we compare E (y j x1 ) versus E (y j x1 ; x2 )? We have seen that
by increasing the conditioning set, the conditional expectation reveals greater detail about the
distribution of y: Is there something more that can be said?
It turns out that there is a simple relationship induced by conditioning. We can think of
the conditional mean E (y j x1 ) as the “explained portion” of y: The remainder y E (y j x1 ) 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 3.7.1 If Ey 2 < 1 then
var (y) var (y E (y j x1 )) var (y E (y j x1 ; x2 ))
Theorem 3.7.1 says that the variance of the di¤erence between y and its conditional mean
(weakly) decreases whenever an additional variable is added to the conditioning information.
The proof of Theorem 3.7.1 is given in Section 3.30.
3.8 CEF Error

The CEF error e is de…ned as the di¤erence 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: (3.11)
In (3.11) 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 de…nition m(x) = E (y j x) and the Conditioning Theorem
E (e j x) = E ((y m(x)) j x)
= E (y j x) E (m(x) j x)
= m(x) m(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 j x)) = E (0) = 0
We state this and some other results formally.
Theorem 3.8.1 Properties of the CEF error

If E jyj < 1 then
1. E (e j x) = 0:
2. E (e) = 0:
3. If E jyjr < 1 for r 1 then E jejr < 1:
4. For any function h (x) such that E jh (x) ej < 1 then E (h (x) e) = 0
The proof of the third result is deferred to Section 3.30:

The fourth result, whose proof is left to Exercise 3.3, says that e is uncorrelated with any
function of the regressors.
The equations
y = m(x) + e
E (e j 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 de…nition.
The condition E (e j x) = 0 is implied by the de…nition of e as the di¤erence between y and the
CEF m (x) : The equation E (e j 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 simpli…cation, but it is not generic feature of the con-
ditional 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 3.6
for the same population as Figure 3.5. 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.
1.0
0.5
0.0
e
−0.5
−1.0
0 10 20 30 40 50
Labor Market Experience (Years)
Figure 3.6: Joint density of CEF error e and experience for white men with education=12.
As a simple example of a case where x and e are mean independent yet dependent, let y = xu
where x and u are independent and Eu = 1: Then
E (y j x) = E (xu j x) = xE (u j x) = x
so the CEF equation is
y =x+e
where
e = x(u 1):
Note that even though e is not independent of x;
E (e j x) = E (x(u 1) j x) = xE ((u 1) j x) = 0
and is thus mean independent.
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 3.8.1.3 implies the following simple but useful result.

Theorem 3.8.2 If Ey 2 < 1 then 2 <1
3.9 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 di¤erence 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 : (3.12)
We can de…ne the best predictor as the function g (x) which minimizes (3.12). 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):
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
where the …rst equality makes the substitution y = m(x) + e and the third equality uses Theorem
3.8.1.4. The right-hand-side after the third equality is minimized by setting g (x) = m (x), yielding
the …nal inequality. The minimum is …nite under the assumption Ey 2 < 1 as shown by Theorem
3.8.2.
We state this formally in the following result.
Theorem 3.9.1 Conditional Mean as Best Predictor

If Ey 2 < 1; then for any predictor g (x),
E (y g (x))2 E (y m (x))2
where m (x) = E (y j x).
3.10 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.
De…nition 3.10.1 If Ey 2 < 1; the conditional variance of y

given x is
2
(x) = var (y j x)
= E (y E (y j x))2 j x
= E e2 j x
Generally, 2 (x) is a non-trivial function of x and can take any form subject to the restriction
p
that it is non-negative. The conditional standard deviation is its square root (x) = 2 (x):
2 2
One way to think about (x) is that it is the conditional mean of e given 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 3.3. The di¤erence between the densities
is not purely a location shift, but is also a di¤erence in spread. Speci…cally, we can see that the
density for men’s log wages is somewhat more spread out than that for women, while the density
for women’s wages is somewhat more peaked. Indeed, the conditional standard deviation for men’s
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 j x = E 2 (x) :
That is, the unconditional error variance is the average conditional variance.
Given the conditional variance, we can de…ne a rescaled error
e
"= : (3.13)
(x)
We can calculate that since (x) is a function of x
e 1
E (" j x) = E jx = E (e j x) = 0
(x) (x)
and
e2 1 2 (x)
var (" j x) = E "2 j x = E 2 (x)
jx = 2 (x)
E e2 j x = 2 (x)
= 1:
Thus " has a conditional mean of zero, and a conditional variance of 1.

Notice that (3.13) can be rewritten as
e = (x)":
and substituting this for e in the CEF equation (3.11), we …nd that
y = m(x) + (x)": (3.14)
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 condi-
tional 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 condi-
tional 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!

3.11 Homoskedasticity and Heteroskedasticity

An important special case obtains when the conditional variance 2 (x) is a constant and inde-
pendent of x. This is called homoskedasticity.
De…nition 3.11.1 The error is homoskedastic if E e2 j 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.
De…nition 3.11.2 The error is heteroskedastic if E e2 j 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 de…nition, the unconditional variance
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. Recall Figure 3.3 and how the
variance of wages varies between men and women.
Some older or introductory textbooks describe heteroskedasticity as the case where “the vari-
ance of e varies across observations”. This is a poor and confusing de…nition. It is more constructive
to understand that heteroskedasticity means that the conditional variance 2 (x) depends on ob-
servables.
Older textbooks also tend to describe homoskedasticity as a component of a correct regression
speci…cation, and describe heteroskedasticity as an exception or deviance. This description has
in‡uenced many generations of economists, but it is unfortunately backwards. The correct view
is that heteroskedasticity is generic and “standard”, while homoskedasticity is unusual and excep-
tional. 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 ho-
moskedasticity assumption when making theoretical investigations into the properties of estimation
and inference methods. The reason is that in many cases homoskedasticity greatly simpli…es 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.
3.12 Regression Derivative

One way to interpret the CEF m(x) = E (y j 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 single regressors, holding the remainder …xed. When a regressor x1
is continuously distributed, we de…ne the marginal e¤ect 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 de…ne the marginal e¤ect as a discrete di¤erence. For example, if x1 is
binary, then the marginal e¤ect of x1 on the CEF is
m(1; x2 ; :::; xk ) m(0; x2 ; :::; xk ):
We can unify the continuous and discrete cases with the notation
8
>
> @
< m(x1 ; :::; xk ); if x1 is continuous
@x1
r1 m(x) =
>
>
: m(1; x ; :::; x ) m(0; x ; :::; x ); if x is binary.
2 k 2 k 1
Collecting the k e¤ects into one k 1 vector, we de…ne the regression derivative of x on y :
2 3
r1 m(x)
6 r2 m(x) 7
6 7
rm(x) = 6 .. 7
4 . 5
rk m(x)
@
When all elements of x are continuous, then we have the simpli…cation rm(x) = m(x), the
@x
vector of partial derivatives.
There are two important points to remember concerning our de…nition of the regression deriv-
ative.
First, the e¤ect 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.
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 rm(x) is the actual e¤ect on the response variable y only if the error e is una¤ected by
the change in the regressor x. We return to a discussion of causal e¤ects in Section 3.28.
3.13 Linear CEF

An important special case is when the CEF m (x) = E (y j x) is linear in x: In this case we can
write the mean equation as
m(x) = x1 1 + x2 2 + + xk k + :
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 coe¢ cient is called the “intercept”. Equivalently, assuming that
the …nal element8 of the vector x is the intercept, then xk = 1. Thus (3.5) has been rede…ned as
the k 1 vector 0 1
x1
B x2 C
B C
B C
x = B ... C : (3.15)
B C
@ xk 1 A
1
8
The order doesn’t matter. It could be any element.
With this rede…nition, then the CEF is
m(x) = x1 1 + x2 2 + + xk k
0
= x (3.16)
where 0 1
1
B C
= @ ... A (3.17)
k
is a k 1 coe¢ cient 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 coe¢ cient vector. That is
rm(x) = :
This is one of the appealing features of the linear CEF model. The coe¢ cients have simple and
natural interpretations as the marginal e¤ects of changing one variable, holding the others constant.
Linear CEF Model
y = x0 + e
E (e j 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 j x) = 0
E e2 j x = 2
3.14 Linear CEF with Nonlinear E¤ects

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 e¤ects.
For example, suppose we have two scalar variables x1 and x2 : The CEF could take the quadratic
form
m(x1 ; x2 ) = x1 1 + x2 2 + x21 3 + x22 4 + x1 x2 5 + 6 : (3.18)
This equation is quadratic in the regressors (x1 ; x2 ) yet linear in the coe¢ cients ( 1 ; :::; 6 ): We
will descriptively call (3.18) a quadratic CEF, and yet (3.18) is also a linear CEF in the sense
of being linear in the coe¢ cients. The key is to understand that (3.18) is quadratic in the variables
(x1 ; x2 ) yet linear in the coe¢ cients ( 1 ; :::; 6 ):
To simplify the expression, we de…ne the transformations x3 = x21 ; x4 = x22 ; x5 = x1 x2 ; and

x6 = 1; and rede…ne the regressor vector as x = (x1 ; :::; x6 ): With this rede…nition,
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 (3.18),
the regression derivative should be de…ned 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 (3.18), the regression derivatives are not a simple coe¢ cient, but are
functions of several coe¢ cients plus the levels of (x1; x2 ): Consequently it is di¢ cult to interpret
the coe¢ cients individually. It is more useful to interpret them as a group.
We typically call 5 the interaction e¤ect. Notice that it appears in both regression derivative
equations, and has a symmetric interpretation in each. If 5 > 0 then the regression derivative of
x1 on y is increasing in the level of x2 (and the regression derivative of x2 on y is increasing in the
level of x1 ); while if 5 < 0 the reverse is true.
3.15 Linear CEF with Dummy Variables

When all regressors are discrete, it turns out the CEF can be written as a linear function of
regressors.
Consider the simplest case of a single binary regressor. A variable is binary if it only takes two
distinct values. (For example, the variable gender.) It is common to call such regressors dummy
variables, and sometimes they are called indicator variables. When there is only a single dummy
regressor the conditional mean can only take two distinct values. For example,
8
< 0 if gender=man
E (y j gender) =
:
1 if gender=woman
To facilitate a mathematical treatment, we typically record dummy variables with the values f0; 1g:
For example
0 if gender=man
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 j x1 ) = + x1
where = 0 and = 1 0 : In this simple regression equation the intercept is equal to the
conditional mean of y for the x1 = 0 subpopulation (men) and the slope is equal to the di¤erence
in the conditional means of the two subpopulations.
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:
8
>
> if x1 = 0 and x2 = 0 (unmarried men)
< 00
01 if x 1 = 0 and x2 = 1 (married men)
E (y j 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 j x1 ; x2 ) = + 1 x1 + 2 x2 + 3 x1 x2
where = 00 ; 1 = 10 00 ; 2 = 01 00 ; and 3 = 11 10 01 + 00 :
We can view the coe¢ cient 1 as the e¤ect of gender on expected log wages for unmarried
wages earners, the coe¢ cient 2 as the e¤ect of marriage on expected log wages for men wage
earners, and the coe¢ cient 3 as the di¤erence between the e¤ects of marriage on expected log
wages among women and among men. Alternatively, it can also be interpreted as the di¤erence
between the e¤ects 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 e¤ect, and describe 3 = 0 as the case when the
interaction e¤ect 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 3.13, we would de…ne the regressor x3 = x1 x2 and the
regressor vector as 0 1
x1
B x2 C
x=B C
@ x3 A :
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 j x1 ; x2 ; x3 ) takes at most 23 = 8 distinct
values and can be written as the linear function
E (y j x1 ; x2 ; x3 ) = + 1 x1 + 2 x2 + 3 x3 + 4 x1 x2 + 5 x1 x3 + 6 x2 x3 + 7 x1 x2 x3
which has eight regressors including the intercept.

In general, if there are p dummy variables x1 ; :::; xp then the CEF E (y j 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 parame-
terizations.
We started this section by saying that the conditional mean is linear whenever all regressors are
discrete, meaning that they take a …nite number of possible values. How can we see this? Take a
categorical variable, 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
8
< 1 if white
x3 = 2 if black
:
3 if 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: 8
< 1 if x3 = 1
E (y j x3 ) = if x3 = 2
: 2
3 if x3 = 3
This is not a linear function of x3 itself, but it can be made so by constructing dummy variables
for two of the three categories. For example
1 if black
x4 =
0 if not black
1 if other
x5 =
0 if not other
In this case, the categorical variable x3 is equivalent to the pair of dummy variables (x4 ; x5 ): The
explicit relationship is 8
< 1 if x4 = 0 and x5 = 0
x3 = 2 if x4 = 1 and x5 = 0
:
3 if x4 = 0 and x5 = 1
Given these transformations, we can write the conditional mean of y as a linear function of x4 and
x5
E (y j x3 ) = E (y j x4 ; x5 ) = + 1 x4 + 2 x5
We can write the CEF as either E (y j x3 ) or E (y j 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 di¤erence that we have not
included the interaction term x4 x5 : This is because the event fx4 = 1 and x5 = 1g is empty by
construction, so x4 x5 = 0 by de…nition.
3.16 Best Linear Predictor

While the conditional mean m(x) = E (y j 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 speci…cation (3.16) as an approx-
imation. In this section we derive a speci…c approximation with a simple interpretation.
Theorem 3.9.1 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 de…ne an approx-
imation 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.
Assumption 3.16.1
1. Ey 2 < 1:
2. E kxk2 < 1:
3. Qxx = E (xx0 ) is positive de…nite.
The …rst two parts of Assumption 3.16.1 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 2 Rk . The mean squared
prediction error is
2
S( ) = E y x0 :
The best linear predictor of y given x, written P(y j x); is found by selecting the vector to
minimize S( ):
De…nition 3.16.1 The Best Linear Predictor of y given x is
P(y j x) = x0
where minimizes the mean squared prediction error

2
S( ) = E y x0 :
The minimizer
= argmin S( ) (3.19)
2Rk
is called the Linear Projection Coe¢ cient.
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 2 0 E (xy) + 0
E xx0 :
The quadratic structure of S( ) means that we can solve explicitly for the minimizer. The …rst-
order condition for minimization (from Appendix A.9) is
@
0= S( ) = 2E (xy) + 2E xx0 : (3.20)
@
Rewriting (3.20) as
2E (xy) = 2E xx0
and dividing by 2, this equation takes the form
Qxy = Qxx (3.21)
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
= Qxx1 Qxy
or
1
= E xx0 E (xy) : (3.22)
It is worth taking the time to understand the notation involved in the expression (3.22). Qxx is a
E(xy)
k k matrix and Qxy is a k 1 column vector. Therefore, alternative expressions such as E(xx 0)
or E (xy) (E (xx0 )) 1 are incoherent and incorrect. We also can now see the role of Assumption
3.16.1.3. It is necessary in order for the solution (3.22) to exist. Otherwise, there would be multiple
solutions to the equation (3.21).
We now have an explicit expression for the best linear predictor:
1
P(y j x) = x0 E xx0 E (xy) :
This expression is also referred to as the linear projection of y on x.

The projection error is
e=y x0 : (3.23)
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: (3.24)
In general we call equation (3.24) or x0 the best linear predictor of y given x; or the linear
projection of y on x. Equation (3.24) 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 3.13 that the linear CEF model is also called the linear regression model.)
An important property of the projection error e is
E (xe) = 0: (3.25)
1
To see this, using the de…nitions (3.23) and (3.22) and the matrix properties AA = I and
Ia = a;
E (xe) = E x y x0
1
= E (xy) E xx0 E xx0 E (xy)
= 0 (3.26)
as claimed.
Equation (3.25) is a set of k equations, one for each regressor. In other words, (3.25) is equivalent
to
E (xj e) = 0 (3.27)
for j = 1; :::; k: As in (3.15), the regressor vector x typically contains a constant, e.g. xk = 1. In
this case (3.27) for j = k is the same as
E (e) = 0: (3.28)
Thus the projection error has a mean of zero when the regressor vector contains a constant. (When
x does not have a constant, (3.28) 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 model.)
It is also useful to observe that since cov(xj ; e) = E (xj e) E (xj ) E (e) ; then (3.27)-(3.28)
together imply that the variables xj and e are uncorrelated.
This completes the derivation of the model. We summarize some of the most important prop-
erties.
Theorem 3.16.1 Properties of Linear Projection Model

Under Assumption 3.16.1,
1. The moments E (xx0 ) and E (xy) exist with …nite elements.
2. The Linear Projection Coe¢ cient (3.19) exists, is unique, and equals
1
= E xx0 E (xy) :
3. The best linear predictor of y given x is

1
P(y j x) = x0 E xx0 E (xy) :
4. The projection error e = y x0 exists and satis…es
E e2 < 1
and
E (xe) = 0:
5. If x contains an constant, then
E (e) = 0:
6. If E kykr < 1 and E kxkr < 1 for r 1 then E jejr < 1:
A complete proof of Theorem 3.16.1 is given in Section 3.30.

It is useful to re‡ect on the generality of Theorem 3.16.1. The only restriction is Assumption
3.16.1. Thus for any random variables (y; x) with …nite variances we can de…ne a linear equation
(3.24) with the properties listed in Theorem 3.16.1. Stronger assumptions (such as the linear CEF
model) are not necessary. In this sense the linear model (3.24) exists quite generally. However,
it is important not to misinterpret the generality of this statement. The linear equation (3.24) is
de…ned 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 xx0 E (xy)
We illustrate projection using three log wage equations introduced in earlier sections.
For our …rst example, we consider a model with the two dummy variables for gender and race
similar to Table 3.1. As we learned in Section 3.15, the entries in this table can be equivalently
expressed by a linear CEF. For simplicity, let’s consider the CEF of ln(wage) as a function of Black
and Female.
E(log(wage) j Black; F emale) = 0:20Black 0:24F emale + 0:10Black F emale + 3:06:

This is a CEF as the variables are dummys and all interactions are included.
Now consider a simpler model omitting the interaction e¤ect. This is the linear projection on
the variables Black and F emale
P(log(wage) j Black; F emale) = 0:15Black 0:23F emale + 3:06:
What is the di¤erence? The projection model suggests a 15% average wage gap for blacks, while
the full CEF shows that the race gap is di¤erentiated by gender: it is 20% for black men (relative
to non-black men) and 10% for black women (relative to non-black women).
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 3.4 and is repeated in Figure 3.7. 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) j Education) = 1:5 + 0:11Education
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 under-
predicts for individuals with lower levels of education. To correct this imbalance we use a linear
spline equation which allows di¤erent rates of return above and below 9 years of education:
P (log(wage) j Education; (Education 9) 1 (Education > 9))

= 2:3 + 0:02Education + :10 (Education 9) 1 (Education > 9)
This equation is displayed in Figure 3.7 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
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.
4.0
●
3.5
●
●
3.0
●
●
● ●
2.5
●
●
● ●
2.0
4 6 8 10 12 14 16 18 20
Years of Education
Figure 3.7: Projections of log(wage) onto Education
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 3.5 and is repeated as the
solid line in Figure 3.8. Superimposed on the …gure are two projections. The …rst (given by the
dot-dashed line) is the linear projection on experience
P(log(wage) j Experience) = 2:5 + 0:011Experience

and the second (given by the dashed line) is the linear projection on experience and its square
P(log(wage) j Experience) = 2:3 + 0:046Experience 0:0007Experience2 :
It is fairly clear from an examination of Figure 3.8 that the …rst linear projection is a poor approx-
imation. 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:
4.0
Conditional Mean
Linear Projection
3.5 Quadratic Projection
3.0
2.5
2.0
0 10 20 30 40 50
Labor Market Experience (Years)
Figure 3.8: Linear and Quadratic Projections of log(wage) onto Experience

Invertibility and Identi…cation
The linear projection coe¢ cient = (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 2 Rk ;
0 0 2
Qxx =E xx0 =E 0
x 0
so Qxx by construction is positive semi-de…nite. The assumption that it is positive de…nite

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-de…nite matrices are invertible if and only if they are positive
de…nite. When Qxx is invertible then = (E (xx0 )) 1 E (xy) exists and is uniquely
de…ned. In other words, in order for to be uniquely de…ned, we must exclude the
degenerate situation of redundant varibles.
Theorem 3.16.1 shows that the linear projection coe¢ cient is identi…ed (uniquely
determined) under Assumptions 3.16.1. The key is invertibility of Qxx . Otherwise, there
is no unique solution to the equation
Qxx = Qxy : (3.29)
When Qxx is not invertible there are multiple solutions to (3.29), all of which yield an
equivalent best linear predictor x0 . In this case the coe¢ cient is not identi…ed as it
does not have a unique value. Even so, the best linear predictor x0 still identi…ed. One
solution is to set
= E xx0 E (xy)
where A denotes the generalized inverse of A (see Appendix A.5).
3.17 Linear Predictor Error Variance

As in the CEF model, we de…ne 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 yx0 + 0
E xx0
= Qyy 2Qyx Qxx1 Qxy + Qyx Qxx1 Qxx Qxx1 Qxy
= Qyy Qyx Qxx1 Qxy
def
= Qyy x (3.30)
One useful feature of this formula is that it shows that Qyy x = Qyy Qyx Qxx1 Qxy equals the
variance of the error from the linear projection of y on x.
3.18 Regression Coe¢ cients

Sometimes it is useful to separate the intercept from the other regressors, and write the linear
projection equation in the format
y = + x0 + e (3.31)
where is the intercept and x does not contain a constant.
Taking expectations of this equation, we …nd
Ey = E + Ex0 + Ee
or
0
y = + x
where y = Ey and x = Ex; since E (e) = 0 from (3.28). Rearranging, we …nd

0
= y x :
Subtracting this equation from (3.31) we …nd

0
y y = (x x) + e; (3.32)
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; (3.32) is also
a linear projection, thus by the formula for the linear projection model,
0 1
= E (x x ) (x x) E (x x) y y
1
= var (x) cov (x; y)
a function only of the covariances9 of x and y:
Theorem 3.18.1 In the linear projection model
y= + x0 + e;
then
0
= y x (3.33)
and
1
= var (x) cov (x; y) : (3.34)
3.19 Regression Sub-Vectors

Let the regressors be partitioned as
x1i
xi = : (3.35)
x2i
9
The covariance matrix between vectors x and z is cov (x; z) = E (x Ex) (z Ez)0 : The (co)variance
matrix of the vector x is var (x) = cov (x; z) = E (x Ex) (x Ex)0 :
We can write the projection of yi on xi as
yi = x0i + ei
= x01i 1 + x02i 2 + ei (3.36)
E (xi ei ) = 0:
In this section we derive formula for the sub-vectors 1 and 2:

Partition Qxx comformably with xi
Q11 Q12 E (x1i x01i ) E (x1i x02i )

Qxx = =
Q21 Q22 E (x2i x01i ) E (x2i x02i )
and similarly Qxy

Q1y E (x1i yi )
Qxy = = :
Q2y E (x2i yi )
By the partitioned matrix inversion formula (A.4)
1
Q11 Q12 def Q11 Q12 Q1112 Q1112 Q12 Q221
Qxx1 = = = : (3.37)
Q21 Q22 Q21 Q22 Q2211 Q21 Q111 Q2211
def def
where Q11 2 = Q11 Q12 Q221 Q21 and Q22 1 = Q22 Q21 Q111 Q12 . Thus
1
=
2
Q1112 Q1112 Q12 Q221 Q1y
=
Q2211 Q21 Q111 Q2211 Q2y
Q1112 Q1y Q12 Q221 Q2y
=
Q2211 Q2y Q21 Q111 Q1y
Q1112 Q1y 2
=
Q2211 Q2y 1
We have shown that
1 = Q1112 Q1y 2
2 = Q2211 Q2y 1
3.20 Coe¢ cient Decomposition

In the previous section we derived formula for the coe¢ cient sub-vectors 1 and 2 : We now
use these formula to give a useful interpretation of the coe¢ cients as obtaining from an interated
projection.
Take equation (3.36) for the case dim(x1i ) = 1 so that 1 2 R:
yi = x1i 1 + x02i 2 + ei : (3.38)
Now consider the projection of x1i on x2i :
x1i = x02i 2 + u1i

E (x2i u1i ) = 0:
From (3.22) and (3.30), 2 = Q221 Q21 and Eu21i = Q11 2 : We can also calculate that
E (u1i yi ) = E x1i 0
2 x2i yi = E (x1i yi ) 0
2 E (x2i yi ) = Q1y Q12 Q221 Q2y = Q1y 2 :
We have found that

E (u1i yi )
1 = Q1112 Q1y 2 =
Eu21i
the coe¢ cient from the simple regression of yi on u1i :
What this means is that in the multivariate projection equation (3.38), the coe¢ cient 1 equals
the projection coe¢ cient from a regression of yi on u1i ; the error from a projection of x1i on the
other regressors x2i : The error u1i can be thought of as the component of x1i which is not linearly
explained by the other regressors. Thus the the coe¢ cient 1 equals the linear e¤ect of x1i on yi ;
after stripping out the e¤ects of the other variables.
There was nothing special in the choice of the variable x1i : So this derivation applies symmet-
rically to all coe¢ cients in a linear projection. Each coe¢ cient equals the simple regression of yi
on the error from a projection of that regressor on all the other regressors. Each coe¢ cient equals
the linear e¤ect of that variable on yi ; after linearly controlling for all the other regressors.
3.21 Omitted Variable Bias

Again, let the regressors be partitioned as in (3.35). Consider the projection of yi on x1i only.
Perhaps this is done because the variables x2i are not observed. This is the equation
yi = x01i 1 + ui (3.39)
E (x1i ui ) = 0
Notice that we have written the coe¢ cient on x1i as 1 rather than 1 and the error as ui rather
than ei : This is because (3.39) is di¤erent than (3.36). Goldberger (1991) introduced the catchy
labels long regression for (3.36) and short regression for (3.39) to emphasize the distinction.
Typically, 1 6= 1 , except in special cases. To see this, we calculate
1
1 = E x1i x01i E (x1i yi )
1
= E x1i x01i E x1i x01i 1+ x02i 2 + ei
1
= 1 + E x1i x01i E x1i x02i 2
= 1 + 2
where
1
= E x1i x01i E x1i x02i
is the coe¢ cient matrix from a projection of x2i on x1i :
Observe that 1 = 1 + 2 6= 1 unless = 0 or 2 = 0: Thus the short and long regressions
have di¤erent coe¢ cients on x1i : They are the same only under one of two conditions. First, if the
projection of x2i on x1i yields a set of zero coe¢ cients (they are uncorrelated), or second, if the
coe¢ cient on x2i in (3.36) is zero. In general, the coe¢ cient in (3.39) is 1 rather than 1 : The
di¤erence 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 the omitted variables
problem. Typically 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.
3.22 Best Linear Approximation

There are alternative ways we could construct a linear approximation x0 to the conditional
mean m(x): In this section we show that one natural approach turns out to yield the same answer
as the best linear predictor.
We start by de…ning the mean-square approximation error of x0 to m(x) as the expected
squared di¤erence between x0 and the conditional mean m(x)
2
d( ) = E m(x) x0 : (3.40)
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 di¤erence d( ) as a density-
weighted average of the function (m(x) x0 )2 ; since
Z
2
d( ) = m(x) x0 fx (x)dx
Rk
where fx (x) is the marginal density of x:

We can then de…ne the best linear approximation to the conditional m(x) as the function x0
obtained by selecting to minimize d( ) :
= argmin d( ): (3.41)
2Rk
Similar to the best linear predictor we are measuring accuracy by expected squared error. The
di¤erence is that the best linear predictor (3.19) selects to minimize the expected squared predic-
tion error, while the best linear approximation (3.41) selects to minimize the expected squared
approximation error.
Despite the di¤erent de…nitions, it turns out that the best linear predictor and the best linear
approximation are identical. By the same steps as in (3.16) plus an application of conditional
expectations we can …nd that
1
= E xx0 E (xm(x)) (3.42)
1
= E xx0 E (xy) (3.43)
(see Exercise 3.18). Thus (3.41) equals (3.19). We conclude that the de…nition (3.41) can be viewed
as an alternative motivation for the linear projection coe¢ cient.
3.23 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 j x) = E (e) = 0
and
E e2 j x = E e2 = 2
which are properties of a homoskedastic linear CEF.

We have shown that when (y; x) are jointly normal, they satisfy a normal linear CEF
y = x0 + e
where
2
e N (0; )
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.
3.24 Regression to the Mean

The term regression originated in an in‡uential paper by Francis Galton published in 1886,
where he examined the joint distribution of the stature (height) of parents and children. E¤ectively,
he was estimating the conditional mean of children’s 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 child’s height is more mediocre than his or her parent’s 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 Galton’s 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 (3.44)
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 (3.33)
= (1 )
so we can write the linear projection (3.44) as
P (y j x) = (1 ) +x :
This shows that the projected height of the child is a weighted average of the population average
height and the parent’s 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 (3.34)
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 child’s height is more mediocre (closer to the population average)
than the parent’s.
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 signi…cant intellectual advance by examing the joint distribu-
tions 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 have
shown that under the assumption of constant (e.g. stable, non-converging) means and variances,
the slope coe¢ cient must be less than one. It cannot be anything else. A slope less than one does
not imply that the variance of 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
(3.44). Since x and e are uncorrelated, it follows that
2
var(y) = var(x) + var(e):
Then var(y) < var(x) if and only if

2 var(e)
<1
var(x)
which is not implied by the simple condition j j < 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
–children’s height will be more average than their parents. The regression fallacy is to incorrectly
conclude that the population is converging. The message 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 pro…ts, and
plotted the average pro…ts 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.
3.25 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
x= +y +e : (3.45)
This is sometimes called the reverse regression. In this equation, the coe¢ cients ; and
error e are de…ned by linear projection. In a stable population we …nd that
= corr(x; y) =
= (1 ) =
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 (3.44), divide through
by and rewrite to …nd the equation
1 1
x= +y e (3.46)
suggesting that the projection of x on y should have a slope coe¢ cient of 1= instead of ; and
intercept of - = rather than : What went wrong? Equation (3.46) is perfectly valid, because it
is a simple manipulation of the valid equation (3.44). The trouble is that (3.46) is not a CEF nor
a linear projection. Inverting a projection (or CEF) does not yield a projection (or CEF). Instead,
(3.45) is a valid projection, not (3.46).
In any event, Galton’s …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.
3.26 Limitations of the Best Linear Predictor

Let’s compare the linear projection and linear CEF models.
From Theorem 3.8.1.4 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 j x) = 0:
To see this in a simple example, suppose we take a normally distributed random variable
x N (0; 1) and set y = x2 : Note that y is a deterministic function of x! Now consider the linear
projection of y on x and an intercept. The intercept and slope may be calculated as
1
1 E (x) E (y)
=
E (x) E x2 E (xy)
1
1 E (x) E x2
=
E (x) E x2 E x3
1
=
0
Thus the linear projection equation takes the form
y= +x +e
where = 1, = 0 and e = x2 1: Observe that E (e) = E x2 1 = 0 and E (xe) = E x3 E (e) =

0; yet E (e j x) = x2 1 6= 0: In this simple example e is a deterministic function of x; yet e and x
are uncorrelated! The point is that a projection error need not be a CEF error.
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 3.9 for a
constructed10 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 di¤erent 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 e¤ect of x on y is di¤erent for individuals in the two
10
The x in Group 1 are N(2; 1) and those in Group 2 are N(4; 1); and the conditional distriubtion of y given x is
N(m(x); 1) where m(x) = 2x x2 =6:
Figure 3.9: Conditional Mean and Two Linear Projections
groups. This conclusion is incorrect because in fact there is no di¤erence in the conditional mean
function. The apparant di¤erence is a by-product of a linear approximation to a non-linear mean,
combined with di¤erent marginal distributions for the conditioning variables.
3.27 Random Coe¢ cient Model

A model which is notationally similar to but conceptually distinct from the linear CEF model
is the linear random coe¢ cient model. It takes the form
y = x0
where the individual-speci…c coe¢ cient is random and independent of x. For example, if x is
years of schooling and y is log wages, then is the individual-speci…c returns to schooling. If
a person obtains an extra year of schooling, is the actual change in their wage. The random
coe¢ cient 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 coe¢ cient equals the regression derivative –the change
in the conditional mean due to a change in the regressors, = rm(x). This is not the e¤ect on an
given individual, it is the e¤ect on the population average. In contrast, in the random coe¢ cient
model, the random vector = rx0 is the true causal e¤ect –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 coe¢ cient model implies a linear
CEF. To see this, let and denote the mean and covariance matrix of :
= E( )
= var ( )
and then decompose the random coe¢ cient as
= +u
where u is distributed independently of x with mean zero and covariance matrix : Then we can
write
E(y j x) = x0 E( j x) = x0 E( ) = x0
so the CEF is linear in x; and the coe¢ cients equal the mean of the random coe¢ cient .
We can thus write the equation as a linear CEF
y = x0 + e (3.47)
where e = x0 u and u = . The error is conditionally mean zero:
E(e j x) = 0:
Furthermore
var (e j x) = x0 var ( )x
= x0 x
so the error is conditionally heteroskedastic with its variance a quadratic function of x.
Theorem 3.27.1 In the linear random coe¢ cient model y = x0 with indepen-
dent of x, E kxk2 < 1; and E k k2 < 1; then
E (y j x) = x0
var (y j x) = x0 x
where = E( ) and = var ( ):
3.28 Causal E¤ects

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 e¤ects of decisions, actions, and policies. For example, we may be
interested in the e¤ect 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
e¤ectiveness 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 e¤ect 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. The causal e¤ect
is typically speci…c to an individual, and also cannot be directly observed.
For example, the causal e¤ect of schooling on wages is the actual di¤erence a person would re-
ceive in wages if we could change their level of education. The causal e¤ect of a medical treatment is
the actual di¤erence in an individual’s health outcome, comparing treatment versus non-treatment.
In both cases the e¤ects are individual and unobservable. For example, suppose that Jennifer would
have earned $10 an hour as a high-school graduate and $20 a 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 e¤ect of schooling is $10 a hour for Jennifer and $4 an hour for George. Furthermore, the
causal e¤ect is unobserved as we only observe the wage corresponding to the actual outcome.
A variable x1 can be said to have a causal e¤ect 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) (3.48)
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 coe¢ cient model
(3.27) studied earlier. We de…ne the causal e¤ect of x1 within this model as the change in y due to
a change in x1 holding the other variables x2 and u constant.
De…nition 3.28.1 In the model (3.48) the the causal e¤ ect of x1 on y is
C(x1 ; x2 ; u) = r1 h (x1 ; x2 ; u) ; (3.49)
the change in y due to a change in x1 ; holding x2 and u constant.
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 person’s abilities, skills,
work ethic, interpersonal connections, and preferences. The causal e¤ect 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 (3.49) is a de…nition, and does not necessarily describe
causality in a fundamental or experimental sense. Perhaps it would be more appropriate to label
(3.49) as a structural e¤ect (the e¤ect 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 e¤ects 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 e¤ects, 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 e¤ect 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 di¤erent
across individuals.
In a sample, we cannot observe both outcomes from the same individual, we only observe the
realized value 8
< y(0) if x1 = 0
y=
:
y(1) if x1 = 1
As the causal e¤ect varies across individuals and is not observable, it cannot be measured on
the individual level. We therefore focus on aggregate causal e¤ects, in particular what is known as
the average causal e¤ect.
De…nition 3.28.2 In the model (3.48) average causal e¤ ect of x1 on y condi-

tional on x2 is
Z
ACE(x1 ; x2 ) = E (C(x1 ; x2 ; u) j x1 ; x2 ) = r1 h (x1 ; x2 ; u) f (u j x1 ; x2 )du
R`
(3.50)
where f (u j x1 ; x2 ) is the conditional density of u given x1 ; x2 .
We can think of the average causal e¤ect ACE(x1 ; x2 ) as the average e¤ect in the general
population.
What is the relationship beteween the average causal e¤ect ACE(x1 ; x2 ) and the regression
derivative r1 m (x1 ; x2 )? Equation (3.48) implies that the CEF is
m(x1 ; x2 ) = E (h (x1 ; x2 ; u) j x1 ; x2 )
Z
= h (x1 ; x2 ; u) f (u j x1 ; x2 )du;
R`
the average causal equation, averaged over the conditional distribution of the observed component
u.
Applying the marginal e¤ect operator, the regression derivative is
Z Z
r1 m(x1 ; x2 ) = r1 h (x1 ; x2 ; u) f (u j x1 ; x2 )du + h (x1 ; x2 ; u) r1 f (u j x1 ; x2 )du
R` Z R`
= ACE(x1 ; x2 ) + h (x1 ; x2 ; u) r1 f (u j x1 ; x2 )du: (3.51)

R`
In general, the average causal e¤ect is not the regression derivative. However, they equal when
the second component in (3.51) is zero. This occurs when r1 f (u j x1 ; x2 ) = 0; that is, when
the conditional density of u given (x1 ; x2 ) does not depend on x1 : The condition is su¢ ciently
important that it has a special name in the treatment e¤ects literature.
De…nition 3.28.3 Conditional Independence Assumption (CIA). Condi-

tional on x2 ; the random variables x1 and u are statistically independent.
The CIA implies f (u j x1 ; x2 ) = f (u j x2 ) does not depend on x1 ; and thus r1 f (u j x1 ; x2 ) = 0:

Thus the CIA implies that r1 m(x1 ; x2 ) = ACE(x1 ; x2 ); the regression derivative equals the average
causal e¤ect.
Theorem 3.28.1 In the structural model (3.48), the Conditional Independence As-
sumption implies
r1 m(x1 ; x2 ) = ACE(x1 ; x2 )
the regression derivative equals the average causal e¤ ect 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 e¤ect. In this case, the CEF has causal economic meaning, giving strong justi…cation
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-e¤ects 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 su¢ cient condition to obtain
the desired result. Full independence implies the CIA and implies that each regression derivative
equal that variable’s average causal e¤ect, but full independence is not necessary in order to be
able to causally interpret a subset of the regressors.
3.29 Existence and Uniqueness of the Conditional Expectation*

In Sections 3.3 and 3.5 we de…ned 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 these are the situations where the conditional mean is easiest to describe and understand.
However, the conditional mean can be de…ned without appealing to the properties of either discrete
or continuous random variables. The conditional mean exists quite generally.
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): The only requirement is that y has a
…nite mean.
Theorem 3.29.1 Existence of the Conditional Mean

If E jyj < 1 then there exists a function m(x) such that for all measurable sets X
E (1 (x 2 X ) y) = E (1 (x 2 X ) m(x)) : (3.52)
The function m(x) is almost everywhere unique, in the sense that if h(x) satis…es
(3.52), then there is a set S such that Pr(S) = 1 and m(x) = h(x) for x 2 S: The
function m(x) is called the conditional mean and is written m(x) = E (y j x) :
See, for example, Ash (1972), Theorem 6.3.3.
The conditional mean m(x) de…ned by (3.52) specializes to (3.7) when (y; x) have a joint density.
The usefulness of de…nition (3.52) is that Theorem 3.29.1 shows that the conditional mean m(x)
exists for all …nite-mean distributions. This de…nition 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.
Theorem 3.29.1 also demonstrates that E jyj < 1 is a su¢ cient condition for identi…cation of the
conditional mean. (Recall, a parameter is identi…ed if it is uniquely determined by the distribution
of the observed variables.)
Theorem 3.29.2 Identi…cation of the Conditional Mean

If E jyj < 1; the conditional mean m(x) = E (y j x) is identi…ed almost everywhere.
Proof of Theorem 3.6.1:

For convenience, assume that the variables have a joint density f (y; x). Since E (y j 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 j x)) = E (y j x) fx (x) dx:
Rk
Substituting in (3.7) and noting that fyjx (yjx) fx (x) = f (y; x) ; we …nd that the above expression
equals Z Z Z Z
yfyjx (yjx) dy fx (x) dx = yf (y; x) dydx = E (y)
Rk R Rk R
the unconditional mean of y:

Again assume that the variables have a joint density. It is useful to observe that
f (y; x1 ; x2 ) f (x1 ; x2 )
f (yjx1 ; x2 ) f (x2 jx1 ) = = f (y; x2 jx1 ) ; (3.53)
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
E (y j x1 ; x2 ) = yf (yjx1 ; x2 ) dy: (3.54)
R
Integrating (3.54) with respect to the conditional density of x2 given x1 , and applying (3.53) we
…nd that
Z
E (E (y j x1 ; x2 ) j x1 ) = E (y j x1 ; x2 ) f (x2 jx1 ) dx2
Rk2
Z Z
= yf (yjx1 ; x2 ) dy f (x2 jx1 ) dx2
k
ZR 2 Z R
= yf (yjx1 ; x2 ) f (x2 jx1 ) dydx2
k
ZR 2 ZR
= yf (y; x2 jx1 ) dydx2
Rk2 R
= E (y j x1 )
as stated.

Z Z
E (g (x) y j x) = g (x) yfyjx (yjx) dy = g (x) yfyjx (yjx) dy = g (x) E (y j x)
R R
This is (3.9). The assumption that E jg (x) yj < 1 is required for the …rst equality to be well-
de…ned. Equation (3.10) follows by applying the Simple Law of Iterated Expectations to (3.9).
Proof of Theorem 3.7.1: The assumption that Ey 2 < 1 implies that all the conditional expec-
tations below exist.
Set z = E(y j x1 ; x2 ). By the conditional Jensen’s inequality (B.16),
(E(z j x1 ))2 E z 2 j x1 :
Taking unconditional expectations, this implies
E (E(y j x1 ))2 E (E(y j x1 ; x2 ))2 :
Similarly,
(Ey)2 E (E(y j x1 ))2 E (E(y j x1 ; x2 ))2 : (3.55)
The variables y; E(y j x1 ) and E(y j x1 ; x2 ) all have the same mean Ey; so the inequality (3.55)
implies that the variances are ranked monotonically:
0 var (E(y j x1 )) var (E(y j x1 ; x2 )) : (3.56)
Next, for = Ey observe that
E (y E(y j x)) (E(y j x) ) = E (y E(y j x)) (E(y j x) )=0
so the decomposition
y =y E(y j x) + E(y j x)
satis…es
var (y) = var (y E(y j x)) + var (E(y j x)) : (3.57)
The monotonicity of the variances of the conditional mean (3.56) applied to the variance decom-
position (3.57) implies the reverse monotonicity of the variances of the di¤erences, completing the
proof.
Proof of Theorem 3.8.1. Applying Minkowski’s Inequality (B.22) to e = y m(x);
(E jejr )1=r = (E jy m(x)jr )1=r (E jyjr )1=r + (E jm(x)jr )1=r < 1;
where the two parts on the right-hand are …nite since E jyjr < 1 by assumption and E jm(x)jr < 1
by the Conditional Expectation Inequality (B.17). The fact that (E jejr )1=r < 1 implies E jejr <
1:
Proof of Theorem 3.16.1. For part 1, by the Expectation Inequality (B.18), (A.6) and Assump-
tion 3.16.1,
E xx0 E xx0 = E kxk2 < 1:
Similarly, using the Expectation Inequality (B.18), the Cauchy-Schwarz Inequality (B.20) and As-
sumption 3.16.1,
1=2 1=2
kE (xy)k E kxyk = E kxk2 Ey 2 < 1:
Thus the moments E (xy) and E (xx0 ) are …nite and well de…ned.
For part 2, the coe¢ cient = (E (xx0 )) 1 E (xy) is well de…ned since (E (xx0 )) 1
exists under
Assumption 3.16.1.
Part 3 follows from De…nition 3.16.1 and part 2.
For part 4, …rst note that
2
Ee2 = E y x0
= Ey 2 2E yx0 + 0
E xx0
1
= Ey 2 2E yx0 E xx0 E (xy)
2
Ey
< 1
The …rst inequality holds because E (yx0 ) (E (xx0 )) 1 E (xy) is a quadratic form and therefore neces-
sarily non-negative. Second, by the Expectation Inequality (B.18), the Cauchy-Schwarz Inequality
(B.20) and Assumption 3.16.1,
1=2 1=2
kE (xe)k E kxek = E kxk2 Ee2 < 1:
It follows that the expectation E (xe) is …nite, and is zero by the calculation (3.26).
For part 6, Applying Minkowski’s Inequality (B.22) to e = y x0 ;
r 1=r
(E jejr )1=r = E y x0
r 1=r
(E jyjr )1=r + E x0
(E jyjr )1=r + (E kxkr )1=r k k
< 1;
the …nal inequality by assumption:

Exercises
Exercise 3.1 Find E (E (E (y j x1 ; x2 ; x3 ) j x1 ; x2 ) j x1 ) :
Exercise 3.2 If E (y j x) = a + bx; …nd E (yx) as a function of moments of x:
Exercise 3.3 Prove Theorem 3.8.1.4 using the law of iterated expectations.
Exercise 3.4 Suppose that the random variables y and x only take the values 0 and 1, and have
the following joint probability distribution
x=0 x=1
y=0 .1 .2
y=1 .4 .3
Find E (y j x) ; E y 2 j x and var (y j x) for x = 0 and x = 1:
Exercise 3.5 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 3.6 Use y = m(x) + e to show that

2
var (y) = var (m(x)) +
Exercise 3.7 Show that the conditional variance can be written as
2
(x) = E y 2 j x (E (y j x))2 :
Exercise 3.8 Suppose that y is discrete-valued, taking values only on the non-negative integers,
and the conditional distribution of y given x is Poisson:
exp ( x0 ) (x0 )j
Pr (y = j j x) = ; j = 0; 1; 2; :::
j!
Compute E (y j x) and var (y j x) : Does this justify a linear regression model of the form y =
x0 + e?
j
Hint: If Pr (y = j) = exp( j! ) ; then Ey = and var(y) = :
Exercise 3.9 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 j x1 ; x2 ) as a linear regression.
Exercise 3.10 True or False. If y = x + e; x 2 R; and E (e j x) = 0; then E x2 e = 0:
Exercise 3.11 True or False. If y = x + e; x 2 R; and E (xe) = 0; then E x2 e = 0:
Exercise 3.12 True or False. If y = x0 + e and E (e j x) = 0; then e is independent of x:
Exercise 3.13 True or False. If y = x0 + e and E(xe) = 0; then E (e j x) = 0:

Exercise 3.14 True or False. If y = x0 + e, E (e j x) = 0; and E e2 j x = 2; a constant, then

e is independent of x:
Exercise 3.15 Let x and y have the joint density f (x; y) = 23 x2 + y 2 on 0 x 1; 0 y 1:

Compute the coe¢ cients of the best linear predictor y = + x + e: Compute the conditional mean
m(x) = E (y j x) : Are the best linear predictor and conditional mean di¤erent?
Exercise 3.16 Let x be a random variable with = Ex and 2 = var(x): De…ne
2 x
g xj ; = :
(x )2 2
Show that Eg (x j m; s) = 0 if and only if m = and s = 2:
Exercise 3.17 Suppose that 0 1

1
x = @ x2 A
x3
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 3.18 Show (3.42)-(3.43), namely that for

2
d( ) = E m(x) x0
then
= argmin d( )
2Rk
1
= E xx0 E (xm(x))
1
= E xx0 E (xy) :
Hint: To show E (xm(x)) = E (xy) use the law of iterated expectations.

