This action might not be possible to undo. Are you sure you want to continue?

BooksAudiobooksComicsSheet Music### Categories

### Categories

### Categories

### Publishers

Scribd Selects Books

Hand-picked favorites from

our editors

our editors

Scribd Selects Audiobooks

Hand-picked favorites from

our editors

our editors

Scribd Selects Comics

Hand-picked favorites from

our editors

our editors

Scribd Selects Sheet Music

Hand-picked favorites from

our editors

our editors

Top Books

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Audiobooks

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Comics

What's trending, bestsellers,

award-winners & more

award-winners & more

Top Sheet Music

What's trending, bestsellers,

award-winners & more

award-winners & more

P. 1

Hansen(2006, Econometrics)|Views: 88|Likes: 0

Published by moimemevollbio

See more

See less

https://www.scribd.com/doc/99833049/Hansen-2006-Econometrics

07/11/2012

text

original

- Introduction
- 1.1 Economic Data
- 1.2 Observational Data
- 1.3 Random Sample
- 1.4 Economic Data
- Matrix Algebra
- 2.1 Terminology
- 2.2 Matrix Multiplication
- 2.3 Trace
- 2.4 Inverse
- 2.5 Eigenvalues
- 2.6 Rank and Positive Deﬁniteness
- 2.7 Matrix Calculus
- 2.8 Determinant
- 2.9 Kronecker Products and the Vec Operator
- Regression and Projection
- 3.1 Conditional Mean
- 3.2 Regression Equation
- 3.3 Conditional Variance
- 3.4 Linear Regression
- 3.5 Best Linear Predictor
- 3.6 Exercises
- Least Squares Estimation
- 4.1 Estimation
- 4.2 Least Squares
- 4.3 Normal Regression Model
- 4.4 Model in Matrix Notation
- 4.5 Projection Matrices
- 4.6 Residual Regression
- 4.7 Bias and Variance
- 4.8 Gauss-Markov Theorem
- 4.9 Semiparametric Eﬃciency
- 4.10 Omitted Variables
- 4.11 Multicollinearity
- 4.12 Inﬂuential Observations
- 4.13 Exercises
- Asymptotic Theory
- 5.1 Inequalities
- 5.2 Weak Law of Large Numbers
- 5.3 Convergence in Distribution
- 5.4 Asymptotic Transformations
- Inference
- 6.1 Sampling Distribution
- 6.2 Consistency
- 6.3 Asymptotic Normality
- 6.4 Covariance Matrix Estimation
- 6.5 Consistency of the White Covariance Matrix Estimate
- 6.6 Alternative Covariance Matrix Estimators
- 6.7 Functions of Parameters
- 6.8 t tests
- 6.9 Conﬁdence Intervals
- 6.10 Wald Tests
- 6.11 F Tests
- 6.12 Normal Regression Model
- 6.13 Problems with Tests of NonLinear Hypotheses
- 6.14 Monte Carlo Simulation
- 6.15 Estimating a Wage Equation
- 6.16 Exercises
- Additional Regression Topics
- 7.1 Generalized Least Squares
- 7.2 Testing for Heteroskedasticity
- 7.3 Forecast Intervals
- 7.4 NonLinear Least Squares
- 7.5 Least Absolute Deviations
- 7.6 Quantile Regression
- 7.7 Testing for Omitted NonLinearity
- 7.8 Irrelevant Variables
- 7.9 Model Selection
- 7.10 Exercises
- The Bootstrap
- 8.1 Deﬁnition of the Bootstrap
- 8.2 The Empirical Distribution Function
- 8.3 Nonparametric Bootstrap
- 8.4 Bootstrap Estimation of Bias and Variance
- 8.5 Percentile Intervals
- 8.6 Percentile-t Equal-Tailed Interval
- 8.7 Symmetric Percentile-t Intervals
- 8.8 Asymptotic Expansions
- 8.9 One-Sided Tests
- 8.10 Symmetric Two-Sided Tests
- 8.11 Percentile Conﬁdence Intervals
- 8.12 Bootstrap Methods for Regression Models
- 8.13 Exercises
- Generalized Method of Moments
- 9.1 Overidentiﬁed Linear Model
- 9.2 GMM Estimator
- 9.3 Distribution of GMM Estimator
- 9.4 Estimation of the Eﬃcient Weight Matrix
- 9.5 GMM: The General Case
- 9.6 Over-Identiﬁcation Test
- 9.7 Hypothesis Testing: The Distance Statistic
- 9.8 Conditional Moment Restrictions
- 9.9 Bootstrap GMM Inference
- 9.10 Exercises
- Empirical Likelihood
- 10.1 Non-Parametric Likelihood
- 10.2 Asymptotic Distribution of EL Estimator
- 10.3 Overidentifying Restrictions
- 10.4 Testing
- 10.5 Numerical Computation
- Endogeneity
- 11.1 Instrumental Variables
- 11.2 Reduced Form
- 11.3 Identiﬁcation
- 11.4 Estimation
- 11.5 Special Cases: IV and 2SLS
- 11.6 Bekker Asymptotics
- 11.7 Identiﬁcation Failure
- 11.8 Exercises
- Univariate Time Series
- 12.1 Stationarity and Ergodicity
- 12.2 Autoregressions
- 12.3 Stationarity of AR(1) Process
- 12.4 Lag Operator
- 12.5 Stationarity of AR(k)
- 12.6 Estimation
- 12.7 Asymptotic Distribution
- 12.8 Bootstrap for Autoregressions
- 12.9 Trend Stationarity
- 12.10 Testing for Omitted Serial Correlation
- 12.11 Model Selection
- 12.12 Autoregressive Unit Roots
- Multivariate Time Series
- 13.1 Vector Autoregressions (VARs)
- 13.2 Estimation
- 13.3 Restricted VARs
- 13.4 Single Equation from a VAR
- 13.5 Testing for Omitted Serial Correlation
- 13.6 Selection of Lag Length in an VAR
- 13.7 Granger Causality
- 13.8 Cointegration
- 13.9 Cointegrated VARs
- Limited Dependent Variables
- 14.1 Binary Choice
- 14.2 Count Data
- 14.3 Censored Data
- 14.4 Sample Selection
- Panel Data
- 15.1 Individual-Eﬀects Model
- 15.2 Fixed Eﬀects
- 15.3 Dynamic Panel Regression
- Nonparametrics
- 16.1 Kernel Density Estimation
- 16.2 Asymptotic MSE for Kernel Estimates
- Probability
- A.1 Foundations
- A.2 Random Variables
- A.3 Expectation
- A.4 Common Distributions
- A.5 Multivariate Random Variables
- A.6 Conditional Distributions and Expectation
- A.7 Transformations
- A.8 Normal and Related Distributions
- A.9 Maximum Likelihood
- Numerical Optimization
- B.1 Grid Search
- B.2 Gradient Methods
- B.3 Derivative-Free Methods

Bruce E. Hansen c °2000, 20061 University of Wisconsin www.ssc.wisc.edu/~bhansen Revised: January 2006 Comments Welcome

This manuscript may be printed and reproduced for individual or instructional use, but may not be printed for commercial purposes.