Chapter 4
The Algebra of Least Squares
4.1 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.
4.2 Least Squares Estimator

In Section 3.16 we derived and discussed the best linear predictor of y given x for a pair
of random variables (y; x) 2 R Rk ; and called this the linear projection model. Applied to
observations from a random sample with observations (yi ; xi : i = 1; :::; n) this model takes the
form
yi = x0i + ei (4.1)
where is de…ned as
= argmin S( ); (4.2)
2Rk
the minimizer of the expected squared error

2
S( ) = E yi x0i ; (4.3)
and has the explicit solution

1
= E xi x0i E (xi yi ) : (4.4)
When a parameter is de…ned as the minimizer of a function as in (4.2), a standard approach
to estimation is to construct an empirical analog of the function, and de…ne the estimator of the
parameter as the minimizer of the empirical function.
The empirical analog of the expected squared error (4.3) is the sample average squared error
n
1X 2
Sn ( ) = yi x0i (4.5)
n
i=1
1
= SSEn ( )
n
where
n
X 2
SSEn ( ) = yi x0i
i=1
is called the sum-of-squared-errors function.
An estimator for is the minimizer of (4.5):
b = argmin Sn ( ):
2Rk
71
CHAPTER 4. THE ALGEBRA OF LEAST SQUARES 72
Figure 4.1: Sum-of-Squared Errors Function
Alternatively, as Sn ( ) is a scale multiple of SSEn ( ); we may equivalently de…ne b as the min-

imizer of SSEn ( ): Hence b is commonly called the least-squares (LS) (or ordinary least
squares (OLS)) estimator of . As discussed in Chapter 2, the hat “^” on b signi…es that it is
an estimator of the parameter : If we want to be explicit about the estimation method, we can
write b ols to signify that it is the OLS estimator.
To visualize the quadratic function Sn ( ), Figure 4.1 displays an example sum-of-squared er-
rors function SSEn ( ) for the case k = 2: The least-squares estimator b is the the pair ( b 1 ; b 2 )
minimizing this function.
4.3 Solving for Least Squares

To solve for b , expand the SSE function to …nd
n
X n
X n
X
SSEn ( ) = yi2 2 0
xi yi + 0
xi x0i
i=1 i=1 i=1
which is quadratic in the vector argument . The …rst-order-condition for minimization of SSEn ( )
is
Xn X n
@
0= SSEn ( b ) = 2 xi yi + 2 xi x0i b : (4.6)
@
i=1 i=1
Pn 0
By inverting the k k matrix i=1 xi xi we …nd an explicit formula for the least-squares estimator
n
! 1 n !
X X
b= xi x 0
xi yi : (4.7)
i
i=1 i=1
This is the natural estimator of the best linear projection coe¢ cient de…ned in (4.2), and can
also be called the linear projection estimator.
Alternatively, equation (4.4) writes the projection coe¢ cient as an explicit function of the
population moments Qxy and Qxx : Their moment estimators are the sample moments
n
b xy = 1X
Q xi yi
n
i=1
n
b xx = 1X
Q xi x0i :
n
i=1
The moment estimator of replaces the population moments in (4.4) with the sample moments:
b 1Q
b = Q b
xx xy
n
! 1 n
!
1X 1X
= xi x0i x i yi
n n
i=1 i=1
n
! 1 n
!
X X
= xi x0i xi yi
i=1 i=1
which is identical with (4.7).
Least Squares Estimation
De…nition 4.3.1 The least-squares estimator b is

b = argmin Sn ( )
2Rk
where
n
1X 2
Sn ( ) = yi x0i
n
i=1
and has the solution

n
! 1 n
!
X X
b= xi x0i xi yi :
i=1 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 least-squares 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 mea-
surement. As viewed by Legendre, (4.1) is a set of n equations with k unknowns. As the
equations cannot be solved exactly, Legendre’s 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 …rst-order conditions (4.6) 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 Legendre’s estimation method.
4.4 Illustration
We illustrate the least-squares estimator in practice with the data set used to generate the
estimates from Chapter 3. 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 ? For this
illustration, we use the sub-sample of non-white married non-military female wages earners with
12 years potential work experience. This sub-sample has 61 observations. Let yi be log wages and
xi be an intercept and years of education. Then
n
1X 3:025
x i yi =
n 47:447
i=1
and
n
1X 1 15:426
xi x0i = :
n 15:426 243
i=1
Thus
1
b = 1 15:426 3:025
15:426 243 47:447
0:626
= : (4.8)
0:156
We often write the estimated equation using the format
\
log(W age) = 0:626 + 0:156 education: (4.9)
An interpretation of the estimated equation is that each year of education is associated with an
16% increase in mean wages.
Equation (4.9) is called a bivariate regression as there are only two variables. A multivari-
ate regression has two or more regressors, and allows a more detailed investigation. Let’s redo
the example, but now including all levels of experience. This expanded sample includes 2454 ob-
servations. Including as regressors years of experience and its square (experience 2 =100) (we divide
by 100 to simplify reporting), we obtain the estimates
\
log(W age) = 1:06 + 0:116 education + 0:010 experience 0:014 experience2 =100: (4.10)
These estimates suggest a 12% increase in mean wages per year of education, holding experience
constant.
4.5 Least Squares Residuals

As a by-product of estimation, we de…ne the …tted or predicted value
yî = x0i b
and the residual

eî = yi yî = yi x0i b : (4.11)
Note that yi = yî + eî and
yi = x0i b + eî : (4.12)
We make a distinction between the error ei and the residual eî : The error ei is unobservable while
the residual eî is a by-product of estimation. These two variables are frequently mislabeled, which
can cause confusion.
Equation (4.6) implies that

n
X
xi eî = 0: (4.13)
i=1
To see this by a direct calculation, using (4.11) and (4.7),

n
X n
X
xi eî = x i yi x0i b
i=1 i=1
Xn n
X
= xi yi xi x0i b
i=1 i=1
n n n
! 1 n
!
X X X X
= xi yi xi x0i xi x0i xi yi
i=1 i=1 i=1 i=1
Xn Xn
= xi yi xi yi
i=1 i=1
= 0:
When xi contains a constant, an implication of (4.13) is

n
1X
eî = 0:
n
i=1
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.
Given the residuals, we can construct an estimator for 2 = Ee2i :
n
1X 2
^2 = eî : (4.14)
n
i=1
4.6 Model in Matrix Notation

For many purposes, including computation, it is convenient to write the model and statistics in
matrix notation. The linear equation (3.24) is a system of n equations, one for each observation.
We can stack these n equations together as
y1 = x01 + e1
y2 = x02 + e2
..
.
yn = x0n + en :
Now de…ne 0 1 0 1 0 1
y1 x01 e1
B y2 C B x02 C B e2 C
B C B C B C
y=B .. C; X=B .. C; e=B .. C:
@ . A @ . A @ . A
yn x0n en
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: (4.15)
Sample sums can also be written in matrix notation. For example

n
X
xi x0i = X 0 X
i=1
Xn
xi yi = X 0 y:
i=1
Therefore
b = X 0X 1
X 0y : (4.16)
The matrix version of (4.12) and estimated version of (4.15) is
y = Xb + e
^;
or equivalently the residual vector is

^= y
e X b:
Using the residual vector, we can write (4.13) as
X 0e
^= 0 (4.18)
and the error variance estimator (4.14) as
^2 = n 1 0
e
ê^ (4.19)
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 0X 1
X 0y
^ = y Xb
e
X 0e
^ = 0
^2 = n 1 0
e
ê^:
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.
4.7 Projection Matrix

De…ne the matrix
1
P = X X 0X X 0:
Observe that
1
P X = X X 0X X 0 X = X:
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
PZ = PX = X X 0X X 0X =X = Z:
As an important example, if we partition the matrix X into two matrices X 1 and X 2 so that
X = [X 1 X 2] ;
then P X 1 = X 1 .
The matrix P is symmetric and idempotent1 . To see that it is symmetric,
1 0
P0 = X X 0X X0
0 1 0
= X0 X 0X (X)0
0 1
= X X 0X X0
0 1
= X (X)0 X 0 X0
= P:
To establish that it is idempotent, the fact that P X = X implies that
1
PP = P X X 0X X0
1
= X X 0X X0
= P:
The matrix P has the property that it creates the …tted values in a least-squares regression:
1
P y = X X 0X X 0y = X b = y
^:
Because of this property, P is also known as the “hat matrix”.

Another useful property is that the trace of P equals the number of columns of X
tr P = k: (4.20)
Indeed,
1
tr P = tr X X 0 X X0
1
= tr X 0X X 0X
= tr (I k )
= k:
1
A matrix P is symmetric if P 0 = P : A matrix P is idempotent if P P = P : See Appendix A.8.
(See Appendix A.4 for de…nition and properties of the trace operator.)
1
The i’th diagonal element of P = X (X 0 X) X 0 is
1
hii = x0i X 0 X xi (4.21)
which is called the leverage of the i’th observation. The hii take values in [0; 1] and sum to k
n
X
hii = k (4.22)
i=1
(See Exercise 4.8).
4.8 Orthogonal Projection

De…ne
M = In P
1
= In X X 0X X0
where I n is the n n identity matrix. Note that
M X = (I n P)X = X PX = 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
MZ = Z P Z = 0:
For example, M X 1 = 0 for any subcomponent X 1 of X, and M P = 0:
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 (4.20) we can calculate
tr M = n k: (4.23)
While P creates …tted values, M creates least-squares residuals:
My = y Py = y Xb = e
^: (4.24)
Another way of writing (4.24) is
y = P y + My = y
^+e
^:
This decomposition is orthogonal, that is
^ = (P y)0 (M y) = y 0 P M y = 0:
^0 e
y
We can also use (4.24) to write an alternative expression for the residual vector. Substituting
y = X + e into e^ = M y and using M X = 0 we …nd
^ = M (X + e) = M e:
e (4.25)
which is free of dependence on the regression coe¢ cient .
Another useful application of (4.24) is to the error variance estimator (4.19)
^2 = n 1 0
e
ê^
1 0
= n y MMy
1 0
= n y M y;
the …nal equality since M M = M . Similarly using (4.25) we …nd
^2 = n 1 0
e M e:
4.9 Regression Components

Partition
X = [X 1 X 2]
and
1
= :
2
Then the regression model can be rewritten as
y = X1 1 + X2 2 + e: (4.26)
0 0 0
The OLS estimator of =( 1; 2) is obtained by regression of y on X = [X 1 X 2 ] and can be
written as
y = Xb + e ^ = X1 b1 + X2 b2 + e ^: (4.27)
b
We are interested in algebraic expressions for 1 and 2 : b
The algebra for the estimator is identical as that for the population coe¢ cients as presented in
Section 3.19.
Partition Qb xx and Qb xy as
2 1 1 0 3
2 3 0
b
Q11 Q12 b X X 1 X X 2
6 n 1 n 1 7
b xx = 4
Q 5=6 7
4 5
b
Q21 Q22 b 1 0 1 0
X X1 X X2
n 2 n 2
and similarly Qxy
2 1 3
2 3 X 01 y
Qb 1y
6 n 7
b xy = 4
Q 5=6 7:
4 5
Qb 2y 1 0
X 2y
n
By the partitioned matrix inversion formula (A.4)
2 3 1 2 11 3 2 3
b 11 Q
Q b 12 Qb b 12
Q Qb 1 b 1 Q
Q b 1
b 12 Q
11 2 11 2 22
b 1=4 5 def 6 7 6 7
Q xx = 4 5=4 5 (4.28)
b
Q21 Q22b Qb 21
b
Q
22
Qb 1
Q21 Q 1
Qb 1
22 1 11 22 1
b 11 2 = Q
where Q b 11 Q b 1Q
b 12 Q b 22 1 = Q
b 21 and Q b 22 b 1Q
b 21 Q
Q b 12 :
22 11
Thus
!
b
b = 1
b
2
" #" #
b 1
Q b 1 Q
Q b b 1 b 1y
Q
11 2 11 2 12 Q22
= 1 b
b
Q22 1 Q b 1
b 21 Q Qb 1 Q2y
11 22 1
1
!
b
Q b
11 2 Q1y 2
= 1
b
Q Qb 2y 1
22 1
Now
b 11 2 = Q
Q b 11 b 1Q
b 12 Q
Q b 21
22
1
1 0 1 0 1 0 1 0
= X 1X 1 X X2 X X2 X X1
n n 1 n 2 n 2
1 0
= X M 2X 1
n 1
where
1
M 2 = In X 2 X 02 X 2 X 02
b 22 1 = 1 X 0 M 1 X 2 where
is the orthogonal projection matrix for X 2 : Similarly Q
n 2
1
M 1 = In X 1 X 01 X 1 X 01
is the orthogonal projection matrix for X 1 : Also
b 1y 2 = Q
Q b 1y b 1Q
b 12 Q
Q b 2y
22
1
1 0 1 0 1 0 1 0
= X 1y X X2 X X2 X y
n n 1 n 2 n 2
1 0
= X M 2y
n 1
b 2y 1 = 1 X 0 M 1 y:
and Q
n 2
Therefore
b = X 0 M 2X 1 1
1 1 X 01 M 2 y (4.29)
and
b = X 0 M 1X 2 1
2 2 X 02 M 1 y : (4.30)
These are algebraic expressions for the sub-coe¢ cient estimates from (4.27).
4.10 Residual Regression

As …rst recognized by Ragnar Frisch, expressions (4.29) and (4.30) can be used to show that
the least-squares estimators b 1 and b 2 can be found by a two-step regression procedure.
Take (4.30). Since M 1 is idempotent, M 1 = M 1 M 1 and thus
b 1
2 = X 02 M 1 X 2 X 02 M 1 y
1
= X 02 M 1 M 1 X 2 X 02 M 1 M 1 y
1
= f0 X
X f f0 e
X
2 2 2 ~1
where
f2 = M 1 X 2
X
and
~1 = M 1 y:
e
Thus the coe¢ cient estimate b 2 is algebraically equal to the least-squares regression of e
~1 on
f
X 2 : Notice that these two are y and X 2 , respectively, premultiplied by M 1 . But we know that
multiplication by M 1 is equivalent to creating least-squares residuals. Therefore e~1 is simply the
least-squares residual from a regression of y on X 1 ; and the columns of Xf2 are the least-squares
residuals from the regressions of the columns of X 2 on X 1 :
We have proven the following theorem.
Theorem 4.10.1 Frisch-Waugh-Lovell

In the model (4.26), the OLS estimator of 2 and the OLS residuals e ^
may be equivalently computed by either the OLS regression (4.27) or via
the following algorithm:
1. Regress y on X 1 ; obtain residuals e

~1 ;
f2 ;
2. Regress X 2 on X 1 ; obtain residuals X
3. Regress e f2 ; obtain OLS estimates b 2 and residuals e

~1 on X ^:
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. Rather, the primary use
is theoretical.
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
is the vector of observed regressors and X 2 = is a vector of ones . In this case,
0 1 0
M2 = I :
Observe that
f1 = M 2 X 1
X
0 1 0
= X1 X1
= X1 X1
and
~ = M 2y
y
0 1 0
= y y
= y y;
which are “demeaned”. The FWL theorem says that b 1 is the OLS estimate from a regression of
yi y on x1i x1 :
n
! 1 n
!
X X
b = (x1i x1 ) (x1i x1 ) 0
(x1i x1 ) (yi y) :
1
i=1 i=1
Thus the OLS estimator for the slope coe¢ cients 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 mod-
els for the analysis of economic problems. Frisch made a number of foundational contribu-
tions to modern economics beyond the Frisch-Waugh-Lovell Theorem, including formalizing
consumer theory, production theory, and business cycle theory.
4.11 Prediction Errors

The least-squares residual eî 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
only using the other observations. We can do this by de…ning the leave-one-out OLS estimator
of as that obtained from the sample of n 1 observations excluding the i’th observation:
0 1 10 1
X X
b
( i) =
@ 1 xj x0j A @
1
xj yj A
n 1 n 1
j6=i j6=i
1
= X 0( i) X ( i) X( i) y ( i) : (4.31)
Here, X ( i) and y ( i) are the data matrices omitting the i’th row. The leave-one-out predicted
value for yi is
y~i = x0i b ( i) ;
and the leave-one-out residual or prediction error is
e~i = yi y~i :
A convenient alternative expression for b ( i) (derived below) is
b 1
( i) =b (1 hii ) 1
X 0X xi eî (4.32)
where hii are the leverage values as de…ned in (4.21).

Using (4.32) we can simplify the expression for the prediction error:
e~i = yi x0i b ( i)
1 1
= yi x0i ^ + (1 hii ) x0i X 0 X xi eî
1
= eî + (1 hii ) hii eî
1
= (1 hii ) eî : (4.33)
A convenient feature of this expression is that it shows that computation of e~i 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
n
2 1X 2
~ = e~i
n
i=1
n
1X 2 2
= (1 hii ) eî :
n
i=1
p
This is also known as the mean squared prediction error. Its square root ~ = ~ 2 is the
prediction standard error.
Proof of Equation (4.32). The Sherman–Morrison formula (A.3) from Appendix A.5 states that
for nonsingular A and vector b
1 1
A bb0 =A 1
+ 1 b0 A 1
b A 1
bb0 A 1
:
This implies
1 1 1 1 1
X 0X xi x0i = X 0X + (1 hii ) X 0X xi x0i X 0 X
and thus
b 1
( i) = X 0X xi x0i X 0y xi yi
1 1
= X 0X X 0y X 0X xi yi
1 0 1 1
+ (1 hii ) XX xi x0i X 0X X 0y xi yi
1 1
= b X 0X xi yi + (1 hii ) 1
X 0X xi x0i b hii yi
1
= b (1 hii ) 1
X 0X xi (1 hii ) yi x0i b + hii yi
1
= b (1 hii ) 1
X 0X xi eî
the third equality making the substitutions b = (X 0 X)

1 1
X 0 y and hii = x0i (X 0 X) xi ; and the
remainder collecting terms.
4.12 In‡uential Observations

Another use of the leave-one-out estimator is to investigate the impact of in‡uential obser-
vations, sometimes called outliers. We say that observation i is in‡uential if its omission from
the sample induces a substantial change in a parameter of interest. From (4.32)-(4.33) we know
that
b b 1 1
( i) = (1 hii ) X 0X xi eî
1
= X 0X xi e~i : (4.34)
By direct calculation of this quantity for each observation i; we can directly discover if a speci…c
observation i is in‡uential for a coe¢ cient estimate of interest.
For a general assessment, we can focus on the predicted values. The di¤erence between the
full-sample and leave-one-out predicted values is
yî y~i = x0i b x0i b ( i)

1
= x0i X 0 X xi e~i
= hii e~i
which is a simple function of the leverage values hii and prediction errors e~i : Observation i is
in‡uential for the predicted value if jhii e~i j is large, which requires that both hii and j~
ei j are large.
One way to think about this is that a large leverage value hii gives the potential for observation
i to be in‡uential. 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 in‡uential as the latter also requires that the prediction error e~i is large.
To determine if any individual observations are in‡uential in this sense, a large number of
diagnostic statistics have been proposed (some names include DFITS, Cook’s Distance, and Welsch
Distance) but as they are not based on statistical theory it is unclear if they are useful for practical
work. Probably the most relevant measure is the change in the coe¢ cient estimates given in
(4.34). The ratio of these changes to the coe¢ cient’s 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 estimated coe¢ cient of interest.
For illustration, consider Figure 4.2 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 26’th 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 26’th observation from the sample.
In this example we can see how the 26’th observation (the “outlier”) greatly tilts the least-squares
…tted line towards the 26’th observation. In fact, the slope coe¢ cient 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 coe¢ cient of
0:41 is meaningful and should raise concern to an applied economist.
10
●
leave−one−out OLS
●
8
● ●
●
●
6
●
● ● ● ● OLS
● ●
●
y
●
4
●
● ●
●
2
●
0
●
●
●
●
2 4 6 8 10
Figure 4.2: Impact of an in‡uential observation on the least-squares estimator
If an observation is determined to be in‡uential, what should be done? As a common cause

of in‡uential observations is data entry error, the in‡uential 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 in‡uential. In
this case it is unclear how to proceed. Some researchers will try to alter the speci…cation to
properly model the in‡uential 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.
4.13 Measures of Fit

When a least-squares regression is reported in applied economics, it is common to see a summary
measure of …t, measuring how well the regressors explain the observed variation in the dependent
variable.
Some common summary measures are based on scaled or Pntransformed estimates of the mean-
P
squared error : These include the sum of squared errors i=1 eî ; the sample variance n 1 ni=1 e^2i =
2 2
q Pn
^ 2 ; the root mean squared error n 1 ^2i (sometimes
i=1 e called the standard error of the re-
P
gression), and the mean prediction error ~ 2 = n1 ni=1 e~2i .
A related and commonly reported statistic is the coe¢ cient of determination or R-squared:
Pn
2 yi y)2
(^ ^2
R = Pni=1 = 1
i=1 (yi y)2 ^ 2y
where
n
1X
^ 2y = (yi y)2
n
i=1
is the sample variance of yi : R2 can be viewed as an estimator of the population parameter
2 var (x0i ) 2
= =1 2
var(yi ) y
where 2y = var(yi ): A high 2 means that forecasts of y using x0 will be quite accurate relative to
the unconditional mean. In this sense R2 can be a useful summary measure for an out-of-sample
forecast or policy experiment.
An alternative estimator of 2 proposed by Theil called R-bar-squared or adjusted R2 is
P
2 (n 1) ni=1 e^2i
R =1 P :
(n k) ni=1 (yi y)2
2
Theil’s estimator R is better estimator of 2 than the unadjusted estimator R2 because it is a
ratio of bias-corrected variance estimates.
2
Unfortunately, the frequent reporting of R2 and R seems to have led to exaggerated beliefs
regarding their usefulness. One mistaken belief is that R2 is a measure of “…t”. This belief is
incorrect, as an incorrectly speci…ed model can still have a reasonably high R2 . For example,
suppose the truth is that xi N (0; 1) and yi = xi + x2i : If we regress yi on xi (incorrectly omitting
x2i ); the best linear predictor is yi = 1 + xi + ei where ei = x2i 1: This is a misspeci…ed regression,
as the true relationship is deterministic! You can also calculate that the population 2 = =(2 + )
which can be arbitrarily close to 1 if is large. For example, if = 8; then R2 ' 2 = :8; or if
= 18 then R2 ' 2 = :9. This example shows that a regression with a high R2 can actually have
poor …t.
Another mistaken belief is that a high R2 is important in order to justify interpretation of the
regression coe¢ cients. This is mistaken as there is no direct association between the level of R2
and the “correctness” of a regression, the accuracy of the coe¢ cient estimates, or the validity of
statistical inferences based on the estimated regression. In contrast, even if the R2 is quite small,
accurate estimates of regression coe¢ cients is quite possible when sample sizes are large.
2
The bottom line is that while R2 and R have appropriate uses, their usefulness should not be
exaggerated.
Henri Theil
2
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 and in‡uential advanced
textbook on econometrics (Theil, 1971).
4.14 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
2
ei j xi N 0; :
This assumption implies

yi j xi N x0i ; 2
:
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 1 1 2
2 0
log L( ; ) = log exp yi xi
(2 2 )1=2 2 2
i=1
n 2 1
= log 2 2
SSEn ( ):
2 2
The maximum likelihood estimator (MLE) ( b mle ; ^ 2mle ) maximizes log L( ; 2 ): Since the latter is
a function of only through the sum of squared errors SSEn ( ); maximizing the likelihood is
identical to minimizing SSEn ( ). Hence
b = b ols ;
mle
the MLE for equals the OLS estimator: Due to this equivalence, the least squares estimator b ols
is often called the MLE.
We can also …nd the MLE for 2 : Plugging b into the log-likelihood we obtain
n
n 1 X
log L b ; 2
= log 2 2
2
e^2i :
2 2
i=1
Maximization with respect to 2 yields the …rst-order condition

n
X
@ n 1
log L b ; ^ 2 = 2 + 2 e^2i = 0:
@ 2 2^ 2 ^2 i=1
Solving for ^ 2 yields the MLE for 2

n
1X 2
^ 2mle = eî
n
i=1
which is the same as the moment estimator (4.14).

It may seem surprising that the MLE b mle is numerically equal to the OLS estimator, despite
emerging from quite di¤erent 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 speci…c 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.
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 Gauss’s approach was reinforced by Laplace’s simultaneous
discovery of the central limit theorem, which provided a justi…cation for viewing random
disturbances as approximately normal.
Exercises
Exercise 4.1 Let y be a random variable with = Ey and 2 = var(y): De…ne
2 y
g y; ; = :
(y )2 2
Pn
Let (^ ; ^ 2 ) be the values such that g n (^ ; ^ 2 ) = 0 where g n (m; s) = n 1
i=1 g (yi ; m; s) : Show that
^ and ^ 2 are the sample mean and variance.
Exercise 4.2 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:
Exercise 4.3 Using matrix algebra, show X 0 e

^ = 0:
^ be the OLS residual from a regression of y on X = [X 1 X 2 ]. Find X 02 e

Exercise 4.4 Let e ^:
Exercise 4.5 Let e ^ be the OLS residual from a regression of y on X: Find the OLS coe¢ cient
from a regression of e
^ on X:
^ = X(X 0 X)
Exercise 4.6 Let y 1 X 0 y: Find the OLS coe¢ cient from a regression of y
^ on X:
Exercise 4.7 Show that if X = [X 1 X 2 ] then P X 1 = X 1 :
Exercise 4.8 Show (4.22), that hii in (4.21) sum to k: (Hint: Use (4.20).)
Exercise 4.9 Show that M is idempotent: M M = M :
Exercise 4.10 Show that tr M = n k:
Exercise 4.11 Show that if X = [X 1 X 2 ] and X 01 X 2 = 0 then P = P 1 + P 2 .
Exercise 4.12 A dummy variable takes on only the values 0 and 1. It is used for categorical
data, such as an individual’s gender. Let d1 and d2 be vectors of 1’s and 0’s, 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 (4.35)
y = d1 1 + d2 2 +e (4.36)
y = + d1 + e (4.37)
Can all three equations (4.35), (4.36), and (4.37) be estimated by OLS? Explain if not.
(a) Compare regressions (4.36) and (4.37). Is one more general than the other? Explain the
relationship between the parameters in (4.36) and (4.37).
(b) Compute 0d and 0d ; where is an n 1 is a vector of ones.

1 2
(c) Letting = ( 1 2 )0 ; write equation (4.36) as y = X + e: Consider the assumption

E(xi ei ) = 0. Is there any content to this assumption in this setting?
Exercise 4.13 Let d1 and d2 be de…ned as in the previous exercise.
(a) In the OLS regression

y = d1 ^ 1 + d2 ^ 2 + u
^;
show that ^ 1 is the sample mean of the dependent variable among the men of the sample
(y 1 ), and that ^ 2 is the sample mean among the women (y 2 ).
(b) Let X (n k) be an additional matrix of regressions. Describe in words the transformations
y = y d1 y 1 d2 y 2
X = X d1 X 1 d2 X 2 :
(c) Compare e from the OLS regresion
y =X e +e
~
with b from the OLS regression
y = d1 ^ 1 + d2 ^ 2 + X b + e
^:
Exercise 4.14 Let b n = (X 0n X n ) X 0n y n denote the OLS estimate when y n is n 1 and X n is

1
n k. A new observation (yn+1 ; xn+1 ) becomes available. Prove that the OLS estimate computed
using this additional observation is
b 1 1
n+1 = bn + 1 X 0n X n xn+1 yn+1 x0n+1 b n :
1+ x0n+1 (X 0n X n ) xn+1
Exercise 4.15 Prove that R2 is the square of the sample correlation between y and y
^:
Exercise 4.16 Show that ~ 2 ^ 2 : Is equality possible?
Exercise 4.17 For which observations will b ( i) = b?
Exercise 4.18 The data …le cps85.dat contains a random sample of 528 individuals from the
1985 Current Population Survey by the U.S. Census Bureau. The …le contains observations on nine
variables, listed in the …le cps85.pdf.
V1 = education (in years)

V2 = region of residence (coded 1 if South, 0 otherwise)
V3 = (coded 1 if nonwhite and non-Hispanic, 0 otherwise)
V4 = (coded 1 if Hispanic, 0 otherwise)
V5 = gender (coded 1 if female, 0 otherwise)
V6 = marital status (coded 1 if married, 0 otherwise)
V7 = potential labor market experience (in years)
V8 = union status (coded 1 if in union job, 0 otherwise)
V9 = hourly wage (in dollars)
Estimate a regression of wage yi on education x1i , experience x2i , and experienced-squared x3i = x22i
(and a constant). Report the OLS estimates.
Let eî be the OLS residual and yî the predicted value from the regression. Numerically calculate
the following:
Pn
(a) i=1 e
î
Pn
(b) i=1 x1i e
î
Pn
(c) i=1 x2i e
î
Pn 2 ^
(d) i=1 x1i e i
Pn 2 ^
(e) i=1 x2i e i
Pn
(f) i=1 y
î eî
Pn
(g) ^2i
i=1 e
(h) R2
Are these calculations consistent with the theoretical properties of OLS? Explain.
Exercise 4.19 Using the data from the previous problem, restimate the slope on education using
the residual regression approach. Regress yi on (1; x2i ; x22i ), regress x1i on (1; x2i ; x22i ), and regress
the residuals on the residuals. Report the estimate from this regression. Does it equal the value
from the …rst OLS regression? Explain.
In the second-stage residual regression, (the regression of the residuals on the residuals), cal-
culate the equation R2 and sum of squared errors. Do they equal the values from the initial OLS
regression? Explain.
Chapter 5
Least Squares Regression
5.1 Introduction
In this chapter we investigate some …nite-sample properties of least-squares applied to a random
sample in the the linear regression model. Throughout this chapter we maintain the following.
Assumption 5.1.1 Linear Regression Model

The observations (yi ; xi ) come from a random sample and satisfy the linear
regression equation
yi = x0i + ei (5.1)
E (ei j xi ) = 0: (5.2)
The variables have …nite second moments
Eyi2 < 1;
E kxi k2 < 1;
and an invertible design matrix
Qxx = E xi x0i > 0:
We will consider both the general case of heteroskedastic regression, where the conditional
variance
E e2i j xi = 2 (xi ) = 2i
is unrestricted, and the specialized case of homoskedastic regression, where the conditional variance
is constant. In the latter case we add the following assumption.
Assumption 5.1.2 Homoskedastic Linear Regression Model

In addition to Assumption 5.1.1,
E e2i j xi = 2
(xi ) = 2
(5.3)
is independent of xi :
91
CHAPTER 5. LEAST SQUARES REGRESSION 92
5.2 Mean of Least-Squares Estimator

In this section we show that the OLS estimator is unbiased in the linear regression model.
Under (5.1)-(5.2) note that
0 1 0 1 0 1
.. .. ..
B . C B . C B 0. C
E (y j X) = B
@ E (y i j X) C =
A @
B E (y i j x ) C B
i A = @ xi
C=X :
A (5.4)
.. .. ..
. . .
Similarly 0 1 0 1
.. ..
B . C B C .
E (e j X) = B C B C
@ E (ei j X) A = @ E (ei j xi ) A = 0: (5.5)
.. ..
. .
By (4.16), conditioning on X, the linearity of expectations, (5.4), and the properties of the
matrix inverse,
1
E bjX = E X 0X X 0y j X
1
= X 0X X 0 E (y j X)
1
= X 0X X 0X
= :
Applying the law of iterated expectations to E b j X = , we …nd that
E b =E E bjX = :
Another way to calculate the same result is as follows. Insert y = X + e into the formula
(4.16) for b to obtain
b = 1
X 0X X 0 (X + e)
1 1
= X 0X X 0X + X 0X X 0e
1
= + X 0X X 0 e: (5.6)
This is a useful linear decomposition of the estimator b into the true parameter and the stochastic
1
component (X 0 X) X 0 e:
Using (5.6), conditioning on X, and (5.5),
1
E b jX = E X 0X X 0e j X
1
= X 0X X 0 E (e j X)
= 0:
Using either derivation, we have shown the following theorem.
Theorem 5.2.1 Mean of Least-Squares Estimator

In the linear regression model (Assumption 5.1.1)
E bjX = (5.7)
and
E( b ) = : (5.8)
Equation (5.8) says that the estimator is unbiased, meaning that the distribution of b is centered
at . Equation (5.7) says that the estimator is conditionally unbiased, which is a stronger result.
It says that b is unbiased for any realization of the regressor matrix X.
5.3 Variance of Least Squares Estimator

In this section we calculate the conditional variance of the OLS estimator.
For any r 1 random vector Z de…ne the r r covariance matrix
var(Z) = E (Z EZ) (Z EZ)0

= EZZ 0 (EZ) (EZ)0
and for any pair (Z; X) de…ne the conditional covariance matrix
var(Z j X) = E (Z E (Z j X)) (Z E (Z j X))0 j X :
The conditional covariance matrix of the n 1 regression error e is the n n matrix
D = E ee0 j X :
The i’th diagonal element of D is
E e2i j X = E e2i j xi = 2
i
while the ij 0 th o¤-diagonal element of D is
E (ei ej j X) = E (ei j xi ) E (ej j xj ) = 0:
where the …rst equality uses independence of the observations (Assumption 1.5.1) and the second
is (5.2). Thus D is a diagonal matrix with i’th diagonal element 2i :
0 2 1
1 0 0
B 0 2 0 C
2 2 B 2 C
D = diag 1 ; :::; n = B . . . .. C : (5.9)
@ . . .
. . . . A
0 0 2
n
In the special case of the linear homoskedastic regression model (5.3), then
E e2i j xi = 2
i = 2
and we have the simpli…cation

2
D = In :
In general, however, D need not necessarily take this simpli…ed form.
For any matrix n r matrix A = A(X),
var(A0 y j X) = var(A0 e j X) = A0 DA: (5.10)
In particular, we can write b = A0 y where A = X (X 0 X)

1
and thus
1 1
var( b j X) = A0 DA = X 0 X X 0 DX X 0 X :
It is useful to note that

n
X
0
X DX = xi x0i 2
i;
i=1
a weighted version of X 0 X.
Rather than working with the variance of the unscaled estimator b ; it will be useful to work
p
with the conditional variance of the scaled estimator n b :
def p
V b = var n b jX
= n var( b j X)
1 1
= n X 0X X 0 DX X 0X
1 1
1 0 1 0 1 0
= XX X DX XX :
n n n
This rescaling might seem rather odd, but it will help provide continuity between the …nite-sample
treatment of this chapter and the asymptotic treatment of later chapters. As we will see in the
next chapter, var( b j X) vanishes as n tends to in…nity, yet V b converges to a constant matrix.
In the special case of the linear homoskedastic regression model, D = I n 2 , so X 0 DX =
X 0 X 2 ; and the variance matrix simpli…es to
1
1 0 2
V b = XX :
n
Theorem 5.3.1 Variance of Least-Squares Estimator

In the linear regression model (Assumption 5.1.1)
p
V b = var n b jX
1 1
1 0 1 0 1 0
= XX X DX XX (5.11)
n n n
where D is de…ned in (5.9).

In the homoskedastic linear regression model (Assumption 5.1.2)
1
1 0 2
V b = XX :
n
5.4 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. 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.
To see this, since E (y j X) = X ; then for any linear estimator e = A0 y we have
E e j X = A0 E (y j X) = A0 X ;
so e is unbiased if (and only if) A0 X = I k : Furthermore, we saw in (5.10) that
var e j X = var A0 y j X = A0 DA = A0 A 2
:
the last equality using the homoskedasticity assumption D = I n 2 . The “best” unbiased linear
estimator is obtained by …nding the matrix A such that A0 A is minimized in the positive de…nite
sense.
Theorem 5.4.1 Gauss-Markov
1. In the homoskedastic linear regression model (Assumption 5.1.2),

the best (minimum-variance) unbiased linear estimator is the least-
squares estimator
b = X 0X 1 X 0y
2. In the linear regression model (Assumption 5.1.1), the best unbiased

linear estimator is
e = X 0D 1 1
X X 0D 1
y (5.12)
The …rst part of the Gauss-Markov theorem is a limited e¢ ciency justi…cation for the least-
squares estimator. The justi…cation is limited because the class of models is restricted to ho-
moskedastic 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,
the least-squares estimator is ine¢ cient. Within the class of linear unbiased estimators the best
estimator is (5.12) and is called the Generalized Least Squares (GLS) estimator. This 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 9.1.
We give a proof of the …rst part of the theorem below, and leave the proof of the second part
for Exercise 5.3.
Proof of Theorem 5.4.1.1. Let A be any n k function of X such that A0 X = I k : The variance
1 2
of the least-squares estimator is (X 0 X) and that of A0 y is A0 A 2 : It is su¢ cient to show
1 1
that the di¤erence A A (X X) is positive semi-de…nite. Set C = A X (X 0 X) : Note that
0 0
X 0 C = 0: Then we calculate that

1 1 0 1 1
A0 A X 0X = C + X X 0X C + X X 0X X 0X
1 1
= C 0C + C 0X X 0X + X 0X X 0C
1 1 1
+ X 0X X 0X X 0X X 0X
= C 0C
The matrix C 0 C is positive semi-de…nite (see Appendix A.7) as required.

5.5 Residuals
What are some properties of the residuals eî = yi x0i b and prediction errors e~i = yi x0i b ( i) ,
at least in the context of the linear regression model?
Recall from (4.25) that we can write the residuals in vector notation as
^ = Me
e
1
where M = I n X (X 0 X) X 0 is the orthogonal projection matrix. Using the properties of
conditional expectation
E (^
e j X) = E (M e j X) = M E (e j X) = 0
and
var (^
e j X) = var (M e j X) = M var (e j X) M = M DM (5.13)
where D is de…ned in (5.9).
We can simplify this expression under the assumption of conditional homoskedasticity
E e2i j xi = 2
:
In this case (5.13) simplies to

2
var (^
e j X) = M :
In particular, for a single observation i; we obtain
ei j X) = E e^2i j X = (1
var (^ hii ) 2
(5.14)
since the diagonal elements of M are 1 hii as de…ned in (4.21). Thus the residuals eî are
heteroskedastic even if the errors ei are homoskedastic.
Similarly, we can write the prediction errors e~i = (1 hii ) 1 eî in vector notation. Set
1 1
M = diagf(1 h11 ) ; ::; (1 hnn ) g:
Then we can write the prediction errors as
~ = M My
e
= M M e:
We can calculate that

E (~
e j X) = M M E (e j X) = 0
and
var (~
e j X) = M M var (e j X) M M = M M DM M
which simpli…es under homoskedasticity to
2
var (~
e j X) = M M M M
2
= M MM :
The variance of the i’th prediction error is then
ei j X) = E e~2i j X
var (~
1 1 2
= (1 hii ) (1 hii ) (1 hii )
1 2
= (1 hii ) :
A residual with constant conditional variance can be obtained by rescaling. The standardized
residuals are
ei = (1 hi ) 1=2 eî ; (5.15)
and in vector notation
e = (e1 ; :::; en )0 = M 1=2
M e:
From our above calculations, under homoskedasticity,
1=2 1=2 2
var (e j X) = M MM
and
var (ei j X) = E e2i j X = 2
(5.16)
and thus these standardized residuals have the same bias and variance as the original errors when
the latter are homoskedastic.
5.6 Estimation of Error Variance

The error variance 2 = Ee2i 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:
n
1X 2
^2 = eî
n
i=1
and equals the MLE in the normal regression model (4.14).

In the linear regression model we can calculate the mean of ^ 2 : From (4.25), the properties of
projection matrices and the trace operator, observe that
1 0 1 1 1 1
^2 = e
ê^ = e0 M M e = e0 M e = tr e0 M e = tr M ee0 :
n n n n n
Then
1
E ^2 j X = tr E M ee0 j X
n
1
= tr M E ee0 j X
n
1
= tr (M D) : (5.17)
n
Adding the assumption of conditional homoskedasticity E e2i j xi = 2; so that D = I n 2; then
(5.17) simpli…es to
1
E ^2 j X = tr M 2
n
2 n k
= ;
n
the …nal equality by (4.23). 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 coe¢ cients to the sample size.
Another way to see this is to use (5.14). Note that

n
1X
E ^2 j X = E e^2i j X
n
i=1
n
1X 2
= (1 hii )
n
i=1
n k 2
=
n
using (4.22).
Since the bias takes a scale form, a classic method to obtain an unbiased estimator is by rescaling
the estimator. De…ne
n
1 X 2
s2 = eî : (5.18)
n k
i=1
By the above calculation,

E s2 j X = 2
(5.19)
so
E s2 = 2
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 esti-
mator constructed with the standardized residuals ei from (5.15) is
n n
2 1X 2 1X 1 2
= ei = (1 hii ) eî :
n n
i=1 i=1
You can show (see Exercise 5.6) that

2 2
E jX = (5.20)
and thus 2 is unbiased for 2 (in the homoskedastic linear regression model).
When the sample sizes are large and the number of regressors small, the estimators ^ 2 ; s2 and
2 are likely to be close.
5.7 Covariance Matrix Estimation Under Homoskedasticity

For inference, we need an estimate of the covariance matrix V b of the least-squares estimator.
In this section we consider the homoskedastic regression model (Assumption 5.1.2).
Under homoskedasticity, the covariance matrix takes the relatively simple form
1
1 0 2
V b = XX
n
which is known up to the unknown scale 2 . In the previous section we discussed three estimators
of 2 : The most commonly used choice is s2 ; leading to the classic covariance matrix estimator
1
0 1 0
Vb b = XX s2 : (5.21)
n
0
Since s2 is conditionally unbiased for 2 , it is simple to calculate that Vb b is conditionally
unbiased for V b under the assumption of homoskedasticity:
1
0 1 0
E Vb b j X = XX E s2 j X
n
1
1 0 2
= XX
n
= V b:
This estimator was the dominant covariance matrix estimator in applied econometrics in pre-
vious generations, and is still the default in most regression packages.
0
If the estimator (5.21) is used, but the regression error is heteroskedastic, it is possible for Vb b
1 1
1 0 1 0 1 0
to be quite biased for the correct covariance matrix V b = XX X DX XX :
n n n
For example, suppose k = 1 and 2i = x2i . The ratio of the true variance of the least-squares
estimator to the expectation of the variance estimator is
1 Pn 4
V b i=1 xi Ex4i Ex4i
= n ' =
0 1 Pn 2 Ex2 2:
E Vb b j X 2 x 2 i Ex2i
n i=1 i
(Notice that we use the fact that 2i = x2i implies 2 = E 2i = Ex2i :) This is the standardized
forth moment (or kurtosis) of the regressor xi : The ratio can be any number greater than one,
0
for example it is 3 if xi N 0; 2 : We conclude that the bias of Vb b can be arbitrarily large.
While this is an extreme and constructed example, the point is that the classic covariance matrix
estimator (5.21) may be quite biased when the homoskedasticity assumption fails.
5.8 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
1 1
1 0 1 0 1 0
V b = XX X DX XX :
n n n
This depends on the unknown matrix D which we can write as
2 2
D = diag 1 ; :::; n
= E ee0 j X
= E diag e21 ; :::; e2n j X :
Thus D is the conditional mean of diag e21 ; :::; e2n ; so the latter is an unbiased estimator for D:
Therefore, if the squared errors e2i were observable, we could construct the unbiased estimator
1 1
ideal 1 0 1 0 1 0
Vb b = XX X diag e21 ; :::; e2n X XX
n n n
n
!
1 0 1
1X 0 2 1 0 1
= XX xi xi ei XX :
n n n
i=1
Indeed,
n
!
ideal 1 0 1
1X 1 0 1
E Vb b jX = XX 0 2
xi xi E ei j X XX
n n n
i=1
n
!
1 0 1
1 X 1 0 1
0 2
= XX xi xi i XX
n n n
i=1
1 1
1 0 1 0 1 0
= XX X DX XX
n n n
= V b
ideal
verifying that Vb b is unbiased for V b
Since the errors are unobserved, V ideal
e2i b is not a feasible estimator. To construct a feasible
estimator we can replace the errors with the least-squares residuals eî ; the prediction errors e~i or
the standardized residuals ei ; e.g.
b = diag e^2 ; :::; e^2 ;
D 1 n
e
D = diag e~1 ; :::; e~2n ;
2
D = diag e21 ; :::; e2n :
Substituting these matrices into the formula for V b we obtain the estimators
1 1
1 0 1 0b 1 0
Vb b = XX X DX XX
n n n
n
!
1 0 1
1X 1 0 1
= XX xi x0i e^2i XX ;
n n n
i=1
1 1
1 0 1 0e 1 0
Ve b = XX X DX XX
n n n
n
!
1 0 1
1X 1 0 1
= XX xi x0i e~2i XX
n n n
i=1
n
!
1 0 1
1 X 2 1 0 1
= XX (1 hii ) xi x0i e^2i XX ;
n n n
i=1
and
1 1
1 0 1 0 1 0
V b = XX X DX XX
n n n
n
!
1 0 1
1X 1 0 1
= XX xi x0i e2i XX
n n n
i=1
n
!
1 0 1
1X 1 0 2 1 0 1
= XX (1 hii ) xi xi eî XX :
n n n
i=1
The estimators Vb b ; Ve b ; and Vare often called robust, heteroskedasticity-consistent, or

b
heteroskedasticity-robust covariance matrix estimators. The estimator Vb b was …rst developed
by Eicker (1963), and introduced to econometrics by White (1980), and is sometimes called the
Eicker-White or White covariance matrix estimator1 . The estimator Ve b was introduced by

Andrews (1991) based on the principle of leave-one-out cross-validation, and the estimator V b was
introduced by Horn, Horn and Duncan (1975) as a reduced-bias covariance matrix estimator.
Since (1 hii ) 2 > (1 hii ) 1 > 1 it is straightforward to show that
Vb b < V b < Ve b : (5.22)
(See Exercise 5.7.) The inequality A < B when applied to matrices means that the matrix B A
is positive de…nite
In general, the bias of the estimators Vb b ; Ve b and V b ; is quite complicated, but they greatly
simplify under the assumption of homoskedasticity (5.3). For example, using (5.14),
n
!
1 1
1 X 1 0 1
E Vb b j X = X 0X xi x0i E e^2i j X XX
n n n
i=1
n
!
1 0 1
1X 1 0 1
= XX xi x0i (1 hii ) 2 XX
n n n
i=1
n
!
1 0 1
2 1 0 1
1X 0 1 0 1
2
= XX XX xi xi hii XX
n n n n
i=1
1
1 0 2
XX
n
= V b:
This calculation shows that Vb b is biased towards zero.

Similarly, (again under homoskedasticity) we can calculate that Ve b is biased away from zero,
speci…cally
1
1 0
E Ve b j X XX 2
(5.23)
n
while the estimator V b is unbiased
1
1 0 2
E V b jX = XX : (5.24)
n
(See Exercise 5.8.)

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.
0
We have introduced four covariance matrix estimators, Vb b ; Vb b ; Ve b ; and V b : Which should
0
you use? The classic estimator Vb b 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 economet-
ric research. Of the three robust estimators, Vb b is the most commonly used, as it is the most
straightforward and familiar. However, Ve b and (in particular) V b are preferred based on their
0
improved bias. Unfortunately, standard regression packages set the classic estimator Vb b as the
default. As Ve b and V b are simple to implement, this should not be a barrier. For example, in
STATA, V b is implemented by selecting “Robust”standard errors and selecting the bias correction
option “1=(1 h)” or using the vce(hc2) option.
1 n
Often, this estimator is rescaled by multiplying by the ad hoc bias adjustment in analogy to the bias-
n k
corrected error variance estimator.
5.9 Standard Errors

A variance estimator such as Vb b is an estimate of the variance of the distribution of b . A
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.
De…nition 5.9.1 A standard error s( b ) for an real-

valued estimator b is an estimate of the standard deviation
of the distribution of b :
When is a vector with estimate b and covariance matrix estimate n b

1V
b, standard errors
for individual elements are the square roots of the diagonal elements of n b
1V
b: That is,
r rh i
^ 1 b
s( j ) = n V ^ = n 1=2
Vb b :
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 the computation of the covariance matrix estimate and standard errors, we return
to the log wage regression (4.9) of Section 4.4. We calculate that s2 = 0:215 and
b = 0:208 3:200
:
3:200 49:961
Therefore the homoskedastic and White covariance matrix estimates are
1
0 1 15:426 10:387 0:659
Vb b = 0:215 =
15:426 243 0:659 0:043
and
1 1
1 15:426 0:208 3:200 1 15:426 7:092 0:445
Vb b = = :
15:426 243 3:200 49:961 15:426 243 0:445 0:029
The standard errors are the square roots
p of the diagonal elements of these matrices.
p For example,
the White standard errors for b 0 are 7:092=61 = 0:341 and that for ^ 1 is :029=61 = :022: A
conventional format to write the estimated equation with standard errors is
\
log(W age) = 0:626 + 0:156 Education:
(:341) (:022)
Alternatively our standard errors could be calculated using Ve b or V b : We report the four
possible standard errors in the following table
q q q q
0
n 1 Vb b n 1 Vb b n 1 Ve b n 1V b
Intercept 0.412 0.341 0.361 0.351
Education 0.026 0.022 0.023 0.022
The homoskedastic standard errors are noticably di¤erent than the others, but the three robust
standard errors are quite close to one another.
5.10 Multicollinearity
If rank(X 0 X) < k; then b is not de…ned2 . This is called strict multicollinearity. This
happens when 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 ): 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 de…nition is not precise, because we have not said what it means
for a matrix to be “near singular”. This is one di¢ culty with the de…nition and interpretation of
multicollinearity.
One implication of near singularity of matrices is that the numerical reliability of the calculations
is reduced. In extreme cases it is possible that the reported calculations will be in error.
A more relevant implication of near multicollinearity is that individual coe¢ cient 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 1
XX= :
n 1
In this case
2 1 2
1 1
var b j X = = 2)
:
n 1 n (1 1
The correlation indexes collinearity, since as approaches 1 the matrix becomes singular. We
can see the e¤ect of collinearity on precision by observing that the variance of a coe¢ cient esti-
1
mate 2 n 1 2 approaches in…nity as approaches 1. Thus the more “collinear” are the
regressors, the worse the precision of the individual coe¢ cient estimates.
What is happening is that when the regressors are highly dependent, it is statistically di¢ cult
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 re‡ected 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 Goldberger’s A Course in Econometrics (1991),
which is reprinted below. To understand his basic point, you should notice how the estimation
1
variance 2 n 1 2 depends equally and symmetrically on the the correlation and the
sample size n.
2
See Appendix A.5 for the de…ntion of the rank of a matrix.
Arthur S. Goldberger
Art Goldberger (1930-2009) was one of the most distinguished members of the Depart-
ment 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 in‡uential graduate econometric textbooks, including including Econometric
Theory (1964), Topics in Regression Analysis (1968), and A Course in Econometrics (1991).
Micronumerosity
Arthur S. Goldberger
A Course in Econometrics (1991), Chapter 23.3
Econometrics texts devote many pages to the problem of multicollinearity in multiple regres-
sion, but they say little about the closely analogous problem of small sample size in estimation
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
satis…ed. 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 su¢ ciently large for the null hypothesis = 0 to be rejected, even
though Vy = 2 =n is large because of micronumerosity. But if the true is small (although
nonzero) the hypothesis = 0 may mistakenly be accepted.
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 in…nity.
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.

In the special case of the normal linear regression model introduced in Section 4.14, we can derive
exact sampling distributions for the least-squares estimator, residuals, and variance estimator.
In particular, under the normality assumption ei j xi N 0; 2 then we have the multivariate
implication
e j 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
b (X 0 X) X 0
1 2 (X 0 X) 1 0
= e N 0; 2M
e
^ M 0
1
where M = I n X (X 0 X) X 0 : Since uncorrelated normal variables are independent, it follows
that b is independent of any function of the OLS residuals including the estimated error variance
s2 or ^ 2 or prediction errors e
~:
The spectral decomposition (see equation (A.5)) of M yields
In k 0
M =H H0
0 0
where H 0 H = I n : Let u = 1H 0e N (0; H 0 H) N (0; I n ) : Then
n^ 2 (n k) s2
2
= 2
1
= 2
^0 e
e ^
1
= 2
e0 M e
1 In k 0
= 2
e0 H H 0e
0 0
In k 0
= u0 u
0 0
2
n k;
a chi-square distribution with n k degrees of freedom.

Furthermore, if standard errors are calculated using the homoskedastic formula (5.21)
h i
1
b b N 0; 2 (X 0 X)
j j j j jj N (0; 1)
= rh i q rh i = q 2 tn k
^
s( j ) 1 2 1
s (X 0 X) n k
2
n k (X 0
X) n k
n k
jj jj
a t distribution with n k degrees of freedom.

Theorem 5.11.1 Normal Regression

In the linear regression model (Assumption 5.1.1) if ei is independent of
xi and distributed N 0; 2 then
b N 0; 2 (X 0 X) 1
n^ 2 (n k)s2 2
2 = 2 n k
^
j j
tn k
s( ^ j )
These are the exact …nite-sample distributions of the least-squares estimator and variance esti-
mators, and are the basis for traditional inference in linear regression.
While elegant, the di¢ culty in applying Theorem 5.11.1 is that the normality assumption is
too restrictive to be empirical plausible, and therefore inference based on Theorem 5.11.1 has no
guarantee of accuracy. We develop a more broadly-applicable 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 student’s
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
student’s t rather than Gosset’s t!
Exercises
1 Pn 0
Exercise 5.1 Explain the di¤erence between n i=1 xi xi and E (xi x0i ) :
Exercise 5.2 True or False. If yi =P xi + ei , xi 2 R; E(ei j xi ) = 0; and eî is the OLS residual
from the regression of yi on xi ; then ni=1 x2i eî = 0:
Exercise 5.3 Prove Theorem 5.4.1.2.
Exercise 5.4 In a linear model

2
y = X + e; E(e j X) = 0; var (e j X) =
with a known function of X , the GLS estimator is
e = X0 1 1
X X0 1
y ;
the residual vector is e

^= y X e ; and an estimate of 2 is
1
s2 = ^0
e 1
^:
e
n k
(a) Find E e j X :
(b) Find var e j X :
1
(c) Prove that e
^ = M 1 e; where M 1 = I X X0 1
X X0 1
:
1
(d) Prove that M 01 1
M1 = 1 1
X X0 1
X X0 1
:
(e) Find E s2 j X :
(f) Is s2 a reasonable estimator for 2?
Exercise 5.5 Let (yi ; xi ) be a random sample with E(y j X) = X : Consider the Weighted
Least Squares (WLS) estimator of
e = X 0W X 1
X 0W y
where W = diag (w1 ; :::; wn ) and wi = xji2 , where xji is one of the xi :
(a) In which contexts would e be a good estimator?
(b) Using your intuition, in which situations would you expect that e would perform better than
OLS?
Exercise 5.6 Show (5.20) in the homoskedastic regression model.
Exercise 5.7 Prove (5.22).
Exercise 5.8 Show (5.23) and (5.24) in the homoskedastic regression model.
Chapter 6
Asymptotic Theory for Least Squares
6.1 Introduction
In the previous chapter 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 dis-
tribution, nor su¢ cient for inference (con…dence 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
f (x; y) = exp (log y log x)2 exp (log x)2
2 xy 2 2
and let ^ be the slope coe¢ cient estimate from a least-squares regression of yi on xi and a constant.
Using simulation methods, the density function of ^ was computed and plotted in Figure 6.1 for
sample sizes of n = 25; n = 100 and n = 800: The vertical line marks the true projection coe¢ cient.
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 coe¢ cient.
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
analytically calculate the exact distribution of ^ except in very special cases. Therefore we typically
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
in…nity. The primary tools of asymptotic theory are the weak law of large numbers (WLLN),
central limit theorem (CLT), and continuous mapping theorem (CMT), which were reviewed in
Chapter 2. With these tools we can approximate the sampling distributions of most econometric
estimators.
It turns out that the asymptotic theory of least-squares estimation applies equally to the pro-
jection model and the linear CEF model, and therefore the results in this chapter will be stated for
the broader projection model (Assumption 1.5.1 and Assumption 3.16.1).
6.2 Consistency of Least-Squares Estimation

In this section we use the weak law of large numbers (WLLN, Theorem 2.6.1 and Theorem 2.7.2)
and continuous mapping theorem (CMT, Theorem 2.9.1) to show that the least-squares estimator
b is consistent for the projection coe¢ cient :
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
109
CHAPTER 6. ASYMPTOTIC THEORY FOR LEAST SQUARES 110
Figure 6.1: Sampling Density of ^
converge in probability to population moments. And third, the CMT states that continuous func-
tions preserve convergence in probability. We now explain each step in brief and then in greater
detail.
First, observe that the OLS estimator
n
! 1 n
!
1 X 1 X
b= xi x0i xi yi = Qb 1Qb
xx xy
n n
i=1 i=1
is a function of the sample moments Qb xx = 1 Pn xi x0 and Q b xy = 1 Pn xi yi :

n i=1 i n i=1
Second, by an application of the WLLN these sample moments converge in probability to the
population moments. Speci…cally, the fact that (yi ; xi ) are mutually independent and identically
distributed (Assumption 1.5.1) implies that any function of (yi ; xi ) is iid, including xi x0i and xi yi :
These variables also have …nite expectations by Theorem 3.16.1.1. Under these conditions, the
WLLN (Theorem 2.7.2) implies that as n ! 1;
X n
b xx = 1
Q
p
xi x0i ! E xi x0i = Qxx (6.1)
n
i=1
and
X n
b xy = 1
Q
p
xi yi ! E (xi yi ) = Qxy : (6.2)
n
i=1
Third, the CMT ( Theorem 2.9.1) allows us to combine these equations to show that b converges
in probability to : Speci…cally, as n ! 1;
b 1Q
b=Q b
xx xy
p
! Qxx1 Qxy
= : (6.3)
p
We have shown that b ! , as n ! 1: In words, the OLS estimator converges in probability to
the projection coe¢ cient 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 xx ; Q
b xy
where g (A; b) = A 1 b is a function of A and b: The function g (A; b) is a continuous function of

A and b at all values of the arguments such that A 1 exists. Assumption 3.16.1 implies that Qxx1
exists and thus g (A; b) is continuous at A = Qxx : This justi…es the application of the CMT in
(6.3).
For a slightly di¤erent demonstration of (6.3), recall that (5.6) implies that
1
b b Q
=Q b (6.4)
xx xe
where
X n
b xe = 1
Q xi ei :
n
i=1
The WLLN and (3.25) imply

b xe p! E (xi ei ) = 0:
Q (6.5)
Therefore
1
b b Q
=Q b
xx xe
p
! Qxx1 0
=0
p
which is the same as b ! .
Theorem 6.2.1 Consistency of Least-Squares

b xx p! Qxx ; Q
Under Assumptions 1.5.1 and 3.16.1, Q b xy p! Qxy ;
p p
b
Q xe ! 0; and b ! as n ! 1:
Theorem 6.2.1 states that the OLS estimator b converges in probability to as n increases,
and thus b is consistent for . In the stochastic order notation, Theorem 6.2.1 can be equivalently
written as
b = + op (1): (6.6)
To illustrate the e¤ect of sample size on the least-squares estimator consider the least-squares
regression
2
ln(W agei ) = 0 + 1 Educationi + 2 Experiencei + 3 Experiencei + ei :
We use the sample of 30,833 white men from the March 2009 CPS. Randomly sorting the observa-
tions, and sequentially estimating the model by least-squares, starting with the …rst 40 observations,
and continuing until the full sample is used, the sequence of estimates are displayed in Figure 6.2.
You can see how the least-squares estimate changes with the sample size, but as the number of
observations increases it settles down to the full-sample estimate ^ 1 = 0:114:
0.15
0.14
0.13
OLS Estimation
0.12
0.11
0.10
0.09
0.08
5000 10000 15000 20000
Number of Observations
Figure 6.2: The least-squares estimator ^ 1 as a function of sample size n
6.3 Consistency of Sample Variance Estimators

Using the methods of Section 6.2 we can show that the estimators ^ 2 and s2 are consistent for
2: (The proof is given in Section 6.18.)
Theorem 6.3.1 Under Assumption 1.5.1 and Assumption 3.16.1,

p p
^ 2 ! 2 and s2 ! 2 as n ! 1:
One implication of this theorem is that multiple estimators can be consistent for the sample
population parameter. While ^ 2 and s2 are unequal in any given application, they are close in
value when n is very large.
6.4 Asymptotic Normality

We started this chapter discussing the need for an approximation to the distribution of the OLS
estimator b : In Section 6.2 we showed that b converges in probability to . Consistency is a useful
…rst step, but in itself does not provide a useful approximation to 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.
p
Take equation (6.4) and multiply it by n: This yields the expression
n
! 1 n
!
p 1 X 1 X
n b = 0
xi xi p xi ei : (6.7)
n n
i=1 i=1
p
This shows that the normalized and centered estimator n b is a function of the sample
1 Pn 0 1 Pn
average n i=1 xi xi and the normalized sample average pn i=1 xi ei : Furthermore, the latter has
mean zero so the central limit theorem (CLT, Theorem 2.8.1) applies.
The product xi ei is iid (since the observations are iid) and mean zero (since E (xi ei ) = 0):
De…ne the k k covariance matrix
= E xi x0i e2i : (6.8)
We require the elements of to be …nite, or equivalently that E kxi ei k2 < 1: Usingkxi ei k2 =
2 2
kxi k ei and the Cauchy-Schwarz Inequality (B.20),
1=2 1=2
E kxi ei k2 = E kxi k2 e2i E kxi k4 Ee4i (6.9)
which is …nite if xi and ei have …nite fourth moments. As ei is a linear combination of yi and xi ;
it is su¢ cient that the observables have …nite fourth moments (Theorem 3.16.1.6).
Assumption 6.4.1 In addition to Assumption 3.16.1, Eyi4 < 1 and

E kxi k4 < 1:
Under Assumption 6.4.1 the CLT (Theorem 2.8.1) can be applied.
Theorem 6.4.1 Under Assumption 1.5.1 and Assumption 6.4.1, as n ! 1

n
1 X d
p xi ei ! N (0; ) (6.10)
n
i=1
where = E xi x0i e2i :
Putting together (6.1), (6.7), and (6.10),

p d
n b ! Qxx1 N (0; )
= N 0; Qxx1 Qxx1
as n ! 1; 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 6.4.2 Asymptotic Normality of Least-Squares Estimator

Under Assumption 1.5.1 and Assumption 6.4.1, as n ! 1
p d
n b ! N (0; V )
where
V = Qxx1 Qxx1 ; (6.11)
Qxx = E (xi x0i ) ; and = E xi x0i e2i :
In the stochastic order notation, Theorem 6.4.2 implies that
b= + Op (n 1=2
) (6.12)
and
b = Op (n 1=2
)
which is stronger than (6.6).

p
The matrix V = avar( b ) is the variance of the asymptotic distribution of n b : Con-
sequently, V is often referred to as the asymptotic covariance matrix of b : The expression
V = Qxx1 Qxx1 is called a sandwich form. It might be worth noticing that there is a di¤erence
between the variance of the asymptotic distribution given in (6.11) and the …nite-sample conditional
variance in the CEF model as given in (5.11):
1 1
1 0 1 0 1 0
V b = XX X DX XX :
n n n
While V and V b are di¤erent, the two are close if n is large. Indeed, as n ! 1
p
V b !V :
There is a special case where and V simplify. We say that ei is a Homoskedastic Pro-
jection Error when
cov(xi x0i ; e2i ) = 0: (6.13)
Condition (6.13) holds in the homoskedastic linear regression model, but is somewhat broader.
Under (6.13) the asymptotic variance formulas simplify as
= E xi x0i E e2i = Qxx 2

(6.14)
V = Qxx1 Qxx1 = Qxx1 2 V 0
(6.15)
In (6.15) we de…ne V 0 = Qxx1 2 whether (6.13) is true or false. When (6.13) is true then V = V 0 ;
otherwise V 6= V 0 : We call V 0 the homoskedastic asymptotic covariance matrix.
Theorem 6.4.2 states that the sampling distribution of the least-squares estimator, after rescal-
ing, is approximately normal when the sample size n is su¢ ciently large. This holds true for all joint
distributions of (yi ; xi ) which satisfy the conditions of Assumption 6.4.1, and is therefore broadly
applicable. Consequently, asymptotic normality is routinely used to approximate the …nite sample
p
distribution of n b :
A di¢ culty is that for any …xed n the sampling distribution of b can be arbitrarily far from the
normal distribution. In Figure 6.1 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
density f (e) = 2 jej ; jej 1: If > 2 the error ei has zero mean and variance =( 2):
As approaches 2, however,q its variance diverges to in…nity. In this context the normalized least-
squares slope estimator n 2 ^ 1 1 has the N(0; 1) asymptotic distibution for any > 2.
q
In Figure 6.3 we display the …nite sample densities of the normalized estimator n 2 ^ 1 1 ;
setting n = 100 and varying the parameter . For = 3:0 the density is very close to the N(0; 1)
Figure 6.3: Density of Normalized OLS estimator with Double Pareto Error
density. As diminishes the density changes signi…cantly, concentrating most of the probability
mass around zero.
Another example is shown in Figure 6.4. Here the model is yi = + ei where
uki E uki
ei = (6.16)
2 1=2
E u2k
i E uki
p
and ui N(0; 1): We show the sampling distribution of n b 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 6.3 and 6.4 is that the N(0; 1) asymptotic approximation is never guaranteed
to be accurate.
6.5 Joint Distribution

Theorem 6.4.2 gives the joint asymptotic distribution of the coe¢ cient estimates. We can use
the result to study the covariance between the coe¢ cient estimates. For example, suppose k = 2
and write the estimates as ( ^ 1 ; ^ 2 ): For simplicity suppose that the regressors are mean zero. Then
we can write
2
1 1 2
Qxx = 2
1 2 2
where 21 and 22 are the variances of x1i and x2i ; and is their correlation. If the error is ho-
moskedastic, then the asymptotic variance matrix for ( ^ 1 ; ^ 2 ) is V 0 = Qxx1 2 : By the formula for
inversion of a 2 2 matrix,
1 2
1 2
Qxx1 = 2 2 (1 2)
2
2 :
1 2 1 2 1
Thus if x1i and x2i are positively correlated ( > 0) then ^ 1 and ^ 2 are negatively correlated (and
vice-versa).
Figure 6.4: Density of Normalized OLS estimator with error process (6.16)
For illustration, Figure 6.5 displays the probability contours of the joint asymptotic distribution
of ^ 1 ^
1 and 2
2
2 when 1 = 2 =
2 2 = 1 and = 0:5: The coe¢ cient estimates are negatively
correlated since the regressors are positively correlated. This means that if ^ 1 is unusually negative,
it is likely that ^ 2 is unusually positive, or conversely. It is also unlikely that we will observe both
^ and ^ unusually large and of the same sign.
1 2
This …nding that the correlation of the regressors is of opposite sign of the correlation of the coef-
…cient 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 f 1; +1g; 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 6.5
when the error is homoskedastic.
Now suppose that the error is heteroskedastic. Speci…cally, suppose that E e2i j x1i = x2i =
5 1
and E e2i j x1i 6= x2i = : You can check that E e2i = 1; E x21i e2i = E x22i e2i = 1 and
4 4
7
E x1i x2i e2i = : Therefore
8
V = Qxx1 Qxx1
2 32 32 3
1 7 1
9 6 1 2 7 6 1 8 7 6 1 2 7
= 4 1 54 7 54 1 5
16 1 1 1
2 8 2
2 3
1
46 1 4 7
= 4 1 5:
3 1
4
Thus the coe¢ cient estimates ^ 1 and ^ 2 are positively correlated (their correlation is 1=4:) The
joint probability contours of their asymptotic distribution is displayed in Figure 6.6. We can see
how the two estimates are positively associated.
Figure 6.5: Contours of Joint Distribution of ( ^ 1 ; ^ 2 ); homoskedastic case
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
estimates.
We can extend the above analysis to study the covariance between coe¢ cient sub-vectors. For
example, partitioning x0i = (x01i ; x02i ) and 0 = 01 ; 02 ; we can write the general model as
yi = x01i 1 + x02i 2 + ei
0 0 0
and the coe¢ cient estimates as b = b 1 ; b 2 : Make the partitions
Q11 Q12 11 12
Qxx = ; = : (6.17)
Q21 Q22 21 22
From (3.37)
Q1112 Q1112 Q12 Q221
Qxx1 =
Q2211 Q21 Q111 Q2211
where Q11 2 = Q11 Q12 Q221 Q21 and Q22 1 = Q22 Q21 Q111 Q12 . Thus when the error is ho-
moskedastic,
cov b 1 ; b 2 = 2
Q1112 Q12 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 = (6.18)
V 21 V 22
where
V 11 = Q1112 11 Q12 Q221 21
1
12 Q22 Q21 + Q12 Q221 1
22 Q22 Q21 Q1112 (6.19)
V 21 = Q2211 21 Q21 Q111 11
1
22 Q22 Q21 + Q21 Q111 1
12 Q22 Q21 Q1112 (6.20)
V 22 = Q2211 22 Q21 Q111 12
1
21 Q11 Q12 + Q21 Q111 1
11 Q11 Q12 Q2211 (6.21)
Unfortunately, these expressions are not easily interpretable.
Figure 6.6: Contours of Joint Distribution of ^ 1 and ^ 2 ; heteroskedastic case
6.6 Uniformly Consistent Residuals*

We have described the least-squares residuals eî as estimates of the errors ei : Are eî consistent
for ei ? Notice that we can write the residual as
eî = yi x0i b
= ei + x0i x0i b
= ei x0i b : (6.22)
p
Since b ! 0 it seems reasonable to guess that eî will be close to ei if n is large.
We can bound the di¤erence in (6.22) using the Schwarz inequality (A.7) to …nd
j^
ei ei j = x0i b kxi k b : (6.23)
To bound (6.23) we can use b = Op (n 1=2 ) from Theorem 6.4.2, but we also need to bound
the random variable kxi k.
The key is Theorem 2.12.1 which shows that E kxi k4 < 1 implies xi = op n1=4 uniformly in
i; or
p
n 1=4 max kxi k ! 0:
1 i n
Applied to (6.23) we obtain
max j^
ei ei j max kxi k b
1 i n 1 i n
= op n1=4 Op (n 1=2
)
1=4
= op (n ):
We have shown the following.

Theorem 6.6.1 Under Assumptions 1.5.1 and 6.4.1, uniformly in 1 i n

1=4
eî = ei + op (n ): (6.24)
What about the squared residuals e^2i ? Squaring the two sides of (6.24) we obtain
2
e^2i = ei + op (n 1=4
)
= e2i + 2ei op (n 1=4
) + op (n 1=2
)
= e2i + op (1) (6.25)
uniformly in 1 i n; since ei = op n1=4 when E jei j4 < 1 by Theorem 2.12.1.
Theorem 6.6.2 Under Assumptions 1.5.1 and 6.4.1, uniformly in 1 i n
e^2i = e2i + op (1)
6.7 Asymptotic Leverage*

Recall the de…nition of leverage from (4.21)
1
hii = x0i X 0 X xi :
These are the diagonal elements of the projection matrix P and appear in the formula for leave-
one-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 (A 1 ) = ( min (A)) 1 :
p p
Since n1 X 0 X ! Qxx > 0 then by the CMT, min n1 X 0 X ! min (Qxx ) > 0: (The latter is
positive since Qxx is positive de…nite and thus all its eigenvalues are positive.) Then by the Trace
Inequality (A.10)
1
hii = x0i X 0 X xi
!
1
1 0 1
= tr XX xi x0i
n n
!
1
1 0 1
max XX tr xi x0i
n n
1
1 0 1
= min XX kxi k2
n n
1
( min (Qxx ) + op (1)) 1 max kxi k2 : (6.26)
n1 i n
Theorem 2.12.1 shows that E kxi k2 < 1 implies

1=2 p
n max kxi k ! 0
1 i n
and thus
p
n 1
max kxi k2 ! 0:
1 i n
It follows that (6.26) is op (1); uniformly in i:
Theorem 6.7.1 Under Assumption 1.5.1 and E kxi k2 < 1, uniformly in

1 i n, hii = op (1):
Theorem (6.7.1) 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 in‡uential, unless one of these conditions is violated.
6.8 Consistent Covariance Matrix Estimation

In Sections 5.7 and 5.8 we introduced estimators of the …nite-sample covariance matrix of the
least-squares estimator in the regression model. In this section we show that these estimators are
consistent for the asymptotic covariance matrix.
First, consider the covariance matrix estimate constructed under the assumption of homoskedas-
ticity:
1
0 1 0 b 1 s2 :
Vb b = XX s2 = Qxx
n
p p
b xx ! Qxx (Theorem 6.2.1), s2 ! 2 (Theorem 6.3.1), and Qxx is invertible (Assumption
Since Q
3.16.1), it follows that
0
b 1 s2 p! Q 1 2 = V 0
Vb b = Q xx xx
0
so that Vb b is consistent for V 0 ; the homoskedastic covariance matrix.
Theorem 6.8.1 Under Assumption 1.5.1 and Assumption 3.16.1,

0 p
Vb b ! V 0 as n ! 1:
Now consider the heteroskedasticity-robust covariance matrix estimators Vb b ; Ve b ; and V b .

Writing
n
b 1X
= xi x0i e^2i ; (6.27)
n
i=1
Xn
e 1 2
= (1 hii ) xi x0i e^2i
n
i=1
and
n
1X 1
= (1 hii ) xi x0i e^2i
n
i=1
as moment estimators for = E xi x0i e2i ; then the covariance matrix estimators are
Vb b b 1 bQ
= Q b 1;
xx xx
Ve b b 1 eQ
= Q b 1;
xx xx
and
V b
b 1 Q
=Q b 1:
xx xx
We can show that b , e , and are consistent for : Combined with the consistency of Q b xx
b e
for Qxx and the invertibility of Qxx we …nd that V b , V b ; and V b converge in probability to
Qxx1 Qxx1 = V : The complete proof is given in Section 6.18.
Theorem 6.8.2 Under Assumption 1.5.1 and Assumption 6.4.1, as n ! 1;

b p! ; e p! ; p p p p
! ; Vb b ! V ; Ve b ! V ; and V b ! V :
6.9 Functions of Parameters

Sometimes we are interested in a lower-dimensional function of the parameter vector =
( 1 ; :::; k ): For example, we may be interested in a single coe¢ cient j or a ratio j = l : In these
cases we can write the parameter of interest as a function of : Let h : Rk ! Rq denote this
function and let
= h( )
denote the parameter of interest. The estimate of is
b = h( b ):
p
By the continuous mapping theorem (Theorem 2.9.1) and the fact b ! we can deduce that
b is consistent for .
Theorem 6.9.1 Under Assumption 1.5.1 and Assumption 3.16.1, if h( ) is con-

p
tinuous at the true value of ; then as n ! 1; b ! :
Furthermore, by the Delta Method (Theorem 2.10.3) we know that b is asymptotically normal.
Theorem 6.9.2 Asymptotic Distribution of Functions of Parameters

Under Assumption 1.5.1 and Assumption 6.4.1, if h( ) is continuously di¤ erentiable at
the true value of ; then as n ! 1;
p d
n b ! N (0; V ) (6.28)
where
V = H0 V H (6.29)
and
@
H = h( )0 :
@
In many cases, the function h( ) is linear:

h( ) = R0
for some k q matrix R: In this case, H = R.
In particular, if R is a “selector matrix”
I
R= (6.30)
0
0 0 0
so that = R0 = 1 for =( 1; 2) ; then
I
V = I 0 V = V 11 ;
0
where V 11 is given in (6.19). Under homoskedasticity the covariance matrix (6.19) simpli…es to
V 011 = Q1112 2
:
We have shown that for the case (6.30) of a subset of coe¢ cients, (6.28) is
p d
n b1 1 ! N (0; V 11 )
with V 11 given in (6.19).
6.10 Asymptotic Standard Errors

How do we estimate the covariance matrix V for b? From (6.29) we see we need estimates of
H and V . We already have an estimate of the latter, Vb b (or Ve b or V b ). To estimate H we
use
c = @ h( b ):
H
@
Putting the parts together we obtain
c 0 Vb b H
Vb = H c
as the covariance matrix estimator for b: As the primary justi…cation for Vb is the asymptotic
approximation (6.28), Vb is often called an asymptotic covariance matrix estimator.
In particular, when h( ) is linear h( ) = R0 then
Vb = R0 Vb b R:
When R takes the form of a selector matrix as in (6.30) then
h i
Vb = Vb 11 = Vb b ;
11
the upper-left block of the covariance matrix estimate Vb b :

When q = 1 (so h( ) is real-valued), the standard error for ^ is the square root of Vb ; that is,
q r
s(^) = n 1=2
Vb = n 1=2 c 0 Vb b H
H c :
This is known as an asymptotic standard error for s(^).

p
The estimator Vb is consistent for V under the conditions of Theorem 6.9.2 since Vb b ! V
by Theorem 6.8.2, and
c = @ h( b )0 p! @ h( )0 = H
H
@ @
p
since b ! and the function @ 0
h( ) is continuous.
@
Theorem 6.10.1 Under Assumption 1.5.1 and Assumption 6.4.1, if h( ) is continuously

di¤ erentiable at the true value of ; then as n ! 1;
p
Vb !V :
6.11 t statistic
Let = h( ) : Rk ! R be any parameter of interest (for example, could be a single element
of ), b its estimate and s(b) its asymptotic standard error. Consider the statistic
b
tn ( ) = : (6.31)
s(b)
Di¤erent writers have called (6.31) a t-statistic, a t-ratio, a z-statistic or a studentized sta-
tistic. We won’t be making such distinctions and will 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.
d
Theorem 6.11.1 tn ( ) ! N (0; 1)
Thus the asymptotic distribution of the t-ratio tn ( ) is the standard normal. Since this dis-
tribution does not depend on the parameters, we say that tn ( ) is asymptotically pivotal. In
special cases (such as the normal regression model, see Section 4.14), 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.
6.12 Con…dence Intervals

A con…dence interval Cn is an interval estimate of 2 R: It is a function of the data and
hence is random. It is designed to cover with high probability. Either 2 Cn or 2 = Cn : Its
coverage probability is Pr( 2 Cn ). The convention is to design con…dence intervals to have
coverage probability approximately equal to a pre-speci…ed target, typically 90% or 95%, or more
generally written as (1 )% for some 2 (0; 1): By reporting a (1 )% con…dence interval Cn ;
we are stating that the true lies in Cn with (1 )% probability across repeated samples.
There is not a unique method to construct con…dence intervals. For example, a simple (yet
silly) interval is
R with probability 1
Cn = b with probability
By construction, if b has a continuous distribution, Pr( 2 Cn ) = 1 ; so this con…dence interval

has perfect coverage, but Cn is uninformative about : This is not a useful con…dence interval.
When we have an asymptotically normal parameter estimate b with standard error s(b); it turns
out that a generally reasonable con…dence interval for takes the form
h i
Cn = b c s(b); b + c s(b) (6.32)
where c > 0 is a pre-speci…ed constant. This con…dence interval is symmetric about the point
estimate b; and its length is proportional to the standard error s(b):
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 = f : jtn ( )j cg = : c c :
s(b)
The coverage probability of this con…dence interval is
Pr ( 2 Cn ) = Pr (jtn ( )j c)
which is generally unknown, but we can approximate the coverage probability by taking the asymp-
totic limit as n ! 1: Since tn ( ) is asymptotically standard normal (Theorem 6.11.1), it follows
that as n ! 1 that
Pr ( 2 Cn ) ! Pr (jZj c) = (c) ( c)
where Z N (0; 1) and (u) = Pr (Z u) is the standard normal distribution function. We call
this the asymptotic coverage probability, and it is a function only of c:
As we mentioned before, the convention is to design the con…dence interval to have a pre-
speci…ed asymptotic coverage probability 1 ; typically 90% or 95%. This means selecting the
constant c so that
(c) ( c) = 1 :
E¤ectively, 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 con…dence interval in applied econometric
practice h i
Cn = b 2s(b); b + 2s(b) :
This is a useful rule-of thumb. This asymptotic 95% con…dence interval Cn is simple to compute
and can be roughly calculated from tables of coe¢ cient 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
meaningless.)
Con…dence intervals are a simple yet e¤ective tool to assess estimation uncertainty. When
reading a set of empirical results, look at the estimated coe¢ cient estimates and the standard
errors. For a parameter of interest, compute the con…dence interval Cn and consider the meaning
of the spread of the suggested values. If the range of values in the con…dence interval are too wide
to learn about ; then do not jump to a conclusion about based on the point estimate alone.
6.13 Regression Intervals

In the linear regression model the conditional mean of yi given xi = x is
m(x) = E (yi j xi = x) = x0 :
In some cases, we want to estimate m(x) at a particular point x: Notice that this is a (linear)
function 0
qof : Letting h( ) = x and = h( ); we see that m(x) b = b = x0 b and H = x; so
s(b) = n 1 x0 Vb x: Thus an asymptotic 95% con…dence interval for m(x) is
q
0b b
1 x0 V
x 2 n x :
Figure 6.7: Wage on Education Regression Intervals
It is interesting to observe that if this is viewed as a function of x; the width of the con…dence set
is dependent on x:
To illustrate, we return to the log wage regression (4.9) of Section 4.4. The estimated regression
equation is
\
log(W age) = x0 b = 0:626 + 0:156x:
where x = Education. The White covariance matrix estimate is
7:092 0:445
Vb b =
0:445 0:029
and the sample size is n = 61: Thus the 95% con…dence interval for the regression takes the form
r
1
0:626 + 0:156x 2 (7:092 0:89x + 0:029x2 ) :
61
The estimated regression and 95% intervals are shown in Figure 6.7. Notice that the con…dence
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 con…dence intervals are especially useful when the
regression includes nonlinear terms. To illustrate, consider the log wage regression (4.10) which
includes experience and its square.
\
log(W age) = 1:06 + 0:116 education + 0:010 experience 0:014 experience2 =100 (6.33)
and has n = 2454 observations. We are interested in plotting the regression estimate and regression
intervals as a function of experience. Since the regression also includes education, in order to plot
the estimates in a simple graph we need to …x education at a speci…c value. We select education=12.
This only a¤ects the level of the estimated regression, since education enters without an interaction.
De…ne the points of evaluation 0 1
1
B 12 C
x=B @
C
A
x
2
x =100
Figure 6.8: Wage on Experience Regression Intervals
where x =experience. The covariance matrix estimate is

0 1
22:92 1:0601 0:56687 0:86626
B 1:0601 0:06454 :0080737 :0066749 C
Vb b = B
@ 0:56687 :0080737
C:
:040736 :075583 A
0:86626 :0066749 :075583 0:14994
Thus the regression interval for education=12, as a function of x =experience is
1:06 + 0:116 12 + 0:010 x 0:014 x2 =100

v 0 10 1
u
u 22:92 1:0601 0:56687 0:86626 1
u B C B C
1u 1 1:0601 0:06454 :0080737 :0066749 12
u 1 12 x x2 =100 B @ 0:56687 :0080737
CB C
50 t 2454 :040736 :075583 A @ x A
0:86626 :0066749 :075583 0:14994 x2 =100
2 p
= 2:452 + 0:010 x :00014 x2 27:592 3:8304 x + 0:23007 x2 0:00616 x3 + 0:0000611 x4
100
The estimated regression and 95% intervals are shown in Figure 6.8. The regression interval
widens greatly for small and large values of experience, indicating considerable uncertainty about
the e¤ect of experience on mean wages for this population. The con…dence bands take a more
complicated shape than in Figure 6.7 due to the nonlinear speci…cation.
6.14 Quadratic Forms

Let = h( ) : Rk ! Rq be any parameter vector of interest, b its estimate and Vb its
covariance matrix estimator. Consider the quadratic form
0 1
Wn ( ) = n b Vb b : (6.34)
When q = 1 then Wn ( ) = tn ( )2 is the square of the t-ratio. When q > 1 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 6.9.2 and Theorem
6.10.1, which show that
p d
n b ! Z N (0; V )
and
p
Vb !V :
It follows that
p 0 1p d
Wn ( ) = n b Vb n b ! Z0 V 1
Z (6.35)
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 6.14.1 If Z N (0; A) with A > 0; q q; then Z 0 A 1

Z 2;
q a chi-square
random variable with q degrees of freedom.
The asymptotic distribution in (6.35) takes exactly this form. It follows that Wn ( ) converges
in distribution to a chi-square random variable.
Theorem 6.14.2 Under Assumption 1.5.1 and Assumption 6.4.1, if h( ) is continuously

di¤ erentiable at the true value of ; then as n ! 1;
d 2
Wn ( ) ! q:
6.15 Con…dence Regions

A con…dence region Cn is a generalization of a con…dence interval to the case 2 Rq with q > 1:
A con…dence region Cn is a set in Rq intended to cover the true parameter value with a pre-selected
probability 1 : Thus an ideal con…dence region has the coverage probability Pr( 2 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 satis…es the conditions of Theorem 6.14.2, a good choice for a
con…dence region is the ellipse
Cn = f : Wn ( ) c1 g :
with c1 the 1 ’th quantile of the 2q distribution. (Thus Fq (c1 )=1 :) These quantiles
can be found from a critical value table for the 2q distribution.
Theorem 6.14.2 implies
2
Pr ( 2 Cn ) ! Pr q c1 =1
which shows that Cn has asymptotic coverage (1 )%:

To illustrate the construction of a con…dence region, consider the estimated regression (6.33) of
the model
\
log(W age) = + education + experience + experience2 =100:
1 2 3
Suppose that the two parameters of interest are the percentage return to education 1 = 100 1 and
the percentage return to experience for individuals with 10 years experience 2 = 100 2 + 20 3 .
(We need to condition on the level of experience since the regression is quadratic in experience.)
These two parameters are a linear transformation of the regression parameters with point estimates
b= 0 100 0 0 b= 11:6
;
0 0 100 20 0:72
and have the covariance matrix estimate
0 1
0 0
0 100 0 0 B 100 0 C
Vb = Vb b B
@ 0 100 A
C
0 0 100 20
0 20
645:4 67:387
=
67:387 165
with inverse
1 0:0016184 0:00066098
Vb = :
0:00066098 0:0063306
Thus the Wald statistic is
0 1
Wn ( ) = n b Vb b
0
11:6 1 0:0016184 0:00066098 11:6 1
= 2454
0:72 2 0:00066098 0:0063306 0:72 2
2 2
= 3:97 (11:6 1) 3:2441 (11:6 1 ) (0:72 2) + 15:535 (0:72 2)
The 90% quantile of the 22 distribution is 4.605 (we use the 22 distribution as the dimension
of is two), so an asymptotic 90% con…dence region for the two parameters is the interior of the
ellipse
2 2
3:97 (11:6 1) 3:2441 (11:6 1 ) (0:72 2) + 15:535 (0:72 2) = 4:605
which is displayed in Figure 6.9. Since the estimated correlation of the two coe¢ cient estimates is
small (about 0.2) the ellipse is close to circular.
6.16 Semiparametric E¢ ciency in the Projection Model

In Section 5.4 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 least-
squares. 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 unsatis-
factory as the projection model is more relevant for empirical application. The question remains:
what is the most e¢ cient estimator of the projection coe¢ cient (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 2.13.2. It follows that the least-squares estimator is semiparametri-
cally e¢ cient 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.
Figure 6.9: Con…dence Region for Return to Experience and Return to Education
To see this, it is worth rephrasing Proposition 2.13.2 with amended notation. Suppose that a pa-
1 Pn
rameter of interest is = g( n) where = Ez i ; for which the moment estimators are b = n i=1 o zi
b 2
and = g(b ): Let L2 (g) = F : E kzk < 1; g (u) is continuously di¤erentiable at u = Ez be
the set of distributions for which b satis…es the central limit theorem.
Proposition 6.16.1 In the class of distributions F 2 L2 (g); b is semi-

parametrically e¢ cient for in the sense that its asymptotic variance equals
the semiparametric e¢ ciency bound.
Proposition 6.16.1 says that under the minimal conditions in which b is asymptotically normal,
then no semiparametric estimator can have a smaller asymptotic variance than b.
To show that an estimator is semiparametrically e¢ cient it is su¢ cient to show that it falls
in the class covered by this Proposition. To show that the projection model falls in this class, we
write = Qxx1 Qxy = g ( ) where = Ez i and z i = (xi x0i ; xi yi ) : The class L2 (g) equals the class
of distributions n o
L4 ( ) = F : Ey 4 < 1; E kxk4 < 1; Exi x0i > 0 :
Proposition 6.16.2 In the class of distributions F 2 L4 ( ); the least-

squares estimator b is semiparametrically e¢ cient for .
The least-squares estimator is an asymptotically e¢ cient estimator of the projection coe¢ cient
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 ine¢ cient. 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 e¢ cient estimation of
the unknown parameter.
From Proposition 6.16.1 we can also deduce that plug-in estimators b = h( b ) are semiparamet-
rically e¢ cient estimators of = h( ) when h is continuously di¤erentiable. We can also deduce
that other parameters estimators are semiparametrically e¢ cient, such as ^ 2 for 2 : To see this,
note that we can write
2 2
= E yi x0i
= Eyi2 2E yi x0i + 0
E xi x0i
= Qyy Qyx Qxx1 Qxy
which is a smooth function of the moments Qyy ; Qyx and Qxx : Similarly the estimator ^ 2 equals
n
1X 2
^2 = eî
n
i=1
b yy
= Q Q b 1Q
b yx Q b
xx xy
Since the variables yi2 ; yi x0i and xi x0i all have …nite variances when F 2 L4 ( ); the conditions of
Proposition 6.16.1 are satis…ed. We conclude:
Proposition 6.16.3 In the class of distributions F 2 L4 ( ); ^ 2 is semi-

parametrically e¢ cient for 2 .
6.17 Semiparametric E¢ ciency in the Homoskedastic Regression

Model*
In Section 6.16 we showed that the OLS estimator is semiparametrically e¢ cient in the projec-
tion model. What if we restrict attention to the classical homoskedastic regression model? Is OLS
still e¢ cient in this class? In this section we derive the asymptotic semiparametric e¢ ciency 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 e¢ cient in this class as well.
Recall that in the homoskedastic regression model the asymptotic variance of the OLS estimator
b for is V 0 = Q 1 2 : Therefore, as described in Section 2.13, it is su¢ cient to …nd a parametric
xx
submodel whose Cramer-Rao bound for estimation of is V 0 : This would establish that V 0 is
the semiparametric variance bound and the OLS estimator b is semiparametrically e¢ cient for :
Let the joint density of y and x be written as f (y; x) = f1 (y j x) f2 (x) ; the product of the
conditional density of y given x and the marginal density of x. Now consider the parametric
submodel
f (y; x j ) = f1 (y j x) 1 + y x0 x0 = 2 f2 (x) : (6.36)
You can check that in this submodel the marginal density of x is f2 (x) and the conditional density
of y given x is f1 (y j x) 1 + (y x0 ) (x0 ) = 2 : To see thatR the latter is a valid conditional
density, observe that the regression assumption implies that yf1 (y j x) dy = x0 and therefore
Z Z Z
0 0 2
f1 (y j x) 1 + y x x = dy = f1 (y j x) dy + f1 (y j x) y x0 dy x0 = 2
= 1:
In this parametric submodel the conditional mean of y given x is

Z
E (y j x) = yf1 (y j x) 1 + y x0 x0 = 2 dy
Z Z
= yf1 (y j x) dy + yf1 (y j x) y x0 x0 = 2
dy
Z Z
2
= yf1 (y j x) dy + y x0 f1 (y j x) x0 = 2
dy
Z
+ y x0 f1 (y j x) dy x0 x0 = 2
= x0 ( + ) ;
R
using the homoskedasticity assumption (y x0 )2 f1 (y j x) dy = 2 : This means that in this
parametric submodel, the conditional mean is linear in x and the regression coe¢ cient is ( ) =
+ :
We now calculate the score for estimation of : Since
@ @ x (y x0 ) = 2
log f (y; x j ) = log 1 + y x0 x0 = 2
=
@ @ 1 + (y x0 ) (x0 ) = 2
the score is
@
log f (y; x j 0 ) = xe= 2 :
s=
@
The Cramer-Rao bound for estimation of (and therefore ( ) as well) is
1 1
E ss0 = 4
E (xe) (xe)0 = 2
Qxx1 = V 0 :
We have shown that there is a parametric submodel (6.36) 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 6.17.1 In the homoskedastic regression model, the semipara-

metric variance bound for estimation of is V 0 = 2 Qxx1 and the OLS
estimator is semiparametrically e¢ cient.
This result is similar to the Gauss-Markov theorem, in that it asserts the e¢ ciency of the least-
squares estimator in the context of the homoskedastic regression model. The di¤erence is that the
Gauss-Markov theorem states that OLS has the smallest variance among the set of unbiased linear
estimators, while Theorem 6.17.1 states that OLS has the smallest asymptotic variance among all
regular estimators. This is a much more powerful statement.