1

Contents

1 Introduction 1.1 Economic Data . . 1.2 Observational Data 1.3 Random Sample . 1.4 Economic Data . . 1 1 1 2 2 4 4 5 6 7 8 8 9 9 10 11 11 12 14 15 15 20 21 21 22 25 26 27 28 30 32 33 34 35

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

2 Matrix Algebra 2.1 Terminology . . . . . . . . . . . . . . . . . 2.2 Matrix Multiplication . . . . . . . . . . . 2.3 Trace . . . . . . . . . . . . . . . . . . . . . 2.4 Inverse . . . . . . . . . . . . . . . . . . . . 2.5 Eigenvalues . . . . . . . . . . . . . . . . . 2.6 Rank and Positive Deﬁniteness . . . . . . 2.7 Matrix Calculus . . . . . . . . . . . . . . . 2.8 Determinant . . . . . . . . . . . . . . . . . 2.9 Kronecker Products and the Vec Operator 3 Regression and Projection 3.1 Conditional Mean . . . . . 3.2 Regression Equation . . . 3.3 Conditional Variance . . . 3.4 Linear Regression . . . . . 3.5 Best Linear Predictor . . 3.6 Exercises . . . . . . . . . 4 Least Squares Estimation 4.1 Estimation . . . . . . . . 4.2 Least Squares . . . . . . . 4.3 Normal Regression Model 4.4 Model in Matrix Notation 4.5 Projection Matrices . . . . 4.6 Residual Regression . . . 4.7 Bias and Variance . . . . 4.8 Gauss-Markov Theorem . 4.9 Semiparametric Eﬃciency 4.10 Omitted Variables . . . . 4.11 Multicollinearity . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

. . . . . . . . . . .

i

4.12 Inﬂuential Observations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.13 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Asymptotic Theory 5.1 Inequalities . . . . . . . . . . 5.2 Weak Law of Large Numbers 5.3 Convergence in Distribution . 5.4 Asymptotic Transformations .

36 38 41 41 42 43 45 46 46 47 48 50 52 53 54 55 56 57 58 60 62 65 67 70 74 74 77 77 78 80 83 84 85 86 89 92 92 93 94 94 96 97 98 99

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

6 Inference 6.1 Sampling Distribution . . . . . . . . . . . . . . . . . . 6.2 Consistency . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Asymptotic Normality . . . . . . . . . . . . . . . . . . 6.4 Covariance Matrix Estimation . . . . . . . . . . . . . . 6.5 Consistency of the White Covariance Matrix Estimate 6.6 Alternative Covariance Matrix Estimators . . . . . . . 6.7 Functions of Parameters . . . . . . . . . . . . . . . . . 6.8 t tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.9 Conﬁdence Intervals . . . . . . . . . . . . . . . . . . . 6.10 Wald Tests . . . . . . . . . . . . . . . . . . . . . . . . 6.11 F Tests . . . . . . . . . . . . . . . . . . . . . . . . . . 6.12 Normal Regression Model . . . . . . . . . . . . . . . . 6.13 Problems with Tests of NonLinear Hypotheses . . . . 6.14 Monte Carlo Simulation . . . . . . . . . . . . . . . . . 6.15 Estimating a Wage Equation . . . . . . . . . . . . . . 6.16 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . 7 Additional Regression Topics 7.1 Generalized Least Squares . . . . . 7.2 Testing for Heteroskedasticity . . . 7.3 Forecast Intervals . . . . . . . . . . 7.4 NonLinear Least Squares . . . . . 7.5 Least Absolute Deviations . . . . . 7.6 Quantile Regression . . . . . . . . 7.7 Testing for Omitted NonLinearity . 7.8 Irrelevant Variables . . . . . . . . . 7.9 Model Selection . . . . . . . . . . . 7.10 Exercises . . . . . . . . . . . . . . 8 The 8.1 8.2 8.3 8.4 8.5 8.6 8.7 8.8

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

. . . . . . . . . .

Bootstrap Deﬁnition of the Bootstrap . . . . . . . . . The Empirical Distribution Function . . . . Nonparametric Bootstrap . . . . . . . . . . Bootstrap Estimation of Bias and Variance Percentile Intervals . . . . . . . . . . . . . . Percentile-t Equal-Tailed Interval . . . . . . Symmetric Percentile-t Intervals . . . . . . Asymptotic Expansions . . . . . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

. . . . . . . .

ii

. . 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9. . . . . . .5 GMM: The General Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11. . . . . . . . . . . .11 8. . . . . . . .5 Numerical Computation . . . . . . .8 Conditional Moment Restrictions . .13 One-Sided Tests . . . 12. . . . . . . . .2 Autoregressions . . . . . . . . . . 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12. . . . . . . . . . . . . . . . . . . . . . . . . .7 Hypothesis Testing: The Distance Statistic 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9. . . . . . . . . . . . .4 Estimation of the Eﬃcient Weight Matrix .8 Exercises . . . . . . . . . . . . . . . . . . 100 101 102 103 104 106 106 107 107 108 109 110 111 112 113 114 117 117 119 120 121 121 124 125 126 127 127 128 129 130 133 135 135 137 137 138 138 139 140 140 141 142 142 143 9 Generalized Method of Moments 9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12. .2 Reduced Form . . . 12.2 Asymptotic Distribution of EL Estimator 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Non-Parametric Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . .10 Exercises . iii . . . . . . . . . . . . . . . 9. . . . 12. . . . . . . . . . . . .5 Stationarity of AR(k) . 9. . . 12. . . . . . . . . . . . 10. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Distribution of GMM Estimator . . . . 9. . . . . . . . . . . . . . . . . . . . Symmetric Two-Sided Tests . . . . . . . . . . . . . . . . . . . . 12 Univariate Time Series 12. . . . . . . . . .1 Overidentiﬁed Linear Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .9 Bootstrap GMM Inference . . . . . . . . . . . . . . . . . . . .10Testing for Omitted Serial Correlation 12. . . .3 Overidentifying Restrictions . . . . . . . . . . . . . . . . . . .10 8. . . . . .1 Instrumental Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 12. . .2 GMM Estimator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Estimation . .3 Identiﬁcation . . . . . . . . . . 12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7 Asymptotic Distribution . . . . . . . . . . . . . . . . . . . . . . . . . .1 Stationarity and Ergodicity . . . . . . . . . . . . . . . . . . . . . . .5 Special Cases: IV and 2SLS 11. . . . . 10. . . . . . . . . . . . . .6 Estimation . . . . . . . . . . . 12. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .4 Lag Operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .12 8. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12. . . . . . . 10. . . . . . . . . . . . . . . . . . . . . . . 11. . . . . . . . . . . . . . . . . . . . . Bootstrap Methods for Regression Models Exercises . . . . . . . . . . . . . . . . . . . .8.12Autoregressive Unit Roots .6 Over-Identiﬁcation Test . . . . . . . . . . . . . . . . . . . . . . . . . . .9 8. . . . . . . 9. . . . . . . . . . . . . . . . . . . . 10 Empirical Likelihood 10. . . . . . . . . . . . . . .7 Identiﬁcation Failure . . . . . . . . . . . . . . . . . . . . . . .8 Bootstrap for Autoregressions . . . . . .6 Bekker Asymptotics . . . . . . . . . . . .9 Trend Stationarity . . . . . 11. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11. . . . . . . . . . . . . . . . . . . . 11. . . . . . . . . . . . . . . . . 11. . 11 Endogeneity 11.11Model Selection . . . .3 Stationarity of AR(1) Process . . . . . . . . . . . Percentile Conﬁdence Intervals . . . . . . . . .