Proof of Theorem 6.3.1. Note that
eî = yi x0i b
= ei + x0i x0i b
= ei x0i b :
Thus 0
e^2i = e2i 2ei x0i b + b xi x0i b (6.37)
and
n
1X 2
^2 = eî
n
i=1
n n
! n
!
1X 2 1X b
0 1X
= ei 2 ei x0i + b xi x0i b
n n n
i=1 i=1 i=1
p 2
!
as n ! 1; the last line using the WLLN and Theorem 6.2.1. Thus ^ 2 is consistent for 2:
Finally, since n=(n k) ! 1 as n ! 1; it follows that as n ! 1;
n p
s2 = ^2 ! 2
:
n k
p
Proof of Theorem 6.8.2. We …rst show b ! : Note that
n
b 1X
= xi x0i e^2i
n
i=1
n n
1X 1X
= xi x0i e2i + xi x0i e^2i e2i : (6.38)
n n
i=1 i=1
We now examine each k k sum on the right-hand-side of (6.38) in turn.

Take the …rst term on the right-hand-side of (6.38). Since xi x0i e2i = kxi k2 e2i ; then by the
Cauchy-Schwarz Inequality (B.20) and Assumption 6.4.1,
1=2
E xi x0i e2i = E kxi k2 e2i E kxi k4 E e4i < 1:
Since this expectation is …nite, we can apply the WLLN (Theorem 2.7.2) to …nd that
n
1X p
xi x0i e2i ! E xi x0i e2i = :
n
i=1
Now take the second term on the right-hand-side of (6.38). By the Triangle Inequality (A.9),
the fact that E kxi k2 < 1 and Theorem 6.6.2,
n n
1X 1X
xi x0i e^2i e2i xi x0i e^2i e2i
n n
i=1 i=1
n
1X
= kxi k2 e^2i e2i
n
i=1
n
1X
kxi k2 max e^2i e2i
n 1 i n
i=1
= Op (1)op (1)
= op (1)
p
Together, we have established that b ! as claimed.
Combined with (6.1) and the invertibilility of Qxx ;
b 1 bQ
Vb b = Q b 1 p! Q 1 Q 1 = V ;
xx xx xx xx
p
from which it follows that Vb b ! V as n ! 1:
p p p
We also need to show that e ! and ! , from which it will follow that Ve b ! V
p e it is su¢ cient to show that e p! . Notice that
and V b ! V : Since b
1X
n h i
e b = xi x0i (1 hii ) 2
1 e^2i
n
i=1
1 Xn h i 1X
n h i
2 2
= xi x0i (1 hii ) 1 e2i + xi x0i (1 hii ) 1 e^2i e2i :
n n
i=1 i=1
Note that Theorem 6.7.1 states max1 i n hii = op (1); and thus by the CMT
2
max (1 hii ) 1 = op (1):
1 i n
Thus
n n
e b 1X 2 1X 2
xi x0i (1 hii ) 1 e2i + xi x0i (1 hii ) 1 e^2i e2i
n n
i=1 i=1
n n
2X 2 1X 2
xi x0i e2i max (1 hii ) 1 + xi x0i max (1 hii ) 1 max e^2i e2i
n 1 i n n 1 i n 1 i n
i=1 i=1
= Op (1)op (1) + Op (1)op (1)op (1)
= op (1)
p p p
Since b ! it follows that e ! and ! :
p d p
Proof of Theorem 6.11.1. By Theorem 6.9.2, n b ! N (0; V ) and Vb^ ! V : Thus
b
tn ( ) =
s(b)
p b
n
= q
Vbb
d N (0; V )
! p
V
= N (0; 1)
The last equality is by the property that linear scales of normal distributions are normal.
Exercises
Exercise 6.1 Take the model yi = x01i 1 +x02i 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: De…ne the ridge
regression estimator
n
! 1 n
!
X X
b= xi x0i + I k xi yi (6.39)
i=1 i=1
where > 0 is a …xed constant. Find the probability limit of b as n ! 1: Is b consistent for ?
Exercise 6.3 For the ridge regression estimator (6.39), set = cn where c > 0 is …xed as n ! 1:
Find the probability limit of b as n ! 1:
Exercise 6.4 Verify some of the calculations reported in Section 6.5. Speci…cally, suppose that
x1i and x2i only take the values f 1; +1g; symmetrically, with
Pr (x1i = x2i = 1) = Pr (x1i = x2i = 1) = 3=8

Pr (x1i = 1; x2i = 1) = Pr (x1i = 1; x2i = 1) = 1=8
5
E e2i j x1i = x2i =
4
2 1
E ei j x1i 6= x2i = :
4
Verify the following:
1. Ex1i = 0
2. Ex21i = 1
1
3. Ex1i x2i =
2
4. E e2i = 1
5. E x21i e2i = 1
7
6. E x1i x2i e2i = :
8
Exercise 6.5 Show (6.18)-(6.21).
Exercise 6.6 The model is
yi = x0i + ei
E (xi ei ) = 0
= E xi x0i e2i :
Find the method of moments estimators ( b ; b ) for ( ; ) :
(a) In this model, are ( b ; b ) e¢ cient estimators of ( ; )?
(b) If so, in what sense are they e¢ cient?

Exercise 6.7 Of the variables (yi ; yi ; xi ) only the pair (yi ; xi ) are observed. In this case, we say
that yi is a latent variable. Suppose
yi = x0i + ei
E (xi ei ) = 0
yi = yi + ui
where ui is a measurement error satisfying
E (xi ui ) = 0
E (yi ui ) = 0
Let b denote the OLS coe¢ cient from the regression of yi on xi :
(a) Is the coe¢ cient from the linear projection of yi on xi ?
(b) Is b consistent for as n ! 1?

p
(c) Find the asymptotic distribution of n b as n ! 1:
p
Exercise 6.8 Find the asymptotic distribution of n ^2 2 as n ! 1:
Exercise 6.9 The model is
yi = xi + ei
E (ei j xi ) = 0
where xi 2 R: Consider the two estimators

Pn
b = x i yi
Pi=1
n 2
i=1 xi
n
e = 1 X yi
:
n xi
i=1
(a) Under the stated assumptions, are both estimators consistent for ?
(b) Are there conditions under which either estimator is e¢ cient?

Chapter 7
Restricted Estimation
7.1 Introduction
In the linear projection model
yi = x0i + ei
E (xi ei ) = 0
a common task is to impose a constraint on the coe¢ cient vector . For example, partitioning
x0i = (x01i ; x02i ) and 0 = 0 0
1 ; 2 ; a typical constraint is an exclusion restriction of the form
2 = 0: In this case the constrained model is
yi = x01i 1 + ei
E (xi ei ) = 0
At …rst glance this appears the same as the linear projection model, but there is one important
di¤erence: the error ei is uncorrelated with the entire regressor vector x0i = (x01i ; x02i ) not just the
included regressor x1i :
In general, a set of q linear constraints on takes the form
R0 =c (7.1)
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 coe¢ cients 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 coe¢ cients 1 = 2 ; and equal and o¤setting coe¢ cients 1 2 = 0:
A typical reason to impose a constraint is that we believe (or have information) that the con-
straint is true. By imposing the constraint we hope to improve estimation e¢ ciency. The goal is
to obtain consistent estimates with reduced variance relative to the unconstrained estimator.
The questions then arise: How should we estimate the coe¢ cient 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.
136
CHAPTER 7. RESTRICTED ESTIMATION 137
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. This estimator is
e = argmin SSEn ( ) (7.3)
R0 =c
where
n
X 2
SSEn ( ) = yi x0i = y0y 2y 0 X + 0
X 0X :
i=1
The estimator e minimizes the sum of squared errors over all such that the restriction (7.1)
holds. We call e the constrained least-squares (CLS) estimator. We follow the convention of
using a tilde “~” rather than a hat “^” to indicate that e is a restricted estimator in contrast to
the unrestricted least-squares estimator b ; and write it as e cls when we want to be clear 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 0
L( ; ) = SSEn ( ) + R0 c (7.4)
2
over ( ; ); where is an s 1 vector of Lagrange multipliers. The …rst-order conditions for
minimization of (7.4) are
@
L( e ; e ) = X 0y + X 0X e + R e = 0 (7.5)
@
and
@
L( ; ) = R0 e c = 0: (7.6)
@
1
Premultiplying (7.5) by R0 (X 0 X) we obtain
1
R0 b + R0 e + R0 X 0 X Re = 0 (7.7)
where b = (X 0 X) X 0 y is the unrestricted least-squares estimator. Imposing R0 e

1
c = 0 from
(7.6) and solving for e we …nd
h i 1
e = R0 X 0 X 1
R R0 b c :
Substuting this expression into (7.5) and solving for e we …nd the solution to the constrained
minimization problem (7.3)
h i 1
e=b 1 1
X 0X R R0 X 0 X R R0 b c : (7.8)
This is a general formula for the CLS estimator. It also can be written as
h i 1
e=b b 1 R R0 Q
Q b 1R R0 b c :
xx xx
7.3 Exclusion Restriction

While (4.4) 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
yi = x01i 1 + x02i 2 + ei (7.9)
the exclusion restriction is 2 = 0; and the constrained equation is
yi = x01i 1 + ei : (7.10)
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
e = x1i x01i x1i yi :
1
i=1 i=1
0 0 0
The CLS estimator of the entire vector = 1; 2 is
e
e= 1 : (7.11)
0
It is not immediately obvious, but (7.8) and (7.11) are algebraically (and numerically) equivalent.
To see this, the …rst component of (7.8) with (7.2) is
" #
1
e = I 0 b Q 1 0 1 0
1
b
xx 0 I Q b
xx 0 I b :
I I
Using (4.28) this equals

1
e
1 = b1 b 12 Q
Q b 22 b
2
= b +Q b 1 Qb b 1b b
1 11 2 12 Q22 Q22 1 2
1 1
= b
Q b b 12 Q
b Q b 2y b 1 Q
+Q b b 1b b 1 b b 1Q
b 21 Q b 1y
11 2 Q1y Q 22 11 2 12 Q22 Q22 1 Q22 1 Q2y Q 11
b 1 Q
= Q b 1y b 1Q
b 12 Q
Q b 1Q
b 21 Q b 1y
11 2 22 11
b 1 Q
= Q b 11 b 1Q
b 12 Q
Q b 1Q
b 21 Q b 1y
11 2 22 11
b 1Q
= Q b 1y
11
which is (7.11) as originally claimed.
7.4 Minimum Distance

The CLS estimator is a special case of a more general class of constrained estimators. To
see this, rewrite the least-squares criterion as follows. Let b be the unconstrained least-squares
estimator, and write the uncontrained least-squares …tted equation as yi = x0i b + eî : Substitute
this equation into SSEn ( ) to obtain

n
X 2
SSEn ( ) = yi x0i
i=1
n
X 2
= x0i b + eî x0i
i=1
n n
!
X 0 X
= e^2i + b xi x0i b
i=1 i=1
0
= n^ 2 + n b b xx b
Q : (7.12)
P
where the third equality uses the fact that ni=1 xi eî = 0: Since the …rst term on the last line does
not depend on it follows that the CLS estimator minimizes the quadratic on the right-side of
(7.12): This is a (squared) weighted Euclidean distance between b and : It is a special case of the
general weighted distance
0
Jn ( ; W n ) = n b Wn1 b
for W n > 0 a k k positive de…nite weight matrix. In summary, we have found that the CLS
estimator can be written as
b 1)
e = argmin Jn ( ; Q
xx
R0 =c
More generally, a minimum distance estimator for is

e
md (W n ) = argmin Jn ( ; W n ) (7.13)
R0 =c
where W n > 0. We have written the estimator as e md (W n ) as it depends upon the weight matrix
W n:
An obvious question is which weight matrix W n is appropriate. We will address this question
after we derive the asymptotic distribution for a general weight matrix.
7.5 Computation
A general method to solve the algebraic problem (7.13) is by the method of Lagrange multipliers.
The Lagrangian is
1
L( ; ) = Jn ( ; W n ) + 0 R0 c
2
which is minimized over ( ; ): The solution is
e 1
md (W n ) =b W n R R0 W n R R0 b c : (7.14)
(See Exercise 7.5.)

1
b
If we set W n = Q xx then (7.14) specializes to the CLS estimator:
e b 1
md (Qxx ) = e cls
In this sense the minimum distance estimator generalizes constrained least-squares.

7.6 Asymptotic Distribution

We …rst show that the class of minimum distance estimators are consistent for the population
parameters when the constraints are valid.
Assumption 7.6.1 R0 = c where R is k q with rank(R) = q:
Theorem 7.6.1 Consistency

Under Assumption 1.5.1, Assumption 3.16.1, Assumption 7.6.1,
p p
and W n ! W > 0; e md (W n ) ! as n ! 1:
Theorem 7.6.1 shows that consistency holds for any weight matrix, so the result includes the
CLS estimator.
Similarly, the constrained estimators are asymptotically normally distributed.
Theorem 7.6.2 Asymptotic Normality

p
Under Assumption 1.5.1, Assumption 6.4.1, Assumption 7.6.1, and W n ! W > 0;
p d
n e md (W n ) ! N (0; V (W )) (7.15)
as n ! 1; where
1 1
V (W ) = V W R R0 W R R0 V V R R0 W R R0 W
1 1
+W R R0 W R R0 V R R0 W R R0 W (7.16)
and
V = Qxx1 Qxx1 :
Theorem 7.6.2 shows that the minimum distance estimator is asymptotically normal for all
positive de…nite weight matrices. The asymptotic variance depends on W . The theorem includes
the CLS estimator as a special case by setting W = Qxx1 :
Theorem 7.6.3 Asymptotic Distribution of CLS Estimator

Under Assumption 1.5.1, Assumption 6.4.1, and Assumption 7.6.1, as n ! 1
p d
n e cls ! N (0; V cls )
where
1 1
V cls = V Qxx1 R R0 Qxx1 R R0 V V R R0 Qxx1 R R0 Qxx1
1 1
+Qxx1 R R0 Qxx1 R R0 V R R0 Qxx1 R R0 Qxx1
7.7 E¢ cient Minimum Distance Estimator

Theorem 7.6.2 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
variance V (W ): This turns out to be W = V has shown in Theorem 7.7.1 below. Since V is
unknown this weight matrix cannot be used for a feasible estimator, but we can replace V with
a consistent estimate Vb and the asymptotic distribution (and e¢ ciency) are unchanged. We call
the minimum distance estimator setting W n = Vb the e¢ cient minimum distance estimator
and takes the form
1
e = b Vb R R0 Vb R R0 b c (7.17)
This estimator has the smallest asymptotic variance in the class of minimum distance estimators,
The asymptotic distribution of (7.17) can be deduced from Theorem 7.6.2.
Theorem 7.7.1 E¢ cient Minimum Distance Estimator

Under Assumption 1.5.1, Assumption 6.4.1, and Assumption 7.6.1, for e de…ned
in (7.17) ,
p d
n e ! N 0; V
as n ! 1; where
1
V =V V R R0 V R R0 V : (7.18)
Since
V V (7.19)
the estimator (7.17) has lower asymptotic variance than the unrestricted estimator.
Furthermor, for any W ;
V V (W ) (7.20)
so (7.17) is asymptotically e¢ cient in the class of minimum distance estimators.
Theorem 7.7.1 shows that the minimum distance estimator with the smallest asymptotic vari-
ance is (7.17). One implication is that the constrained least squares estimator is generally inef-
…cient. The interesting exception is the case of conditional homoskedasticity, in which case the
optimal weight matrix is W = V 1 = 2Q
xx so in this case CLS is an e¢ cient minimum dis-
tance estimator. Otherwise when the error is conditionally heteroskedastic, there are asymptotic
e¢ ciency gains by using minimum distance rather than least squares.
The fact that CLS is generally ine¢ cient is counter-intuitive and requires some re‡ection 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
our statistical analysis has shown that this is not the e¢ cient estimation method. Instead, the
e¢ cient 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.
Inequality (7.19) shows that the e¢ cient minimum distance estimator e has a smaller asymptotic
variance than the unrestricted least squares estimator b : This means that estimation is more e¢ cient
by imposing correct restrictions when we use the minimum distance method.
7.8 Exclusion Restriction Revisited

We return to the example of estimation with a simple exclusion restriction. The model is
yi = x01i 1 + x02i 2 + ei
with the exclusion restriction 2 = 0: We have introduced three estimators of 1: The …rst is
unconstrained least-squares applied to (7.9), which can be written as
b 1 Q
b =Q b
1 11 2 1y 2 :
From Theorem 6.28 and equation (6.19) its asymptotic variance is
avar( b 1 ) = Q1112 11 Q12 Q221 21

1
12 Q22 Q21 + Q12 Q221 1
22 Q22 Q21 Q1112 :
The second estimator of 1 is the CLS estimator, which can be written as
e
1;cls
b 1Q
=Q b 1y :
11
Its asymptotic variance can be deduced from Theorem 7.6.3, but it is simpler to apply the CLT
directly to show that
avar( e 1;cls ) = Q111 11 Q111 : (7.21)
The third estimator of 1 is the e¢ cient minimum distance estimator. Applying (7.17), it equals
1
e = b1 Vb 12 Vb 22 b 2 (7.22)
1;md
where we have partitioned " #

Vb 11 Vb 12
Vb = :
Vb 21 Vb 22
From Theorem 7.7.1 its asymptotic variance is
avar( e 1;md ) = V 11 V 12 V 221 V 21 : (7.23)
In general, the three estimators are di¤erent, and they have di¤erent asymptotic variances.
It is quite instructive to compare the asymptotic variances of the CLS and unconstrained least-
squares estimators to assess whether or not the constrained estimator is necessarily more e¢ cient
than the unconstrained estimator.
First, consider the case of conditional homoskedasticity. In this case the two covariance matrices
simplify to
avar( b 1 ) = 2 Q1112
and
avar( e 1;cls ) = 2
Q111 :
If Q12 = 0 (so x1i and x2i are orthogonal) then these two variance matrices equal and the two
estimators have equal asymptotic e¢ ciency. Otherwise, since Q12 Q221 Q21 0; then Q11 Q11
Q12 Q221 Q21 ; and consequently
1
Q111 2
Q11 Q12 Q221 Q21 2
:
This means that under conditional homoskedasticity, e 1;cls has a lower asymptotic variance matrix
than b 1 : Therefore in this context, constrained least-squares is more e¢ cient than unconstrained
least-squares. This is consistent with our intuition that imposing a correct restriction (excluding
an irrelevant regressor) improves estimation e¢ ciency.
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 let’s use the simple heteroskedastic example from Section 6.5. In that example,
1 7 3
Q11 = Q22 = 1; Q12 = ; 11 = 22 = 1; and 12 = : We can calculate that Q11 2 = and
2 8 4
2
avar( b 1 ) = (7.24)
3
avar( e 1;cls ) = 1 (7.25)
5
avar( e 1;md ) = : (7.26)
8
Thus the restricted least-squares estimator e 1 has a larger variance than the unrestricted least-
squares estimator b 1 ! The minimum distance estimator has the smallest variance of the three, as
expected.
What we have found is that when the estimation method is least-squares, deleting the irrelevant
variable x2i can actually decrease the precision of estimation of 1 ; or equivalently, adding the
irrelevant variable x2i can actually improve the precision of the estimation.
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 (7.10) to be less e¢ cient for estimation of 1 than
least-squares applied to the long regression (7.9), 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 e¢ ciency.
It turns out that a more re…ned answer is appropriate. Constrained estimation is desirable,
but not constrained least-squares estimation. While least-squares is asymptotically e¢ cient for
estimation of the unconstrained projection model, it is not an e¢ cient estimator of the constrained
projection model.
7.9 Variance and Standard Error Estimation

The asymptotic covariance matrix (7.18) may be estimated by replacing V with a consistent
estimates such as Vb . This variance estimator is then
1
Vb = Vb Vb R R0 Vb R R0 Vb : (7.27)
We can calculate standard errors for any linear combination h0 e so long as h does not lie in
the range space of R. A standard error for h0 e is
1=2
s(h0 e ) = n h Vb h
1 0
:
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 (7.28)
where r : Rk ! Rq : This includes the linear constraints (7.1) as a special case. An example of
(7.28) which cannot be written as (7.1) is 1 2 = 1; or r( ) = 1 2 1:
The minimum distance estimator of subject to (7.28) solves the minimization problem
e = argmin Jn ( ) (7.29)
r( )=0
where 0 1
Jn ( ) = n b Vb b :
The solution minimizes the Lagrangian

1 0
L( ; ) = Jn ( ) + r( ) (7.30)
2
over ( ; ):
Computationally, there is no explicit expression for the solution e so it must be found numer-
ically. Computational methods are based on the method of quadratic programming and are not
reviewed here.
@
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.
Theorem 7.10.1 Under Assumption 1.5.1, Assumption 6.4.1, and Assumption

7.10.1, for e de…ned in (7.29) ,
p d
n e ! N 0; V
as n ! 1; where
1
V =V V R R0 V R R0 V
The asymptotic variance matrix can be estimated by

1
Vb = Vb Vb R b 0 Vb R
b R b b 0 Vb
R
where
b = @ r( e )0 :
R
@
Standard errors for the elements of e are the square roots of the diagonal elements of n b
1V :

Proof of Theorem 7.7.1, Equation (7.20). 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 R R0 V R?
C 0V C =
R0? V R R0? V R?
0 0
= 0
0 R? V R?
and
R0 V (W )R R0 V (W )R?
C 0 V (W )C =
R0? V (W )R R0? V (W )R?
0 0
= 1 1 :
0 R0? V R? + R0? W R (R0 W R) R0 V R (R0 W R) R0 W R?
Thus
C 0 V (W ) V C = C 0 V (W )C C 0V C
0 0
= 1 1 0
0 R0? W R (R0 W R) R V R (R0 W R)
0
R0 W R?
Since C is invertible it follows that V (W ) V 0 which is (7.20).
Proof of Theorem 7.10.1. For simplicity, we assume that the constrained estimator is consistent
e p! . This can be shown with more e¤ort, 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
the line segment joining e and such that
@
rj ( e ) = rj ( ) + rj ( j)
0 e : (7.31)
@
Let Rn be the k q matrix
@ @ @
Rn = r1 ( 1) r2 ( 2) rq ( q) :
@ @ @
p p p
Since e ! it follows that j ! , and by the CMT, Rn ! R: Stacking the (7.31), we obtain
r( e ) = r( ) + Rn0 e :
Since r( e ) = 0 by construction and r( ) = 0 by Assumption 7.6.1, this implies
0 = Rn0 e : (7.32)
The …rst-order condition for (7.30) is

1
Vb b e =R
e e:
Premultiplying by R 0 Vb ; inverting, and using (7.32), we …nd

1 1
e = R 0 Vb R
e Rn0 b e = R 0 Vb R
e Rn0 b :
n n
Thus
1
e = I Vb R
e R 0 Vb H
f Rn0 b :
n
From Theorem 6.4.2 and Theorem 6.8.2 we …nd

p 1 p
n e = I Vb R e R 0 Vb R
n
e Rn0 n b
d 1
! I V R R0 V R R0 N (0; V )
= N 0; V :
Exercises
Exercise 7.1 In the model y = X 1 1 + X 2 2 + e; show directly from de…nition (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 de…nition (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:
Exercise 7.4 Verify that for e de…ned in (7.8) that R0 e = c:
Exercise 7.5 Verify (7.14).
b 1 equals the CLS

Exercise 7.6 Verify that the minimum distance estimator e with W n = Q xx
estimator.
Exercise 7.7 Prove Theorem 7.6.1.
Exercise 7.8 Prove Theorem 7.6.2.
Exercise 7.9 Prove Theorem 7.6.3. (Hint: Use that CLS is a special case of Theorem 7.6.2.)
1
Exercise 7.10 Verify that (7.18) is V (W ) with W = V :
Exercise 7.11 Prove (7.19). Hint: Use (7.18).
Exercise 7.12 Verify (7.21), (7.22) and (7.23)
Exercise 7.13 Verify (7.24), (7.25), and (7.26).

Chapter 8
Testing
8.1 t tests
The t-test is routinely used to test hypotheses on . A simple null and composite hypothesis
takes the form
H0 : = 0
H1 : 6= 0
where 0 is some pre-speci…ed value. A t-test rejects H0 in favor of H1 when jtn ( 0 )j is large. By
“large” we mean that the observed value of the t-statistic would be unlikely if H0 were true.
Formally, we …rst pick an asymptotic signi…cance level . We then …nd z =2 ; the upper =2
quantile of the standard normal distribution which has the property that if Z N(0; 1) then
Pr jZj > z =2 = :
For example, z:025 = 1:96 and z:05 = 1:645: A test of asymptotic signi…cance rejects H0 if
jtn j > z =2 : Otherwise the test does not reject, or “accepts” H0 :
The asymptotic signi…cance level is because Theorem 6.11.1 implies that
Pr (reject H0 j H0 true) = Pr jtn j > z =2 j = 0

! Pr jZj > z =2 = :
The rejection/acceptance dichotomy is associated with the Neyman-Pearson approach to hypothesis

testing.
While there is no objective scienti…c basis for choice of signi…cance level ; the common practice
is to set = :05 or 5%. This implies a critical value of z:025 = 1:96 2. When jtn j > 2 it is common
to say that the t-statistic is statistically signi…cant. and if jtn j < 2 it is common to say that
the t-statistic is statistically insigni…cant. It is helpful to remember that this is simply a way of
saying “Using a t-test, the hypothesis that = 0 can [cannot] be rejected at the asymptotic 5%
level.”
A related statistic is the asymptotic p-value, which can be interpreted as a measure of the
evidence against the null hypothesis. The asymptotic p-value of the statistic tn is
pn = p(tn )
where p(t) is the tail probability function
p(t) = Pr (jZj > jtj) = 2 (1 (jtj)) :
If the p-value pn is small (close to zero) then the evidence against H0 is strong.
147
CHAPTER 8. TESTING 148
An equivalent statement of a Neyman-Pearson test is to reject at the % level if and only if

pn < : Signi…cance tests can be deduced directly from the p-value since for any ; pn < if
and only if jtn j > z =2 : The p-value is more general, however, in that the reader is allowed to pick
the level of signi…cance , in contrast to Neyman-Pearson rejection/acceptance reporting where
the researcher picks the signi…cance level. (However, the Neyman-Pearson approach requires the
reader to select the signi…cance level before observing the p-value.)
Another helpful observation is that the p-value function is a unit-free transformation of the
d
t statistic. That is, under H0 ; pn ! U[0; 1]; so the “unusualness” of the test statistic can be
compared to the easy-to-understand uniform distribution, regardless of the complication of the
distribution of the original test statistic. To see this fact, note that the asymptotic distribution of
jtn j is F (x) = 1 p(x): Thus
Pr (1 pn u) = Pr (1 p(tn ) u)
= Pr (F (tn ) u)
1
= Pr jtn j F (u)
1
! F F (u) = u;
d d
establishing that 1 pn ! U[0; 1]; from which it follows that pn ! U[0; 1]:
8.2 t-ratios
Some applied papers (especially older ones) report “t-ratios”for each estimated coe¢ cient. For
a coe¢ cient these are
^
tn = tn (0) = ;
s(^)
the ratio of the coe¢ cient estimate to its standard error, and equal the t-statistic for the test of
the hypothesis H0 : = 0: Such papers often discuss the “signi…cance” of certain variables or
coe¢ cients, or describe “which regressors have a signi…cant e¤ect on y” by noting which t-ratios
exceed 2 in absolute value.
This is very poor econometric practice, and should be studiously avoided. It is 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 coe¢ cient 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.
Instead, when a coe¢ cient is of interest, it is constructive to focus on the point estimate,
its standard error, and its con…dence interval. The point estimate gives our “best guess” for the
value. The standard error is a measure of precision. The con…dence 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 con…dence interval describe the bounds on the likely
possibilities. If the con…dence interval embraces too broad a set of values for ; then the dataset
is not su¢ ciently informative to render inferences about : On the other hand if the con…dence
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 widely-seen statement “the t-ratio is
highly signi…cant” has little interpretive value.
The above discussion requires that the researcher knows what the coe¢ cient means (in terms
of the economic problem) and can interpret values and magnitudes, not just signs. This is critical
for good applied econometric practice.
8.3 Wald Tests

Sometimes = h( ) is a q 1 vector, and it is desired to test the joint restrictions simultane-
ously. We have the null and alternative
H0 : = 0
H1 : 6= 0:
A commonly used test of H0 against H1 is the Wald statistic (6.34) evaluated at the null hypothesis
0 1
Wn = n b 0 Vb b 0 : (8.1)
Typically, we have b = h( b ) with asymptotic covariance matrix estimate
c 0 Vb H
Vb = H c
where
c = @ h( b ):
H
@
Then
0 1
Wn = n h( b ) 0
c 0 Vb H
H c h( b ) 0 :
When h is a linear function of ; h( ) = R0 ; then the Wald statistic simpli…es to

0 1
Wn = n R 0 b 0 R0 Vb R R0 b 0 :
As shown in Theorem 6.14.2, when = then Wn ! 2; a chi-square random variable with

0 q
q degrees of freedom.
Theorem 8.3.1 Under Assumption 1.5.1, Assumption 6.4.1,

d 2
rank(H ) = q; and H0 ; then Wn ! q .
An asymptotic Wald test rejects H0 in favor of H1 if Wn exceeds 2q ( ); the upper- quantile

of the 2q distribution. For example, 21 (:05) = 3:84 = z:025
2 : The Wald test fails to reject if W is
n
2
less than q ( ): As with t-tests, it is conventional to describe a Wald test as “signi…cant” if Wn
exceeds the 5% critical value.
Notice that the asymptotic distribution in Theorem 8.3.1 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 = p(Wn ); where p(x) = Pr 2q x is the tail probability
function of the 2q distribution. The Wald test rejects at the % level if and only if pn < ; and
pn is asymptotically U[0; 1] under H0 : In applied work it is good practice to report the p-value of
a Wald statistic, as it helps readers intrepret the magnitude of the statistic.
8.4 Minimum Distance Tests

The Wald test (8.1) measures the distance between the unrestricted estimate b and the null
hypothesis 0 . A minimum distance test measures the distance between b and the restricted
estimate e of the previous chapter. Recall that under the restriction
h( ) = 0
the e¢ cient minimum distance estimate solves the minimization problem

e = argmin Jn ( )
h( )= 0
where 0 1
Jn ( ) = n b Vb b :
The minimum distance test statistic of H0 against H1 is
Jn = Jn ( e ) = min Jn ( )
h( )= 0
or more simply
0 1
Jn = n b e Vb b e :
An asymptotic test rejects H0 in favor of H1 if Jn exceeds 2q ( ); the upper- quantile of the 2q

distribution. Otherwise the test does not reject H0 :
When h( ) is linear it turns out that Jn = Wn ; so the Wald and minimum distance tests are
equal. When h( ) is non-linear then the two tests are di¤erent.
The chi-square critical value is justi…ed by the following theorm.
Theorem 8.4.1 UnderAssumption 1.5.1, Assumption 6.4.1, rank(H ) =

d 2.
q; and H0 ; then Jn ! q
8.5 F Tests
Take the linear model
y = X1 1 + X2 2 +e
where X 1 is n k1 ; X 2 is n k2 ; k = k1 + k2 ; and the null hypothesis is
H0 : 2 = 0:
0
In this case, = 2; and there are q = k2 restrictions. Also h( ) = R0 is linear with R =
I
a selector matrix. We know that the Wald statistic takes the form
0 1
Wn = nb Vb b
0 1
= n b 2 R0 Vb R b :
2
Now suppose that covariance matrix is computed under the assumption of homoskedasticity, so
0 1
that Vb is replaced with Vb = s2 n 1 X 0 X : We de…ne the “homoskedastic” Wald statistic
0 0 1
Wn0 = nb Vb b
0 1
= n b 2 R0 Vb 0 R b :
2
What we show in this section is that this Wald statistic can be written very simply using the
formula
~0 e
e ~ e ^0 e
^
Wn0 = (n k) 0 (8.2)
e
ê ^
where
1
~= y
e X 1 e 1; e = X0 X1
1 1 X 01 y
are from OLS of y on X 1 ; and
1
^= y
e X b; b = X 0X X 0y
are from OLS of y on X = (X 1 ; X 2 ):

The elegant feature about (8.2) is that it is directly computable from the standard output
from two simple OLS regressions, as the sum of squared errors is a typical output from statistical
packages. This statistic is typically reported as an “F-statistic” which is de…ned as
Wn0 ~0 e
e ~ e ^0 e
^ =k2
Fn = = 0 :
k2 e
ê ^=(n k)
1
While it should be emphasized that equality (8.2) only holds if Vb 0 = s2 n 1 X 0 X ; still this
formula often …nds good use in reading applied papers. Because of this connection we call (8.2) the
F form of the Wald statistic. (We can also call Wn0 a homoskedastic form of the Wald statistic.)
We now derive expression (8.2). First, note that by partitioned matrix inversion (A.4)
1
0 0 1 0 X 01 X 1 X 01 X 2 1
R XX R=R R = X 02 M 1 X 2
X 02 X 1 X 02 X 2
where M 1 = I X 1 (X 01 X 1 ) 1X 0 :
1 Thus
1 1 1
R0 Vb 0 R =s 2
n 1
R0 X 0 X R =s 2
n 1
X 02 M 1 X 2
and
0 1
Wn0 = n b 2 R0 Vb 0 R b
2
b 0 (X 0 M 1 X 2 ) b
2 2 2
= :
s2
To simplify this expression further, note that if we regress y on X 1 alone, the residual is
~ = M 1 y: Now consider the residual regression of e
e f2 = M 1 X 2 : By the FWL theorem,
~ on X
0
f b
~ = X2 2 + e
e f
^ and X 2 e^ = 0: Thus
0
~0 e
e ~ = f2 b 2 + e
X ^ f2 b 2 + e
X ^
0
= b 2Xf0 X
f b ^0 e
2 2 2+e ^
0
= b X M 1X 2 b + e
0
2 ^0 e
2 ^; 2
or alternatively,
b 0 X 0 M 1X 2 b = e
~0 e
~ ^0 e
e ^:
2 2 2
Also, since
1
s2 = (n k) ^0 e
e ^
we conclude that
~0 e
e ~ e ^0 e
^
Wn0 = (n k) 0
e
ê ^
as claimed.
In many statistical packages, when an OLS regression is estimated, an “F-statistic”is reported.
This is Fn when X 1 is a vector of ones, so H0 is an intercept-only model. This special F statistic is
testing the hypothesis that all slope coe¢ cients (all coe¢ cients other than the intercept) are 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 su¢ ciently large that this F statistic is
nearly always highly signi…cant. 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.

Now let us partition =( 1; 2) and consider tests of the linear restriction
H0 : 2 =0
H1 : 2 6= 0
in the normal regression model. In parametric models, a good test statistic is the likelihood ratio,
which is twice the di¤erence in the log-likelihood function evaluated under the null and alternative
hypotheses. The estimator under the alternative is the unrestricted estimator ( b 1 ; b 2 ; ^ 2 ) discussed
above. The Gaussian log-likelihood at these estimates is
n 1 0
log L( b 1 ; b 2 ; ^ 2 ) = log 2 ^ 2 e
ê^
2 2^ 2
n n n
= log ^ 2 log (2 ) :
2 2 2
The MLE under the null hypothesis is the restricted estimates ( e 1 ; 0; ~ 2 ) where e 1 is the OLS
estimate from a regression of yi on x1i only, with residual variance ~ 2 : The log-likelihood of this
model is
n n n
log L( e 1 ; 0; ~ 2 ) = log ~ 2 log (2 ) :
2 2 2
The LR statistic for H0 against H1 is
LRn = 2 log L( b 1 ; b 2 ; ^ 2 ) log L( e 1 ; 0; ~ 2 )

= n log ~ 2 log ^ 2
~2
= n log :
^2
By a …rst-order Taylor series approximation
~2 ~2
LRn = n log 1 + 1 'n 1 = Wn0 :
^2 ^2
the homoskedastic Wald statistic. This shows that the two statistics (LRn and Wn0 ) can be numer-
ically close. It also shows that the homoskedastic Wald statistic for linear hypotheses can also be
interpreted as an appropriate likelihood ratio statistic under normality.
8.7 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
2
ei N(0; )
Figure 8.1: Wald Statistic as a function of s
and consider the hypothesis

H0 : = 1:
Let ^ and ^ 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
r
H0 (r) : =1
r r 1
for any positive integer r: Letting h( ) = ; and noting H = r ; we …nd that the standard
Wald test for H0 (r) is
2
^r 1
Wn (r) = n 2r 2 :
^ 2 r2 ^
While the hypothesis r = 1 is una¤ected by the choice of r; the statistic Wn (r) varies with r: This
is an unfortunate feature of the Wald statistic.
To demonstrate this e¤ect, we have plotted in Figure 8.1 the Wald statistic Wn (r) as a function
of r; setting n= 2 = 10: The increasing solid line is for the case ^ = 0:8: The decreasing dashed
line is for the case ^ = 1:6: It is easy to see that in each case there are values of r for which the
test statistic is signi…cant relative to asymptotic critical values, while there are other values of r
for which the test statistic is insigni…cant. This is distressing since the choice of r is arbitrary and
irrelevant to the actual hypothesis.
d
Our …rst-order asymptotic theory is not useful to help pick r; as Wn (r) ! 21 under H0 for any
r: 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 arti…cial datasets, to which we apply the statistical tools of interest. This
produces random draws from the statistic’s 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 (r) > 3:84 j = 1) : Given the simplicity of the model, this probability depends only on
r; n; and 2 : In Table 2.1 we report the results of a Monte Carlo simulation where we vary these
three parameters. The value of r is varied from 1 to 10, n is varied among 20, 100 and 500, and
is varied among 1 and 3. Table 4.1 reports the simulation estimate of the Type I error probability
from 50,000 random samples. Each row of the table corresponds to a di¤erent value of r –and thus
corresponds to a particular choice of test statistic. The second through seventh columns contain the
Type I error probabilities for di¤erent combinations of n and . These probabilities are calculated
as the percentage of the 50,000 simulated Wald statistics Wn (r) which are larger than 3.84. The
null hypothesis r = 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 devia-
tions 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 r leads
to a test with unacceptable Type I error probabilities.
In Table 4.1 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 r increases,
which is not surprising given the dependence of Wn (r) on r as shown in Figure 8.1.
Table 4.1
Type I error Probability of Asymptotic 5% Wn (r) Test
=1 =3
r 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
In this example it is not surprising that the choice r = 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 (8.3)

E (xi ei ) = 0
and the hypothesis

1
H0 : =r
2
where r is a known constant. Equivalently, de…ne = 1 = 2 ; so the hypothesis can be stated as

H0 : = r:
Let b = ( ^ 0 ; ^ 1 ; ^ 2 ) be the least-squares estimates of (8.3), let Vb be an estimate of the
asymptotic covariance matrix for b and set ^ = ^ 1 = ^ 2 . De…ne
0 1
0
B C
B C
B 1 C
B C
Hc1 = BB ^2 C
C
B C
B C
B ^ C
@ A 1
^2
2
1=2
so that the standard error for ^ is s(^) = n ^ 01 Vb
1H ^
H : In this case a t-statistic for H0 is
1
^
^
1
r
2
t1n = :
s(^)
An alternative statistic can be constructed through reformulating the null hypothesis as
H0 : 1 r 2 = 0:
A t-statistic based on this formulation of the hypothesis is

^ r ^2
1
t2n = 1=2
:
n b
1H 0 V H2
2
where 0
1
0
H2 = @ 1 A :
r
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:
Table 4.2
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)
2 t1n t2n t1n t2n t1n t2n t1n t2n
.10 .47 .06 .00 .06 .28 .05 .00 .05
.25 .26 .06 .00 .06 .15 .05 .00 .05
.50 .15 .06 .00 .06 .10 .05 .00 .05
.75 .12 .06 .00 .06 .09 .05 .00 .05
1.00 .10 .06 .00 .06 .07 .05 .02 .05
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 4.2. 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 di¤erent sampling behavior.
The common message from both examples is that Wald statistics are sensitive to the algebraic
formulation of the null hypothesis. The 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.
8.8 Monte Carlo Simulation

In the previous section 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
interest, for example an estimator ^ or a t-statistic (^ )=s(^): The exact distribution of Tn is
Gn (u; F ) = Pr (Tn u j 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 dis-
tribution 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
n independent random pairs (yi ; xi ) ; i = 1; :::; n; are drawn from the distribution F using
the computer’s random number generator.
The statistic Tn = Tn ((y1 ; x1 ) ; :::; (yn ; xn ) ; ) is calculated on this pseudo data.
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 b0 th experiment result in the draw Tnb ; b = 1; :::; B: These results are stored.
They constitute a random sample of size B from the distribution of Gn (u; F ) = Pr (Tnb u) =
Pr (Tn u j F ) :
From a random sample, we can estimate any feature of interest using (typically) a method of
moments estimator. For example:
Suppose we are interested in the bias, mean-squared error (MSE), or variance of the distribution
of ^ : We then set Tn = ^ ; run the above experiment, and calculate
B B
\^) = 1 X 1 X^
Bias( Tnb = b
B B
b=1 b=1
XB B
1 1 X ^ 2
M\
SE(^) = (Tnb )2 = b
B B
b=1 b=1
2
\
var(^) = M\
SE(^) \^)
Bias(
Suppose we are interested in the Type I error associated with an asymptotic 5% two-sided t-test.
We would then set Tn = ^ =s(^) and calculate
B
^ 1 X
P = 1 (Tnb 1:96) ; (8.4)
B
b=1
the percentage of the simulated t-ratios which exceed the asymptotic 5% critical value.
Suppose we are interested in the 5% and 95% quantile of Tn = ^: We then compute the 5% and
95% sample quantiles of the sample fTnb g: 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 fTnb g
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 50’th sorted value and the 95% sample quantile is the 950’th 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.
In particular, it is simple to make inferences about rejection probabilities from statistical
tests, such as the percentage estimate reported in (8.4). The random variable 1 (Tnb 1:96) is
iid Bernoulli, equalling 1 with probability p = E1 (Tp nb 1:96) : The average (8.4) is therefore an
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
p value. Forpexample, if we are assessing
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.
8.9 Estimating a Wage Equation

We again return to our wage equation. We use the sample of wage earners from the March 2004
Current Population Survey, excluding military. For the dependent variable we use the natural log
of wages so that coe¢ cients may be interpreted as semi-elasticities. For regressors we include years
of education, potential work experience, experience squared, and dummy variable indicators for
the following: married, female, union member, immigrant, hispanic, and non-white. Furthermore,
we included a dummy variable for state of residence (including the District of Columbia, this adds
50 regressors). The available sample is 18,808 so the parameter estimates are quite precise and
reported in Table 4.1, excluding the coe¢ cients on the state dummy variables.
Table 4.1 displays the parameter estimates in a standard format. The Table clearly states the
estimation method (OLS), the dependent variable (log(Wage)), and the regressors are clearly la-
beled. Parameter estimates are both reported for the coe¢ cients of interest (the coe¢ cients on the
state dummy variables are omitted) and standard errors are reported for all reported coe¢ cient es-
timates. In addition to the coe¢ cient 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( ^ )
Intercept 1.027 .032
Education .101 .002
Experience .033 .001
Experience2 :00057 .00002
Married .102 .008
Female :232 .007
Union Member .097 .010
Immigrant :121 .013
Hispanic :102 .014
Non-White :070 .010
^ .4877
Sample Size 18,808
Note: Equation also includes state dummy variables.
As a general rule, it is best to always report standard errors along with parameter estimates
(as done in Table 4.1). This allows readers to assess the precision of the parameter estimates, and
form con…dence intervals and t-tests on individual coe¢ cients if desired. For example, if you are
interested in the di¤erence in mean wages between men and women, you can read from the table
that the estimated coe¢ cient on the Female dummy variable is 0:232; implying a mean wage
di¤erence of 23%. To assess the precision, you can see that the standard error for this coe¢ cient
estimate is 0.007. This implies a 95% asymptotic con…dence interval for the coe¢ cient estimate of
[ :246; :218]. This means that we have estimated the di¤erence in mean wages between men and
women to lie between 22% and 25%. I interpret this as a precise estimate because there is not an
important di¤erence between the lower and upper bound.
Instead of reporting standard errors, some empirical researchers report t-ratios for each pa-
rameter estimate. “t-ratios” are t-statistics which test the hypothesis that the coe¢ cient equals
zero. An example is reported in Table 4.2. In this example, all the t-ratios are highly signi…cant,
ranging in magnitude from 9.3 to 50. What we learn from these statistics is that these coe¢ cients
are non-zero, but not much more. In a sample of this size this …nding is rather uninteresting;
consequently the reporting of t-ratios is a waste of space. Again consider the male-female wage
di¤erence. Table 4.2 reports that the t-ratio is 33, enabling us to reject the hypothesis that the
coe¢ cient is zero. But how precise is the reported estimate of a wage gap of 23%? It is hard to
assess from a quick reading of Table 4.2 Standard errors are much more useful, for they enable for
quick and easy assessment of the degree of estimation uncertainty.
Table 4.2
OLS Estimates of Linear Equation for Log(Wage)
Improper Reporting: t-ratios replacing standard errors
^ t
Intercept 1.027 32
Education .101 50
Experience .033 33
Experience2 :00057 28
Married .102 12.8
Female :232 33
Union Member .097 9.7
Immigrant :121 9:3
Hispanic :102 7:3
Non-White :070 7
Returning to the estimated wage equation, one might question whether or not the state dummy
variables are relevant. Computing the Wald statistic (8.1) that the state coe¢ cients are jointly zero,
we …nd Wn = 550: Alternatively, re-estimating the model with the 50 state dummies excluded, the
restricted standard deviation estimate is ~ = :4945: The F form of the Wald statistic (8.2) is
~2 :49452
Wn = n 1 = 18; 808 1 = 528:
^2 :48772
Notice that the two statistics are close, but not equal. Using either statistic the hypothesis is easily
rejected, as the 1% critical value for the 250 distribution is 76.
Another interesting question which can be addressed from these estimates is the maximal impact
of experience on mean wages. Ignoring the other coe¢ cients, we can write this e¤ect as
2
log(W age) = 2 Experience + 3 Experience +
Our question is: At which level of experience do workers achieve the highest wage? In this
quadratic model, if 2 > 0 and 3 < 0 the solution is
2
= :
2 3
From Table 4.1 we …nd the point estimate