. . . . . . . 160 16. . . . . . . . . . . . . . . . . . . . . . . . . .8 Cointegration . . . . . . . . 13. . . . . . . . . . . . . . . . . . . .5 Testing for Omitted Serial Correlation 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13. . . . . . . . . . . . . . . .6 Selection of Lag Length in an VAR . . . . . . . . . . . . . . . . . . . . . . . .2 Gradient Methods . . . . 13. . . 14. . . . . . . . . . . 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .1 Grid Search . . . . . . . . . . . . . . . . . . . . . . . . . . .5 Multivariate Random Variables . . . . . . . . . . . . . . . . . . . . . . iv . . . . . . . . . . . . . A. . . .7 Transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .2 Fixed Eﬀects . . . . . . . . . A. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. . . . . . . . . . . 162 A Probability A. . . . . . . . . . . A. . . . . . . . . . . . . . . .6 Conditional Distributions and Expectation . . 13. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B Numerical Optimization B. . . . . .4 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . A. . . . . . 157 15. . . . . . . . . . . . . . . 13. . . . .4 Common Distributions .1 Vector Autoregressions (VARs) . . .2 Asymptotic MSE for Kernel Estimates . . . . . . . . . . . . . . . . . . .1 Kernel Density Estimation . . . . . . . . . . . . . . . . .2 Estimation . . . . . . .8 Normal and Related Distributions . . . . . . . . . . . . . . . . . . . . . . B. . . . . . . . . . . . . . . . . . . . . . .4 Single Equation from a VAR . . . . . . A. . . . . . . . .13 Multivariate Time Series 13. . . . . . . . . . . . . . . . . . .3 Derivative-Free Methods . . . . . . . . . . . . . . . . . .9 Maximum Likelihood . . . . . . . . . . . . . . . . 14. . . . . A. . . . . . . . . . . . . .1 Individual-Eﬀects Model . . . . . . . . . . . . . . 165 165 167 168 169 171 173 175 175 178 183 183 184 185 . . . .2 Random Variables .3 Censored Data . .3 Dynamic Panel Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .3 Restricted VARs . . . . . . . . . . . . . . . 14 Limited Dependent Variables 14. . . . . . . 145 145 146 146 147 147 148 148 149 149 151 151 153 153 154 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .7 Granger Causality . . . . .2 Count Data . . . . A. . . . . . . . . . . . . . . . . . . . . . 14. 15 Panel Data 157 15. . . . . . . . . . . . . . 157 15.1 Foundations . . . . . . .1 Binary Choice . . . . . . 159 16 Nonparametrics 160 16. . . . .9 Cointegrated VARs . . . B. . . . . . . . . . . . . . . . . . .3 Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13.

most economic data is observational. Surveys are a typical source for cross-sectional data. 1 . However. even immoral! Instead. and panel. To measure the returns to schooling. These data sets consist surveys of a set of individuals. Panel data combines elements of cross-section and time-series. Econometric theory concerns the study and development of tools and methods for applied econometric applications. and assess the joint dependence. We can measure the joint distribution of these variables. For example.2 Observational Data A common econometric question is to quantify the impact of one set of variables on another variable. households. or corporations. The individuals surveyed may be persons. The diﬀerences between the groups could be attributed to the diﬀerent levels of education. a concern in labor economics is the returns to schooling — the change in earnings induced by increasing a worker’s education. an experiment might randomly divide children into groups. household or corporation) is surveyed on multiple occasions. holding other variables constant. Time-series data is indexed by time. There are three major types of economic data sets: cross-sectional. But we cannot infer causality. 1. as we are not able to manipulate one variable to see the direct eﬀect on the other. This type of data is characterized by serial dependence. timeseries.1 Economic Data An econometric study requires data for analysis. Cross-sectional data sets are characterized by mutually independent observations. prices and interest rates. experiments such as this are infeasible. and then follow the children’s wage path as they mature and enter the labor force. Ideally. To continue the above example. The quality of the study will be largely determined by the data available. what we observe (through data collection) is the level of a person’s education and their wage. Each individual (person. Applied econometrics concerns the application of these tools to economic data. Typical examples include macroeconomic aggregates. we would use experimental data to answer these questions. Another issue of interest is the earnings gap between men and women. mandate diﬀerent levels of education to the diﬀerent groups. repeated over time.Chapter 1 Introduction Econometrics is the study of estimation and inference for economic models using economic data. They are distinguished by the dependence structure across observations. 1.

an econometrician has data {(y1 . High ability individuals do better in school. a person’s level of education is (at least partially) determined by that person’s choices and their achievement in education. . Some other excellent data sources are listed below. . We call these observations the sample. We will return to a discussion of some of these issues in Chapter 11. Rather it means that the pair (yi . Causal inference requires identiﬁcation. Bureau of Labor Statistics: http://www. available at http://rfe.stlouisfed.census. This discussion means that causality cannot be infered from observational data alone. or iid. This is an alternative explanation for an observed positive correlation between educational levels and wages. xn )} = {(yi . the development of the internet has provided a convenient forum for dissemination of economic data.. x2 ) . taking on only the values 0 and 1. In this case the data are independent and identically distributed. The point is that multiple explanations are consistent with a positive correlation between schooling levels and education. Sometimes the independent part of the label iid is misconstrued. and their high ability is the fundamental reason for their high wages. xi ) have a distribution F which we call the population.gov/releases/ National Bureau of Economic Research: http://www. x1 ) .edu/Data/index. If the data is cross-sectional (each observation is a diﬀerent individual) it is often reasonable to assume they are mutually independent.. household or ﬁrm). xi ) : i = 1. many regressors in econometric practice are binary.. (yi .gov/ Federal Reserve Bank of St..federalreserve. (yn . .nber. and this is based on strong assumptions. 1. and therefore choose to attain higher levels of education.org/fred2/ Board of Governors of the Federal Reserve System: http://www. and the goal of statistical inference is to learn about features of F from the sample..For example. At this point in our analysis it is unimportant whether the observations yi and xi come from continuous or discrete distributions. The distribution F is unknown. Knowledge of the joint distibution cannot distinguish between these explanations.html. We call this a random sample. n} where each pair {yi ... xj ) for i 6= j.g. It is not a statement about the relationship between yi and xi .4 Economic Data Fortunately for economists. This abstraction can a source of confusion as it does not correspond to a physical population in the real world. xi } ∈ R × Rk is an observation on an individual (e. and are typically called dummy variables..gov/econ/www/ 2 . Many large-scale economic datasets are available without charge from governmental agencies.. This “population” is inﬁnitely large. xi ) . (y2 .bls. These factors are likely to be aﬀected by their personal abilities and attitudes towards work. it is reasonable to model each observation as a random draw from the same probability distribution. The random variables (yi . An excellent starting point is the Resources for Economists Data Links. If the data is randomly gathered.org/ US Census: http://www. 1. For example. The fact that a person is highly educated suggests a high level of ability.. Louis: http://research.3 Random Sample Typically.wustl. xi ) is independent of the pair (yj .

apdi.bea. Bureau of Economic Analysis: http://www.doc.sipp.census.gov/sipp/ Panel Study of Income Dynamics (PSID): http://psidonline.net/imf/ 3 .census.edu/ U.isr.compustat.umich.gov/ CompuStat: http://www.bls.S.htm Survey of Income and Program Participation (SIPP): http://www.com/www/ International Financial Statistics (IFS): http://ifs.gov/cps/cpsmain.Current Population Survey (CPS): http://www.

A vector a is a k × 1 list of numbers. If r = 1 or k = 1 then A is a vector. a vector a is an element of Euclidean k space. . We write this as ⎞ ⎛ a1 ⎜ a2 ⎟ ⎟ ⎜ a=⎜ . . ⎥ ⎣ . then A is a scalar. ⎠ . A matrix can be written as a set of column vectors or as a set of row vectors. . 2. If r = k = 1. That is.1 Terminology Equivalently. typically arranged in a column. If k = 1 then a is a scalar. . .Chapter 2 Matrix Algebra This chapter reviews the essential components of matrix algebra. α0 k where ⎢ ⎢ ai = ⎢ ⎣ 4 ⎡ a1i a2i . written as ⎤ ⎡ a11 a12 · · · a1r ⎢ a21 a22 · · · a2r ⎥ ⎥ ⎢ A=⎢ . . ⎦ . ak . ak1 ak2 · · · akr By convention aij refers to the i0 th row and j 0 th column of A. . ⎟ ⎝ . ⎥ = [aij ] . . ⎡ 0 ⎤ α1 ⎢ α0 ⎥ ¤ ⎢ 2 ⎥ £ A = a1 a2 · · · ar = ⎢ . A matrix A is a k × r rectangular array of numbers. ⎦ ⎣ . hence a ∈ Rk . aki ⎤ ⎥ ⎥ ⎥ ⎦ A scalar a is a single number.

is obtained by ﬂipping the matrix on its diagonal. then a0 is a 1 × k row vector. are conformable. . . . which implies aij = aji . A square matrix is diagonal if the only non-zero elements appear on the diagonal. . so that aij = 0 if i 6= j. .are column vectors and α0 = j are row vectors. . and this is the only case where this product is deﬁned. . Then ⎡ 0 ⎤ a1 ⎢ a0 ⎥ £ ¤ ⎢ 2 ⎥ AB = ⎢ . ⎦ . vectors and/or scalars. ⎤ ⎡ a11 a21 · · · ak1 ⎢ a12 a22 · · · ak2 ⎥ ⎥ ⎢ B = A0 = ⎢ . then we deﬁne their product AB by writing A as a set of row vectors and B as a set of column vectors (each of length r). ⎥ . or the product AB. a1r a2r · · · akr £ aj1 aj2 · · · ajr ¤ Thus bij = aji for all i and j. . a0 k ⎡ 0 ⎤ a1 b1 a0 b2 · · · a0 bs 1 1 ⎢ a0 b1 a0 b2 · · · a0 bs ⎥ 2 2 ⎢ 2 ⎥ = ⎢ . then their inner product is a0 b = a1 b1 + a2 b2 + · · · + ak bk = aj bj When the number of columns of A equals the number of rows of B. Note that if A is k × r. ⎦ ⎣ . The transpose of a matrix. then A0 is r × k. . denoted B = A0 . . . ⎥ . . A square matrix is symmetric if A = A0 . . 5 Note that a0 b = b0 a. If a is a k × 1 vector. . ⎦ ⎣ . Ak1 Ak2 · · · Akr where the Aij denote matrices. 2. A partitioned matrix takes the form ⎤ ⎡ A11 A12 · · · A1r ⎢ A21 A22 · · · A2r ⎥ ⎥ ⎢ A=⎢ .2 Matrix Multiplication k X j=1 If a and b are both k × 1. ⎥ . we say that A and B. A matrix is square if k = r. a0 b1 a0 b2 · · · k k a0 bs k . . . If A is k × r and B is r × s. . ⎦ ⎣ . . A square matrix is upper (lower) diagonal if all elements below (above) the diagonal equal zero. ⎥ b1 b2 · · · bs ⎣ .

. for k × r A and r × k B we have tr (AB) = tr (BA) tr (A + B) = tr (A) + tr (B) 6 . We say that two vectors a and b are orthogonal if a0 b = 0.An alternative way to write the matrix product is ∙ ¸∙ A11 A12 B11 AB = A21 A22 B21 ∙ A11 B11 + A12 B21 = A21 B11 + A22 B21 and £ to use matrix partitions. 0 0 ··· 1 2. The columns of a k × r matrix A. For example. ⎥ . r ≤ k. . ¸ B12 B22 ¸ A11 B12 + A12 B22 A21 B12 + A22 B22 ⎡ B1 B2 . A square matrix A is called orthogonal if A0 A = Ik . Also. An important diagonal matrix is the identity matrix. . . which has ones on the diagonal. are said to be orthogonal if A0 A = Ir . then AIr = A and Ik A = A. ⎤ ⎥ ⎥ ⎥ ⎦ AB = A1 A2 · · · Ar Br = A1 B1 + A2 B2 + · · · + Ar Br r X = Ar Br j=1 ¤⎢ ⎢ ⎢ ⎣ Important properties are that if A is k × r. . . . A k × k identity matrix is denoted as ⎡ ⎤ 1 0 ··· 0 ⎢ 0 1 ··· 0 ⎥ ⎢ ⎥ Ik = ⎢ . .3 Trace The trace of a k × k square matrix A is the sum of its diagonal elements tr (A) = k X i=1 aii Some straightforward properties for square matrices A and B are tr (cA) = c tr (A) ¡ ¢ tr A0 = tr (A) tr (Ik ) = k. ⎦ ⎣ .

or is nonsingular. If A − BD−1 C and D − CA−1 B are non-singular. The Moore-Penrose generalized inverse A− satisﬁes (2. a0 bk k ⎤ ⎥ ⎥ ⎥ ⎦ a0 bi i b0 ai i = = tr (BA) . then A−1 = A.1) . 2. . . .3) plus the following three conditions A− AA− = A− AA− is symmetric A− A is symmetric For any matrix A. if A is an orthogonal matrix.3) The matrix A− is not necessarily unique. . the Moore-Penrose generalized inverse A− exists and is unique. 7 Also. The following fact about inverting partitioned matrices is sometimes useful.2) C D D − CA−1 B − D − CA−1 B CA−1 (AC)−1 = C −1 A−1 ¡ ¢−1 −1 (A + C)−1 = A−1 A−1 + C −1 C ¡ −1 ¢ −1 −1 −1 −1 −1 A − (A + C) = A A +C A ¡ ¢ −1 (A + BCD) = A−1 − A−1 BC C + CDA−1 BC CDA−1 (2. This matrix is called the inverse of A and is denoted by A−1 .4 Inverse A k × k matrix A has full rank.The can be seen since ⎢ ⎢ tr (AB) = tr ⎢ ⎣ = k X i=1 k X i=1 ⎡ a0 b1 a0 b2 · · · 1 1 a0 b1 a0 b2 · · · 2 2 . (2. . if there is no c 6= 0 such that Ac = 0. . 0b 0b ak 1 ak 2 · · · a0 bk 1 a0 bk 2 . we can still deﬁne a generalized inverse A− as a matrix which satisﬁes AA− A = A. For non-singular A and C. (2. In this case there exists a unique matrix B such that AB = BA = Ik . some properties include AA−1 = A−1 A = Ik ¡ ¢−1 ¡ −1 ¢0 = A0 A Even if a matrix A does not possess an inverse. then # ¡ ¢−1 ¡ ¢−1 ∙ ¸−1 " A − BD−1 C A B − A − BD−1 C BD−1 ¡ ¡ ¢−1 ¢−1 = . .