^
^= 2
= 28:69:
2^ 3
Using the Delta Method, we can calculate a standard error of s(^) = :40; implying a 95% con…dence
interval of [27:9; 29:5]:
However, this is a poor choice, as the coverage probability of this con…dence interval is one
minus the Type I error of the hypothesis test based on the t-test. In Section 8.7 we discovered
that such t-tests have very poor Type I error rates. Instead, we found better Type I error rates by
reformulating the hypothesis as a linear restriction. These t-statistics take the form
^ + 2^
2 3
tn ( ) = 1=2
h0 V^ h
where
1
h =
2
and V^ is the covariance matrix for ( ^ 2 ^ 3 ):

In the present context we are interested in forming a con…dence interval, not testing a hypothesis,
so we have to go one step further. Our desired con…dence interval will be the set of parameter values
which are not rejected by the hypothesis test. This is the set of such that jtn ( )j 1:96: Since
tn ( ) is a non-linear function of ; there is not a simple expression for this set, but it can be found
numerically quite easily. This set is [27:0; 29:5]: Notice that the upper end of the con…dence interval
is the same as that from the delta method, but the lower end is substantially lower.
Exercises
2
Exercise 8.1 Prove that if an additional regressor X k+1 is added to X; Theil’s adjusted R
increases if and only if jtk+1 j > 1; where tk+1 = ^ k+1 =s( ^ k+1 ) is the t-ratio for ^ k+1 and
1=2
s( ^ k+1 ) = s2 [(X 0 X) 1
]k+1;k+1
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
columns. Let b 1 and b 2 be the OLS estimates of 1 and 2 . For simplicity, you may assume that
both samples have the same number of observations n:
p b b
(a) Find the asymptotic distribution of n 2 1 ( 2 1) as n ! 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 Investment to Capital Ratio (multiplied by 100).
Qi Total Market Value to Asset Ratio (Tobin’s Q).
Ci Cash Flow to Asset Ratio.
Di 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 con…dence intervals for the coe¢ cients.
(c) This regression is related to Tobin’s q theory of investment, which suggests that investment
should be predicted solely by Qi : Thus the coe¢ cient on Qi should be positive and the others
should be zero. Test the joint hypothesis that the coe¢ cients on Ci and Di are zero. Test the
hypothesis that the coe¢ cient on Qi is zero. Are the results consistent with the predictions
of the theory?
(d) Now try a non-linear (quadratic) speci…cation. Regress Ii on Qi ; Ci , Di ; Q2i ; Ci2 , Di2 ; Qi Ci ;
Qi Di ; Ci Di : Test the joint hypothesis that the six interaction and quadratic coe¢ cients 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):
(a) First estimate an unrestricted Cobb-Douglass speci…cation
log T Ci = 1 + 2 log Qi + 3 log P Li + 4 log P Ki + 5 log P Fi + ei : (8.5)
Report parameter estimates and standard errors. You should obtain the same OLS estimates
as in Hayashi’s equation (1.7.7), but your standard errors may di¤er.
(b) What is the economic meaning of the restriction H0 : 3 + 4 + 5 = 1?
(c) Estimate (8.5) by constrained least-squares imposing 3+ 4+ 5 = 1. Report your parameter

estimates and standard errors.
(d) Estimate (8.5) by e¢ cient 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
Additional Regression Topics
9.1 Generalized Least Squares

In the projection model, we know that the least-squares estimator is semi-parametrically e¢ cient
for the projection coe¢ cient. However, in the linear regression model
yi = x0i + ei
E (ei j xi ) = 0;
the least-squares estimator is ine¢ cient. The theory of Chamberlain (1987) can be used to show
that in this model the semiparametric e¢ ciency bound is obtained by the Generalized Least
Squares (GLS) estimator (5.12) introduced in Section 5.4.1. The GLS estimator is sometimes
called the Aitken estimator. The GLS estimator (9.1) is infeasible since the matrix D is unknown.
A feasible GLS (FGLS) estimator replaces the unknown D with an estimate D ^ = diagf^ 21 ; :::; ^ 2n g:
We now discuss this estimation problem.
Suppose that we model the conditional variance using the parametric form
2
i = 0 + z 01i 1
0
= zi;
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 = e2i : Then
E ( i j xi ) = 0 + z 01i 1
and we have the regression equation
i = 0 + z 01i 1 + i (9.1)
E( i j xi ) = 0:
This regression error i is generally heteroskedastic and has the conditional variance
var ( i j xi ) = var e2i j xi

2
= E e2i E e2i j xi j xi
2
= E e4i j xi E e2i j xi :
Suppose ei (and thus i) were observed. Then we could estimate by OLS:

1 p
b = Z 0Z Z0 !
163
CHAPTER 9. ADDITIONAL REGRESSION TOPICS 164
and p d
n (b ) ! N (0; V )
where
1 1
V = E z i z 0i E z i z 0i 2
i E z i z 0i : (9.2)
While ei is not observed, we have the OLS residual eî = yi x0i b = ei x0i ( b ): Thus
i ^ i
= e^2i e2i
= 2ei x0i b + (b )0 xi x0i ( b ):
And then
n n n
1 X 2X p 1X p
p zi i = z i ei x0i n b + zi( b )0 xi x0i ( b ) n
n n n
i=1 i=1 i=1
p
!0
Let
1
e = Z 0Z Z0^ (9.3)
be from OLS regression of î on z i : Then
p p 1 1
n (e ) = n (b )+ n Z 0Z n 1=2
Z0
d
! N (0; V ) (9.4)
Thus the fact that i is replaced with î is asymptotically irrelevant. We call (9.3) the skedastic
regression, as it is estimating the conditional variance of the regression of yi on xi : We have shown
that is consistently estimated by a simple procedure, and hence we can estimate 2i = z 0i by
~ 2i = ~ 0 z i : (9.5)
Suppose that ~ 2i > 0 for all i: Then set
e = diagf~ 2 ; :::; ~ 2 g
D 1 n
and
1 1 1
e = X 0D
e X e
X 0D y:
This is the feasible GLS, or FGLS, estimator of : Since there is not a unique speci…cation 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 speci…cation
(9.1), there is no guarantee that ~ 2i > 0 for all i: If ~ 2i < 0 for some i; then the FGLS estimator
is not well de…ned. Furthermore, if ~ 2i 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
i = max[~ 2i ; c^ 2 ]
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 speci…ed, then FGLS is asymp-
totically equivalent to GLS. As the proof is tricky, we just state the result without proof.
Theorem 9.1.1 If the skedastic regression is correctly speci…ed,

p p
n e GLS e
F GLS ! 0;
and thus p d
n e F GLS ! N (0; V ) ;
where
2 1
V = E i xi x0i :
Examining the asymptotic distribution of Theorem 9.1.1, the natural estimator of the asymp-
totic variance of e is
n
! 1
0 1 X 1 0~ 1 1
e
V = 2 0
~ i xi xi = XD X :
n n
i=1
0
which is consistent for V as n ! 1: This estimator Ve is appropriate when the skedastic
regression (9.1) is correctly speci…ed.
It may be the case that 0 z i is only an approximation to the true conditional variance 2i =
E(ei j xi ). In this case we interpret 0 z i as a linear projection of e2i on z i : ~ should perhaps be
2
called a quasi-FGLS estimator of : Its asymptotic variance is not that given in Theorem 9.1.1.
Instead,
1 1 2 1 1
0
V = E zi xi x0i E 0
zi 2 0
i xi xi E 0
zi xi x0i :
V takes a sandwich form similar to the covariance matrix of the OLS estimator. Unless 2i = 0 z i ,
0
Ve is inconsistent for V .
0
An appropriate solution is to use a White-type estimator in place of Ve : This may be written
as
n
! 1 n
! n
! 1
1 X 1 X 1 X
Ve = ~ i 2 xi x0i ~ i 4 e^2i xi x0i ~ i 2 xi x0i
n n n
i=1 i=1 i=1
1 1
1 0e 1 1 0e 1
bDe 1 1 0e 1
= XD X XD D X XD X
n n n
where D b = diagf^ e21 ; :::; e^2n g: This is estimator is robust to misspeci…cation of the conditional vari-
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 speci…cation 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
and FGLS estimators will converge in probability to di¤erent limits as they will be estimating two
di¤erent projections. The FGLS probability limit will depend on the particular function selected for
the skedastic regression. The point is that the e¢ ciency gains from FGLS are built on the stronger
assumption of a correct conditional mean, and the cost is a loss of robustness to misspeci…cation.
9.2 Testing for Heteroskedasticity

The hypothesis of homoskedasticity is that E e2i j xi = 2, or equivalently that
H0 : 1 =0
in the regression (9.1). We may therefore test this hypothesis by the estimation (9.3) and con-
structing 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 (9.2)
for ~ simpli…es to
1
V = E z i z 0i E 2i : (9.6)
Hence the standard test of H0 is a classic F (or Wald) test for exclusion of all regressors from
the skedastic regression (9.3). The asymptotic distribution (9.4) and the asymptotic variance (9.6)
under independence show that this test has an asymptotic chi-square distribution.
Theorem 9.2.1 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 di¤erences between popular
tests are which transformations of xi enter z i : Motivated by the form of the asymptotic variance
of the OLS estimator b ; White (1980) proposed that the test for heteroskedasticity be based on
setting z i to equal all non-redundant elements of xi ; its squares, and all cross-products. Breusch-
Pagan (1979) proposed what might appear to be a distinct test, but the only di¤erence is that they
allowed for general choice of z i ; and replaced E 2i with 2 4 which holds when ei is N 0; 2 : If
this simpli…cation 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 scienti…c 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.
9.3 Forecast Intervals

For a given value of xi = x; we may want to forecast (guess) yi out-of-sample. A reasonable
rule is the conditional mean m(x) as it is the mean-square-minimizing forecast. A point forecast is
the estimated conditional mean m(x)
^ = x0 b . We would also like a measure of uncertainty for the
forecast.
The forecast error is eî = yi m(x)
^ = ei x0 b : As the out-of-sample error ei is
independent of the in-sample estimate ^ ; this has variance
0
e2i = E e2i j xi = x + x0 E b
E^ b x
2 1 0
= (x) + n x V x:
Assuming E e2i j xi = 2; the natural estimate of this variance is ^ 2 + n 1 x0 Vb x; so a standard

q
error for the forecast is s^(x) = ^ 2 + n 1 x0 Vb x: Notice that this is di¤erent from the standard
error for the conditional mean. If we have an estimate of the conditional

q variance function, e.g.
~ (x) = e z from (9.5), then the forecast standard error is s^(x) = ~ (x) + n 1 x0 Vb x
2 0 2
It would appear natural to conclude that an asymptotic 95% forecast interval for yi is
h i
x0 b 2^
s(x) ;
but this turns out to be incorrect. In general, the validity of an asymptotic con…dence interval is
based on the asymptotic normality of the studentized ratio. In the present case, this would require
the asymptotic normality of the ratio
ei x0 b
:
s^(x)
But no such asymptotic approximation can be made. The only special exception is the case where
ei 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 ei given
xi = x; which is a much more di¢ culth task. Perhaps
i due to this di¢ culty, many applied forecasters
use the simple approximate interval x b 2^
0 s(x) despite the lack of a convincing justi…cation.
9.4 NonLinear Least Squares

In some cases we might use a parametric regression function m (x; ) = E (yi j xi = x) which
is a non-linear function of the parameters : We describe this setting as non-linear regression.
Examples of nonlinear regression functions include
x
m (x; ) = 1 + 2
1 + 3x
m (x; ) = 1 + 2x
3
m (x; ) = 1 + 2 exp( 3 x)
0
m (x; ) = G(x ); G known
0 0 x2 3
m (x; ) = 1 x1 + 2 x1
4
m (x; ) = 1 + 2x + 3 (x 4 ) 1 (x > 4 )
0 0
m (x; ) = 1 x1 1 (x2 < 3) + 2 x1 1 (x2 > 3)
In the …rst …ve examples, m (x; ) is (generically) di¤erentiable in the parameters : In the …nal
two examples, m is not di¤erentiable with respect to 4 and 3 which alters some of the analysis.
When it exists, let
@
m (x; ) = m (x; ) :
@
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.
The least squares estimator b minimizes the normalized sum-of-squared-errors
n
1X
Sn ( ) = (yi m (xi ; ))2 :
n
i=1
When the regression function is nonlinear, we call this the nonlinear least squares (NLLS)
estimator. The NLLS residuals are eî = yi m xi ; b :
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
Since sum-of-squared-errors function Sn ( ) is not quadratic, b must be found by numerical
methods. See Appendix E. When m(x; ) is di¤erentiable, then the FOC for minimization are
n
X
0= m xi ; b eî : (9.7)
i=1
Theorem 9.4.1 Asymptotic Distribution of NLLS Estimator

If the model is identi…ed and m (x; ) is di¤ erentiable with respect to ,
p d
n b ! N (0; V )
1 1
V = E m i m0 i E m i m0 i e2i E m i m0 i
where m i = m (xi ; 0 ):
Based on Theorem 9.4.1, an estimate of the asymptotic variance V is

n
! 1 n
! n
! 1
1 X 1 X 1 X
Vb = m^ im^0i m ^ 0 i e^2i
^ im m ^0i
^ im
n n n
i=1 i=1 i=1
^ i = m (xi ; b) and eî = yi m(xi ; b):

where m
Identi…cation is often tricky in nonlinear regression models. Suppose that
0 0
m(xi ; ) = 1zi + 2 xi ( )
where xi ( ) is a function of xi and the unknown parameter : Examples include xi ( ) = xi ;

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

0
yi = 1zi + ei
and both 2 and have dropped out. This means that under H0 ; is not identi…ed. 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. 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 9.4.1 (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
The …rst step is to show that ^ ! 0 : Proving that nonlinear estimators are consistent is more
challenging than for linear estimators. We sketch the main argument. The idea is that ^ minimizes
the sample criterion function Sn ( ); which (for any ) converges in probability to the mean-squared
error function E (yi m (xi ; ))2 : Thus it seems reasonable that the minimizer ^ will converge in
probability to 0 ; the minimizer of E (yi m (xi ; ))2 . 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 di¤erentiabel in : For a complete treatment of the theory of optimization
estimators see Newey and McFadden (1994).
p
Since ^ ! 0 ; ^ is close to 0 for n large, so the minimization of Sn ( ) only needs to be
examined for close to 0 : Let
yi0 = ei + m0 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
= ei m i( 0)
= yi0 m 0
i :
Hence the sum of squared errors function is

n
X n
X 2
Sn ( ) = (yi m (xi ; ))2 ' yi0 m0 i
i=1 i=1
and the right-hand-side is the SSE function for a linear regression of yi0 on m i : Thus the NLLS
estimator b has the same asymptotic distribution as the (infeasible) OLS regression of yi0 on m i ;
which is that stated in the theorem.
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 de…nition 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 ) = :5: Two useful facts about
the median are that
= argmin E jy j (9.8)
and
E sgn (y )=0
where
1 if u 0
sgn (u) =
1 if u < 0
is the sign function.
These facts and de…nitions motivate three estimators ofP : The …rst de…nition is the 50th
empirical quantile. The second is the value which minimizes n1 ni=1 jyi j ; and the third de…nition
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 let’s consider the conditional median of y given a random vector x: Let m(x) = med (y j x)
denote the conditional median of y given x: The linear median regression model takes the form
yi = x0i + ei
med (ei j xi ) = 0
In this model, the linear function med (yi j 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
Pr(yi x0 j xi = x) = Pr(yi > x0 j xi = x) = :5
E (sgn (ei ) j xi ) = 0
E (xi sgn (ei )) = 0
= min E jyi x0i j
These facts motivate the following estimator. Let

n
1X
LADn ( ) = yi x0i
n
i=1
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
1X
xi sgn yi x0i b = 0: (9.9)
n
i=1
The LAD estimator has an asymptotic normal distribution.
Theorem 9.5.1 Asymptotic Distribution of LAD Estimator

When the conditional median is linear in x
p d
n b ! N (0; V )
where
1 1 1
V = E xi x0i f (0 j xi ) Exi x0i E xi x0i f (0 j xi )
4
and f (e j x) is the conditional density of ei given xi = x:
The variance of the asymptotic distribution inversely depends on f (0 j x) ; the conditional

density of the error at its median. When f (0 j 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 j x) = f (0) and the asymptotic variance simpli…es
1
(Exi x0i )
V = (9.10)
4f (0)2
This simpli…cation is similar to the simpli…cation of the asymptotic covariance of the OLS estimator
under homoskedasticity.
Computation of standard error for LAD estimates typically is based on equation (9.10). The
main di¢ culty 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 18. 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 :
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 jyi x0i j :
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 di¤erentiable in
.) It follows that ^ ; the minimizer of LADn ( ); converges in probability to 0 ; the minimizer of
E jyi x0i j.
P
Since sgn (a) = 1 2 1 (a 0) ; (9.9) is equivalent to g n ( b ) = 0; where g n ( ) = n 1 ni=1 g i ( )
and g i ( ) = xi (1 2 1 (yi x0i )) : Let g( ) = Eg i ( ). We need three preliminary results. First,
by the central limit theorem (Theorem 2.8.1)
n
X
p 1=2 d
n (g n ( 0) g( 0 )) = n gi( 0) ! N 0; Exi x0i
i=1
since Eg i ( 0 )g i ( 0 )0 = Exi x0i : Second using the law of iterated expectations and the chain rule of
di¤erentiation,
@ @
g( ) = Exi 1 2 1 yi x0i
@ 0 @ 0
@
= 2 0 E xi E 1 ei x0i x0i 0 j xi
@
" Z 0 #
xi x0i 0
@
= 2 0 E xi f (e j xi ) de
@ 1
= 2E xi x0i f x0i x0i 0 j xi

so
@
g( ) = 2E xi x0i f (0 j xi ) :
@ 0
Third, by a Taylor series expansion and the fact g( ) = 0
@
g( b ) ' g( ) b :
@ 0
Together
1
p @ p
n b 0 ' g( 0) ng( ^ )
@ 0
1p
= 2E xi x0i f (0 j xi ) n g( ^ ) gn( ^ )
1 1p
' E xi x0i f (0 j xi ) n (g n ( 0) g( 0 ))
2
d 1 1
! E xi x0i f (0 j xi ) 0
N 0; Exi xi
2
= N (0; V ) :
p
The third line follows from an asymptotic empirical process argument and the fact that b ! 0.
9.6 Quantile Regression

Quantile regression has become quite popular in recent econometric practice. For 2 [0; 1] the
’th quantile Q of a random variable with distribution function F (u) is de…ned as
Q = inf fu : F (u) g
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
Q = argmin E (U ): (9.11)
where (q) is the piecewise linear function
q (1 ) q<0
(q) = (9.12)
q q 0
= q( 1 (q < 0)) :
This generalizes representation (9.8) for the median to all quantiles.

For the random variables (yi ; xi ) with conditional distribution function F (y j x) the conditional
quantile function q (x) is
Q (x) = inf fy : F (y j x) g:
Again, when F (y j x) is continuous and strictly monotonic in y, then F (Q (x) j 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 computa-
tional convenience it is typical to assume that they are (approximately) linear in x (after suitable
transformations). This linear speci…cation assumes that Q (x) = 0 x where the coe¢ cients
vary across the quantiles : We then have the linear quantile regression model
yi = x0i + ei
where ei is the error de…ned to be the di¤erence between yi and its ’th conditional quantile x0i :
By construction, the ’th conditional quantile of ei is zero, otherwise its properties are unspeci…ed
without further restrictions.
Given the representation (9.11), the quantile regression estimator b for solves the mini-
mization problem
b = argmin S ( )
n
where
n
1X
Sn ( ) = yi x0i
n
i=1
and (q) is de…ned in (9.12).

Since the quanitle regression criterion function Sn ( ) does not have an algebraic solution,
numerical 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.
Theorem 9.6.1 Asymptotic Distribution of the Quantile Regres-

sion Estimator
When the ’th conditional quantile is linear in x
p d
n b ! N (0; V ) ;
where
1 1
V = (1 ) E xi x0i f (0 j xi ) Exi x0i E xi x0i f (0 j xi )
and f (e j 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 j xi ) = f (0) ; the unconditional density of
ei at 0, and we have the simpli…cation
(1 ) 1
V = 2 E xi x0i :
f (0)
A recent monograph on the details of quantile regression is Koenker (2005).
9.7 Testing for Omitted NonLinearity

If the goal is to estimate the conditional expectation E (yi j xi ) ; it is useful to have a general
test of the adequacy of the speci…cation.
One simple test for neglected nonlinearity is to add nonlinear functions of the regressors to the
regression, and test their signi…cance using a Wald test. Thus, if the model yi = x0i b + eî has been
…t by OLS, let z i = h(xi ) denote functions of xi which are not linear functions of xi (perhaps
squares of non-binary regressors) and then …t yi = x0i e +z 0i ~ + e~i by OLS, and form a Wald statistic
for = 0:
Another popular approach is the RESET test proposed by Ramsey (1969). The null model is
yi = x0i + ei
which is estimated by OLS, yielding predicted values yî = x0i b : Now let
0 2 1
yî
B .. C
zi = @ . A
yîm
be an (m 1)-vector of powers of yî : Then run the auxiliary regression
yi = x0i e + z 0i e + e~i (9.13)
by OLS, and form the Wald statistic Wn for = 0: It is easy (although somewhat tedious) to
d 2
show that under the null hypothesis, Wn ! m 1 : Thus the null
is rejected at the % level if Wn
exceeds the upper % tail critical value of the 2
m 1 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(x0i ) + ei
where G( ) is a smooth “link”function. To see why this is the case, note that (9.13) may be written
as
2 3 m
yi = x0 e + x0 b ~ + x0 b ~ +
i i 1 i x0 b 2 ~ + e~i i m 1
which has essentially approximated G( ) by a m’th order polynomial.
9.8 Model Selection

In earlier sections we discussed the costs and bene…ts of inclusion/exclusion of variables. How
does a researcher go about selecting an econometric speci…cation, when economic theory does not
provide complete guidance? This is the question of model selection. It is important that the model
selection question be well-posed. For example, the question: “What is the right model for y?”
is not well-posed, because it does not make clear the conditioning set. In contrast, the question,
“Which subset of (x1 ; :::; xK ) enters the regression function E (yi j x1i = x1 ; :::; xKi = xK )?”is well
posed.
In many cases the problem of model selection can be reduced to the comparison of two nested
models, as the larger problem can be written as a sequence of such comparisons. We thus consider
the question of the inclusion of X 2 in the linear regression
y = X1 1 + X2 2 + e;
where X 1 is n k1 and X 2 is n k2 : This is equivalent to the comparison of the two models
M1 : y = X1 1 + e; E (e j X 1 ; X 2 ) = 0
M2 : y = X1 1 + X2 2 + e; E (e j X 1 ; X 2 ) = 0:
Note that M1 M2 : To be concrete, we say that M2 is true if 2 6= 0:

To …x notation, models 1 and 2 are estimated by OLS, with residual vectors e ^1 and e
^2 ; estimated
2 2
variances ^ 1 and ^ 2 ; etc., respectively. To simplify some of the statistical discussion, we will on
occasion use the homoskedasticity assumption E e2i j x1i ; x2i = 2 :
A model selection procedure is a data-dependent rule which selects one of the two models. We
c There are many possible desirable properties for a model selection procedure.
can write this as M.
One useful property is consistency, that it selects the true model with probability one if the sample
is su¢ ciently large. A model selection procedure is consistent if
c = M1 j M1
Pr M ! 1
c = M2 j M2
Pr M ! 1
However, this rule only makes sense when the true model is …nite dimensional. If the truth is
in…nite dimensional, it is more appropriate to view model selection as determining the best …nite
sample approximation.
A common approach to model selection is to base the decision on a statistical test such as
the Wald Wn : The model selection rule is as follows. For some critical level ; let c satisfy
Pr 2k2 > c = : Then select M1 if Wn c ; else select M2 .
A major problem with this approach is that the critical level is indeterminate. The rea-
soning which helps guide the choice of in hypothesis testing (controlling Type I error) is not
relevant for model selection. That is, if is set to be a small number, then Pr M c = M1 j M1
1 but Pr Mc = M2 j M2 could vary dramatically, depending on the sample size, etc. An-
other problem is that if is held …xed, then this model selection procedure is inconsistent, as
Pr Mc = M1 j M1 ! 1 < 1:
Another common approach to model selection is to use a selection criterion. One popular choice
is the Akaike Information Criterion (AIC). The AIC under normality for model m is
km
AICm = log ^ 2m + 2 : (9.14)
n
where ^ 2m is the variance estimate for model m; and km is the number of coe¢ cients in the
model. The AIC can be derived as an estimate of the Kullback Leibler information distance
K(M) = E (log f (y j X) log f (y j X; M)) between the true density and the model density. The
expectation is taken with respect to the true density. The rule is to select M1 if AIC1 < AIC2 ;
else select M2 : AIC selection is inconsistent, as the rule tends to over…t. Indeed, since under M1 ;
d
LRn = n log ^ 21 log ^ 22 ' Wn ! 2
k2 ; (9.15)
then
c = M1 j M1
Pr M = Pr (AIC1 < AIC2 j M1 )
k1 k1 + k2
= Pr log(^ 21 ) + 2 < log(^ 22 ) + 2 j M1
n n
= Pr (LRn < 2k2 j M1 )
2
! Pr k2 < 2k2 < 1:
While many criterions similar to the AIC have been proposed, the most popular is one proposed
by Schwarz based on Bayesian arguments. His criterion, known as the BIC, is
km
BICm = log ^ 2m + log(n) : (9.16)
n
Since log(n) > 2 (if n > 8); the BIC places a larger penalty than the AIC on the number of
estimated parameters and is more parsimonious.
In contrast to the AIC, BIC model selection is consistent. Indeed, since (9.15) holds under M1 ;
LRn p
! 0;
log(n)
so
c = M1 j M1
Pr M = Pr (BIC1 < BIC2 j M1 )
=Pr (LRn < log(n)k2 j M1 )
LRn
= Pr < k2 j M1
log(n)
! Pr (0 < k2 ) = 1:
Also under M2 ; one can show that

LRn p
! 1;
log(n)
thus
c = M2 j M2 LRn
Pr M = Pr > k2 j M2
log(n)
! 1:
We have discussed model selection between two models. The methods extend readily to the
issue of selection among multiple regressors. The general problem is the model
yi = 1 x1i + 2 x2i + + K xKi + ei ; E (ei j xi ) = 0

and the question is which subset of the coe¢ cients are non-zero (equivalently, which regressors
enter the regression).
There are two leading cases: ordered regressors and unordered.
In the ordered case, the models are
M1 : 1 6= 0; 2 = 3 = = K =0
M2 : 1 6= 0; 2 6= 0; 3 = = K =0
..
.
MK : 1 6= 0; 2 6= 0; : : : ; K 6= 0:
which are nested. The AIC selection criteria estimates the K models by OLS, stores the residual
variance ^ 2 for each model, and then selects the model with the lowest AIC (9.14). Similarly for
the BIC, selecting based on (9.16).
In the unordered case, a model consists of any possible subset of the regressors fx1i ; :::; xKi g;
and the AIC or BIC in principle can be implemented by estimating all possible subset models.
However, there are 2K such models, which can be a very large number. For example, 210 = 1024;
and 220 = 1; 048; 576: In the latter case, a full-blown implementation of the BIC selection criterion
would seem computationally prohibitive.
Exercises
Exercise 9.1 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 coe¢ cient estimates
and standard errors.
(b) Consider augmenting the model by squares and/or cross-products of the conditioning vari-
ables. 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 di¤erent for men and women. Interpret.
(e) Test whether the error variance is di¤erent 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.
(h) Do the OLS and FGLS estimates di¤er greatly? Note any interesting di¤erences.
(i) Compare the estimated standard errors. Note any interesting di¤erences.
Exercise 9.2 In the homoskedastic regression model y = X + e with E(ei j xi ) = 0 and E(e2i j
xi ) = 2 ; suppose ^ is the OLS estimate of with covariance matrix V^ ; based on a sample of
2 2
size n: Let ^ be the estimate of : 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 ( ^ ; V^ ; ^ 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 9.3 Suppose that yi = g(xi ; )+ei with E (ei j xi ) = 0; ^ is the NLLS estimator, and V^ is
the estimate of var ^ : You are interested in the conditional mean function E (yi j xi = x) = g(x)
at some x: Find an asymptotic 95% con…dence interval for g(x):
Exercise 9.4 For any predictor g(xi ) for yi ; the mean absolute error (MAE) is
E jyi g(xi )j :
Show that the function g(x) which minimizes the MAE is the conditional median m (x) = med(yi j
xi ):
Exercise 9.5 De…ne

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 (9.11).
Exercise 9.7 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 : (9.17)
(a) Following Nerlove, add the variable (log Qi )2 to the regression. Do so. Assess the merits of
this new speci…cation using (i) a hypothesis test; (ii) AIC criterion; (iii) BIC criterion. Do
you agree with this modi…cation?
(b) Now try a non-linear speci…cation. Consider model (9.17) plus the extra term 6zi; where
1
z i = log Qi (1 + exp ( (log Qi 7 ))) :
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 or you like) values of 7 in this range. For each value of 7 ; calculate z i
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 ).

Chapter 10
The Bootstrap
10.1 De…nition of the Bootstrap

Let F denote a distribution function for the population of observations (yi ; xi ) : Let
Tn = Tn ((y1 ; x1 ) ; :::; (yn ; xn ) ; F )
be a statistic of interest, for example an estimator ^ or a t-statistic ^ =s(^): Note that we

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 j 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!1 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 di¤erent ap-
proximation. 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
Gn (u) = Gn (u; Fn ): (10.1)
We call Gn 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-
tribution is called the bootstrap data. The statistic Tn = Tn ((y1 ; x1 ) ; :::; (yn ; xn ) ; Fn ) constructed
on this sample is a random variable with distribution Gn : 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
179
CHAPTER 10. THE BOOTSTRAP 180
sample moment:
n
1X
Fn (y; x) = 1 (yi y) 1 (xi x) : (10.2)
n
i=1
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 2.6.1), Fn (y; x) !
F (y; x) : Furthermore, by the CLT (Theorem 2.8.1),
p d
n (Fn (y; x) F (y; x)) ! N (0; F (y; x) (1 F (y; x))) :
To see the e¤ect 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.
Figure 10.1: Empirical Distribution Functions
The EDF is a valid discrete probability distribution which puts probability mass 1=n at each
pair (yi ; xi ), i = 1; :::; n: Notationally, it is helpful to think of a random pair (yi ; xi ) with the
distribution Fn : That is,
Pr(yi y; xi x) = Fn (y; x):
We can easily calculate the moments of functions of (yi ; xi ) :
Z
Eh (yi ; xi ) = h(y; x)dFn (y; x)
n
X
= h (yi ; xi ) Pr (yi = yi ; xi = xi )
i=1
n
1X
= h (yi ; xi ) ;
n
i=1
the empirical sample average.

10.3 Nonparametric Bootstrap

The nonparametric bootstrap is obtained when the bootstrap distribution (10.1) is de…ned
using the EDF (10.2) as the estimate Fn of F:
Since the EDF Fn is a multinomial (with n support points), in principle the distribution Gn could
be calculated by direct methods. However, as there are 2nn 1 possible samples f(y1 ; x1 ) ; :::; (yn ; xn )g;
such a calculation is computationally infeasible. The popular alternative is to use simulation to ap-
proximate the distribution. The algorithm is identical to our discussion of Monte Carlo simulation,
with the following points of clari…cation:
The sample size n used for the simulation is the same as the sample size.
The random vectors (yi ; xi ) are drawn randomly from the empirical distribution. This is
equivalent to sampling a pair (yi ; xi ) randomly from the sample.
The bootstrap statistic Tn = Tn ((y1 ; x1 ) ; :::; (yn ; xn ) ; Fn ) is calculated for each bootstrap sam-
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 su¢ ces.
When the statistic Tn is a function of F; it is typically through dependence on a parameter.
For example, the t-ratio ^ =s(^) depends on : As the bootstrap statistic replaces F with
Fn ; it similarly replaces with n ; the value of implied by Fn : Typically n = ^; the parameter
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
sample f(y1 ; x1 ) ; :::; (yn ; xn )g will necessarily have some ties and multiple values, which is generally
not a problem.
10.4 Bootstrap Estimation of Bias and Variance

The bias of ^ is n = E(^ 0 ): Let Tn ( ) =
^ : Then n = E(Tn ( 0 )): The bootstrap
counterparts are ^ = ^((y1 ; x1 ) ; :::; (yn ; xn )) and Tn = ^ n =
^ ^: The bootstrap estimate
of n is
n = E(Tn ):
If this is calculated by the simulation described in the previous section, the estimate of n is
B
1 X
^n = Tnb
B
b=1
XB
1 ^ ^
= b
B
b=1
= ^ ^:
If ^ is biased, it might be desirable to construct a biased-corrected estimator (one with reduced

bias). Ideally, this would be
~=^ n;
but n is unknown. The (estimated) bootstrap biased-corrected estimator is

~ = ^ ^n
= ^ (^ ^)
= 2^ ^ :
Note, in particular, that the biased-corrected estimator is not ^ : Intuitively, the bootstrap makes
the following experiment. Suppose that ^ is the truth. Then what is the average value of ^
calculated from such samples? The answer is ^ : If this is lower than ^; this suggests that the
estimator is downward-biased, so a biased-corrected estimator of should be larger than ^; and the
best guess is the di¤erence between ^ and ^ : Similarly if ^ is higher than ^; then the estimator is
upward-biased and the biased-corrected estimator should be lower than ^.
Let Tn = ^: The variance of ^ is
Vn = E(Tn ETn )2 :
Let Tn = ^ : It has variance

Vn = E(Tn ETn )2 :
The simulation estimate is
B
^ 1 X ^ ^
2
Vn = b :
B
b=1
A bootstrap
q standard error for ^ is the square root of the bootstrap estimate of variance,
s (^) = V^n :
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 con…dence 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 approxi-
mation is suspected. It appears superior to calculate bootstrap con…dence 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.]
Let qn ( ) denote the quantile function of the true sampling distribution, and qn ( ) = qn ( ; Fn )
denote the quantile function of the bootstrap distribution. Note that this function will change
depending on the underlying statistic Tn whose distribution is Gn :
Let Tn = ^; an estimate of a parameter of interest. In (1 )% of samples, ^ lies in the region
[qn ( =2); qn (1 =2)]: This motivates a con…dence interval proposed by Efron:
C1 = [qn ( =2); qn (1 =2)]:
This is often called the percentile con…dence interval.

Computationally, the quantile qn ( ) is estimated by q^n ( ); the ’th sample quantile of the
simulated statistics fTn1 ; :::; TnB g; as discussed in the section on Monte Carlo simulation. The
(1 )% Efron percentile interval is then [^ qn ( =2); q^n (1 =2)]:
The interval C1 is a popular bootstrap con…dence 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 de…ne
= f ( ) as the parameter of interest for a monotonically increasing function f; then percentile
method applied to this problem will produce the con…dence interval [f (qn ( =2)); f (qn (1 =2))];
which is a naturally good property.
However, as we show now, C1 is in a deep sense very poorly motivated.
It will be useful if we introduce an alternative de…nition C1 . Let Tn ( ) = ^ and let qn ( )
be the quantile function of its distribution. (These are the original quantiles, with subtracted.)
Then C1 can alternatively be written as
C1 = [^ + qn ( =2); ^ + q (1
n =2)]:
This is a bootstrap estimate of the “ideal” con…dence interval
C10 = [^ + qn ( =2); ^ + qn (1 =2)]:
The latter has coverage probability
Pr 0 2 C10 = Pr ^ + qn ( =2) 0
^ + qn (1 =2)
= Pr qn (1 =2) ^ 0 qn ( =2)
= Gn ( qn ( =2); F0 ) Gn ( qn (1 =2); F0 )
which generally is not 1 ! There is one important exception. If ^ 0 has a symmetric distribution,
then Gn ( u; F0 ) = 1 Gn (u; F0 ); so
Pr 0 2 C10 = 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
and this idealized con…dence interval is accurate. Therefore, C10 and C1 are designed for the case
that ^ has a symmetric distribution about 0 :
When ^ does not have a symmetric distribution, C1 may perform quite poorly.
However, by the translation invariance argument presented above, it also follows that if there
exists some monotonically increasing transformation f ( ) such that f (^) is symmetrically distributed
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.
Let Tn ( ) = ^ . Then
1 = Pr (qn ( =2) Tn ( 0 ) qn (1 =2))

= Pr ^ qn (1 =2) 0
^ qn ( =2) ;
so an exact (1 )% con…dence interval for 0 would be
C20 = [^ qn (1 =2); ^ qn ( =2)]:
This motivates a bootstrap analog
C2 = [^ qn (1 =2); ^ qn ( =2)]:
Notice that generally this is very di¤erent from the Efron interval C1 ! They coincide in the special
case that Gn (u) is symmetric about ^; but otherwise they di¤er.
Computationally, this interval can be estimated from a bootstrap simulation by sorting the
bootstrap statistics Tn = ^ ^ ; which are centered at the sample estimate ^: These are sorted
to yield the quantile estimates q^n (:025) and q^n (:975): The 95% con…dence interval is then [^
q^ (:975); ^ q^ (:025)]:
n n
This con…dence interval is discussed in most theoretical treatments of the bootstrap, but is not
widely used in practice.
10.6 Percentile-t Equal-Tailed Interval

Suppose we want to test H0 : = 0 against H1 : < 0 at size : We would set Tn ( ) =
^ =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
qn ( ); and the test rejects if Tn ( 0 ) < qn ( ):
Similarly, if the alternative is H1 : > 0 ; the bootstrap test rejects if Tn ( 0 ) > qn (1 ):
Computationally, these critical values can be estimated from a bootstrap simulation by sorting
the bootstrap t-statistics Tn = ^ ^ =s(^ ): Note, and this is important, that the bootstrap test
statistic is centered at the estimate ^; and the standard error s(^ ) is calculated on the bootstrap
sample. These t-statistics are sorted to …nd the estimated quantiles q^n ( ) and/or q^n (1 ):
Let Tn ( ) = ^ =s(^). Then taking the intersection of two one-sided intervals,
1 = Pr (qn ( =2) Tn ( 0 ) qn (1 =2))

= Pr qn ( =2) ^ 0 =s(^) qn (1 =2)
= Pr ^ s(^)qn (1 =2) 0
^ s(^)qn ( =2) ;
so an exact (1 )% con…dence interval for 0 would be
C30 = [^ s(^)qn (1 =2); ^ s(^)qn ( =2)]:
This motivates a bootstrap analog
C3 = [^ s(^)qn (1 =2); ^ s(^)qn ( =2)]:
This is often called a percentile-t con…dence 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, dis-
cussed above.
10.7 Symmetric Percentile-t Intervals

Suppose we want to test H0 : = 0 against H1 : 6= 0 at size : We would set Tn ( ) =
^ =s(^) and reject H0 in favor of H1 if jTn ( 0 )j > c; where c would be selected so that
Pr (jTn ( 0 )j > c) = :
Note that
Pr (jTn ( 0 )j < c) = Pr ( c < Tn ( 0 ) < c)

= Gn (c) Gn ( c)
Gn (c);
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 jTn ( 0 )j :

The bootstrap estimate is qn ( ); the 1 quantile of the distribution of jTn j ; or the number
which solves the equation
Gn (qn ( )) = Gn (qn ( )) Gn ( qn ( )) = 1 :
Computationally, qn ( ) is estimated from a bootstrap simulation by sorting the bootstrap t-

statistics jTn j = ^ ^ =s(^ ); and taking the upper % quantile. The bootstrap test rejects if
jTn ( 0 )j > qn ( ):
Let
C4 = [^ s(^)qn ( ); ^ + s(^)qn ( )];
where qn ( ) is the bootstrap critical value for a two-sided hypothesis test. C4 is called the symmetric
percentile-t interval. It is designed to work well since
Pr ( 0 2 C4 ) = Pr ^ s(^)qn ( ) 0
^ + s(^)q ( )
n
= Pr (jTn ( 0 )j < qn ( ))
' Pr (jTn ( 0 )j < 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
Wn ( ) = n ^ V^ ^
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
rejects if Wn qn ( ); where qn ( ) is the (1 )% quantile of the distribution of
0 1
Wn = n ^ ^ V^ ^ ^ :
Computationally, the critical value qn ( ) is found as the quantile from simulated values of Wn :
Note in the simulation that the Wald statistic is a quadratic form in ^ ^ ; not ^ 0 :
[This is a typical mistake made by practitioners.]
10.8 Asymptotic Expansions

Let Tn 2 R be a statistic such that
d 2
Tn ! N(0; ): (10.3)
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!1
or
u
Gn (u; F ) = + o (1) : (10.4)
u
While (10.4) says that Gn converges to as n ! 1; it says nothing, however, about the rate
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.
De…nition 10.8.1 an = o(1) if an ! 0 as n ! 1
De…nition 10.8.2 an = O(1) if jan j is uniformly bounded.
De…nition 10.8.3 an = o(n r) if nr jan j ! 0 as n ! 1.
Basically, an = O(n r ) if it declines to zero like n r :

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),

u 1 1 3=2
Gn (u; F ) = + g1 (u; F ) + g2 (u; F ) + O(n )
n1=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 di¤ erentiable 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 1=2
Gn (u; F ) +n g1 (u; F ):
Since the derivative of g1 is odd, the density function is skewed. To a third order of approximation,
u 1=2 1
Gn (u; F ) +n g1 (u; F ) + n g2 (u; F )
which adds a symmetric non-normal component to the approximate density (for example, adding
leptokurtosis).
p
[Side Note: When Tn = n Xn = ; a standardized sample mean, then
1
g1 (u) = 3 u2 1 (u)
6
1 1
g2 (u) = 4 u3 3u + 2
3 u5 10u3 + 15u (u)
24 72
where (u) is the standard normal pdf, and
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 indi¤erence
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 con…dence regions based on an asymptotically normal t-ratio Tn . An asymptotic test is based
on (u):
To the second order, the exact distribution is
1 1
Pr (Tn < u) = Gn (u; F0 ) = (u) + g1 (u; F0 ) + O(n )
n1=2
since = 1: The di¤erence is
1 1
(u) Gn (u; F0 ) = g1 (u; F0 ) + O(n )
n1=2
1=2
= O(n );
so the order of the error is O(n 1=2 ):

A bootstrap test is based on Gn (u); which from Theorem 10.8.1 has the expansion
1 1
Gn (u) = Gn (u; Fn ) = (u) + g1 (u; Fn ) + O(n ):
n1=2
Because (u) appears in both expansions, the di¤erence between the bootstrap distribution and
the true distribution is
1 1
Gn (u) Gn (u; F0 ) = (g1 (u; Fn ) g1 (u; F0 )) + O(n ):
n1=2
p
Since Fn converges to F at rate n; and g1 is continuous with respect to F; the di¤erence
p
(g1 (u; Fn ) g1 (u; F0 )) converges to 0 at rate n: Heuristically,
@
g1 (u; Fn ) g1 (u; F0 ) g1 (u; F0 ) (Fn F0 )
@F
= O(n 1=2 );
@
The “derivative” @F g1 (u; F ) is only heuristic, as F is a function. We conclude that
1
Gn (u) Gn (u; F0 ) = O(n );
or
1
Pr (Tn u) = Pr (Tn u) + O(n );
which is an improved rate of convergence over the asymptotic test (which converged at rate
O(n 1=2 )). This rate can be used to show that one-tailed bootstrap inference based on the t-
ratio achieves a so-called asymptotic re…nement –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
jyj has distribution function
H(u) = H(u) H( u)
since
Pr (jyj u) = Pr ( u y u)
= Pr (y u) Pr (y u)
= H(u) H( u):
For example, if Z N(0; 1); then jZj has distribution function
(u) = (u) ( u) = 2 (u) 1:
Similarly, if Tn has exact distribution Gn (u; F ); then jTn j has the distribution function
Gn (u; F ) = Gn (u; F ) Gn ( u; F ):
d d
A two-sided hypothesis test rejects H0 for large values of jTn j : Since Tn ! Z; then jTn j !
jZj : 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 3=2
( u) + 1=2 g1 ( u; F ) + g2 ( u; F ) + O(n )
n n
2
= (u) + g2 (u; F ) + O(n 3=2 ); (10.5)
n
where the simpli…cations are because g1 is even and g2 is odd. Hence the di¤erence between the
asymptotic distribution and the exact distribution is
2 3=2 1
(u) Gn (u; F0 ) = g2 (u; F0 ) + O(n ) = O(n ):
n
The order of the error is O(n 1 ):

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
2 3=2
Gn (u) = Gn (u; Fn ) = (u) + g2 (u; Fn ) + O(n ):
n
Thus the di¤erence between the bootstrap and exact distributions is
2
Gn (u) Gn (u; F0 ) = (g2 (u; Fn ) g2 (u; F0 )) + O(n 3=2 )
n
= O(n 3=2 );
p
the last equality because Fn converges to F0 at rate n; and g2 is continuous in F: Another way
of writing this is
Pr (jTn j < u) = Pr (jTn j < u) + O(n 3=2 )
so the error from using the bootstrap distribution (relative to the true unknown distribution) is
O(n 3=2 ): This is in contrast to the use of the asymptotic distribution, whose error is O(n 1 ): Thus
a two-sided bootstrap test also achieves an asymptotic re…nement, similar to a one-sided test.
A reader might get confused between the two simultaneous e¤ects. 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. Two-
sided tests will have more accurate size (Reported Type I error), but one-sided tests might have
more power against alternatives of interest. Con…dence 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 con…dence intervals is unclear, and needs to be determined on a case-by-case basis.
10.11 Percentile Con…dence Intervals

p
To evaluate the coverage rate of the percentile interval, set Tn = n ^ 0 : We know that
d
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
Gn (u; F ) = + O(n 1=2 );
p
where = V ; and for the bootstrap
u 1=2
Gn (u) = Gn (u; Fn ) = + O(n );
^
where ^ = V (Fn ) is the bootstrap estimate of : The di¤erence is
u u 1=2
Gn (u) Gn (u; F0 ) = + O(n )
^
u u 1=2
= (^ ) + O(n )
1=2
= O(n )
Hence the order of the error is O(n 1=2 ):

p
The good news is that the percentile-type methods (if appropriately used) can yield n-
convergent asymptotic inference. Yet these methods do not require the calculation of standard
errors! This means that in contexts where standard errors are not available or are di¢ cult 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 con…dence region. Therefore if standard errors are available,
it is unclear if there are any bene…ts 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 Gn (u) = Gn (u; Fn ); where Fn is the EDF.
Any other consistent estimate of F may be used to de…ne 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 ine¢ cient in contexts where
more is known about F: We discuss bootstrap methods appropriate for the linear regression model
yi = x0i + ei
E (ei j xi ) = 0:
The non-parametric bootstrap resamples the observations (yi ; xi ) from the EDF, which implies
yi = xi 0 ^ + ei
E (xi ei ) = 0
but generally
E (ei j xi ) 6= 0:
The bootstrap distribution does not impose the regression assumption, and is thus an ine¢ cient
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 con-
struct 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 su¢ cient to sample the xi and ei independently, and then create
yi = xi 0 ^ + ei : There are di¤erent ways to impose independence. A non-parametric method
is to sample the bootstrap errors ei randomly from the OLS residuals f^ e1 ; :::; e^n g: A parametric
method is to generate the bootstrap errors ei from a parametric distribution, such as the normal
ei N(0; ^ 2 ):
For the regressors xi , a nonparametric method is to sample the xi randomly from the EDF
or sample values fx1 ; :::; xn g: A parametric method is to sample xi from an estimated parametric
distribution. A third approach sets xi = 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 j xi ) = 0: Unfortunately this is a harder problem.
One proposal which imposes the regression condition without independence is the Wild Boot-
strap. The idea is to construct a conditional distribution for ei so that
E (ei j xi ) = 0
E ei 2 j xi = e^2i
E ei 3 j xi = e^3i :
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
p ! ! p
1+ 5 5 1
Pr ei = eî = p
2 2 5
p ! ! p
1 5 5+1
Pr ei = eî = p
2 2 5
For each xi ; you sample ei using this two-point distribution.

Exercises
Exercise 10.1 Let Fn (x) denote the EDF of a random sample. Show that
p d
n (Fn (x) F0 (x)) ! N (0; F0 (x) (1 F0 (x))) :
Exercise 10.2 Take a random sample fy1 ; :::; yn g 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
fy1 ; :::; yn g be a random sample from the empirical distribution function and let Tn = y n be its
sample mean. Find the bootstrap moments ETn and var (Tn ) :
Exercise 10.3 Consider the following bootstrap procedure for a regression of yi on xi : Let ^
^ = y X ^ the OLS residuals.
denote the OLS estimator from the regression of y on X, and e
(a) Draw a random vector (x ; e ) from the pair f(xi ; eî ) : i = 1; :::; ng : That is, draw a random
integer i0 from [1; 2; :::; n]; and set x = xi0 and e = eî0 . Set y = x 0 ^ + e : Draw (with
replacement) n such vectors, creating a random bootstrap data set (y ; X ):
(b) Regress y on X ; yielding OLS estimates ^ and any other statistic of interest.
Show that this bootstrap procedure is (numerically) identical to the non-parametric boot-
strap.
Exercise 10.4 Consider the following bootstrap procedure. Using the non-parametric bootstrap,
generate bootstrap samples, calculate the estimate ^ on these samples and then calculate
Tn = (^ ^)=s(^);
where s(^) is the standard error in the original data. Let qn (:05) and qn (:95) denote the 5% and
95% quantiles of Tn , and de…ne the bootstrap con…dence interval
h i
C = ^ s(^)qn (:95); ^ s(^)qn (:05) :
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
Tn = ^=s(^) > c where c is picked so that Type I error is : You do this as follows. Using the non-
parametric bootstrap, you generate bootstrap samples, calculate the estimates ^ on these samples
and then calculate
Tn = ^ =s(^ ):
Let qn (:95) denote the 95% quantile of Tn . You replace c with qn (:95); and thus reject H0 if
Tn = ^=s(^) > qn (:95): What is wrong with this procedure?
Exercise 10.6 Suppose that in an application, ^ = 1:2 and s(^) = :2: Using the non-parametric
bootstrap, 1000 samples are generated from the bootstrap distribution, and ^ is calculated on each
sample. The ^ are sorted, and the 2.5% and 97.5% quantiles of the ^ are .75 and 1.3, respectively.
(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 data…le 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% con…dence intervals for the regression
coe¢ cients using both the asymptotic normal approximation and the percentile-t bootstrap.
Chapter 11
Generalized Method of Moments
11.1 Overidenti…ed Linear Model

Consider the linear model
yi = x0i + ei
= x01i 1 + x02i 2 + ei
E (xi ei ) = 0
where x1i is k 1 and x2 is r 1 with ` = k + r: We know that without further restrictions, an

asymptotically e¢ cient 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 = x01i 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 e¢ cient, as there are ` restrictions in E (xi ei ) = 0; while 1 is
of dimension k < `. This situation is called overidenti…ed. 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 (11.1)
where 0 is the true value of : In our previous example, g(y; z; ) = z (y x01 ): 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 = x0i + 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
identi…ed, otherwise it is overidenti…ed. The variables xi may be components and functions of
z i ; but this is not required. This model falls in the class (11.1) by setting
g(y; x; z; 0) = z (y x0 ) (11.2)
193
CHAPTER 11. GENERALIZED METHOD OF MOMENTS 194
11.2 GMM Estimator

De…ne the sample analog of (11.2)
n n
1X 1X 1
gn( ) = gi( ) = z i yi x0i = Z 0y Z 0X : (11.3)
n n n
i=1 i=1
The method of moments estimator for is de…ned as the parameter value which sets g n ( ) = 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 de…ne an estimator which sets g n ( )
“close” to zero.
For some ` ` weight matrix W n > 0; let
Jn ( ) = n g n ( )0 W n g n ( ):
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 ( ).
De…nition 11.2.1 b GM M = argmin Jn ( ) :
Note that if k = `; then g n ( b ) = 0; and the GMM estimator is the method of moments
estimator: The …rst order conditions for the GMM estimator are
@
0 = Jn ( b )
@
@
= 2 g n ( b )0 W n g n ( b )
@
1 0 1 0
= 2 X Z Wn Z y Xb
n n
so
2 X 0Z W n Z 0X b = 2 X 0Z W n Z 0y
which establishes the following.
Proposition 11.2.1
b 1
GM M = X 0Z W n Z 0X X 0Z W n Z 0y :
While the estimator depends on W n ; the dependence is only up to scale, for if W n is replaced
by cW n for some c > 0; b GM M does not change.
11.3 Distribution of GMM Estimator

p
Assume that W n ! W > 0: Let
Q = E z i x0i
and
= E z i z 0i e2i = E g i g 0i ;
where g i = z i ei : Then
1 0 1 0 p
X Z Wn ZX ! Q0 W Q
n n
and
1 0 1 d
X Z Wn p Z 0e ! Q0 W N (0; ) :
n n
We conclude:
Theorem 11.3.1 Asymptotic Distribution of GMM Estimator

p d
n b ! N (0; V ) ;
where
1 1
V = Q0 W Q Q0 W W Q Q0 W Q :
In general, GMM estimators are asymptotically normal with “sandwich form”asymptotic vari-
ances.
1
The optimal weight matrix W 0 is one which minimizes V : This turns out to be W 0 = :
The proof is left as an exercise. This yields the e¢ cient GMM estimator:
b = X 0Z 1 1
Z 0X X 0Z 1
Z 0 y:
Thus we have
Theorem 11.3.2 Asymptotic Distribution of E¢ cient GMM Es-

timator p d 1
n b ! N 0; Q0 1 Q :
1 p
W0 = is not known in practice, but it can be estimated consistently. For any W n ! W 0 ;
b
we still call the e¢ cient GMM estimator, as it has the same asymptotic distribution.
By “e¢ cient”, 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 e¢ cient, 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
can have a smaller asymptotic variance than G0 1 G where G = E @@ 0 g i ( ): Since the GMM
estimator has this asymptotic variance, it is semiparametrically e¢ cient.
This result shows that in the linear model, no estimator has greater asymptotic e¢ ciency than
the e¢ cient linear GMM estimator. No estimator can do better (in this …rst-order asymptotic
sense), without imposing additional assumptions.
11.4 Estimation of the E¢ cient Weight Matrix

Given any weight matrix W n > 0; the GMM estimator b is consistent yet ine¢ cient. For
1
example, we can set W n = I ` : In the linear model, a better choice is W n = (Z 0 Z) : Given
any such …rst-step estimator, we can de…ne the residuals eî = yi x0i b and moment equations
î = z i eî = g(yi ; xi ; z i ; b ): Construct
g
n
1X
gn = gn( b ) = î ;
g
n
i=1
î = g
g î gn;
and de…ne ! !
n 1 n 1
1X 1X
Wn = g î 0
î g = g ^0i
î g g n g 0n : (11.4)
n n
i=1 i=1
p 1
Then W n ! = W 0 ; and GMM using W n as the weight matrix is asymptotically e¢ cient.
A common alternative choice is to set
n
! 1
1X
Wn = ^0i
î g
g
n
i=1
which uses the uncentered moment conditions. Since Eg i = 0; these two estimators are asymptot-
ically equivalent under the hypothesis of correct speci…cation. 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. Eg i 6= 0; so the uncentered
estimator will contain an undesirable bias term and the power of the test will be adversely a¤ected.
A simple solution is to use the centered moment conditions to construct the weight matrix, as in
(11.4) above.
Here is a simple way to compute the e¢ cient GMM estimator for the linear model. First, set
W n = (Z 0 Z) 1 , estimate b using this weight matrix, and construct the residual eî = yi x0i b :
Then set gî = z i eî ; and let g
^ be the associated n ` matrix: Then the e¢ cient GMM estimator is
1 1 1
b = X 0Z g
^0 g
^ ng n g 0n Z 0X X 0Z g
^0 g
^ ng n g 0n Z 0 y:
In most cases, when we say “GMM”, we actually mean “e¢ cient GMM”. There is little point in
using an ine¢ cient GMM estimator when the e¢ cient estimator is easy to compute.
An estimator of the asymptotic variance of ^ can be seen from the above formula. Set
1 1
Vb = n X 0 Z g
^0 g
^ ng n g 0n Z 0X :
Asymptotic standard errors are given by the square roots of the diagonal elements of Vb :
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
1 X
J( ) = n g n ( )0 g i ( )g i ( )0 g n ( ):
n
i=1
where
gi ( ) = gi( ) gn( )
The b which minimizes this function is called the continuously-updated GMM estimator, and
was introduced by L. Hansen, Heaton and Yaron (1996).
The estimator appears to have some better properties than traditional GMM, but can be nu-
merically tricky to obtain in some cases. This is a current area of research in econometrics.
11.5 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. Identi…cation requires l k = dim( ): The GMM estimator
minimizes
J( ) = n g n ( )0 W n g n ( )
where
n
1X
gn( ) = gi( )
n
i=1
and !
n 1
1X
Wn = g ^0i
î g g n g 0n ;
n
i=1
with gî = g i ( e ) constructed using a preliminary consistent estimator e , perhaps obtained by …rst
setting W n = I: Since the GMM estimator depends upon the …rst-stage estimator, often the weight
matrix W n is updated, and then b recomputed. This estimator can be iterated if needed.
Theorem 11.5.1 Distribution of Nonlinear GMM Estimator

Under general regularity conditions,
p d 1
n b ! N 0; G0 1
G ;
where
1
= E g i g 0i
and
@
G=E g ( ):
@ 0 i
The variance of b may be estimated by
1
Vb ^0 ^
= G
1
^
G
where X
^ =n 1
g î 0
î g
i
and
X @
^=n 1 ^
G 0 g i ( ):
i
@
The general theory of GMM estimation and testing was exposited by L. Hansen (1982).
11.6 Over-Identi…cation Test

Overidenti…ed models (` > k) are special in the sense that there may not be a parameter value
such that the moment condition
Eg(yi ; xi ; z i ; ) = 0
holds. Thus the model –the overidentifying restrictions –are testable.
For example, take the linear model yi = 01 x1i + 02 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 = 01 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
both E (x1i (yi x01i 1 )) = 0 and E (x2i (yi x01i 1 )) = 0 hold simultaneously. In this sense an
exclusion restriction can be seen as an overidentifying restriction.
p
Note that g n ! Eg i ; 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 0n W n g n
1
= n2 g 0n g
^0 g
^ ng n g 0n gn:
is a quadratic form in g n ; and is thus a natural test statistic for H0 : Eg i = 0:
Theorem 11.6.1 (Sargan-Hansen). Under the hypothesis of correct spec-

i…cation, and if the weight matrix is asymptotically e¢ cient,
d
Jn = Jn ( b ) ! 2
` k:
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 restric-
tions. 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
overidenti…cation 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-identi…ed models are estimated by GMM, it is customary to report the J statistic
as a general test of model adequacy.
11.7 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 ( )
and those under H0 are

e = argmin J( ):
h( )=0
The two minimizing criterion functions are Jn ( b ) and Jn ( e ): The GMM distance statistic is the
di¤erence
Dn = Jn ( e ) Jn ( b ):
Proposition 11.7.1 If the same weight matrix W n is used for both null
and alternative,
1. D 0
d 2
2. D ! r
3. If h is linear in ; then D equals the Wald statistic.
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 alter-
native, 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 compu-
tation of alternative models.
11.8 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 ( ) j z i ) = 0
where ei ( ) is some s 1 function of the observation and the parameters. In many cases, s = 1.
It turns out that this conditional moment restriction is much more powerful, and restrictive,
than the unconditional moment restriction discussed above.
Our linear model yi = x0i + ei with instruments z i falls into this class under the stronger
assumption E (ei j z i ) = 0: Then ei ( ) = yi x0i :
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 x0i , while in a nonlinear
regression model ei ( ) = yi g(xi ; ): In a joint model of the conditional mean and variance
8
< yi x0i
ei ( ; ) = :
: 0 2 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
satis…es Eg i ( ) = 0 and hence de…nes a GMM estimator. The obvious problem is that the class of
functions is in…nite. Which should be selected?
This is equivalent to the problem of selection of the best instruments. If xi 2 R is a valid

instrument satisfying E (ei j xi ) = 0; then xi ; x2i ; x3i ; :::; etc., are all valid instruments. Which
should be used?
One solution is to construct an in…nite 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 Cham-
berlain (1987). Take the case s = 1: Let
@
Ri = E ei ( ) j z i
@
and
2
i = E ei ( )2 j z i :
Then the “optimal instrument” is
2
Ai = i Ri
so the optimal moment is
g i ( ) = Ai ei ( ):
Setting g i ( ) to be this choice (which is k 1; so is just-identi…ed) 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 x0i ; note that
Ri = E (xi j z i )
and
2
i = E e2i j z i ;
so
2
Ai = i E (xi j z i ) :
In the case of linear regression, xi = z i ; so Ai = i 2 z i : Hence e¢ cient GMM is GLS, as we
discussed earlier in the course.
In the case of endogenous variables, note that the e¢ cient 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 e¢ cient
instrument is also inversely proportional to the conditional variance of ei : This is the same as the
GLS estimator; namely that improved e¢ ciency can be obtained if the observations are weighted
inversely to the conditional variance of the errors.
11.9 Bootstrap GMM Inference

Let b be the 2SLS or GMM estimator of . Using the EDF of (yi ; z i ; xi ), we can apply the
bootstrap methods discussed in Chapter 10 to compute estimates of the bias and variance of b ;
and construct con…dence 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 re…nement when the model is over-identi…ed, jeopardizing the
theoretical justi…cation for percentile-t methods. Furthermore, the bootstrap applied J test will
yield the wrong answer.
The problem is that in the sample, b is the “true”value and yet g n ( ^ ) 6= 0: Thus according to
random variables (yi ; z i ; xi ) drawn from the EDF Fn ;
E gi b = g n ( ^ ) 6= 0:
This means that (yi ; z i ; xi ) do not satisfy the same moment conditions as the population distrib-
ution.
A correction suggested by Hall and Horowitz (1996) can solve the problem. Given the bootstrap
sample (y ; Z ; X ); de…ne the bootstrap GMM criterion
0
Jn ( ) = n gn( ) gn( ^ ) W n gn( ) gn( ^ )
where g n ( b ) is from the in-sample data, not from the bootstrap data.
Let b minimize Jn ( ); and de…ne all statistics and tests accordingly. In the linear model, this
implies that the bootstrap estimator is
b = X 0Z W Z 0X 1
n n X 0Z W n Z 0y Z 0e
^ :
where e^ = y X b are the in-sample residuals. The bootstrap J statistic is Jn ( b ):

Brown and Newey (2002) have an alternative solution. They note that we can sample from
the observations with the empirical likelihood probabilities pî described in Chapter 12. Since
Pn
î g i b = 0; this sampling scheme preserves the moment conditions of the model, so no
i=1 p
recentering or adjustments is needed. Brown and Newey argue that this bootstrap procedure will
be more e¢ cient than the Hall-Horowitz GMM bootstrap.
Exercises
Exercise 11.1 Take the model
yi = x0i + ei
E (xi ei ) = 0
e2i = z 0i + i
E (z i i ) = 0:
Find the method of moments estimators ^ ; ^ for ( ; ) :
Exercise 11.2 Take the single equation
y = X +e
E (e j Z) = 0
1
Assume E e2i j z i = 2: Show that if ^ is estimated by GMM with weight matrix W n = (Z 0 Z) ;
then p d 1
n b ! N 0; 2
Q0 M 1
Q
where Q = E (z i x0i ) and M = E (z i z 0i ) :
Exercise 11.3 Take the model yi = x0i + ei with E (z i ei ) = 0: Let eî = yi x0i b where b is
consistent for (e.g. a GMM estimator with arbitrary weight matrix). De…ne the estimate of the
optimal GMM weight matrix
n
! 1
1X
Wn = z i z 0i e^2i :
n
i=1
p 1
Show that W n ! where = E z i z 0i e2i :
Exercise 11.4 In the linear model estimated by GMM with general weight matrix W ; the asymp-
totic variance of ^ GM M is
1 1
V = Q0 W Q Q0 W W Q Q0 W Q
1 1
(a) Let V 0 be this matrix when W = : Show that V 0 = Q0 1
Q :
(b) We want to show that for any W ; V V 0 is positive semi-de…nite (for then V 0 is the smaller
1
possible covariance matrix and W = is the e¢ cient 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 11.5 The equation of interest is
yi = g(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 e¢ cient

GMM estimator for .
Exercise 11.6 In the linear model y = X + e with E(xi ei ) = 0; a Generalized Method of

Moments (GMM) criterion function for is de…ned as
1 1
Jn ( ) = (y X )0 X b X 0 (y X ) (11.5)
n
P
where b = n1 ni=1 xi x0i e^2i ; eî = yi x0i ^ are the OLS residuals, and b = (X 0 X)
1
X 0 Y is LS: The
GMM estimator of ; subject to the restriction h( ) = 0; is de…ned as
e = argmin Jn ( ):
h( )=0
The GMM test statistic (the distance statistic) of the hypothesis h( ) = 0 is
D = Jn ( e ) = min Jn ( ): (11.6)
h( )=0
(a) Show that you can rewrite Jn ( ) in (11.5) as

0 1
Jn ( ) = n b V^ b
thus e is the same as the minimum distance estimator.

(b) Show that in this setting, the distance statistic D in (11.6) equals the Wald statistic.
Exercise 11.7 Take the linear model
yi = x0i + ei
E (z i ei ) = 0:
and consider the GMM estimator ^ of : Let

1
Jn = ng n ( b )0 b gn( b )
d 2
denote the test of overidentifying restrictions. Show that Jn ! ` k as n ! 1 by demonstrating
each of the following:
1
(a) Since > 0; we can write = CC 0 and = C0 1
C 1
0 1
(b) Jn = n C 0 g n ( b ) C0 ^ C C 0gn( b )
(c) C 0 g n ( b ) = D n C 0 g n ( 0) where
1
1 0 1 0 b 1 1 0 1 0 b 1
Dn = I ` C0 ZX XZ ZX XZ C0 1
n n n n
1 0
gn( 0) = Z e:
n
p 1
(d) D n ! I ` R (R0 R) R0 where R = C 0 E (z i x0i )
d
(e) n1=2 C 0 g n ( 0) !X N (0; I ` )
d 1
(f) Jn ! X 0 I ` R (R0 R) R0 X
1
(g) X 0 I ` R (R0 R) R0 X 2 :
` k
1
Hint: I ` R (R0 R) R0 is a projection matrix.
Chapter 12
Empirical Likelihood
12.1 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
Xn
pi = 1: (12.1)
i=1
Since each observation is observed once in the sample, the log-likelihood function for this multino-
mial distribution is
n
X
log L (p1 ; :::; pn ) = log(pi ): (12.2)
i=1
First let us consider a just-identi…ed 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
(12.1). This is equivalent to maximizing
n n
!
X X
log(pi ) pi 1
i=1 i=1
where is a Lagrange multiplier. The n …rst order conditions are 0 = pi 1 : Combined with the
1
constraint (12.1) we …nd that the MLE is pi = n yielding the log-likelihood n log(n):
Now consider the case of an overidenti…ed model with moment condition
Eg i ( 0) =0
where g is ` 1 and is k 1 and for simplicity we write g i ( ) = 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
X n
pi g i ( ) = 0 (12.3)
i=1
The empirical likelihood estimator is the value of which maximizes the multinomial log-
likelihood (12.2) subject to the restrictions (12.1) and (12.3).
204
CHAPTER 12. EMPIRICAL LIKELIHOOD 205
The Lagrangian for this maximization problem is

n n
! n
X X X
0
L ( ; p1 ; :::; pn ; ; ) = log(pi ) pi 1 n pi g i ( )
i=1 i=1 i=1
where and are Lagrange multipliers. The …rst-order-conditions of L with respect to pi , ; and
are
1
= + n 0gi ( )
pi
n
X
pi = 1
i=1
n
X
pi g i ( ) = 0:
i=1
Multiplying the …rst equation by pi , summing over i; and using the second and third equations, we
…nd = n and
1
pi = :
n 1 + 0gi ( )
Substituting into L we …nd
n
X
0
R( ; ) = n log (n) log 1 + gi ( ) : (12.4)
i=1
For given ; the Lagrange multiplier ( ) minimizes R ( ; ) :
( ) = argmin R( ; ): (12.5)
This minimization problem is the dual of the constrained maximization problem. The solution
(when it exists) is well de…ned 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 (pro…le)
empirical log-likelihood function for .
R( ) = R( ; ( ))
n
X
= n log (n) log 1 + ( )0 g i ( )
i=1
The EL estimate ^ is the value which maximizes R( ); or equivalently minimizes its negative
^ = argmin [ R( )] (12.6)
Numerical methods are required for calculation of ^ (see Section 12.5).

As a by-product of estimation, we also obtain the Lagrange multiplier ^ = ( ^ ); probabilities
1
pî = 0
:
n 1 + ^ gi ^
and maximized empirical likelihood

n
X
R( ^ ) = log (^
pi ) : (12.7)
i=1
12.2 Asymptotic Distribution of EL Estimator

De…ne
@
Gi ( ) = g ( ) (12.8)
@ 0 i
G = EGi ( 0)
0
= E gi ( 0) gi ( 0)
and
1
V = G0 1
G (12.9)
1
V = G G0 1
G G0 (12.10)
For example, in the linear model, Gi ( ) = z i x0i ; G = E (z i x0i ), and = E z i z 0i e2i :
Theorem 12.2.1 Under regularity conditions,

p d
n ^ 0 ! N (0; V )
p d
n^ ! 1
N (0; V )
p
where V and V are de…ned in (12.9) and (12.10), and n ^ 0 and
p ^
n are asymptotically independent.
The theorem shows that asymptotic variance V for ^ is the same as for e¢ cient GMM. Thus
the EL estimator is asymptotically e¢ cient.
Chamberlain (1987) showed that V is the semiparametric e¢ ciency bound for in the overi-
denti…ed 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 e¢ cient estimator for .
Proof of Theorem 12.2.1. ( ^ ; ^ ) jointly solve
@
n
X gi ^
0 = R( ; ) = 0
(12.11)
@ 1 + ^ gi ^
i=1
0
@
n
X Gi ^
0 = R( ; ) = 0
: (12.12)
@ 1 + ^ gi ^
i=1
P P P
Let Gn = n1 ni=1 Gi ( 0 ) ; g n = n1 ni=1 g i ( 0 ) and n = n1 ni=1 g i ( 0
0) gi ( 0) :
Expanding (12.12) around = 0 and = 0 = 0 yields
0 ' G0n ^ 0 : (12.13)
Expanding (12.11) around = 0 and = 0 = 0 yields
0' gn Gn ^ 0 + n
^ (12.14)
Premultiplying by G0n n
1
and using (12.13) yields
0 ' G0n n
1
gn G0n n
1
Gn ^ 0 + G0n n
1
n
^
= G0n n
1
gn G0n n
1
Gn ^ 0
Solving for ^ and using the WLLN and CLT yields

p 1 1p
n ^ 0 ' G0n n 1 Gn G0n n ng n (12.15)
d 1
! G0 1
G G0 1
N (0; )
= N (0; V )
Solving (12.14) for ^ and using (12.15) yields

p 1 p
n ^ ' n 1 I Gn G0n n
1
Gn G0n n
1
ng n (12.16)
d 1 1
! I G G0 1
G G0 1
N (0; )
1
= N (0; V )
Furthermore, since
1
G0 I 1
G G0 1
G G0 = 0
p p ^
n ^ 0 and n are asymptotically uncorrelated and hence independent.
12.3 Overidentifying Restrictions

In a parametric likelihood context, tests are based on the di¤erence in the log likelihood func-
tions. The same statistic can be constructed for empirical likelihood. Twice the di¤erence between
the unrestricted empirical log-likelihood n log (n) and the maximized empirical log-likelihood for
the model (12.7) is
Xn
0
LRn = 2 log 1 + ^ g i ^ : (12.17)
i=1
d 2 :
Theorem 12.3.1 If Eg i ( 0) = 0 then LRn ! ` k
The EL overidenti…cation test is similar to the GMM overidenti…cation test. They are asymp-
totically …rst-order equivalent, and have the same interpretation. The overidenti…cation 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 12.3.1. First, by a Taylor expansion, (12.15), and (12.16),

n
1 X p
p gi ^ ' n g n + Gn ^ 0
n
i=1
1 p
' I Gn G0n n
1
Gn G0n n
1
ng n
p
' n n^:
Second, since log(1 + u) ' u u2 =2 for u small,

n
X 0
LRn = 2 log 1 + ^ g i ^
i=1
n n
0X X 0
' 2^ gi ^ ^0 gi ^ gi ^ ^
i=1 i=1
0
' n^ n
^
d
! N (0; V )0 1
N (0; V )
2
= ` k
where the proof of the …nal equality is left as an exercise.
12.4 Testing
Let the maintained model be
Eg i ( ) = 0 (12.18)
where g is ` 1 and is k 1: By “maintained” we mean that the overidentfying restrictions
contained in (12.18) 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
R( ~ ) = max R( ):
h( )=0
Fundamentally, the restricted EL estimator ~ is simply an EL estimator with ` k+a overidentifying

restrictions, so there is no fundamental change in the distribution theory for ~ relative to ^ : To test
the hypothesis h( ) while maintaining (12.18), the simple overidentifying restrictions test (12.17)
is not appropriate. Instead we use the di¤erence in log-likelihoods:
LRn = 2 R( ^ ) R( ~ ) :
This test statistic is a natural analog of the GMM distance statistic.
d 2:
Theorem 12.4.1 Under (12.18) and H0 : h( ) = 0; LRn ! a
The proof of this result is more challenging and is omitted.

12.5 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 (12.4). De…ne
gi ( )
gi ( ; ) =
1 + 0gi ( )
Gi ( )0
Gi ( ; ) =
1 + 0gi ( )
The …rst derivatives of (12.4) are

n
X
@
R = R( ; ) = gi ( ; )
@
i=1
Xn
@
R = R( ; ) = Gi ( ; ) :
@
i=1
The second derivatives are

n
X
@2
R = 0R( ; )= g i ( ; ) g i ( ; )0
@ @ i=1
n
X
@2 Gi ( )
R = 0R( ; )= g i ( ; ) G i ( ; )0
@ @ i=1
1 + 0gi ( )
0 1
@2
@2
n
X @ @ 0 g i ( )0
= @Gi ( ; ) Gi ( ; )0 A
R 0R( ; )= 0
@ @ i=1
1+ gi ( )
Inner Loop
The so-called “inner loop” solves (12.5) for given : The modi…ed Newton method takes a
quadratic approximation to Rn ( ; ) yielding the iteration rule
1
j+1 = j (R ( ; j )) R ( ; j) : (12.19)
where > 0 is a scalar steplength (to be discussed next). The starting value 1 can be set to the
zero vector. The iteration (12.19) is continued until the gradient R ( ; j ) is smaller than some
prespeci…ed tolerance.
E¢ cient convergence requires a good choice of steplength : One method uses the following
quadratic approximation. Set 0 = 0; 1 = 12 and 2 = 1: For p = 0; 1; 2; set
1
p = j p (R ( ; j )) 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 (12.19).
A complication is that must be constrained so that 0 pi 1 which holds if

0
n 1+ gi ( ) 1 (12.20)
for all i: If (12.20) fails, the stepsize needs to be decreased.

Outer Loop
The outer loop is the minimization (12.6). This can be done by the modi…ed Newton method
described in the previous section. The gradient for (12.6) is
@ @ 0
R = R( ) = R( ; ) = R + 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 (12.6) is
@
R = R( )
@ @ 0
@ 0
= R ( ; ( )) + R ( ; ( ))
@ 0
= R ( ; ( )) + R0 + 0
R + 0
R
0 1
= R 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 13
Endogeneity
We say that there is endogeneity in the linear model y = x0i + ei if is the parameter of
interest and E(xi ei ) 6= 0: This cannot happen if is de…ned by linear projection, so requires a
structural interpretation. The coe¢ cient must have meaning separately from the de…nition of a
conditional mean or linear projection.
Example: Measurement error in the regressor. Suppose that (yi ; xi ) are joint random
variables, E(yi j xi ) = xi 0 is linear, is the parameter of interest, and xi is not observed. Instead
we observe xi = xi + ui where ui is an k 1 measurement error, independent of yi and xi : Then
yi = xi 0 + ei
= (xi ui )0 + ei
= x0i + vi
where
vi = ei u0i :
The problem is that
E (xi vi ) = E (xi + ui ) ei u0i = E ui u0i 6= 0
if 6= 0 and E (ui u0i ) 6= 0: It follows that if ^ is the OLS estimator, then
^ p! = E xi x0i
1
E ui u0i 6= :
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

qi = 2 pi + e2i :
e1i
Assume that ei = is iid, Eei = 0; 1 + 2 = 1 and Eei e0i = 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,
1 1 qi e1i
=
1 2 pi e2i
211
CHAPTER 13. ENDOGENEITY 212
so
1
qi 1 1 e1i
=
pi 1 2 e2i
2 1 e1i
=
1 1 e2i
2 e1i + 1 e2i
= :
(e1i e2i )
The projection of qi on pi yields
qi = pi + "i
E (pi "i ) = 0
where
E (pi qi ) 2 1
= =
E p2i 2
p
Hence if it is estimated by OLS, ^ ! ; which does not equal either 1 or 2: This is called
simultaneous equations bias.
13.1 Instrumental Variables

Let the equation of interest be
yi = x0i + ei (13.1)
where xi is k 1; and assume that E(xi ei ) 6= 0 so there is endogeneity. We call (13.1) the
structural equation. In matrix notation, this can be written as
y = X + e: (13.2)
Any solution to the problem of endogeneity requires additional information which we call in-
struments.
De…nition 13.1.1 The ` 1 random vector z i is an instrumental vari-

able for (13.1) 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
x1i k1
xi = (13.3)
x2i k2
where E(x1i ei ) = 0 yet E(x2i ei ) 6= 0: We call x1i exogenous and x2i endogenous. By the above
de…nition, x1i is an instrumental variable for (13.1); so should be included in z i : So we have the
partition
x1i k1
zi = (13.4)
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 coe¢ cients
in the equation.
The model is just-identi…ed if ` = k (i.e., if `2 = k2 ) and over-identi…ed if ` > k (i.e., if
`2 > k2 ):
We have noted that any solution to the problem of endogeneity requires instruments. This does
not mean that valid instruments actually exist.
13.2 Reduced Form

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 z 0i E z i x0i
be the ` k matrix of coe¢ cients from a projection of xi on z i ; and de…ne

0
ui = xi zi
as the projection error. Then the reduced form linear relationship between xi and z i is
0
xi = z i + ui : (13.5)
In matrix notation, we can write (13.5) as
X =Z +U (13.6)
where U is n k:
By construction,
E(z i u0i ) = 0;
so (13.5) is a projection and can be estimated by OLS:
x = z^ + u
^
^ = z z 1 z0x :
0
Substituting (13:6) into (13.2), we …nd
y = (Z + U ) +e
= Z + v; (13.7)
where
= (13.8)
and
v=U + e:
Observe that
E (z i vi ) = E z i u0i + E (z i ei ) = 0:
Thus (13.7) is a projection equation and may be estimated by OLS. This is
y = Z^ + v
^;
^ = Z Z 1 Z 0y
0
The equation (13.7) is the reduced form for y: (13.6) and (13.7) 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 ^ ; ^

13.3 Identi…cation
The structural parameter relates to ( ; ) through (13.8). The parameter is identi…ed,
meaning that it can be recovered from the reduced form, if
rank ( ) = k: (13.9)
1
Assume that (13.9) holds. If ` = k; then = : If ` > k; then for any W > 0; =
0 1 0
( W ) W :
If (13.9) is not satis…ed, then cannot be recovered from ( ; ) : Note that a necessary (although
not su¢ cient) condition for (13.9) is ` k:
Since Z and X have the common variables X 1 ; we can rewrite some of the expressions. Using
(13.3) and (13.4) 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
(13.6) can be rewritten as
X1 = Z1
X2 = Z1 12 + Z2 22 + U 2: (13.10)
is identi…ed if rank( ) = k; which is true if and only if rank( 22 ) = k2 (by the upper-diagonal
structure of ): Thus the key to identi…cation of the model rests on the `2 k2 matrix 22 in
(13.10).
13.4 Estimation
The model can be written as
yi = x0i + ei
E (z i ei ) = 0
or
Eg i ( ) = 0
g i ( ) = z i yi x0i :
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:
13.5 Special Cases: IV and 2SLS

If the model is just-identi…ed, so that k = `; then the formula for GMM simpli…es. We …nd that
b = 1
X 0 ZW n Z 0 X X 0 ZW n Z 0 y
1 1
= Z 0X W n 1 X 0Z X 0 ZW n Z 0 y
1
= Z 0X Z 0y
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-identi…ed
case, the weight matrix places no role. This is also the MME estimator of ; and the EL estimator.
1
Another interpretation stems from the fact that since = ; we can construct the Indirect
Least Squares (ILS) estimator:
1
b = b b
1 1 1
= Z 0Z Z 0X Z 0Z Z 0y
1 1
= Z 0X Z 0Z Z 0Z Z 0y
1
= Z 0X Z 0y :
which again is the IV estimator.
Recall that the optimal weight matrix is an estimate of the inverse of = E z i z 0i e2i : In the
2
special case that E ei j z i = 2 (homoskedasticity), then 0
= E (z i z i ) 2 / E (z i z 0i ) suggesting
1
the weight matrix W n = (Z 0 Z) : Using this choice, the GMM estimator equals
1 1 1
b = X 0Z Z 0Z Z 0X X 0Z Z 0Z Z 0y
2SLS
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 e¢ cient GMM, but otherwise it is
ine¢ cient.
It is useful to observe that writing
1
P = Z Z 0Z Z0
c = PX = Zb
X
then the 2SLS estimator is

b = 1
X 0P X X 0P y
1
= c0 X
X c c0 y:
X
The source of the “two-stage” name is since it can be computed as follows

1
First regress X on Z; vis., b = (Z 0 Z) c = Z b = P X:
(Z 0 X) and X
1
Second, regress y on X; c0 X
c vis., b = X c c0 y:
X
c Recall, X = [X 1 ; X 2 ] and Z = [X 1 ; Z 2 ]: Then

It is useful to scrutinize the projection X:
h i
c = X
X c1 ; X
c2
= [P X 1 ; P X 2 ]
= [X 1 ; P X 2 ]
h i
= X 1; X c2 ;
c2 : So only the
since X 1 lies in the span of X: Thus in the second stage, we regress y on X 1 and X
endogenous variables X 2 are replaced by their …tted values:
c2 = Z 1 b 12 + Z 2 b 22 :
X
13.6 Bekker Asymptotics

Bekker (1994) used an alternative asymptotic framework to analyze the …nite-sample bias in
the 2SLS estimator. Here we present a simpli…ed version of one of his results. In our notation, the
model is
y = X +e (13.11)
X = Z +U (13.12)
= (e; U )
E ( j Z) = 0
0
E jZ = S
As before, Z is n l so there are l instruments.

First, let’s analyze the approximate bias of OLS applied to (13.11). Using (13.12),
1 0 0
E Xe = E (xi ei ) = E (z i ei ) + E (ui ei ) = s21
n
and
1 0
E XX = E xi x0i
n
0
= E z i z 0i + E ui z 0i + 0
E z i u0i + E ui u0i
0
= Q + S 22
where Q = E (z i z 0i ) : Hence by a …rst-order approximation

1
1 0 1 0
E ^ OLS E XX E Xe
n n
0 1
= Q + S 22 s21 (13.13)
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 0P y :
1
Let P = Z (Z 0 Z) Z 0 . By the spectral decomposition of an idempotent matrix, P = H H 0
where = diag (I l ; 0) : Let Q = H 0 S 1=2 which satis…es EQ0 Q = I n and partition Q = (q 01 Q02 )
where q 1 is l 1: Hence
1 0 1 1=20
E P jZ = S E Q0 Q j Z S 1=2
n n
1 1=20 1 0
= S E q q S 1=2
n n 1 1
l 1=20 1=2
= S S
n
= S
where
l
= :
n
Using (13.12) and this result,
1 1 1
E X 0P e = E 0
Z 0e + E U 0 P e = s21 ;
n n n
and
1 1
E X 0P X = 0
E z i z 0i + 0
E (z i ui ) + E ui z 0i + E U 0P U
n n
0
= Q + S 22 :
Together
1
1 0 1 0
E ^ 2SLS E X PX E X Pe
n n
0 1
= Q + S 22 s21 : (13.14)
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 (13.14)
approaches that in (13.13), 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 ! 1 (so that the number of instruments goes to in…nity proportionately with sample
size) then the expression in (13.14) is the probability limit of ^ 2SLS
13.7 Identi…cation Failure

Recall the reduced form equation
X2 = Z1 12 + Z2 22 + U 2:
The parameter fails to be identi…ed if 22 has de…cient rank. The consequences of identi…cation
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
and 22 = = E (zi xi ) =Ezi2 : We see that is identi…ed if and only if 6= 0; which occurs
when E (xi zi ) 6= 0. Thus identi…cation 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
n
1 X d
p zi ei ! N1 N 0; E zi2 e2i (13.15)
n
i=1
n n
1 X 1 X d
p zi x i = p zi ui ! N2 N 0; E zi2 u2i (13.16)
n n
i=1 i=1
therefore Pn
p1
n i=1 zi ei d N1
^ = Pn ! Cauchy;
1 N2
i=1 zi xi
p
n
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).
Suppose that identi…cation 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
1=2
22 =n C;
where C is a full rank matrix. The n 1=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
= n 1=2 c:
Then (13.15) is una¤ected, but (13.16) instead takes the form
n n n
1 X 1 X 2 1 X
p z i xi = p zi + p zi ui
n n n
i=1 i=1 i=1
n n
1X 2 1 X
= zi c + p zi ui
n n
i=1 i=1
d
! Qc + N2
therefore
^ d N1
! :
Qc + N2
As in the case of complete identi…cation failure, we …nd that ^ is inconsistent for and the
^
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 identi…cation failure. Once again, the
equation to focus on is the reduced form
X2 = Z1 12 + Z2 22 + U2
and identi…cation 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 su¢ cient for identi…cation. It is not even su¢ cient 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 de…nitive proof that 22 has full rank. Unfortunately, tests of de…cient rank are di¢ cult 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 de…cient rank.
Exercises
1. Consider the single equation model
yi = xi + ei ;
where yi and zi are both real-valued (1 1). Let ^ denote the IV estimator of using as an
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 = x0i + ei
E (ei j xi ) = 0
suppose 2i = E e2i j X i is known. Show that the GLS estimator of can be written as an
IV estimator using some instrument z i : (Find an expression for z i :)
3. Take the linear model

y = X + e:
Let the OLS estimator for be ^ and the OLS residual be e
^ = y X ^.
Let the IV estimator for using some instrument Z be ~ and the IV residual be e
~ = y X ~.
~0 e
If Z is indeed endogeneous, will IV “…t” better than OLS, in the sense that e ~<e^0 e
^; at
least in large samples?
4. The reduced form between the regressors xi and instruments z i takes the form
0
xi = z i + ui
or
X =Z +U
where X i 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 de…ned by the population moment condition
E z i u0i = 0
1
Show that the method of moments estimator for is ^ = (Z 0 Z) (Z 0 X) :
5. In the structural model
y = X +e
X = Z +U
with l k; l k; we claim that is identi…ed (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
identi…ed.
6. Take the linear model
yi = x i + e i
E (ei j xi ) = 0:
where xi and are 1 1:

(a) Show that E (xi ei ) = 0 and E x2i ei = 0: Is z i = (xi x2i )0 a valid instrumental variable
for estimation of ?
(b) De…ne the 2SLS estimator of ; using z i as an instrument for xi : How does this di¤er
from OLS?
(c) Find the e¢ cient GMM estimator of based on the moment condition
E (z i (yi xi )) = 0:
Does this di¤er from 2SLS and/or OLS?
7. Suppose that price and quantity are determined by the intersection of the linear demand and
supply curves
Demand : Q = a0 + a1 P + a2 Y + e1
Supply : Q = b0 + b1 P + b2 W + e2
where income (Y ) and wage (W ) are determined outside the market. In this model, are the
parameters identi…ed?
8. The data …le card.dat is taken from David Card “Using Geographic Variation in College
Proximity to Estimate the Return to Schooling”in Aspects of Labour Market Behavior (1995).
There are 2215 observations with 29 variables, listed in card.pdf. We want to estimate a wage
equation
2
log(W age) = 0 + 1 Educ + 2 Exper + 3 Exper + 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),
f atheduc (the education, in years, of the father) and motheduc (the education, in years,
of the mother).
(d) Re-estimate the model by e¢ cient 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 overidenti…cation.
(f) Discuss your …ndings.
Chapter 14
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 yt 1 are not independent,
so classical assumptions are not valid.
We can separate time series into two categories: univariate (yt 2 R is scalar); and multivariate
(yt 2 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).
14.1 Stationarity and Ergodicity
De…nition 14.1.1 fyt g is covariance (weakly) stationary if
E(yt ) =
is independent of t; and
cov (yt ; yt k) = (k)
is independent of t for all k: (k) is called the autocovariance function.
(k) = (k)= (0) = corr(yt ; yt k)
is the autocorrelation function.
De…nition 14.1.2 fyt g is strictly stationary if the joint distribution of

(yt ; :::; yt k ) is independent of t for all k:
De…nition 14.1.3 A stationary time series is ergodic if (k) ! 0 as

k ! 1.
221
CHAPTER 14. UNIVARIATE TIME SERIES 222
The following two theorems are essential to the analysis of stationary time series. There proofs
are rather di¢ cult, however.
Theorem 14.1.1 If yt is strictly stationary and ergodic and xt =

f (yt ; yt 1 ; :::) is a random variable, then xt is strictly stationary and er-
godic.
Theorem 14.1.2 (Ergodic Theorem). If yt is strictly stationary and er-

godic and E jyt j < 1; then as T ! 1;
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
^ (k) = (yt ^ ) (yt k ^) :
T
t=1
The sample autocorrelation

^ (k)
^(k) = :
^ (0)
Theorem 14.1.3 If yt is strictly stationary and ergodic and Eyt2 < 1;

then as T ! 1;
p
1. ^ ! E(yt );
p
2. ^ (k) ! (k);
p
3. ^(k) ! (k):
Proof of Theorem 14.1.3. Part (1) is a direct consequence of the Ergodic theorem. For Part
(2), note that
T
1X
^ (k) = (yt ^ ) (yt k ^)
T
t=1
XT T T
1 1X 1X
= yt yt k yt ^ yt k^ + ^2:
T T T
t=1 t=1 t=1
By Theorem 14.1.1 above, the sequence yt yt k is strictly stationary and ergodic, and it has a …nite
mean by the assumption that Eyt2 < 1: Thus an application of the Ergodic Theorem yields
T
1X p
yt yt k ! E(yt yt k ):
T
t=1
Thus
p 2 2 2 2
^ (k) ! E(yt yt k) + = E(yt yt k) = (k):
p
Part (3) follows by the continuous mapping theorem: ^(k) = ^ (k)=^ (0) ! (k)= (0) = (k):
14.2 Autoregressions
In time-series, the series f:::; y1 ; y2 ; :::; yT ; :::g are jointly random. We consider the conditional
expectation
E (yt j Ft 1 )
where Ft 1 = fyt 1 ; yt 2 ; :::g is the past history of the series.
An autoregressive (AR) model speci…es that only a …nite number of past lags matter:
E (yt j Ft 1) = E (yt j yt 1 ; :::; yt k ) :
A linear AR model (the most common type used in practice) speci…es linearity:
E (yt j Ft 1) = + 1 yt 1 + 2 yt 1 + + k yt k :
Letting
et = yt E (yt j Ft 1) ;
then we have the autoregressive model
yt = + 1 yt 1 + 2 yt 1 + + k yt k + et
E (et j Ft 1) = 0:
The last property de…nes a special time-series process.
De…nition 14.2.1 et is a martingale di¤ erence sequence (MDS) if