If λi is an eigenvalue of A. then A − λi Ik is singular so there exists a non-zero vector hi such that (A − λi Ik ) hi = 0 The vector hi is called a latent vector or characteristic vector or eigenvector of A corresponding to λi . λ−1 . The matrix B need not be unique. If A is symmetric. We call B a matrix square root of A. c0 Ac = α0 a ≥ 0 where α = Gc. One way to construct B is to use the spectral decomposition A = HΛH 0 where Λ is diagonal.2. • If A has distinct characteristic roots. If A is also symmetric (most idempotent matrices are) then all its characteristic roots equal either zero or one. which are not necessarily distinct and may be real or complex.. We say that X is n × k. We say that a square matrix A is positive semi-deﬁnite if for all non-zero c. Let λi and hi . We now state some useful properties. This is written as A ≥ 0. 1 2 k The decomposition A = HΛH 0 is called the spectral decomposition of a matrix. λ−1 .) If A is positive deﬁnite. c0 Ac ≥ 0. then A is positive semi-deﬁnite. Q • det(A) = k λi i=1 Pk • tr(A) = i=1 λi • A is non-singular if and only if all its characteristic roots are non-zero. and the characteristic roots are all real. A square matrix A is idempotent if AA = A. then A = HΛH 0 and H 0 AH = Λ.. The characteristic equation of a square matrix A is The left side is a polynomial of degree k in λ so it has exactly k roots. there exists a nonsingular matrix P such that A = P −1 ΛP and P AP −1 = Λ. • The characteristic roots of A−1 are λ−1 . 2. Then A = AA = HΛH 0 HΛH 0 = HΛ2 H 0 . 8 . and let H = [h1 · · · hk ].. then A is non-singular and A−1 exists.6 Rank and Positive Deﬁniteness The rank of a square matrix is the number of its non-zero characteristic roots. A−1 > 0. then A > 0 if and only if all its characteristic roots are positive.. Let Λ be a diagonal matrix with the characteristic roots in the diagonal. Furthermore. • If A is symmetric. (For any c 6= 0. If A > 0 we can ﬁnd a matrix B such that A = BB 0 . i = 1. If A = G0 G. . c0 Ac > 0. .5 Eigenvalues det (A − λIk ) = 0. k < n. has full rank k if there is no non-zero c such that Xc = 0. They are called the latent roots or characteristic roots or eigenvalues of A. and then set B = HΛ1/2 . This is written as A > 0. X 0 X is symmetric and positive deﬁnite. k denote the k eigenvalues and eigenvectors of a square matrix A. In this case... We say that A is positive deﬁnite if for all non-zero c. note that we can write A = HΛH 0 where H is orthogonal and Λ contains the (real) characteristic roots. To see this.

Additionally. .. and let επ = +1 if this count is even and επ = −1 if the count is odd. tr(A) = rank(A). ⎠ .. .. If A is 2 × 2. jk ) denote a permutation of (1... The vector derivative is ∂ ⎜ g(x) = ⎝ ∂x ∂ ∂x1 g(x) and ⎟ .. ∂ g(x) ∂xk ··· ∂ ∂xk g(x) ⎞ ³ ∂ g(x) = ∂x0 Some properties are now summarized. .... k) ....4) H0 M =H 0 0 with H 0 H = In ..By the uniqueness of the characteristic roots.. Then X det A = επ a1j1 a2j2 · · · akjk π Some properties include • det A = det A0 • det (αA) = αk det A • det(AB) = (det A) (det B) ¢ ¡ • det A−1 = (det A)−1 9 . xk ) : Rk → R. then its determinant is det A = a11 a22 − a12 a21 . i Hence they must equal either 0 or 1. we deduce that Λ2 = Λ and λ2 = λi for i = 1. .. Let π = (j1 . xk ) be k × 1 and g(x) = g(x1 . • • • • ∂ ∂x ∂ ∂x0 ∂ ∂x ´ . (a0 x) = ∂ ∂x (x0 a) = a (Ax) = A (x0 Ax) = (A + A0 ) x (x0 Ax) = A + A0 ∂2 ∂x∂x0 2.7 Matrix Calculus ⎛ ∂ ∂x1 g(x) Let x = (x1 . There are k! such permutations. It follows that the spectral decomposition of A takes the form ¸ ∙ In−k 0 (2. .. There is a unique count of the number of inversions of the indices of such permutations (relative to the natural order (1. k. For a general k × k matrix A = [aij ] . . 2. we can deﬁne the determinant as follows. .8 Determinant The determinant is deﬁned for square matrices.. k))..

am1 B am2 B · · · amn B Some important properties are now summarized. ⎠ ⎝ . . . The vec of A. denoted by vec (A) . ⎟ . . denoted A ⊗ B. The Kronecker product of A and B. . .9 Kronecker Products and the Vec Operator Let A = [a1 a2 · · · an ] = [aij ] be m × n. an Let B be any matrix. • If A is orthogonal. • (A + B) ⊗ C = A ⊗ C + B ⊗ C • (A ⊗ B) (C ⊗ D) = AC ⊗ BD • A ⊗ (B ⊗ C) = (A ⊗ B) ⊗ C • (A ⊗ B)0 = A0 ⊗ B 0 • tr (A ⊗ B) = tr (A) tr (B) • If A is m × m and B is n × n. . then det A = 2.• det ∙ A B C D ¸ • det A 6= 0 if and only if A is nonsingular. is the matrix ⎡ ⎤ a11 B a12 B a1n B ⎢ a21 B a22 B · · · a2n B ⎥ ⎢ ⎥ A⊗B =⎢ ⎥. These results hold for matrices for which all matrix multiplications are conformable. is the mn × 1 vector ⎞ ⎛ a1 ⎜ a2 ⎟ ⎟ ⎜ vec (A) = ⎜ . ⎣ ⎦ . det(A ⊗ B) = (det (A))n (det (B))m • (A ⊗ B)−1 = A−1 ⊗ B −1 • If A > 0 and B > 0 then A ⊗ B > 0 • vec (ABC) = (C 0 ⊗ A) vec (B) • tr (ABCD) = vec (D0 )0 (C 0 ⊗ A) vec (B) 10 . . then det A = ±1 ¢ ¡ = (det D) det A − BD−1 C if det D 6= 0 Qk i=1 aii • If A is triangular (upper or lower). . .

⎟ . the conditional density of yi given xi .1. To illustrate. m(x) can take any form. the conditioning variable. education and experience. . In this context we partition the observations into the pair (yi . (3. The data are from the 2004 Current Population Survey 1 11 .1. An important summary measure is the conditional mean Z ∞ yf (y | x) dy.1 by the arrows drawn to the x-axis. we can focus on f (y | x) . ⎠ ⎝ . While it is easy to observe that the two densities are unequal.1 displays the density1 of hourly wages for men and women. xki 3. gender. and that for women is $20. You can see that the right tail of then density is much thicker than the left tail. In the example presented in Figure 3.1) xi = ⎜ . We list the elements of xi in the vector ⎞ ⎛ x1i ⎜ x2i ⎟ ⎟ ⎜ (3. We call yi the dependent variable. Take a closer look at the density functions displayed in Figure 3. the mean wage for men is $27. from the population of white non-military wage earners with a college degree and 10-15 years of potential work experience.Chapter 3 Regression and Projection The most commonly applied econometric tool is regression.1 Conditional Mean To study how the distribution of yi varies with the variables xi in the population. Figure 3. These are asymmetric (skewed) densities. These are conditional density functions — the density of hourly wages conditional on race.73. holding the other variables constant. it is useful to have numerical measures of the diﬀerence. or the covariates. In this context it is often convenient These are nonparametric density estimates using a Gaussian kernel with the bandwidth selected by crossvalidation. These are indicated in Figure 3.22. We call xi alternatively the regressor. and exists so long as E |yi | < ∞. which is a common feature of wage distributions. See Chapter 16.2) m(x) = E (yi | xi = x) = −∞ In general. the mean is not necessarily a good summary of the central tendency. The two density curves show the eﬀect of gender on the distribution of wages. xi ) where yi is a scalar (real-valued) and xi is a vector. When a distribution is skewed. This is used when the goal is to quantify the impact of one set of variables on another variable.