E (et j Ft 1 ) = 0:
Regression errors are naturally a MDS. Some time-series processes may be a MDS as a conse-
quence 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 di¢ cult 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
Ft 1 : Thus for k > 0; E (yt k et ) = 0:
14.3 Stationarity of AR(1) Process

A mean-zero AR(1) is
yt = yt 1 + et :
Assume that et is iid, E(et ) = 0 and Ee2t = 2 < 1:
By back-substitution, we …nd
2
yt = et + et 1 + et 2 + :::
X1
k
= et k:
k=0
Loosely speaking, this series converges if the sequence ke gets small as k ! 1: This occurs
t k
when j j < 1:
Theorem 14.3.1 If and only if j j < 1 then yt is strictly stationary and

ergodic.
We can compute the moments of yt using the in…nite sum:

1
X
k
Eyt = E (et k) =0
k=0
1
X 2
2k
var(yt ) = var (et k) = 2
:
1
k=0
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 = + Eyt 1;
and solving for Eyt = Eyt 1 we …nd Eyt = =(1 ):
14.4 Lag Operator

An algebraic construct which is useful for the analysis of autoregressive models is the lag oper-
ator.
De…nition 14.4.1 The lag operator L satis…es Lyt = yt 1:
De…ning L2 = LL; we see that L2 yt = Lyt 1 = yt 2: In general, Lk yt = yt k:

The AR(1) model can be written in the format
yt yt 1 = et
or
(1 L) yt = et :
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 14.3.1, an AR(1) is stationary i¤ j j < 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).
14.5 Stationarity of AR(k)

The AR(k) model is
yt = 1 yt 1 + 2 yt 2 + + k yt k + et :
Using the lag operator,
2 k
yt 1 Lyt 2 L yt k L yt = et ;
or
(L)yt = et
where
2 k
(L) = 1 1L 2L kL :
We call (L) the autoregressive polynomial of yt :
The Fundamental Theorem of Algebra says that any polynomial can be factored as
1 1 1
(z) = 1 1 z 1 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 j j denote the modulus of a complex number :
Theorem 14.5.1 The AR(k) is strictly stationary and ergodic if and only
if j j 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.
14.6 Estimation
Let
0
xt = 1 yt 1 yt 2 yt k
0
= 1 2 k :
Then the model can be written as

yt = x0t + et :
The OLS estimator is
^ = X 0X 1
X 0 y:
To study ^ ; it is helpful to de…ne the process ut = xt et : Note that ut is a MDS, since
E (ut j Ft 1) = E (xt et j Ft 1) = xt E (et j Ft 1) = 0:
By Theorem 14.1.1, it is also strictly stationary and ergodic. Thus

T T
1X 1X p
xt et = ut ! E (ut ) = 0: (14.1)
T T
t=1 t=1
The vector xt is strictly stationary and ergodic, and by Theorem 14.1.1, so is xt x0t : Thus by the
Ergodic Theorem,
T
1X p
xt x0t ! E xt x0t = Q:
T
t=1
Combined with (14.1) and the continuous mapping theorem, we see that
T
! 1 T
!
^= 1X 1X p
+ xt x0t xt et !Q 1
0 = 0:
T T
t=1 t=1
We have shown the following:
Theorem 14.6.1 If the AR(k) process yt is strictly stationary and ergodic

p
and Eyt2 < 1; then ^ ! as T ! 1:
14.7 Asymptotic Distribution
Theorem 14.7.1 MDS CLT. If ut is a strictly stationary and ergodic

MDS and E (ut u0t ) = < 1; then as T ! 1;
T
1 X d
p ut ! N (0; ) :
T t=1
Since xt et is a MDS, we can apply Theorem 14.7.1 to see that

T
1 X d
p xt et ! N (0; ) ;
T t=1
where
= E(xt x0t e2t ):
Theorem 14.7.2 If the AR(k) process yt is strictly stationary and ergodic

and Eyt4 < 1; then as T ! 1;
p d
T ^ ! N 0; Q 1
Q 1
:
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.
14.8 Bootstrap for Autoregressions

In the non-parametric bootstrap, we constructed the bootstrap sample by randomly resampling
from the data values fyt ; xt g: This creates an iid bootstrap sample. Clearly, this cannot work in a
time-series application, as this imposes inappropriate independence.
Brie‡y, there are two popular methods to implement bootstrap resampling for time-series data.
Method 1: Model-Based (Parametric) Bootstrap.
1. Estimate ^ and residuals e^t :
2. Fix an initial condition (y k+1 ; y k+2 ; :::; y0 ):
3. Simulate iid draws ei from the empirical distribution of the residuals f^

e1 ; :::; e^T g:
4. Create the bootstrap series yt by the recursive formula
yt = ^ + ^ 1 yt 1 + ^ 2 yt 2 + + ^ k yt k + et :
This construction imposes homoskedasticity on the errors ei ; which may be di¤erent 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 model-
misspeci…cation.
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.
14.9 Trend Stationarity
yt = 0 + 1t + St (14.2)
St = 1 St 1 + 2 St 2 + + k St l + et ; (14.3)
or
yt = 0 + 1t + 1 yt 1 + 2 yt 1 + + k yt k + et : (14.4)
There are two essentially equivalent ways to estimate the autoregressive parameters ( 1 ; :::; k ):
You can estimate (14.4) by OLS.
You can estimate (14.2)-(14.3) sequentially by OLS. That is, …rst estimate (14.2), get the
residual S^t ; and then perform regression (14.3) replacing St with S^t : This procedure is some-
times called Detrending.
The reason why these two procedures are (essentially) the same is the Frisch-Waugh-Lovell
theorem.
Seasonal E¤ects
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 di¤erencing operator. If s is the number of seasons (typically s = 4 or
s = 12);
s yt = yt yt s ;
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.
14.10 Testing for Omitted Serial Correlation

For simplicity, let the null hypothesis be an AR(1):
yt = + yt 1 + ut : (14.5)
We are interested in the question if the error ut is serially correlated. We model this as an AR(1):
ut = ut 1 + et (14.6)
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 (14.5) and (14.6), we take (14.5) and lag the equation once:
yt 1 = + yt 2 + ut 1:
We then multiply this by and subtract from (14.5), to …nd
yt yt 1 = + yt 1 yt 1 + ut ut 1;
or
yt = (1 ) + ( + ) yt 1 yt 2 + 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 coe¢ cient on yt 2 is zero.
An appropriate test of H0 against H1 is therefore a Wald test that the coe¢ cient on yt 2 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 coe¢ cients on yt k 1 ; :::; yt k m are jointly zero. An appropriate test is
the Wald test of this restriction.
14.11 Model Selection

What is the appropriate choice of k in practice? This is a problem of model selection.
One approach to model selection is to choose k based on a Wald tests.
Another is to minimize the AIC or BIC information criterion, e.g.
2k
AIC(k) = log ^ 2 (k) + ;
T
where ^ 2 (k) is the estimated residual variance from an AR(k)
One ambiguity in de…ning 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 best 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 (uni…ed) sample.
14.12 Autoregressive Unit Roots

The AR(k) model is
(L)yt = + et
k
(L) = 1 1L kL :
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 yt 1 + 1 yt 1 + + k 1 yt (k 1) + et : (14.7)
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 (14.7) is
(1 L) 0L 1 (L L2 ) k 1 (L
k 1
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 yt 1 + + k 1 yt (k 1) + et ;
which is an AR(k-1) in the …rst-di¤erence 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):
Since 0 is the parameter of a linear regression, the natural test statistic is the t-statistic for
H0 from OLS estimation of (14.7). 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 signi…cance 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 14.12.1 Dickey-Fuller Theorem.

Assume 0 = 0: As T ! 1;
d
T ^ 0 ! (1 1 2 k 1 ) DF
^0
ADF = ! DFt :
s(^ 0 )
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
conventional rate of T 1=2 : 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
tabulated, and may be used for testing the hypothesis H0 : Note: The standard error s(^ 0 ) is the
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 insu¢ cient 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 + 2t + 0 yt 1 + 1 yt 1 + + k 1 yt (k 1) + et : (14.8)
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 non-
stationary, 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 (14.8).
If a time trend is included, the test procedure is the same, but di¤erent critical values are
required. The ADF test has a di¤erent distribution when the time trend has been included, and a
di¤erent 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 15
Multivariate Time Series
A multivariate time series y t is a vector process m 1. Let Ft 1 = (y t 1 ; y t 2 ; :::) be all lagged

information at time t: The typical goal is to …nd the conditional expectation E (y t j Ft 1 ) : Note
that since y t is a vector, this conditional expectation is also a vector.
15.1 Vector Autoregressions (VARs)

A VAR model speci…es that the conditional mean is a function of only a …nite number of lags:
E (y t j Ft 1) = E yt j yt 1 ; :::; y t k :
A linear VAR speci…es that this conditional mean is linear in the arguments:
E yt j yt 1 ; :::; y t k = a0 + A1 y t 1 + A2 y t 2 + Ak y t k:
Observe that a0 is m 1,and each of A1 through Ak are m m matrices.

De…ning the m 1 regression error
et = y t E (y t j Ft 1) ;
we have the VAR model

y t = a0 + A1 y t 1 + A2 y t 2 + Ak y t k + et
E (et j Ft 1) = 0:
Alternatively, de…ning the mk + 1 vector
0 1
1
B yt C
B 1 C
B C
xt = B y t 2 C
B .. C
@ . A
yt k
and the m (mk + 1) matrix

A= a0 A1 A2 Ak ;
then
y t = Axt + et :
The VAR model is a system of m equations. One way to write this is to let a0j be the jth row
of A. Then the VAR system can be written as the equations
Yjt = a0j xt + ejt :
Unrestricted VARs were introduced to econometrics by Sims (1980).
231
CHAPTER 15. MULTIVARIATE TIME SERIES 232
15.2 Estimation
Consider the moment conditions
E (xt ejt ) = 0;
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
^j = (X 0 X)
a 1
X 0yj :
An alternative way to compute this is as follows. Note that
^0j = y 0j X(X 0 X)
a 1
:
^ we …nd
And if we stack these to create the estimate A;
0 1
y 01
B y0 C
^ B 2 C
A = B . C X(X 0 X) 1
@ .. A
y 0m+1
= Y 0 X(X 0 X) 1
;
where
Y = y1 y2 ym
the T m matrix of the stacked y 0t :
This (system) estimator is known as the SUR (Seemingly Unrelated Regressions) estimator,
and was originally derived by Zellner (1962)
15.3 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 equa-
tions. The GMM framework gives a convenient method to impose such restrictions on estimation.
15.4 Single Equation from a VAR

Often, we are only interested in a single equation out of a VAR system. This takes the form
yjt = a0j xt + et ;
and xt consists of lagged values of yjt and the other ylt0 s: In this case, it is convenient to re-de…ne
the variables. Let yt = yjt ; and z t be the other variables. Let et = ejt and = aj : Then the single
equation takes the form
yt = x0t + et ; (15.1)
and h i
0
xt = 1 yt 1 yt k z 0t 1 z 0t k :
This is just a conventional regression with time series data.

15.5 Testing for Omitted Serial Correlation

Consider the problem of testing for omitted serial correlation in equation (15.1). Suppose that
et is an AR(1). Then
yt = x0t + et
et = et 1 + ut (15.2)
E (ut j Ft 1) = 0:
Then the null and alternative are
H0 : =0 H1 : 6= 0:
Take the equation yt = x0t + et ; and subtract o¤ the equation once lagged multiplied by ; to get
yt yt 1 = x0t + et x0t 1 + et 1
= x0t xt 1 + et et 1;
or
yt = yt 1 + x0t + x0t 1 + ut ; (15.3)
which is a valid regression model.
So testing H0 versus H1 is equivalent to testing for the signi…cance of adding (yt 1 ; xt 1 ) 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 signi…cance 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 = x0t + 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 (15.2) is a special case of (15.3), under the restriction = :
This restriction, which is called a common factor restriction, may be tested if desired. If valid,
the model (15.2) 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 (15.2) is uncommon in recent applications.
15.6 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 testing-
based approach (using, for example, the Wald statistic), or an information criterion approach. The
formula for the AIC and BIC are
p
AIC(k) = log det ^ (k) + 2
T
p log(T )
BIC(k) = log det ^ (k) +
T
XT
^ (k) = 1 et (k)0
^t (k)^
e
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.
15.7 Granger Causality

Partition the data vector into (y t ; z t ): De…ne the two information sets
F1t = yt; yt 1 ; y t 2 ; :::

F2t = yt; zt; yt 1 ; z t 1 ; y t 2 ; z t 2 ; ; :::
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 j F1;t 1) = E (y t j F2;t 1) :
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 de…nition 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 de…nition
of causality was developed by Granger (1969) and Sims (1972).
In a linear VAR, the equation for y t is
yt = + 1yt 1 + + k yt k + z 0t 1 1 + + z 0t k k + et :
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 j Ft 1 ) 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 time-series econo-
metrics, and co-winner in 2003 of the Nobel Memorial Prize in Economic Sciences (along
with Robert Engle). In addition to formalizing the de…nition of causality known as Granger
causality, he invented the concept of cointegration, introduced spectral methods into econo-
metrics, and formalized methods for the combination of forecasts.
15.8 Cointegration
The idea of cointegration is due to Granger (1981), and was articulated in detail by Engle and
Granger (1987).
De…nition 15.8.1 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.
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 0
1) speci…es that the spread (the di¤erence in
returns) is stationary. If y = (log(Consumption) log(Income))0 ; then = (1 1)0 speci…es
that log(Consumption=Income) 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
(Y1t ; Y2t ) and = ( 1 2 ) and normalize 1 = 1: Then ^ 2 = (y 02 y 2 ) 1 y 02 y 1 ! 2 : Furthermore
d
this estimation is super-consistent: T ( ^ 2 ) ! Limit; as …rst shown by Stock (1987). This
2
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
residual z^t = ^ y t ; from OLS of y1t on y2t : Their justi…cation was Stock’s result that ^ is super-
consistent under H1 : Under H0 ; however, ^ is not consistent, so the ADF critical values are not
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 a¤ected by the presence of the time trend. The asymptotic distribution was worked out
in B. Hansen (1992).
15.9 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
yt = yt 1 + D(L) y t 1 + et
where
= A(1)
= I + A1 + A2 + + Ak :
Theorem 15.9.1 Granger Representation Theorem

0
y t is cointegrated with m r if and only if rank( ) = r and =
where is m r, rank ( ) = r:
Thus cointegration imposes a restriction upon the parameters of a VAR. The restricted model
can be written as
0
yt = yt 1 + D(L) y t 1 + et
yt = zt 1 + D(L) y t 1 + et :
If is known, this can be estimated by OLS of y t on z t 1 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 di¢ culty is that is not identi…ed without normalization. When r = 1; we typically just
normalize one element to equal unity. When r > 1; this does not work, and di¤erent authors have
adopted di¤erent identi…cation 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 16
Limited Dependent Variables
A “limited dependent variable”y is one which takes a “limited”set of values. The most common
cases are
Binary: y 2 f0; 1g
Multinomial: y 2 f0; 1; 2; :::; kg
Integer: y 2 f0; 1; 2; :::g
Censored: y 2 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.
16.1 Binary Choice

The dependent variable yi 2 f0; 1g: This represents a Yes/No outcome. Given some regressors
xi ; the goal is to describe Pr (yi = 1 j xi ) ; as this is the full conditional distribution.
The linear probability model speci…es that
Pr (yi = 1 j xi ) = x0i :
As Pr (yi = 1 j xi ) = E (yi j xi ) ; this yields the regression: yi = x0i + ei which can be estimated by
OLS. However, the linear probability model does not impose the restriction that 0 Pr (yi j 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 j xi ) = F x0i
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 + e u) 1 :
237
CHAPTER 16. LIMITED DEPENDENT VARIABLES 238
Normal: F (u) = (u):
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 = x0i + ei
ei F()
1 if yi > 0
yi = :
0 otherwise
For then
Pr (yi = 1 j xi ) = Pr (yi > 0 j xi )

= Pr x0i + ei > 0 j xi
= Pr ei > x0i j xi
= 1 F x0i
= F x0i :
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)1 y
; y = 0; 1:
In the Binary choice model, yi is conditionally Bernoulli with Pr (yi = 1 j xi ) = pi = F (x0i ) : Thus
the conditional density is
f (yi j xi ) = pyi i (1 p i )1 yi
yi
= F x0i (1 F x0i )1 yi
:
Hence the log-likelihood function is

n
X
log L( ) = log f (yi j xi )
i=1
Xn
yi
= log F x0i (1 F x0i )1 yi
i=1
n
X
= yi log F x0i + (1 yi ) log(1 F x0i )
i=1
X X
= log F x0i + log(1 F x0i ):
yi =1 yi =0
The MLE ^ is the value of which maximizes log L( ): Standard errors and test statistics are
computed by asymptotic approximations. Details of such calculations are left to more advanced
courses.
16.2 Count Data

If y 2 f0; 1; 2; :::g; a typical approach is to employ Poisson regression. This model speci…es that
k
exp ( i) i
Pr (yi = k j xi ) = ; k = 0; 1; 2; :::
k!
i = exp(x0i ):
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
n
X n
X
log L( ) = log f (yi j xi ) = exp(x0i ) + yi x0i log(yi !) :
i=1 i=1
The MLE is the value ^ which maximizes log L( ):

Since
E (yi j xi ) = i = exp(x0i )
is the conditional mean, this motivates the label Poisson “regression.”
Also observe that the model implies that
var (yi j xi ) = i = exp(x0i );
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.
16.3 Censored Data

The idea of “censoring” is that some data above or below a threshold are mis-reported at the
threshold. Thus the model is that there is some latent process yi with unbounded support, but we
observe only
yi if yi 0
yi = : (16.1)
0 if yi < 0
(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
yi = x0i + ei
2
ei iid N(0; )
with the observed variable yi generated by the censoring equation (16.1). This model (now called
the Tobit) speci…es 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
estimates E (yi j xi ) ; not E (yi j xi ) = x0i ; and the latter is of interest. Thus OLS will be biased
for the parameter of interest :
[Note: it is still possible to estimate E (yi j xi ) by LS techniques. The Tobit framework postu-
lates that this is not inherently interesting, that the parameter of is de…ned by an alternative
statistical structure.]
Consistent estimation will be achieved by the MLE. To construct the likelihood, observe that
the probability of being censored is
Pr (yi = 0 j xi ) = Pr (yi < 0 j xi )

= Pr x0i + ei < 0 j xi
ei x0i
= Pr < j xi
x0i
= :
The conditional distribution function above zero is Gaussian:

Z y
1 z x0i
Pr (yi = y j xi ) = dz; y > 0:
0
Therefore, the density function can be written as

1(y=0) 1(y>0)
x0i 1 z x0i
f (y j xi ) = ;
where 1 ( ) is the indicator function.

Hence the log-likelihood is a mixture of the probit and the normal:
n
X
log L( ) = log f (yi j xi )
i=1
X x0i X yi x0i
1
= log + log :
yi =0 yi >0
The MLE is the value ^ which maximizes log L( ):
16.4 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 di¤erent from the population
of interest. For example, if you ask for volunteers for an experiment, and they wish to extrapolate
the e¤ects of the experiment on a general population, you should worry that the people who
volunteer may be systematically di¤erent 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 e¤ect of
“training” on the general population, not just on the volunteers.
A simple sample selection model can be written as the latent model
yi = x0i + e1i
Ti = 1 z 0i + 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
e0i 1
N 0; 2 :
e1i
It is presumed that we observe fxi ; z i ; Ti g for all observations.

Under the normality assumption,
e1i = e0i + vi ;
where vi is independent of e0i N(0; 1): A useful fact about the standard normal distribution is
that
(x)
E (e0i j 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 fTi = 1g: However,
E (e1i j Ti = 1; z i ) = E e1i j fe0i > z 0i g; z i

= E e0i j fe0i > z 0i g; z i + E vi j fe0i > z 0i g; z i
= z 0i ;
which is non-zero. Thus

e1i = z 0i + ui ;
where
E (ui j Ti = 1; z i ) = 0:
Hence
yi = x0i + z 0i + ui (16.2)
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 0i ^ ):
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.
The estimator can also work quite poorly if (z 0i ^ ) does not have much in-sample variation.
This can happen if the Probit equation does not “explain”much about the selection choice. Another
potential problem is that if z i = xi ; then (z 0i ^ ) can be highly collinear with xi ; so the second
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
this is valid, it will ensure that (z 0i ^ ) is not collinear with xi ; and hence improve the second stage
estimator’s precision.
Chapter 17
Panel Data
A panel is a set of observations on individuals, collected over time. An observation is the pair
fyit ; xit g; where the i subscript denotes the individual, and the t subscript denotes time. A panel
may be balanced :
fyit ; xit g : t = 1; :::; T ; i = 1; :::; n;
or unbalanced :
fyit ; xit g : For i = 1; :::; n; t = ti ; :::; ti :
17.1 Individual-E¤ects Model

The standard panel data speci…cation is that there is an individual-speci…c e¤ect which enters
linearly in the regression
yit = x0it + 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 (17.1)
If this condition fails, then OLS is inconsistent. (17.1) fails if the individual-speci…c unobserved
e¤ect ui is correlated with the observed explanatory variables xit : This is often believed to be
plausible if ui is an omitted variable.
If (17.1) is true, however, OLS can be improved upon via a GLS technique. In either event,
OLS appears a poor estimation choice.
Condition (17.1) is called the random e¤ ects hypothesis. It is a strong assumption, and most
applied researchers try to avoid its use.
17.2 Fixed E¤ects

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 8
< 1 if i = j
dij = ;
:
0 else
242
CHAPTER 17. PANEL DATA 243
and 0 1
di1
B C
di = @ ... A ;
din
an n 1 dummy vector with a “1” in the i0 th place. Let
0 1
u1
B .. C
u=@ . A:
un
Then note that

ui = d0i u;
and
yit = x0it + d0i u + eit : (17.2)
Observe that
E (eit j xit ; di ) = 0;
so (17.2) is a valid regression, with di as a regressor along with xi :
OLS on (17.2) 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,
^ = 1
X 0 (I P D) X X 0 (I P D) y
1
= X 0X X 0y ;
where
y = y D(D 0 D) 1
D0 y
X = X D(D 0 D) 1
D 0 X:
Since the regression of yit on di is a regression onto individual-speci…c dummies, the predicted value
from these regressions is the individual speci…c mean y i ; and the residual is the demean value
yit = yit yi:
The …xed e¤ects estimator ^ is OLS of yit on xit , the dependent variable and regressors in deviation-
from-mean form.
CHAPTER 17. PANEL DATA 244
Another derivation of the estimator is to take the equation
yit = x0it + ui + eit ;
and then take individual-speci…c means by taking the average for the i0 th individual:
ti ti ti
1 X 1 X 1 X
yit = x0it + ui + eit
Ti t=t Ti t=t Ti t=t
i i i
or
y i = x0i + ui + ei :
Subtracting, we …nd
yit = xit0 + eit ;
which is free of the individual-e¤ect ui :
17.3 Dynamic Panel Regression

A dynamic panel regression has a lagged dependent variable
yit = yit 1 + x0it + ui + eit : (17.3)
This is a model suitable for studying dynamic behavior of individual agents.

Unfortunately, the …xed e¤ects estimator is inconsistent, at least if T is held …nite as n ! 1:
This is because the sample mean of yit 1 is correlated with that of eit :
The standard approach to estimate a dynamic panel is to combine …rst-di¤erencing with IV or
GMM. Taking …rst-di¤erences of (17.3) eliminates the individual-speci…c e¤ect:
yit = yit 1 + x0it + eit : (17.4)
However, if eit is iid, then it will be correlated with yit 1 :

2
E ( yit 1 eit ) = E ((yit 1 yit 2 ) (eit eit 1 )) = E (yit 1 eit 1 ) = e:
So OLS on (17.4) 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 yt 2 is uncorrelated with
eit : Thus values of yit k ; k 2, are valid instruments.
Hence a valid estimator of and is to estimate (17.4) by IV using yt 2 as an instrument for
yt 1 (which is just identi…ed). Alternatively, GMM using yt 2 and yt 3 as instruments (which is
overidenti…ed, 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 di¤erent for each equation. This is
conveniently organized by the GMM principle, as this enables the moments from the di¤erent time-
periods 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 18
Nonparametrics
18.1 Kernel Density Estimation

d
Let X be a random variable with continuous distribution F (x) and density f (x) = dx F (x):
The goal is to estimate
P f (x) from a random sample (X1 ; :::; X n g While F (x) can be estimated by
the EDF F^ (x) = n 1 ni=1 1 (Xi x) ; we cannot de…ne dx d ^
F (x) since F^ (x) is a step function. The
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 speci…c …xed number, and then see how the method generalizes to
estimating the entire function.
De…nition 18.1.1 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

1 u2
K(u) = p exp
2 2
the Epanechnikov
3
4 1 u2 ; juj 1
K(u) =
0 juj > 1
and the Biweight or Quartic
15 2
16 1 u2 ; juj 1
K(u) =
0 juj > 1
In practice, the choice between these three rarely makes a meaningful di¤erence 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
n
1X
f^(x) = Kh (Xi x) :
n
i=1
245
CHAPTER 18. NONPARAMETRICS 246
This estimator is the average of a set of weights. If a large number of the observations Xi are near
x; then the weights are relatively large and f^(x) is larger. Conversely, if only a few Xi are near x;
then the weights are small and f^(x) is small. The bandwidth h controls the meaning of “near”.
Interestingly, f^(x) is a valid density. That is, f^(x) 0 for all x; and
Z Z n n Z n Z
1 1
1X 1X 1
1X 1
f^(x)dx = Kh (Xi x) dx = Kh (Xi x) dx = K (u) du = 1
1 1 n n 1 n 1
i=1 i=1 i=1
where the second-to-last equality makes the change-of-variables u = (Xi x)=h:

We can also calculate the moments of the density f^(x): The mean is
Z n Z
1
1X 1
xf^(x)dx = xKh (Xi x) dx
1 n 1
i=1
n
X 1 Z
1
= (Xi + uh) K (u) du
n 1
i=1
Xn Z n Z
1 1
1X 1
= Xi K (u) du + h uK (u) du
n 1 n 1
i=1 i=1
Xn
1
= Xi
n
i=1
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
1
1X 1
x f^(x)dx =
2
x2 Kh (Xi x) dx
1 n 1
i=1
Xn Z 1
1
= (Xi + uh)2 K (u) du
n 1
i=1
Xn n Z n Z
1 2X 1
1X 2 1
= Xi2 + Xi h K(u)du + h u2 K (u) du
n n 1 n 1
i=1 i=1 i=1
Xn
1
= Xi2 + h2 2
K
n
i=1
where Z 1
2
K = u2 K (u) du
1
is the variance of the kernel. It follows that the variance of the density f^(x) is
Z Z n n
!2
1 1 2
1X 2 1X
x2 f^(x)dx xf^(x)dx = Xi + h2 2
K Xi
1 1 n n
i=1 i=1
2
= ^ + h2 2K
Thus the variance of the estimated density is in‡ated by the factor h2 2

K relative to the sample
moment.
18.2 Asymptotic MSE for Kernel Estimates

For …xed x and bandwidth h observe that
Z 1 Z 1 Z 1
EKh (X x) = Kh (z x) f (z)dz = Kh (uh) f (x + hu)hdu = K (u) f (x + hu)du
1 1 1
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
Z 1
1
EKh (X x) ' K (u) f (x) + f 0 (x)hu + f 00 (x)h2 u2 du
1 2
Z 1 Z 1 Z 1
1
= f (x) K (u) du + f 0 (x)h K (u) udu + f 00 (x)h2 K (u) u2 du
1 1 2 1
1
= f (x) + f 00 (x)h2 2K :
2
The bias of f^(x) is then
n
1X 1
Bias(x) = Ef^(x) f (x) = EKh (Xi x) f (x) = f 00 (x)h2 2
K:
n 2
i=1
We see that the bias of f^(x) at x depends on the second derivative f 00 (x): The sharper the derivative,
the greater the bias. Intuitively, the estimator f^(x) smooths data local to Xi = x; so is estimating
a smoothed version of f (x): The bias results from this smoothing, and is larger the greater the
curvature in f (x):
We now examine the variance of f^(x): Since it is an average of iid random variables, using
…rst-order Taylor approximations and the fact that n 1 is of smaller order than (nh) 1
1
var (x) = var (Kh (Xi x))
n
1 1
= EKh (Xi x)2 (EKh (Xi x))2
n n
Z 1
1 z x 2 1
' K f (z)dz f (x)2
nh2 1 h n
Z 1
1
= K (u)2 f (x + hu) du
nh 1
Z
f (x) 1
' K (u)2 du
nh 1
f (x) R(K)
= :
nh
R1
where R(K) = 1 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)
AM SEh (x) = f 00 (x)2 h4 4
K + :
4 nh
A global measure of precision is the asymptotic mean integrated squared error (AMISE)
Z
h4 4K R(f 00 ) R(K)
AM ISEh = AM SEh (x)dx = + : (18.1)
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 ! 1: That is, h must tend to zero, but at a slower rate
than n 1 :
Equation (18.1) is an asymptotic approximation to the MSE. We de…ne 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)
AM ISEh = h3 4 00
K R(f ) =0
dh nh2
yielding
1=5
R(K) 1=2
h0 = 4 R(f 00 ) n : (18.2)
K
This solution takes the form h0 = cn 1=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(n 1=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 di¢ cult problem, and there is a
large and continuing literature on the subject. The asymptotically optimal choice given in (18.2)
depends on R(K); 2K ; and R(f 00 ): The …rst two are determined by the kernel function. Their
values for the three functions introduced in the previous section are given here.
R
2 = 1 u2 K (u) du
R1
K K 1 R(K) = 1 K (u)2 du
p
Gaussian 1 1/(2 )
Epanechnikov 1=5 1=5
Biweight 1=7 5=7
An obvious di¢ culty 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 Silverman’s Rule-of-Thumb. It
uses formula (18.2) but replaces R(f 00 ) with ^ 5 R( 00 ); where is the N(0; 1) distribution and ^ 2 is
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
p
from normal, the rule-of-thumb h can be quite ine¢ cient. 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.
Gaussian Kernel: hrule = 1:06^ n 1=5
Epanechnikov Kernel: hrule = 2:34^ n 1=5
Biweight (Quartic) Kernel: hrule = 2:78^ n 1=5
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
f^(x): There are other approaches, but implementation can be delicate. I now discuss some of these
choices. The plug-in approach is to estimate R(f 00 ) in a …rst step, and then plug this estimate into
the formula (18.2). This is more treacherous than may …rst appear, as the optimal h for estimation
of the roughness R(f 00 ) is quite di¤erent 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
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
0 1
a1
B a2 C
B C
a=B . C
.
@ . A
ak
Equivalently, a vector a is an element of Euclidean k space, written as a 2 Rk : If k = 1 then a is

a scalar.
A matrix A is a k r rectangular array of numbers, written as
2 3
a11 a12 a1r
6 a21 a22 a2r 7
6 7
A=6 . .. .. 7
4 .. . . 5
ak1 ak2 akr
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,
2 3
1
6 2 7
6 7
A= a1 a2 ar =6 . 7
.
4 . 5
k
where 2 3
a1i
6 a2i 7
6 7
ai = 6 .. 7
4 . 5
aki
are column vectors and
j = aj1 aj2 ajr
250
APPENDIX A. MATRIX ALGEBRA 251
are row vectors.

The transpose of a matrix, denoted A0 ; is obtained by ‡ipping the matrix on its diagonal.
Thus 2 3
a11 a21 ak1
6 a12 a22 ak2 7
6 7
A0 = 6 . . .. 7
4 .. .. . 5
a1r a2r akr
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
2 3
1 0 0
6 0 1 0 7
6 7
Ik = 6 . . .. 7 :
.
4 . . . . 5
0 0 1
A partitioned matrix takes the form

2 3
A11 A12 A1r
6 A21 A22 A2r 7
6 7
A=6 . .. .. 7
4 .. . . 5
Ak1 Ak2 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 de…ne the sum
A + B = (aij + bij ) :
Matrix addition follows the communtative 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 de…ne their product as
Ac = cA = (aij c) :
If a and b are both k 1; then their inner product is

k
X
a0 b = a1 b1 + a2 b2 + + ak bk = aj bj :
j=1
Note that a0 b = b0 a: We say that two vectors a and b are orthogonal if a0 b = 0:

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 de…ned.
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 de…ned as
2 0 3
a1
6 a0 7
6 2 7
AB = 6 . 7 b1 b2 bs
4 .. 5
a0k
2 0 3
a1 b1 a01 b2 a01 bs
6 a0 b1 a0 b2 a02 bs 7
6 2 2 7
= 6 . .. .. 7 :
4 . . . . 5
a0k b1 a0k b2 a0k bs
Matrix multiplication is not communicative: 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
A11 B 11 + A12 B 21 A11 B 12 + A12 B 22

= :
A21 B 11 + A22 B 21 A21 B 12 + A22 B 22
As another example,
2 3
B1
6 B2 7
6 7
AB = A1 A2 Ar 6 .. 7
4 . 5
Br
= A1 B 1 + A2 B 2 + + Ar B r
Xr
= 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
k
X
tr (A) = aii :
i=1
Some straightforward properties for square matrices A and B and real c are
tr (cA) = c tr (A)
tr A0 = tr (A)
tr (A + B) = tr (A) + tr (B)
tr (I k ) = k:
Also, for k r A and r k B we have
tr (AB) = tr (BA) : (A.1)
Indeed,
2 3
a01 b1 a01 b2 a01 bk
6 a02 b1 a02 b2 a02 bk 7
6 7
tr (AB) = tr 6 .. .. .. 7
4 . . . 5
0 0
ak b1 ak b2 a0k bk
k
X
= a0i bi
i=1
Xk
= b0i ai
i=1
= 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 A 1
called the inverse of A which satis…es
1 1
AA =A A = Ik:
For non-singular A and C; some important properties include
1 1
AA = A A = Ik
1 0 0 1
A = A
1 1 1
(AC) = C A
1 1 1 1 1 1
(A + C) = A A +C C
1 1 1 1 1 1
A (A + C) = A A +C A
Also, if A is an orthogonal matrix, then A 1 = A:

Another useful result for non-singular A is known as the Woodbury matrix identity
1 1 1 1 1 1
(A + BCD) =A A BC C + CDA BC CDA : (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
A bb0 =A 1
+ 1 b0 A 1
b A 1
bb0 A 1
: (A.3)
The following fact about inverting partitioned matrices is quite useful.

1
A11 A12 A11 A12 A1112 A1112 A21 A221
= = (A.4)
A21 A22 A21 A22 A2211 A21 A111 A2211
where A11 2 = A11 A12 A221 A21 and A22 1 = A22 A21 A111 A12 : There are alternative algebraic
representations for the components. For example, using the Woodbury matrix identity you can
show the following alternative expressions
A11 = A111 + A111 A12 A2211 A21 A111

A22 = A221 + A221 A21 A1112 A12 A221
A12 = A111 A21 A2211
A21 = A221 A21 A112
1
Even if a matrix A does not possess an inverse, we can still de…ne the Moore-Penrose gen-
eralized inverse A as the matrix which satis…es
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
then
A11 0
A = :
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 de…nition is rarely needed. However, we present
the de…nition 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 de…ned as X
det A = " a1j1 a2j2 akjk :
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)

1 1
det A = (det A)
A B 1
det = (det D) det A BD C if det D 6= 0
C D
det A 6= 0 if and only if A is nonsingular.
Qk
If A is triangular (upper or lower), then det A = i=1 aii
If A is orthogonal, then det A = 1
A.7 Eigenvalues
The characteristic equation of a 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 corre-

sponding 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) = ki=1 i
P
tr(A) = ki=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.
1 1 1 1
The characteristic roots of A are 1 ; 2 ; ..., k :
The matrix H has the orthonormal properties H 0 H = I and HH 0 = I.

1 1
H = H 0 and (H 0 ) =H
A.8 Positive De…niteness

We say that a k k symmetric square matrix A is positive semi-de…nite if for all c 6= 0;
c0 Ac 0: This is written as A 0: We say that A is positive de…nite 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-de…nite. (For any c 6= 0; c0 Ac =

0 0 where = Gc:) If G has full rank, then A is positive de…nite.
1 1
If A is positive de…nite, then A is non-singular and A exists. Furthermore, A > 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 non-

negative diagonal elements. All diagonal elements of are strictly positive if (and only if)
A > 0:
1
If A > 0 then A 1 =H H 0:
If A 0 and rank (A) = r < k then A = H H 0 where A is the Moore-Penrose

generalized inverse, and = diag 1 1 ; 2 1 ; :::; k 1 ; 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 :
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-de…nite. 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 2i = 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
Ik r 0
A=H H0 (A.5)
0 0
with H 0 H = I k . Additionally, tr(A) = rank(A):
A.9 Matrix Calculus

Let x = (x1 ; :::; xk ) be k 1 and g(x) = g(x1 ; :::; xk ) : Rk ! R: The vector derivative is
0 @ 1
@x1 g (x)
@ B .. C
g (x) = @ . A
@x @
@xk g (x)
and
@ @ @
g (x) = @x1 g (x) @xk g (x) :
@x0
Some properties are now summarized.
@ @
@x (a0 x) = @x (x0 a) = a
@
@x0 (Ax) = A
@
@x (x0 Ax) = (A + A0 ) x
@2
@x@x0 (x0 Ax) = A + A0
A.10 Kronecker Products and the Vec Operator

Let A = [a1 a2 an ] be m n: The vec of A; denoted by vec (A) ; is the mn 1 vector
0 1
a1
B a2 C
B C
vec (A) = B . C :
@ .. A
an
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
2 3
a11 B a12 B a1n B
6 a21 B a22 B a2n B 7
6 7
A B=6 .. .. .. 7:
4 . . . 5
am1 B am2 B amn B
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 B0
tr (A B) = tr (A) tr (B)
If A is m m and B is n n; det(A B) = (det (A))n (det (B))m
1 1 1
(A B) =A B
If A > 0 and B > 0 then A B>0
vec (ABC) = (C 0 A) vec (B)
0
tr (ABCD) = vec (D 0 ) (C 0 A) vec (B)
A.11 Vector and Matrix Norms and Inequalities

The Euclidean norm of an m 1 vector a is
1=2
kak = a0 a
m
!1=2
X
= a2i :
i=1
The Euclidean norm of an m n matrix A is

kAk = kvec (A)k
1=2
= tr A0 A
0 11=2
Xm X n
= @ a2ij A :
i=1 j=1
A useful calculation is for any m 1 vectors a and b, using (A.1),

0 1=2 1=2
ab0 = tr ba0 ab = b0 ba0 a = kak kbk
and in particular
aa0 = kak2 (A.6)
Some useful inequalities are now given:
Schwarz Inequality: For any m 1 vectors a and b,

a0 b kak kbk : (A.7)
Schwarz Matrix Inequality: For any m n matrices A and B;
0
AB kAk kBk : (A.8)
Triangle Inequality: For any m n matrices A and B;
kA + Bk kAk + kBk : (A.9)
Trace Inequality. For any m m matrices A and B such that A is symmetric and B 0
tr (AB) max (A) tr (B) (A.10)
where max (A) is the largest eigenvalue of A.
Proof of Schwarz Inequality: First, suppose that kbk = 0: Then b = 0 and both ja0 bj = 0 and
1 0
kak kbk = 0 so the inequality is true. Second, suppose that kbk > 0 and de…ne 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 :
Rearranging, this implies that
2
a0 b a0 a b0 b :
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 de…nition of the matrix Euclidean norm and the Schwarz
inequality
a01 b1 a01 b2
A0 B = a02 b1 a02 b2
.. .. ..
. . .
ka1 k kb1 k ka1 k kb2 k
ka2 k kb1 k ka2 k kb2 k
.. .. ..
. . .
0 11=2
X n X n
= @ kai k2 kbj k2 A
i=1 j=1
n
!1=2 n
!1=2
X X
= kai k2 kbi k2
i=1 i=1
0 11=2 0 11=2
Xn X
m Xn X
m
= @ a2ji A @ kbji k2 A
i=1 j=1 i=1 j=1
= kAk kBk
Proof of Triangle Inequality: Let a = vec (A) and b = vec (B) . Then by the de…nition of the
matrix norm and the Schwarz Inequality
kA + Bk2 = ka + bk2
= a0 a + 2a0 b + b0 b
a0 a + 2 a0 b + b0 b
kak2 + 2 kak kbk + kbk2
= (kak + kbk)2
= (kAk + kBk)2
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. De…ne
C = H 0 BH which has non-negative diagonal elements Cjj since B is positive semi-de…nite. Then
m
X m
X
tr (AB) = tr ( C) = j Cjj max j Cjj = max (A) tr (C)
j
j=1 j=1
where the inequality uses the fact that Cjj 0: But note that
tr (C) = tr H 0 BH = tr HH 0 B = tr (B)
since H is orthonormal. Thus tr (AB) max (A) tr (B) as stated.

Appendix B
Probability
B.1 Foundations
The set S of all possible outcomes of an experiment is called the sample space for the exper-
iment. Take the simple example of tossing a coin. There are two outcomes, heads and tails, so
we can write S = fH; T g: If two coins are tossed in sequence, we can write the four outcomes as
S = fHH; HT; T H; T T g:
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 = fHH; HT g;
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 fHH; HT g and fT Hg are disjoint. Furthermore, if the sets
A1 ; A2 ; ::: are pairwise disjoint and [1i=1 Ai = S; then the collection A1 ; A2 ; ::: is called a partition
of S:
The following are elementary set operations:
Union: A [ B = fx : x 2 A or x 2 Bg:
Intersection: A \ B = fx : x 2 A and x 2 Bg:
Complement: Ac = fx : x 2 = Ag:
The following are useful properties of set operations.
Communtatitivity: 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
DeMorgan’s Laws: (A [ B) = A \ B ; (A \ B) = Ac [ B c :
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 ; 2 B , A 2 B implies Ac 2 B, and
A1 ; A2 ; ::: 2 B implies [1i=1 Ai 2 B. A simple example is f;; Sg 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 de…ne probabilities for events contained in B.
We now can give the axiomatic de…nition of probability. Given S and B, a probability function
Pr satis…es Pr(S) P1= 1; Pr(A) 0 for all A 2 B, and if A1 ; A2 ; ::: 2 B are pairwise disjoint, then
Pr ([1 i=1 A i ) = i=1 Pr(A i ):
Some important properties of the probability function include the following
Pr (;) = 0
Pr(A) 1
Pr (Ac ) = 1 Pr(A)
260
APPENDIX B. PROBABILITY 261
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)
Bonferroni’s Inequality: Pr(A \ B) Pr(A) + Pr(B) 1
Boole’s Inequality: Pr (A [ B) Pr(A) + Pr(B)
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 coe¢ cients which are
de…ned for nonnegative integers n and r; n r; as
n n!
= :
r r! (n 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 With
Replacement Replacement
n!
Ordered (n r)! nr
n n+r 1
Unordered r r
In the lottery example, if counting is unordered and without replacement, the number of po-
tential combinations is 49
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 j B) = :
Pr(B)
For any B; the conditional probability function is a valid probability function where S has been
replaced by B: Rearranging the de…nition, we can write
Pr(A \ B) = Pr (A j 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 j 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
fAi : i 2 Ig; !
\ Y
Pr Ai = Pr (Ai ) :
i2I i2I
Theorem 1 (Bayes’ Rule). For any set B and any partition A1 ; A2 ; ::: of the sample space, then
for each i = 1; 2; :::
Pr (B j Ai ) Pr(Ai )
Pr (Ai j B) = P1
j=1 Pr (B j Aj ) Pr(Aj )
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 de…ne 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! 1 F (x) = 0 and limx!1 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
Pr (X = j ) = j ; j = 1; :::; r (B.3)
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 2 R so the representation (B.3) is unavailable. Instead, we represent the relative
probabilities by the probability density function (PDF)
d
f (x) = F (x)
dx
so that Z x
F (x) = f (u)du
1
and Z b
Pr (a X b) = f (u)du:
a
These expressions only make sense if F (x) is di¤erentiable. While there are examples of continuous
random variables which do not possess a PDF, these cases are unusualRand are typically ignored.
1
A function f (x) is a PDF if and only if f (x) 0 for all x 2 R and 1 f (x)dx:
B.3 Expectation
For any measurable real function g; we de…ne the mean or expectation Eg(X) as follows. If
X is discrete,
Xr
Eg(X) = g( j ) j ;
j=1
and if X is continuous Z 1
Eg(X) = g(x)f (x)dx:
1
The latter is well de…ned and …nite if
Z 1
jg(x)j f (x)dx < 1: (B.4)
1
If (B.4) does not hold, evaluate

Z
I1 = g(x)f (x)dx
g(x)>0
Z
I2 = g(x)f (x)dx
g(x)<0
If I1 = 1 and I2 < 1 then we de…ne Eg(X) = 1: If I1 < 1 and I2 = 1 then we de…ne

Eg(X) = 1: If both I1 = 1 and I2 = 1 then Eg(X) is unde…ned.
Since E (a + bX) = a + bEX; we say that expectation is a linear operator.
For m > 0; we de…ne the m0 th moment of X as EX m and the m0 th central moment as
E (X EX)m :
Two special
p moments are the mean = EX and variance 2 = E (X )2 = EX 2 2 : We
call = 2 the standard deviation of X: We can also write 2 = var(X). For example, this
2
allows the convenient expression var(a + bX) = b 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 jXjm < 1 then
dm
M( ) = E (X m )
d m =0
which is why it is called the moment generating function.
More generally, the characteristic function (CF) of X is
C( ) = E exp (i X)
p
where i = 1 is the imaginary unit. The CF always exists, and when E jXjm < 1
dm
C( ) = im E (X m ) :
d m =0
The Lp norm, p 1; of the random variable X is
kXkp = (E jXjp )1=p :
B.4 Gamma Function

The gamma function is de…ned for > 0 as
Z 1
1
( )= x exp ( x) :
0
It satis…es the property
(1 + ) = ( )
so for positive integers n;
(n) = (n 1)!
Special values include
(1) = 1
and
1
= 1=2 :
2
Sterling’s formula is an expansion for the its logarithm
1 1 1 1 1
log ( ) = log(2 ) + log z+ 3
+ 5
+
2 2 12 360 1260
B.5 Common Distributions

For reference, we now list some important discrete distribution function.
Bernoulli
Pr (X = x) = px (1 p)1 x
; x = 0; 1; 0 p 1
EX = p
var(X) = p(1 p)
Binomial
n x
Pr (X = x) = p (1 p)n x
; x = 0; 1; :::; n; 0 p 1
x
EX = np
var(X) = np(1 p)
Geometric
Pr (X = x) = p(1 p)x 1
; x = 1; 2; :::; 0 p 1
1
EX =
p
1 p
var(X) =
p2
Multinomial
n!
Pr (X1 = x1 ; X2 = x2 ; :::; Xm = xm ) = px1 px2 pxmm ;
x1 !x2 ! xm ! 1 2
x1 + + xm = n;
p1 + + pm = 1
EXi = pi
var(Xi ) = npi (1 pi )
cov (Xi ; Xj ) = npi pj
Negative Binomial
(r + x) r
Pr (X = x) = p (1 p)x 1
; x = 0; 1; 2; :::; 0 p 1
x! (r)
r (1 p)
EX =
p
r (1 p)
var(X) =
p2
Poisson
x
exp ( )
Pr (X = x) = ; x = 0; 1; 2; :::; >0
x!
EX =
var(X) =
We now list some important continuous distributions.

Beta
( + ) 1 1
f (x) = x (1 x) ; 0 x 1; > 0; >0
( ) ( )
=
+
var(X) =
( + + 1) ( + )2
Cauchy
1
f (x) = ; 1<x<1
(1 + x2 )
EX = 1
var(X) = 1
Exponential
1 x
f (x) = exp ; 0 x < 1; >0
EX =
2
var(X) =
Logistic
exp ( x)
f (x) = ; 1 < x < 1;
(1 + exp ( x))2
EX = 0
2
var(X) =
3
Lognormal
!
1 (log x )2
f (x) = p exp 2
; 0 x < 1; >0
2 x 2
2
EX = exp + =2
2 2
var(X) = exp 2 + 2 exp 2 +
Pareto
f (x) = +1
; x < 1; > 0; >0
x
EX = ; >1
1
2
var(X) = ; >2
( 1)2 ( 2)
Uniform
1
f (x) = ; a x b
b a
a+b
EX =
2
(b a)2
var(X) =
12
Weibull
1 x
f (x) = x exp ; 0 x < 1; > 0; >0
1= 1
EX = 1+
2= 2 2 1
var(X) = 1+ 1+
Gamma
1 1 x
f (x) = x exp ; 0 x < 1; > 0; >0
( )
EX =
2
var(X) =
Chi-Square
1 x
f (x) = xr=2 1
exp ; 0 x < 1; r>0
(r=2)2r=2 2
EX = r
var(X) = 2r
Normal
!
1 (x )2 2
f (x) = p exp 2
; 1 < x < 1; 1< < 1; >0
2 2
EX =
2
var(X) =
Student t
r+1
x2 ( r+1
2 )
2
f (x) = p r 1+ ; 1 < x < 1; r>0
r 2
r
EX = 0 if r > 1
r
var(X) = if r > 2
r 2
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
@2
f (x; y) = F (x; y):
@x@y
For a Borel measurable set A 2 R2 ;

Z Z
Pr ((X < Y ) 2 A) = f (x; y)dxdy
A
For any measurable function g(x; y);

Z 1 Z 1
Eg(X; Y ) = g(x; y)f (x; y)dxdy:
1 1
The marginal distribution of X is
FX (x) = Pr(X x)
= lim F (x; y)
y!1
Z x Z 1
= f (x; y)dydx
1 1
so the marginal density of X is

Z 1
d
fX (x) = FX (x) = f (x; y)dy:
dx 1
Similarly, the marginal density of Y is

Z 1
fY (y) = f (x; y)dx:
1
The random variables X and Y are de…ned 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 ) : (B.5)
if the expectations exist. For example, if X and Y are independent then
E(XY ) = EXEY:
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 ))
0 0
= E exp X E exp Y
= MX ( )MY ( ): (B.6)
The covariance between X and Y is
cov(X; Y ) = XY = E ((X EX) (Y EY )) = EXY EXEY:
The correlation between X and Y is

XY
corr (X; Y ) = XY = :
x Y
The Cauchy-Schwarz Inequality implies that
j XY j 1: (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)
@k
f (x) = F (x):
@x1 @xk
For a measurable function g : Rk ! Rs ; we de…ne the expectation

Z
Eg(X) = g(x)f (x)dx
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
var Xi = var (X i )
i=1 i=1
B.7 Conditional Distributions and Expectation

The conditional density of Y given X = x is de…ned as
f (x; y)
fY jX (y j x) =
fX (x)
if fX (x) > 0: One way to derive this expression from the de…nition of conditional probability is
@
fY jX (y j x) = lim Pr (Y y j x X x + ")
@y "!0
@ Pr (fY yg \ fx X x + "g)
= lim
@y "!0 Pr(x X x + ")
@ F (x + "; y) F (x; y)
= lim
@y "!0 FX (x + ") FX (x)
@
@ @x F (x+ "; y)
= lim
@y "!0 fX (x + ")
@2
@x@y F (x; y)
=
fX (x)
f (x; y)
= :
fX (x)
The conditional mean or conditional expectation is the function

Z 1
m(x) = E (Y j X = x) = yfY jX (y j x) dy:
1
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 de…ne the conditional variance of Y given X = x as
2
(x) = var (Y j X = x)
= E (Y m(x))2 j X = x
= E Y2 jX =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 j X) = m(X) and var (Y j X) = 2 (X): For
example, if E (Y j X = x) = + 0 x; then E (Y j X) = + 0 X; a transformation of X:
The following are important facts about conditional expectations.
Simple Law of Iterated Expectations:
E (E (Y j X)) = E (Y ) (B.8)
Proof :
E (E (Y j X)) = E (m(X))
Z 1
= m(x)fX (x)dx
1
Z 1Z 1
= yfY jX (y j x) fX (x)dydx
1 1
Z 1 Z 1
= yf (y; x) dydx
1 1
= E(Y ):
Law of Iterated Expectations:
E (E (Y j X; Z) j X) = E (Y j X) (B.9)
Conditioning Theorem. For any function g(x);

E (g(X)Y j X) = g (X) E (Y j X) (B.10)
Proof : Let
h(x) = E (g(X)Y j X = x)
Z 1
= g(x)yfY jX (y j x) dy
1
Z 1
= g(x) yfY jX (y j x) dy
1
= g(x)m(x)
where m(x) = E (Y j X = x) : Thus h(X) = g(X)m(X), which is the same as E (g(X)Y j X) =
g (X) E (Y j X) :
B.8 Transformations
Suppose that X 2 Rk with continuous distribution function FX (x) and density fX (x): Let
Y = g(X) where g(x) : Rk ! Rk is one-to-one, di¤erentiable, and invertible. Let h(y) denote the
inverse of g(x). The Jacobian is
@
J(y) = det h(y) :
@y 0
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)
= Pr (X h(Y ))
= FX (h(Y )) :
Taking the derivative, the density of Y is
d d
fY (y) = 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
d
fY (y) = fX (h(Y )) h(y):
dy
We can write these two cases jointly as
fY (y) = fX (h(Y )) jJ(y)j : (B.11)
This is known as the change-of-variables formula. This same formula (B.11) holds for k > 1; but
its justi…cation 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 [0,1):
Since fX (x) = 1 for 0 x 1 (B.11) shows that
fY (y) = exp( y); 0 y 1;
an exponential density.
B.9 Normal and Related Distributions

The standard normal density is
1 x2
(x) = p exp ; 1 < x < 1:
2 2
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;
10
and EX = 945:
If Z is standard normal and X = + Z; then using the change-of-variables formula, X has
density !
1 (x )2
f (x) = p exp ; 1 < x < 1:
2 2 2
which is the univariate normal density. The mean and variance of the distribution are and
2 ; and it is conventional to write X N ; 2 .
For x 2 Rk ; the multivariate normal density is
1 (x )0 1
(x )
f (x) = k=2 1=2
exp ; x 2 Rk :
(2 ) det ( ) 2
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 i 0 0
=2 ;
respectively.
If X 2 Rk is multivariate normal and the elements of X are mutually uncorrelated, then
= diagf 2j g is a diagonal matrix. In this case the density function can be written as
!!
2 2 2 2
1 (x1 1) = 1 + + (xk k) = k
f (x) = exp
(2 )k=2 1 k
2
k 2
!
Y 1 xj j
= 1=2
exp 2
j=1 (2 ) j 2 j
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 2r .
Theorem B.9.3 If Z N (0; A) with A > 0; q q; then Z 0 A 1

Z 2:
q
2
p
Theorem B.9.4 Let Z N (0; 1) and Q r be independent. Then Tr = Z= Q=r is distributed
as student’s t with r degrees of freedom.
Proof of Theorem B.9.1. By the change-of-variables formula, the density of Y = a + BX is

!
0 1
1 (y Y) Y (y Y)
f (y) = k=2 1=2
exp ; y 2 Rk :
(2 ) det ( Y ) 2
1=2
where Y = a+B and Y = B B 0 ; where we used the fact that det (B B 0 ) = 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
1
E exp (tQ) = r r=2
xr=2 1 exp (tx) exp ( x=2) dy = (1 2t) r=2
0 2 2
R1
where the second equality uses the fact that 0 y a 1 exp ( by) dy = b a (a); which can be found
by applying change-of-variables to the gamma function. Our goal is to calculate the MGF of
r=2
Q = X 0 X and show that it equals (1 2t) P ; which will establish that Q 2.
r
0 r 2
Note that we can write Q = X X = j=1 Zj where the Zj are independent N (0; 1) : The
distribution of each of the Zj2 is
p
Pr Zj2 y = 2 Pr (0 Zj y)
Z py
1 x2
= 2 p exp dx
0 2 2
Z y
1 s
= 1 1=2
s 1=2 exp ds
0 2 2
2
1 p
using the change–of-variables s = x2 and the fact 2 = : Thus the density of Zj2 is
1 1=2 x
f1 (x) = 1 x exp
2 21=2 2
2 1=2
which is the 1 and by our above calculation has the MGF of E exp tZj2 = (1 2t) :
Pr
Since the Zj2 are mutually independent, (B.6) implies that the MGF of Q = 2
j=1 Zj is
h ir
1=2 r=2 2
(1 2t) = (1 2t) ; which is the MGF of the r density as desired:
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 A 1 = C 10 C 1 and
1 1 10 1
C Z N 0; C AC = N 0; C CC 0 C 10
= N (0; I q ) :
Thus
0
Z 0A 1
Z = Z 0C 10
C 1
Z= C 1
Z C 1
Z 2
q:
Proof of Theorem B.9.4. Using the simple law of iterated expectations, Tr has distribution
function
!
Z
F (x) = Pr p x
Q=r
( r )
Q
= E Z x
r
" r !#
Q
= E Pr Z x jQ
r
r !
Q
= E x
r
Thus its density is
!r
d Q
f (x) = E x
dx r
r !r !
Q Q
= E x
r r
Z r !
1
1 qx2 q 1 r=2 1
= p exp r q exp ( q=2) dq
0 2 2r r 2 2r=2
r+1
x2 ( r+1
2 )
2
= p r 1+
r 2
r
which is that of the student t with r degrees of freedom.
B.10 Inequalities
Jensen’s Inequality
Pm (…nite form). If g( ) : R ! R is convex, then for any non-negative weights
aj such that j=1 aj = 1; and any real numbers xj
0 1
Xm m
X
g@ aj xj A aj g (xj ) : (B.12)
j=1 j=1
In particular, setting aj = 1=m; then

0 1
m
X m
1 1 X
g@ xj A g (xj ) : (B.13)
m m
j=1 j=1
Loève’s cr Inequality. For r > 0;

r
m
X m
X
aj cr jaj jr (B.14)
j=1 j=1
where cr = 1 when r 1 and cr = mr 1 when r 1:
Jensen’s Inequality (probabilistic form). If g( ) : Rm ! R is convex, then for any random

vector x for which E kxk < 1 and E jg (x)j < 1;
g(E(x)) E (g (x)) : (B.15)
Conditional Jensen’s Inequality. If g( ) : Rm ! R is convex, then for any random vectors

(y; x) for which E kyk < 1 and E kg (y)k < 1;
g(E(y j x)) E (g (y) j x) : (B.16)
Conditional Expectation Inequality. For any r such that E jyjr < 1; then
E jE(y j x)jr E jyjr < 1: (B.17)
Expectation Inequality. For any random matrix Y for which E kY k < 1;
kE(Y )k E kY k : (B.18)
Hölder’s Inequality. If p > 1 and q > 1 and p1 + 1q = 1; then for any random m n matrices X
and Y;
E X 0Y (E kXkp )1=p (E kY kq )1=q : (B.19)
Cauchy-Schwarz Inequality. For any random m n matrices X and Y;

1=2 1=2
E X 0Y E kXk2 E kY k2 : (B.20)
Matrix Cauchy-Schwarz Inequality. Tripathi (1999). For any random x 2 Rm and y 2 R` ,
Eyx0 Exx0 Exy 0 Eyy 0 (B.21)
Minkowski’s Inequality. For any random m n matrices X and Y;
(E kX + Y kp )1=p (E kXkp )1=p + (E kY kp )1=p (B.22)
Liapunov’s Inequality. For any random m n matrix X and 1 r p;
(E kXkr )1=r (E kXkp )1=p (B.23)
Markov’s Inequality (standard form). For any random vector x and non-negative function
g(x) 0;
1
Pr(g(x) > ) Eg(x): (B.24)
Markov’s Inequality (strong form). For any random vector x and non-negative function
g(x) 0;
1
Pr(g(x) > ) E (g (x) 1 (g(x) > )) : (B.25)
Chebyshev’s Inequality. For any random variable x;

var (x)
Pr(jx Exj > ) 2
: (B.26)
Proof of Jensen’s Inequality (B.12). By the de…nition of convexity, for any 2 [0; 1]
g ( x1 + (1 ) x2 ) g (x1 ) + (1 ) g (x2 ) : (B.27)
This implies
0 1 0 1
Xm m
X aj
g@ aj xj A = g @a1 g (x1 ) + (1 a1 ) xj A
1 a1
j=1 j=2
0 1
m
X
a1 g (x1 ) + (1 a1 ) g @ bj xj A :
j=2
Pm
where bj = aj =(1 a1 ) and j=2 bj = 1: By another application of (B.27) this is bounded by
0 0 11 0 1
m
X m
X
a1 g (x1 )+(1 a1 ) @b2 g(x2 ) + (1 b2 )g @ cj xj AA = a1 g (x1 )+a2 g(x2 )+(1 a1 ) (1 b2 )g @ cj xj A
j=2 j=2
where cj = bj =(1 b2 ): By repeated application of (B.27) we obtain (B.12).
Proof of Loève’s cr Inequality. For r 1 this is simply a rewriting of the …nite form Jensen’s
Pm
inequality (B.13) with g(u) = ur : For r < 1; de…ne bj = jaj j = j=1 jaj j : The facts that 0 bj 1
r
and r < 1 imply bj bj and thus
m
X Xm
1= bj brj
j=1 j=1
which implies 0 1r
Xm m
X
@ jaj jA jaj jr :
j=1 j=1
The proof is completed by observing that

0 1r 0 1r
Xm m
X
@ aj A @ jaj jA :
j=1 j=1
Proof of Jensen’s Inequality (B.15). 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:
0
Applying expectations, Eg(x) a + b Ex = g(Ex); as stated.
Proof of Conditional Jensen’s Inequality. The same as the proof of (B.15), but using condi-
tional expectations. The conditional expectations exist since E kyk < 1 and E kg (y)k < 1:
Proof of Conditional Expectation Inequality. As the function jujr is convex for r 1, the
Conditional Jensen’s inequality implies
jE(y j x)jr E (jyjr j x) :
Taking unconditional expectations and the law of iterated expectations, we obtain
E jE(y j x)jr EE (jyjr j x) = E jyjr < 1

as required.
Proof of Expectation Inequality. By the Triangle inequality, for 2 [0; 1];
k U 1 + (1 )U 2 k kU 1 k + (1 ) kU 2 k
which shows that the matrix norm g(U ) = kU k is convex. Applying Jensen’s Inequality (B.15) we
…nd (B.18).
1 1
Proof of Hölder’s Inequality. Since p + q = 1 an application of Jensen’s Inequality (B.12)
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

u v
u1=p v 1=q +
p q
and this inequality holds for any u > 0 and v > 0:
Set u = kXkp =E kXkp and v = kY kq =E kY kq : Note that Eu = Ev = 1: By the matrix Schwarz
Inequality (A.8), kX 0 Y k kXk kY k. Thus
E kX 0 Y k E (kXk kY k)
p 1=p q 1=q
(E kXk ) (E kY k ) (E kXkp )1=p (E kY kq )1=q
= E u1=p v 1=q
u v
E+
p q
1 1
= +
p q
= 1;
which is (B.19).
Proof of Cauchy-Schwarz Inequality. Special case of Hölder’s with p = q = 2:
Proof of Matrix Cauchy-Schwarz Inequality. De…ne e = y (Eyx0 ) (Exx0 ) x: Note that

Eee0 0 is positive semi-de…nite. We can calculate that
Eee0 = Eyy 0 Eyx0 Exx0 Exy 0 :
Since the left-hand-side is positive semi-de…nite, so is the right-hand-side, which means Eyy 0
(Eyx0 ) (Exx0 ) Exy 0 as stated.
Proof of Liapunov’s Inequality. The function g(u) = up=r is convex for u > 0 since p r: Set
u = kXkr : By Jensen’s inequality, g (Eu) Eg (u) or
(E kXkr )p=r E (kXkr )p=r = E kXkp :
Raising both sides to the power 1=p yields (E kXkr )1=r (E kXkp )1=p as claimed.
Proof of Minkowski’s Inequality. Note that by rewriting, using the triangle inequality (A.9),
and then Hölder’s Inequality to the two expectations
E kX + Y kp = E kX + Y k kX + Y kp 1
E kXk kX + Y kp 1
+ E kY k kX + Y kp 1
1=q 1=q
(E kXkp )1=p E kX + Y kq(p 1)
+ (E kY kp )1=p E kX + Y kq(p 1)
= (E kXkp )1=p + (E kY kp )1=p E (kX + Y kp )(p 1)=p
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 )(p 1)=p ;
we obtain (B.22).
Proof of Markov’s Inequality. Let F denote the distribution function of x: Then

Z
Pr (g(x) ) = dF (u)
fg(u) g
Z
g(u)
dF (u)
fg(u) g
Z
1
= 1 (g(u) > ) g(u)dF (u)
1
= E (g (x) 1 (g(x) > ))
the inequality using the region of integration fg(u) > g: This establishes the strong form (B.25).
Since 1 (g(x) > ) 1; the …nal expression is less than 1 E (g(x)) ; establishing the standard
form (B.24).
Proof of Chebyshev’s Inequality. De…ne y = (x Ex)2 and note that Ey = var (x) : The events
fjx Exj > g and y > 2 are equal, so by an application Markov’s inequality we …nd
2 2 2
Pr(jx Exj > ) = Pr(y > ) E (y) = var (x)
as stated.
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 j ) where F is a known distribution function and 2 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
fn (y 1 ; :::; y n j ) = f (y i j ) :
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
n
X
log L( ) = log f (y i j ) :
i=1
The likelihood score is the derivative of the log-likelihood, evaluated at the true parameter
value.
@
Si = log f (y i j 0 ) :
@
We also de…ne the Hessian
@2
H= E log f (y i j 0 ) (B.28)
@ @ 0
and the outer product matrix
= E S i S 0i : (B.29)
We now present three important features of the likelihood.
Theorem B.11.1
@
E log f (y j ) =0 (B.30)
@ = 0
ES i = 0 (B.31)
and
H= I (B.32)
The matrix I is called the information, and the equality (B.32) is called the information
matrix equality.
The maximum likelihood estimator (MLE) ^ is the parameter value which maximizes the
likelihood (equivalently, which maximizes the log-likelihood). We can write this as
^ = argmax log L( ): (B.33)
2
In some simple cases, we can …nd an explicit expression for ^ as a function of the data, but these
cases are rare. More typically, the MLE ^ must be found by numerical methods.
To understand why the MLE ^ is a natural estimator for the parameter observe that the
standardized log-likelihood is a sample average and an estimator of E log f (y i j ) :
n
1 1X p
log L( ) = log f (y i j ) ! E log f (y i j ) :
n n
i=1
As the MLE ^ maximizes the left-hand-side, we can see that it is an estimator of the maximizer of
the right-hand-side. The …rst-order condition for the latter problem is
@
0= E log f (y i j )
@
which holds at = 0 by (B.30). This suggests that ^ is an estimator of 0 : In. fact, under
p
conventional regularity conditions, ^ is consistent, ^ ! 0 as n ! 1: Furthermore, we can derive
its asymptotic distribution.
p d
Theorem B.11.2 Under regularity conditions, n ^ 0 ! N 0; I 1
.
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.
Theorem B.11.3 Cramer-Rao Lower Bound. If e is an unbiased reg-

ular estimator of ; then var(e) (nI) :
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
p e 1
n 0 is bounded below by I : In other words, the best possible asymptotic variance among
all (regular) estimators is I 1 : An estimator is called asymptotically e¢ cient if its asymptotic
variance equals this lower bound. Theorem B.11.2 shows that the MLE has this asymptotic variance,
and is thus asymptotically e¢ cient.
Theorem B.11.4 The MLE is asymptotically e¢ cient in the sense that

its asymptotic variance equals the Cramer-Rao Lower Bound.
Theorem B.11.4 gives a strong endorsement for the MLE in parametric models.
Finally, consider functions of parameters. If = g( ) then the MLE of is b = g(b):
This is because maximization (e.g. (B.33)) is una¤ected by parameterization and transformation.
Applying the Delta Method to Theorem B.11.2 we conclude that
p p d
n b ' G0 n b ! N 0; G0 I 1
G (B.34)
where G = @@ g( 0 ): By Theorem B.11.4, b is an asymptotically e¢ cient estimator for . The

asymptotic variance G0 I 1 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
, and the MLE b = g(b) is asymptotically e¢ cient.
Proof of Theorem B.11.1. To see (B.30);

Z
@ @
E log f (y j ) = log f (y j ) f (y j 0 ) dy
@ = 0
@ = 0
Z
@ f (y j 0 )
= f (y j ) dy
@ f (y j ) = 0
Z
@
= f (y j ) dy
@ = 0
@
= 1 = 0:
@ = 0
Equation (B.31) follows by exchanging integration and di¤erentiation
@ @
E log f (y j 0) = E log f (y j 0) = 0:
@ @
Similarly, we can show that !
@2
@ @
f (y j 0 )
0
E = 0:
f (y j 0 )
By direction computation,
@2 @ @ 0
@2 @ @
f (y j 0 )
0
@ f (y j 0) @ f (y j 0)
0 log f (y j 0) = 2
@ @ f (y j 0 ) f (y j 0)
@2
@ @
f (y j 0 )
0 @ @ 0
= log f (y j 0) log f (y j 0) :
f (y j 0 ) @ @
Taking expectations yields (B.32).
Proof of Theorem B.11.2 Taking the …rst-order condition for maximization of log L( ), and
making a …rst-order Taylor series expansion,
@
0 = log L( )
@ =^
n
X @
= log f y i j ^
@
i=1
n
X n
X
@ @2 ^
= log f (y i j 0) + 0 log f (y i j n) 0 ;
@ @ @
i=1 i=1
where n lies on a line segment joining ^ and 0 : (Technically, the speci…c value of n varies by
row in this expansion.) Rewriting this equation, we …nd
n
! 1 n
!
X @2 X
^ 0 = log f (y i j n) Si
0
i=1
@ @ i=1
where S i are the likelihood scores. Since the score S i is mean-zero (B.31) with covariance matrix
(equation B.29) an application of the CLT yields
n
1 X d
p S i ! N (0; ) :
n
i=1
The analysis of the sample Hessian is somewhat more complicated due to the presence of n :
2 p
Let H( ) = @ @@ 0 log f (y i ; ) : If it is continuous in ; then since n ! 0 it follows that
p
H( n) ! H and so
n n
1 X @2 1X @2
0 log f (y i ; n) = 0 log f (y i ; n) H( n) + H( n)
n @ @ n @ @
i=1 i=1
p
!H
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,
p d
n ^ 0 !H 1
N (0; ) = N 0; H 1
H 1
= N 0; I 1
;
the …nal equality using Theorem B.11.1 .
Proof of Theorem B.11.3. Let Y = (y 1 ; :::; y n ) be the sample, and set

n
X
@
S= log fn (Y ; 0) = Si
@
i=1
which by Theorem (B.11.1) has mean zero and variance nI: Write the estimator e = e (Y ) as a
function of the data. Since e is unbiased for any ;
Z
= E = e (Y ) f (Y ; ) dY :
e
Di¤erentiating with respect to and evaluating at 0 yields

Z Z
@ @
I = e (Y ) 0 f (Y ; ) dY = e (Y ) 0 log f (Y ; ) f (Y ; 0 ) dY = E eS 0 = E e 0 S0
@ @
the …nal equality since E (S) = 0

By the matrix Cauchy-Schwarz inequality (B.21), E e 0 S 0 = I; and var (S) = E (SS 0 ) =
nI;
0
var e = E e 0
e 0
0
E e 0 S 0 E SS 0 E S e 0
= E SS 0
= (nI)
as stated.
Appendix C
Numerical Optimization
Many econometric estimators are de…ned by an optimization problem of the form

^ = argmin Q( ) (C.1)
2
where the parameter is 2 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
for given ; but ^ is not available in closed form. In this case, numerical methods are required to
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; :::; Gg. At each point evaluate the criterion function
and …nd the gridpoint which yields the smallest value of the criterion, which is (^ |) where |^ =
argmin0 j G Q( (j)): This value (^ ^
|) is the gridpoint estimate of : If the grid is su¢ ciently …ne to
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 re…ned by a two-step execution. For
an error bound of " pick G so that G2 = (b a)=" For the …rst step de…ne an equally spaced
grid on the region [a; b] with G gridpoints, which is { (j) = a + j(b a)=G : j = 0; :::; Gg:
At each point evaluate the criterion function and let |^ = argmin0 j G Q( (j)). For the second
step de…ne an equally spaced grid on [ (^ | 1); (^ | + 1)] with G gridpoints, which is { 0 (k) =
| 1) + 2k(b a)=G2 : k = 0; :::; Gg: Let k^ = argmin0 k G Q( 0 (k)): The estimate of ^ is
(^
k^ . The advantage of the two-step method over a one-step grid search is that the number of
p
function evaluations has been reduced from (b a)=" to 2 (b a)=" which can be substantial. The
disadvantage is that if the function Q( ) is irregular, the …rst-step grid may not bracket ^ which
thus would be missed.
C.2 Gradient Methods

Gradient Methods are iterative methods which produce a sequence i : i = 1; 2; ::: which
are designed to converge to ^: All require the choice of a starting value 1; and all require the
282
APPENDIX C. NUMERICAL OPTIMIZATION 283
computation of the gradient of Q( )
@
g( ) = Q( )
@
and some require the Hessian
@2
H( ) = 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
Q( + j ") Q( )
gj ( ) ' :
"
Similarly,
Q( + j" + k ")
Q( + k ") Q( + j ") + Q( )
gjk ( ) '
"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 Newton’s method which is based on a quadratic
approximation. By a Taylor’s expansion for close to ^
0 = g(^) ' g( ) + H( ) ^
which implies
^= H( ) 1
g( ):
This suggests the iteration rule
î+1 = i H( i ) 1
g( i ):
where
One problem with Newton’s method is that it will send the iterations in the wrong direction if
H( i ) is not positive de…nite. One modi…cation to prevent this possibility is quadratic hill-climbing
which sets
î+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 de…nite.
Another productive modi…cation is to add a scalar steplength i : In this case the iteration
rule takes the form
i+1 = i Di gi i (C.2)
where g i = g( i ) and D i = H( i ) 1 for Newton’s 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 gi )
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 + D i 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 284
Newton’s 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
gi = gi gi 1
i = i i 1
and : The DFP method sets

0
i i Di 1 g i g 0i D i 1
Di = Di 1 + 0 + :
i gi g 0i D i 1 g i
The BFGS methods sets

0 0 0
i i i i i g 0i D i 1 Di 1 gi i
Di = Di 1 + 0 2 gi0 D i 1 gi + 0 + 0 :
i gi 0
i gi i gi i gi
For any of the gradient methods, the iterations continue until the sequence has converged in
some sense. This can be de…ned by examining whether j i i 1 j ; jQ ( i ) Q ( i 1 )j or jg( i )j
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 de…ned when Q( ) is non-di¤erentiable. 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 Go¤e,
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 prin-
ciple.” 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 dis-
tributed 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 coe¢ cients
in a structural equation,” Econometrica, 25, 77-83.
[12] Bekker, P.A. (1994): “Alternative approximations to the distributions of instrumental vari-
able 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] Breusch, T.S. and A.R. Pagan (1979): “The Lagrange multiplier test and its application to
model speci…cation in econometrics,” Review of Economic Studies, 47, 239-253.
[17] Brown, B. W. and Whitney K. Newey (2002): “GMM, e¢ cient bootstrapping, and improved
inference ,” Journal of Business and Economic Statistics.
285
BIBLIOGRAPHY 286
[18] 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.
[19] Casella, George and Roger L. Berger (2002): Statistical Inference, 2nd Edition, Duxbury
Press.
[20] Chamberlain, Gary (1987): “Asymptotic e¢ ciency in estimation with conditional moment
restrictions,” Journal of Econometrics, 34, 305-334.
[21] Choi, In and Peter C.B. Phillips (1992): “Asymptotic and …nite sample distribution theory for
IV estimators and tests in partially identi…ed structural equations,”Journal of Econometrics,
51, 113-150.
[22] Chow, G.C. (1960): “Tests of equality between sets of coe¢ cients in two linear regressions,”
Econometrica, 28, 591-603.
[23] Cragg, John (1992): “Quasi-Aitken Estimation for Heterskedasticity of Unknown Form"
Journal of Econometrics, 54, 179-201.
[24] Davidson, James (1994): Stochastic Limit Theory: An Introduction for Econometricians.
Oxford: Oxford University Press.
[25] Davison, A.C. and D.V. Hinkley (1997): Bootstrap Methods and their Application. Cambridge
University Press.
[26] 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.
[27] Donald Stephen G. and Whitney K. Newey (2001): “Choosing the number of instruments,”
Econometrica, 69, 1161-1191.
[28] Dufour, J.M. (1997): “Some impossibility theorems in econometrics with applications to
structural and dynamic models,” Econometrica, 65, 1365-1387.
[29] Efron, Bradley (1979): “Bootstrap methods: Another look at the jackknife,” Annals of Sta-
tistics, 7, 1-26.
[30] Efron, Bradley (1982): The Jackknife, the Bootstrap, and Other Resampling Plans. Society
for Industrial and Applied Mathematics.
[31] Efron, Bradley and R.J. Tibshirani (1993): An Introduction to the Bootstrap, New York:
Chapman-Hall.
[32] Eicker, F. (1963): “Asymptotic normality and consistency of the least squares estimators for
families of linear regressions,” Annals of Mathematical Statistics, 34, 447-456.
[33] Engle, Robert F. and Clive W. J. Granger (1987): “Co-integration and error correction:
Representation, estimation and testing,” Econometrica, 55, 251-276.
[34] Frisch, Ragnar (1933): “Editorial,” Econometrica, 1, 1-4.
[35] Frisch, Ragnar and F. Waugh (1933): “Partial time regressions as compared with individual
trends,” Econometrica, 1, 387-401.
[36] Gallant, A. Ronald and D.W. Nychka (1987): “Seminonparametric maximum likelihood es-
timation,” Econometrica, 55, 363-390.
[37] Gallant, A. Ronald and Halbert White (1988): A Uni…ed Theory of Estimation and Inference
for Nonlinear Dynamic Models. New York: Basil Blackwell.
BIBLIOGRAPHY 287
[38] Galton, Francis (1886): “Regression Towards Mediocrity in Hereditary Stature,”The Journal
of the Anthropological Institute of Great Britain and Ireland, 15, 246-263.
[39] Goldberger, Arthur S. (1991): A Course in Econometrics. Cambridge: Harvard University

Press.
[40] Go¤e, W.L., G.D. Ferrier and J. Rogers (1994): “Global optimization of statistical functions
with simulated annealing,” Journal of Econometrics, 60, 65-99.
[41] Gauss, K.F. (1809): “Theoria motus corporum coelestium,” in Werke, Vol. VII, 240-254.
[42] Granger, Clive W. J. (1969): “Investigating causal relations by econometric models and
cross-spectral methods,” Econometrica, 37, 424-438.
[43] Granger, Clive W. J. (1981): “Some properties of time series data and their use in econometric
speci…cation,” Journal of Econometrics, 16, 121-130.
[44] Granger, Clive W. J. and Timo Teräsvirta (1993): Modelling Nonlinear Economic Relation-
ships, Oxford University Press, Oxford.
[45] Gregory, A. and M. Veall (1985): “On formulating Wald tests of nonlinear restrictions,”
Econometrica, 53, 1465-1468,
[46] Haavelmo, T. (1944): “The probability approach in econometrics,” Econometrica, supple-

ment, 12.
[47] Hall, A. R. (2000): “Covariance matrix estimation and the power of the overidentifying
restrictions test,” Econometrica, 68, 1517-1527,
[48] Hall, P. (1992): The Bootstrap and Edgeworth Expansion, New York: Springer-Verlag.
[49] 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.
[50] Hall, P. and J.L. Horowitz (1996): “Bootstrap critical values for tests based on Generalized-
Method-of-Moments estimation,” Econometrica, 64, 891-916.
[51] Hahn, J. (1996): “A note on bootstrapping generalized method of moments estimators,”

Econometric Theory, 12, 187-197.
[52] Hamilton, James D. (1994) Time Series Analysis.
[53] Hansen, Bruce E. (1992): “E¢ cient estimation and testing of cointegrating vectors in the
presence of deterministic trends,” Journal of Econometrics, 53, 87-121.
[54] Hansen, Bruce E. (1996): “Inference when a nuisance parameter is not identi…ed under the
null hypothesis,” Econometrica, 64, 413-430.
[55] Hansen, Bruce E. (2006): “Edgeworth expansions for the Wald and GMM statistics for non-
linear restrictions,” Econometric Theory and Practice: Frontiers of Analysis and Applied
Research, edited by Dean Corbae, Steven N. Durlauf and Bruce E. Hansen. Cambridge Uni-
versity Press.
[56] Hansen, Lars Peter (1982): “Large sample properties of generalized method of moments
estimators, Econometrica, 50, 1029-1054.
[57] 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.
BIBLIOGRAPHY 288
[58] Hausman, J.A. (1978): “Speci…cation tests in econometrics,” Econometrica, 46, 1251-1271.
[59] Heckman, J. (1979): “Sample selection bias as a speci…cation error,” Econometrica, 47, 153-
161.
[60] Horowitz, Joel (2001): “The Bootstrap,” Handbook of Econometrics, Vol. 5, J.J. Heckman
and E.E. Leamer, eds., Elsevier Science, 3159-3228.
[61] Imbens, G.W. (1997): “One step estimators for over-identi…ed generalized method of moments
models,” Review of Economic Studies, 64, 359-383.
[62] Imbens, G.W., R.H. Spady and P. Johnson (1998): “Information theoretic approaches to
inference in moment condition models,” Econometrica, 66, 333-357.
[63] Jarque, C.M. and A.K. Bera (1980): “E¢ cient tests for normality, homoskedasticity and
serial independence of regression residuals, Economic Letters, 6, 255-259.
[64] Johansen, S. (1988): “Statistical analysis of cointegrating vectors,” Journal of Economic

Dynamics and Control, 12, 231-254.
[65] Johansen, S. (1991): “Estimation and hypothesis testing of cointegration vectors in the pres-
ence of linear trend,” Econometrica, 59, 1551-1580.
[66] Johansen, S. (1995): Likelihood-Based Inference in Cointegrated Vector Auto-Regressive Mod-

els, Oxford University Press.
[67] Johansen, S. and K. Juselius (1992): “Testing structural hypotheses in a multivariate cointe-
gration analysis of the PPP and the UIP for the UK,”Journal of Econometrics, 53, 211-244.
[68] Kitamura, Y. (2001): “Asymptotic optimality and empirical likelihood for testing moment
restrictions,” Econometrica, 69, 1661-1672.
[69] Kitamura, Y. and M. Stutzer (1997): “An information-theoretic alternative to generalized

method of moments,” Econometrica, 65, 861-874..
[70] Koenker, Roger (2005): Quantile Regression. Cambridge University Press.
[71] Kunsch, H.R. (1989): “The jackknife and the bootstrap for general stationary observations,”
Annals of Statistics, 17, 1217-1241.
[72] Kwiatkowski, D., P.C.B. Phillips, P. Schmidt, and Y. Shin (1992): “Testing the null hypoth-
esis 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.
[73] Lafontaine, F. and K.J. White (1986): “Obtaining any Wald statistic you want,” Economics
Letters, 21, 35-40.
[74] Lehmann, E.L. and George Casella (1998): Theory of Point Estimation, 2nd Edition,
Springer.
[75] Lehmann, E.L. and Joseph P. Romano (2005): Testing Statistical Hypotheses, 3rd Edition,
Springer.
[76] Li, Qi and Je¤rey Racine (2007) Nonparametric Econometrics.
[77] Lovell, M.C. (1963): “Seasonal adjustment of economic time series,”Journal of the American
Statistical Association, 58, 993-1010.
BIBLIOGRAPHY 289
[78] 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.
[79] MacKinnon, James G. and Halbert White (1985): “Some heteroskedasticity-consistent covari-
ance matrix estimators with improved …nite sample properties,”Journal of Econometrics, 29,
305-325.
[80] Magnus, J. R., and H. Neudecker (1988): Matrix Di¤ erential Calculus with Applications in
Statistics and Econometrics, New York: John Wiley and Sons.
[81] Mann, H.B. and A. Wald (1943). “On stochastic limit and order relationships,” The Annals
of Mathematical Statistics 14, 217–226.
[82] Muirhead, R.J. (1982): Aspects of Multivariate Statistical Theory. New York: Wiley.
[83] Nelder, J. and R. Mead (1965): “A simplex method for function minimization,” Computer
Journal, 7, 308-313.
[84] Newey, Whitney K. (1990): “Semiparametric e¢ ciency bounds,” Journal of Applied Econo-
metrics, 5, 99-135.
[85] Newey, Whitney K. and Daniel L. McFadden (1994): “Large Sample Estimation and Hy-
pothesis Testing,” in Robert Engle and Daniel McFadden, (eds.) Handbook of Econometrics,
vol. IV, 2111-2245, North Holland: Amsterdam.
[86] Newey, Whitney K. and Kenneth D. West (1987): “Hypothesis testing with e¢ cient method
of moments estimation,” International Economic Review, 28, 777-787.
[87] Owen, Art B. (1988): “Empirical likelihood ratio con…dence intervals for a single functional,”
Biometrika, 75, 237-249.
[88] Owen, Art B. (2001): Empirical Likelihood. New York: Chapman & Hall.
[89] Park, Joon Y. and Peter C. B. Phillips (1988): “On the formulation of Wald tests of nonlinear
restrictions,” Econometrica, 56, 1065-1083,
[90] Phillips, Peter C.B. (1989): “Partially identi…ed econometric models,” Econometric Theory,
5, 181-240.
[91] Phillips, Peter C.B. and Sam Ouliaris (1990): “Asymptotic properties of residual based tests
for cointegration,” Econometrica, 58, 165-193.
[92] Politis, D.N. and J.P. Romano (1996): “The stationary bootstrap,”Journal of the American
Statistical Association, 89, 1303-1313.
[93] Potscher, B.M. (1991): “E¤ects of model selection on inference,” Econometric Theory, 7,
163-185.
[94] Qin, J. and J. Lawless (1994): “Empirical likelihood and general estimating equations,” The
Annals of Statistics, 22, 300-325.
[95] Ramsey, J. B. (1969): “Tests for speci…cation errors in classical linear least-squares regression
analysis,” Journal of the Royal Statistical Society, Series B, 31, 350-371.
[96] Rudin, W. (1987): Real and Complex Analysis, 3rd edition. New York: McGraw-Hill.
[97] Said, S.E. and D.A. Dickey (1984): “Testing for unit roots in autoregressive-moving average
models of unknown order,” Biometrika, 71, 599-608.
BIBLIOGRAPHY 290
[98] Shao, J. and D. Tu (1995): The Jackknife and Bootstrap. NY: Springer.
[99] Sargan, J.D. (1958): “The estimation of economic relationships using instrumental variables,”
Econometrica, 2 6, 393-415.
[100] Shao, Jun (2003): Mathematical Statistics, 2nd edition, Springer.
[101] 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.
[102] Shin, Y. (1994): “A residual-based test of the null of cointegration against the alternative of
no cointegration,” Econometric Theory, 10, 91-115.
[103] Silverman, B.W. (1986): Density Estimation for Statistics and Data Analysis. London: Chap-
man and Hall.
[104] Sims, C.A. (1972): “Money, income and causality,”American Economic Review, 62, 540-552.
[105] Sims, C.A. (1980): “Macroeconomics and reality,” Econometrica, 48, 1-48.
[106] Staiger, D. and James H. Stock (1997): “Instrumental variables regression with weak instru-
ments,” Econometrica, 65, 557-586.
[107] Stock, James H. (1987): “Asymptotic properties of least squares estimators of cointegrating
vectors,” Econometrica, 55, 1035-1056.
[108] Stock, James H. (1991): “Con…dence intervals for the largest autoregressive root in U.S.
macroeconomic time series,” Journal of Monetary Economics, 28, 435-460.
[109] Stock, James H. and Jonathan H. Wright (2000): “GMM with weak identi…cation,” Econo-
metrica, 68, 1055-1096.
[110] Theil, H. (1953): “Repeated least squares applied to complete equation systems,”The Hague,
Central Planning Bureau, mimeo.
[111] Theil, H. (1971): Principles of Econometrics, New York: Wiley.
[112] Tobin, James (1958): “Estimation of relationships for limited dependent variables,” Econo-
metrica, 2 6, 24-36.
[113] Tripathi, Gautam (1999): “A matrix extension of the Cauchy-Schwarz inequality,”Economics

Letters, 63, 1-3.
[114] van der Vaart, A.W. (1998): Asymptotic Statistics, Cambridge University Press.
[115] 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.
[116] Wang, J. and E. Zivot (1998): “Inference on structural parameters in instrumental variables
regression with weak instruments,” Econometrica, 66, 1389-1404.
[117] White, Halbert (1980): “A heteroskedasticity-consistent covariance matrix estimator and a

direct test for heteroskedasticity,” Econometrica, 48, 817-838.
[118] White, Halbert (1984): Asymptotic Theory for Econometricians, Academic Press.
[119] Wooldridge, Je¤rey M. (2002) Econometric Analysis of Cross Section and Panel Data, MIT
Press.
BIBLIOGRAPHY 291
[120] Zellner, Arnold. (1962): “An e¢ cient method of estimating seemingly unrelated regressions,
and tests for aggregation bias,”Journal of the American Statistical Association, 57, 348-368.

Econometrics

Uploaded by

Document Information

Original Description:

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

Uploaded by

Copyright:

Available Formats

ECONOMETRICS

This Revision: January 13, 2011

3 Conditional Expectation and Projection 28

3.13 Linear CEF . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4 The Algebra of Least Squares 71

5 Least Squares Regression 91

6 Asymptotic Theory for Least Squares 109

7 Restricted Estimation 136

9 Additional Regression Topics 163

9.5 Least Absolute Deviations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 169

10 The Bootstrap 179

11 Generalized Method of Moments 193

12 Empirical Likelihood 204

14 Univariate Time Series 221

15 Multivariate Time Series 231

16 Limited Dependent Variables 237

17 Panel Data 242

A Matrix Algebra 250

C Numerical Optimization 282

1.1 What is Econometrics?

1.2 The Probability Approach to Econometrics

1.3 Econometric Terms and Notation

The i’th observation is the set (yi ; xi ; z i ):

1.4 Observational Data

1.5 Standard Data Structures

Some contemporary econometric applications combine elements of cross-section, time-series,

In the random sampling framework, we think of an individual observation (yi ; xi ; z i ) as a re-

1.6 Sources for Economic Data

Bureau of Labor Statistics

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

U.S. Bureau of Economic Analysis

International Financial Statistics

The American Economic Review states:

The Journal of Political Economy states:

1.7 Econometric Software

1.8 Reading the Manuscript

2.2 Population and Sample Mean

This is the average value of y in the population.

Assumption 2.2.1 The observations fy1 ; :::; yn g are a

2.3 Sample Mean is Unbiased

De…nition 2.3.1 An estimator b for is unbiased if Eb = .

2.5 Convergence in Probability

De…nition 2.5.1 A random variable zn 2 R converges in probability

lim Pr (jzn zj ) = 1: (2.1)

2.6 Weak Law of Large Numbers

Theorem 2.6.1 Weak Law of Large Numbers (WLLN)

The proof of Theorem 2.6.1 is presented in Section 2.15.

Theorem 2.6.2 Under Assumption 2.2.1 and E jyj < 1; b = y is consis-

Almost Sure Convergence and the Strong Law*

Convergence in probability is sometimes called weak convergence. A related con-

De…nition 2.6.2 A random variable zn 2 R converges almost surely

Pr lim jzn zj = 1: (2.4)

Theorem 2.6.3 Strong Law of Large Numbers (SLLN)

2.7 Vector-Valued Moments

2.8 Convergence in Distribution

De…nition 2.8.1 Let z n be a random vector with distribution Fn (u) = Pr (z n u) : We

Theorem 2.8.1 Central Limit Theorem (CLT). If E kyk2 < 1 then

2.9 Functions of Moments

where g : Rm ! Rk : As one example, the geometric mean of wages w is