To illustrate.2) evaluated at the observed value of xi : ei = yi − m(xi ).2 Regression Equation The regression error ei is deﬁned to be the diﬀerence between yi and its conditional mean (3.36. then comparisons become more complicated. For this reason. and a Ph. Figure 3.2 shows the density of log hourly wages for the same population. 2 (3. with mean log hourly wages drawn in with the arrows.30. This is a more robust measure of the typical wage gap between men and women than the diﬀerence in the untransformed wage means. Assuming for simplicity that this is the true joint distribution.Figure 3.A. Of particular interest to graduate students may be the observation that diﬀerence between a B.1: Wage Densities for White College Grads with 10-15 Years Work Experience to transform the data by taking the (natural) logarithm. degree in mean log hourly wages is 0. wage regressions typically use log wages as a dependent variable rather than the level of wages. The conditional expectation function is close to linear. this yields the formula yi = m(xi ) + ei . 3. The main point to be learned from Figure 3. the solid line displays the conditional expectation of log wages varying with education. When the distribution of the control variable is continuous. implying an average 36% diﬀerence in wage levels.3 displays a scatter plot2 of log wages against education levels. The diﬀerence in the log mean wage between men and women is 0.D. By construction.5. which implies a 30% average wage diﬀerence for this population. Figure 3.3) White non-military male wage earners with 10-15 years of potential work experience. The comparison in Figure 3.3 is how the conditional expectation describes an important feature of the conditional distribution. the dashed line is a linear projection approximation which will be discussed in the Section 3.1 is facilitated by the fact that the control variable (gender) is discrete. 12 .

2. The equations yi = m(xi ) + ei E (ei | xi ) = 0. To show the ﬁrst statement. E(ei ) = 0. not a model. It is important to understand that this is a framework. E(xi ei ) = 0. The remaining parts of the Theorem are left as an exercise. restrictions on the permissible class of regression functions m(x). because no restrictions have been placed on the joint distribution of the data.2. E (ei | xi ) = E ((yi − m(xi )) | xi ) = m(xi ) − m(xi ) = E (yi | xi ) − E (m(xi ) | xi ) = 0. E (ei | xi ) = 0. E (h(xi )ei ) = 0 for any function h (·) . by the deﬁnition of ei and the linearity of conditional expectations. most typically. 3.2: Log Wage Densities Theorem 3.1 Properties of the regression error ei 1. These equations hold true by deﬁnition. 4. A regression model imposes further restrictions on the joint distribution. 13 . are often stated jointly as the regression framework.Figure 3.

when σ 2 (x) is a constant so that ¢ ¡ (3. i . The right-hand-side is minimized by setting g(x) = m(x). in the sense of achieving the lowest mean squared error.Figure 3. 3.3. Given the random variable xi .4) E e2 | xi = σ 2 i 14 While the conditional mean is a good measure of the location of a conditional distribution. the conditional variance of yi is σ 2 = σ 2 (xi ). In contrast. The expected squared error using this prediction function is E (yi − g(xi ))2 = E (ei + m(xi ) − g(xi ))2 = Ee2 + 2E (ei (m(xi ) − g(xi ))) + E (m(xi ) − g(xi ))2 i = Ee2 + E (m(xi ) − g(xi ))2 i ≥ Ee2 i where the second equality uses Theorem 3. let g(x) be an arbitrary predictor of yi given xi = x.2.3: Conditional Mean of Wages Given Education The conditional mean also has the property of being the the best predictor of yi . Thus the mean squared error is minimized by the conditional mean.3 Conditional Variance Generally. and can take any form.1. subject to the restriction p that it is non-negative. it does not provide information about the spread of the distribution. A common measure of the dispersion is the conditional variance ¡ ¢ σ 2 (x) = V ar (yi | xi = x) = E e2 | xi = x . σ 2 (x) is a non-trivial function of x. The conditional standard deviation is its square root σ(x) = σ 2 (x). To see this. In the general case i where σ 2 (x) depends on x we say that the error ei is heteroskedastic.

it is more realistic to view the linear speciﬁcation (3. Some textbooks inappropriately describe heteroskedasticity as the case where “the variance of ei varies across observation i”.we say that the error ei is homoskedastic. where x1i is the ﬁrst element of the vector xi deﬁned in (3.3) is when the conditional mean function m(x) is linear in x (or linear in functions of x). take the conditional wage densities displayed in Figure 3. In this case (3.5 Best Linear Predictor While the conditional mean m(x) = E (yi | xi = x) is the best predictor of yi among all functions of xi .1.5. Thus (3.1).6) as an approximation. ⎠ . i (3. we can write it as m(x) = x0 β = β 1 + x2i β 2 + · · · + xki β k where ⎞ β1 ⎜ . An important special case is homoskedastic linear regression model yi = x0 β + ei i E (ei | xi ) = 0 ¢ ¡ E e2 | xi = σ 2 . we assume that x1i = 1. As an example. So while men have higher average wages.9) 3.8) is called the linear regression model. they are also somewhat more dispersed. ⎟ β=⎝ . 3. βk ⎛ (3.7) is a k × 1 parameter vector. 15 . Instead.1) has been redeﬁned as the k × 1 vector ⎛ ⎞ 1 ⎜ x2i ⎟ ⎜ ⎟ (3. which we derive in this section. We call this the “constant” or “intercept”.4 Linear Regression An important special case of (3. Equivalently. The conditional standard deviation for men is 12. ⎟ . and the linear assumption of the previous section is empirically unlikely to accurate.6) (3. its functional form is typically unknown. This concept is less helpful than deﬁning heteroskedasticity as the dependence of the conditional variance on the observables xi . Equation (3. ⎠ .1 and that for women is 10. Notationally. xki When m(x) is linear in x.3) can be writen as yi = x0 β + ei i E (ei | xi ) = 0. it is convenient to augment the regressor vector xi by listing the number “1” as an element.5) xi = ⎜ .8) (3. ⎝ .

i 4.10).12) This completes the derivation of the linear projection model.5. otherwise they are distinct. In order for β to be uniquely deﬁned.which is quadratic in β. Given the deﬁnition of β in (3. written E (xi x0 ) > 0. This occurs when redundant variables are included in xi .1 1. It is invertible if and only if it is positive deﬁnite. Equivalently.7) is 0= Solving for β we ﬁnd ¡ ¢ ∂ S(β) = −2E (xi yi ) + 2E xi x0 β. E (xi x0 ) is invertible. For any β ∈ Rk a linear predictor for yi is x0 β with expected squared prediction i error ¢2 ¡ S(β) = E yi − x0 β i ¢ ¢ ¡ ¡ 2 = Eyi − 2E yi x0 β + β 0 E xi x0 β. α0 E (xi x0 ) α = E (α0 xi )2 ≥ 0 so the matrix E (xi x0 ) is by construction i i positive semi-deﬁnite. i which requires that for all non-zero α. The error is i ei = yi − x0 β. i 16 . 2 2.10) It is worth taking the time to understand the notation involved in this expression. i ∂β ¢¢−1 ¡ ¡ E (xi yi ) . alternative expressions such as E xixi ) or E (xi yi ) (E (xi x0 ))−1 i ( i 0i ) are incoherent and incorrect. this situation must be excluded. Observe i that for any non-zero α ∈ Rk . Eyi < ∞. We now summarize the assumptions necessary for its derivation and list the implications in Theorem 3.10) exists and is unique as long as the k ×k matrix E (xi x0 ) is invertible. α0 E (xi x0 ) α = E (α0 xi )2 > 0.1.5. Assumption 3. 3. β = E xi x0 i In the linear projection model the coeﬃcient β is deﬁned so that the function x0 β is the i best linear predictor of yi . i (3. Rewriting. Therefore. E (xi x0 ) is a mai E(x y trix and E (xi yi ) is a vector. The vector (3. As before. i i (3. there cannot i exist a non-zero vector α such that α0 xi = 0 identically. E (x0 xi ) < ∞. i (3. x0 β is the best linear predictor for yi . The best linear predictor is obtained by selecting β to minimize S(β). by “best” we mean the predictor function with lowest mean squared error. we obtain a decomposition of yi into linear predictor and error yi = x0 β + ei .11) Notice that the error from the linear prediction equation ei is equal to the error from the regression equation when (and only when) the conditional mean is linear in xi . xi contains an intercept. The ﬁrst-order condition for minimization (from Section 2.

1.11) are well deﬁned.4 we know that the regression error has the property E (xi ei ) = 0.5.13) The two equations (3. 17 . By deﬁnition. suppose that for xi ∈ R that Ex3 = 0 and ei = x2 .4 guarantees that the solution β exits.5.1. Figure 3.1 Under Assumption 3.5. (3.5.1.1.1. Using the deﬁnitions (3.1. i i i Since E (xi ei ) = 0 we say that xi and ei are orthogonal.1. Assumption 3.12) and (3. Since ei is mean zero by (3.1 are weak. Equation (3.2. E (xi ei ) = 0 and E (ei ) = 0.1.13) implies that xi and ei are uncorrelated.14) holds.4 is required to ensure that β is uniquely deﬁned.10) ¡ ¡ ¢¢ E (xi ei ) = E xi yi − x0 β i ¡ ¢¡ ¡ ¢¢−1 E (xi yi ) = E (xi yi ) − E xi x0 E xi x0 i i = 0.5.14) Proof.5.10) are deﬁned.8)-(3. Then i i Exi e2 = Ex3 = 0 yet E (ei | xi ) = x2 6= 0.2 and 3.12) can be alternatively interpreted as a projection decomposition. x0 β is the i projection of yi on xi since the error ei is orthogonal with xi . Let’s compare it with the linear regression model (3.14) follows from (3.Theorem 3.9).3 are called regularity conditions. Furthermore.5.1 is employed to guarantee that (3.1.5. Assumption 3.10) and (3. (3.5. The ﬁnite variance Assumptions 3.14).4: Hourly Wage as a Function of Experience The conditions listed in Assumption 3. Since from Theorem 3.1. the orthogonality restriction (3. Assumption 3.5.1. For example.13) and Assumption 3.2 and 3.11) and (3. ¥ (3. This means that the equation (3.5.13) summarize the linear projection model. However.3 ensure that the moments in (3. the converse is not true as the projection error does not necessarily satisfy E (ei | xi ) = 0. Assumption 3. it follows that linear regression is a special case of the projection model.

In this example it is a much better approximation to the conditional mean than the linear projection.4 we display as well this quadratic projection. and under-predicts for the rest. there are no changes in our methods or analysis. the dashed line is the linear projection of log wages on eduction.12) with the properties listed in Theorem 3. Figure 3. and the straight dashed line is the linear projection. In the example just presented. In this case (3. in many economic models the parameter β may be deﬁned within the model. This defect in linear projection can be partially corrected through a careful selection of regressors.4 displays the relationship3 between mean log hourly wages and labor market experience. Figure 3. It over-predicts wages for young and old workers. In this case the linear projection is a poor approximation to the conditional mean. since the resulting function is quadratic in experience. However. it misses the strong downturn in expected wages for those above 35 years work experience (equivalently.5. it is important to not misinterpret the generality of this statement. No additional assumptions are required. In this example the linear projection is a close approximation to the conditional mean.5 for a 3 In the population of Caucasian non-military male wage earners with 12 years of education. Most importantly. for those over 53 in age).12) is deﬁned by projection and the associated coeﬃcient deﬁnition (3. The linear equation (3.10) may not hold and the implications of Theorem 3. Other than the redeﬁnition of the regressor vector. we can augment the regressor vector xi to include both experience and experience2 .5: Conditional Mean and Two Linear Projections Another defect of linear projection is that it is sensitive to the marginal distribution of the regressors when the conditional mean is non-linear. for any pair (yi .5. These structural models require alternative estimation methods. Returning to the joint distribution displayed in Figure 3. The solid line is the conditional mean.1 may be false. and are discussed in Chapter 11.1. In Figure 1. A projection of log wages on these two variables can be called a quadratic projection. xi ) we can deﬁne a linear projection equation (3.We have shown that under mild regularity conditions.10). In other cases the two may be quite diﬀerent. In contrast.3. We illustrate the issue in Figure 3. 18 .

1) and those in Group 2 are N(4. combined with diﬀerent marginal distributions for the conditioning variables. and Group 1 has a lower mean value of xi than Group 2. The data are divided in two — Group 1 and Group 2 — which have diﬀerent marginal distributions for the regressor xi . 4 19 . The separate linear projections of yi on xi for these two groups is displayed in the Figure with the dashed lines. This conclusion is incorrect because is fact there is no diﬀerence in the conditional mean between the two groups.constructed4 joint distribution of yi and xi . The apparant diﬀerence is a by-product of a linear approximation to a non-linear mean. and the conditional distriubtion of yi given xi is N(m(xi ). A defect with linear projection is that it leads to the incorrect conclusion that the eﬀect of xi on yi is diﬀerent for individuals in the two Groups. The xi in Group 1 are N(2. 1) where m(x) = 2x − x2 /6. 1). These two projections are distinct approximations to the conditional mean. The solid line is the non-linear conditional mean of yi given xi .

9. 3. xi ∈ R. True or False. Deﬁne µ ¶ ¢ ¡ x−µ 2 g x. Take the bivariate linear projection model yi = β 0 + β 1 xi + ei E (ei ) = 0 E (xi ei ) = 0 Deﬁne µy = Eyi . Compute the conditional mean m(x) = E (yi | xi = x) . then E(x2 ei ) = 0. µx = Exi . Compute 2 the coeﬃcients of the linear projection yi = β 0 + β 1 xi + ei . σ 2 = V ar(xi ). Prove parts 2. then ei is independent of xi .3. If yi = x0 β + ei . then E(ei | xi ) = 0. 2. ¡ ¢ 4. and E(e2 | xi ) = σ 2 . Are they diﬀerent? Hint: If P (Y = j) = 5. Compute E(yi | xi = x) and V ar(yi | xi = x). and have the following joint probability distribution X=0 X=1 Y =0 . True or False.1. Does this justify a linear regression model of the form yi = x0 β + εi ? i then EY = λ and V ar(Y ) = λ. 3 and 4 of Theorem 3. True or False. . Suppose that Y and X only take the values 0 and 1. µ. Suppose that yi is discrete-valued.6 Exercises 1. 2.1 . taking values only on the non-negative integers. If yi = x0 β + ei and E(ei | xi ) = 0. i 11. Show that x y β 1 = σ xy /σ 2 and β 0 = µy − β 1 µx . (x − µ)2 − σ 2 Show that Eg (X. xi ∈ R. s) = 0 if and only if m = µ and s = σ 2 . If yi = xi β + ei . i 8. e−λ λj j! . then ei is i i independent of xi . and E(xi ei ) = 0. 1. j! 0 j = 0.4 . i 7. If yi = x0 β + ei and E(xi ei ) = 0. 20 . True or False. If yi = xi β + ei . and the conditional distribution of yi given xi is Poisson: P (yi = k | xi = x) = e−x β (x0 β)j . Let xi and yi have the joint density f (x.. σ = .2. i 10.3 Find E(Y | X = x). x 6. E(ei | xi ) = 0. E(Y 2 | X = x) and V ar(Y | X = x). 0 ≤ y ≤ 1. and E(ei | xi ) = 0. a constant.2 Y =1 . then E(x2 ei ) = 0. True or False. yi ). σ 2 = V ar(yi ) and σ xy = Cov(xi . y) = 3 x2 + y 2 on 0 ≤ x ≤ 1. Let X be a random variable with µ = EX and σ 2 = V ar(X). m..

8.3) with the sample moments: ³ ¡ ¢´−1 ˆ ˆ ˆ β = E xi x0 E (xi yi ) i Ã n !−1 n 1X 1X 0 xi xi xi yi = n n i=1 i=1 Ã n !−1 n X X = xi x0 xi yi .2) can be written in the parametric 0 β)) = 0.3) In Sections 4.Chapter 4 Least Squares Estimation This chapter explores estimation and inference in the linear projection model yi = x0 β + ei i E (xi ei ) = 0 ¡ ¡ ¢¢−1 E (xi yi ) β = E xi x0 i (4. we narrow the focus to the linear regression model.3) writes the projection coeﬃcient β as an explicit function of population moments E (xi yi ) and E (xi x0 ) . 4. but for most of the chapter we retain the broader focus on the projection model. The function g(β) can be estimated by form g(β) = E (xi (yi − xi g (β) = ˆ ¢ 1X ¡ xi yi − x0 β . i It follows that the moment estimator of β replaces the population moments in (4.4) i i=1 i=1 ˆ Another way to derive β is as follows. i n i=1 n 21 .1 Estimation Equation (4. Their moment estimators are the sample moments i ˆ E (xi yi ) = ¢ ¡ ˆ E xi x0 = i 1X xi yi n 1 n i=1 n X i=1 n xi x0 .2) (4. (4.7 and 4.1) (4. Observe that (4.

0.117 Educationi . 30 = .40 µ ¶µ ¶ 34. An interpretation of the estimated equation is that each year of education is associated with an 11.37 µ ¶ 1.14 14.4).71 = −2. These are white male wage earners from the March 2004 Current Population Survey.7% increase in mean wages. 4. with 10-15 years of potential work experience. i 22 .83 µ We often write the estimated equation using the format \ log(W agei ) = 1.ˆ This is a set of k equations which are linear in β.95 xi yi = 42.4).3.14 14. To illustrate. Then µ ¶ n 1X 2.95 42.170 37.83 ¶−1 µ ¶ .94 −2.40 n i=1 1 n Thus ˆ β = n X i=1 xi x0 i = µ 1 14. 40 0. The estimator β is the value which jointly sets these equations equal to zero: 0 = g (β) ˆ ˆ n ´ 1X ³ ˆ xi yi − x0 β = i n i=1 (4. n X i=1 2 yi − 2β 0 n X i=1 xi yi + β 0 n X i=1 xi x0 β. Let yi be log wages and xi be an intercept and years of education. ¶ 2. 40 2.30 + 0. This sample has 988 observations. consider the data used to generate Figure 3.14 205.117 1 14. excluding military.14 205. Deﬁne the sum-of-squared errors (SSE) function n X¡ ¢2 yi − x0 β Sn (β) = i i=1 = This is a quadratic function of β.2 Least Squares There is another classic motivation for the estimator (4.5) = n n 1X 1X ˆ xi yi − xi x0 β i n n i=1 i=1 whose solution is (4.

4). Figure 4. It is important to understand the distinction between the error ei and the residual ei . The error is unobservable.The Ordinary Least Squares (OLS) estimator is the value of β which minimizes Sn (β). Following convention we will call β the OLS estimator of β.2 displays the contour lines of the same function — horizontal slices at equally spaced heights. which can cause confusion.7) gives the ﬁrst-order conditions for minimization: 0 = ∂ ˆ Sn (β) ∂β n n X X ˆ = −2 xi yi + 2 xi x0 β i i=1 i=1 ˆ whose solution is (4. Since the function Sn (β) is a quadratic function of β. while the residual is a by-product of estimation.1: Sum-of-Squared Errors Function As a by-product of OLS estimation. Figure 4.1 displays an example sum—of-squared errors function Sn (β) for the case k = 2. To visualize the sum-of-squared errors function. we deﬁne the predicted value ˆ yi = x0 β ˆ i and the residual ˆ ei = yi − yi ˆ ˆ = yi − x0 β. i ˆ ˆ Note that yi = yi + ei . Matrix calculus (see Appendix 2. the contour lines are ellipses. These two ˆ variables are frequently mislabeled. 23 . Figure 4.

ˆ n i=1 n Since xi contains a constant.6) An alternative estimator uses the formula n 1 X 2 s = ei . and hold true for all linear regression estimates.Figure 4.7.2: Sum-of-Squared Error Function Contours Equation (4. 24 . ˆ n i=1 Thus the residuals have a sample mean of zero and the sample correlation between the regressors and the residual is zero. one implication is that n 1X ei = 0. These are algebraic results. ˆ σ = ˆ n 2 i=1 n (4. The error variance σ 2 = Ee2 is also a parameter of interest. It measures the variation in the i “unexplained” part of the regression. ˆ n−k 2 i=1 (4.7) A justiﬁcation for the latter choice will be provided in Section 4. Its method of moments estimator is the sample average 1X 2 ei .5) implies that 1X xi ei = 0.

3 Normal Regression Model Another motivation for the least-squares estimator can be obtained from the normal regression model. σ 2 ) maximize Ln (β. and therefore is expected to be a better estimator of ρ2 than the unadjusted estimator R2 . Hence the MLE for β equals the OLS estimator. σ 2 ). maximizing the likelihood is identical to minimizing Sn (β). It is an ˆ inappropriate label as the value of R2 does not help interpret the parameter estimates β or test statistics concerning β. Since Ln (β. Pn y2 ˆ σ2 ˆ R2 = Pn i=1 i 2 = 1 − 2 σy ˆ i=1 (yi − y) σ2 = ˆy 1X (yi − y)2 n i=1 n is the sample variance of yi . n−1 i=1 Theil’s estimator R is a ratio of adjusted variance estimators.where A measure of the explained variation relative to the total variation is the coeﬃcient of determination or R-squared. 4. where likelihood methods can be used for estimation. σ ) = 2σ (2πσ 2 )1/2 i=1 ¡ ¢ n 1 = − log 2πσ 2 − 2 Sn (β) 2 2σ ˆ ˆ The MLE (β. ˆ Plugging β into the log-likelihood we obtain n ¡ ¢ n 1 X 2 ˆ ei ˆ Ln (β. σ 2 ). Instead. testing. σ 2 ) = − log 2πσ 2 − 2 2 2σ i=1 25 . it should be viewed as an estimator of the population parameter ρ2 = σ2 V ar (x0 β) i =1− 2 V ar(yi ) σy where σ 2 = V ar(yi ). The R2 is frequently mislabeled as a measure of “ﬁt”. This is a parametric model. This is the linear regression model with the additional assumption that the error ei is independent of xi and has the distribution N (0. and distribution theory. σ 2 ) is a function of β only through the sum of squared errors Sn (β). An alternative estimator of ρ2 proposed by Theil called “R-bar-squared” is y R =1− where σ2 = ˜y 2 n 2 s2 σ2 ˜y 1 X (yi − y)2 . The log-likelihood function for the normal regression model is Ã ¶! µ n X ¢2 1 1 ¡ 2 0 log exp − 2 yi − xi β Ln (β.

6). en ⎞ Observe that Y and e are n × 1 vectors. n or equivalently Y = Xβ + e. ⎠ xi x0 = X 0 X i xi yi = X 0 Y. and sample error variance can be written as ¢−1 ¡ 0 ¢ ¡ ˆ β = X 0X XY ˆ e = Y − Xβ ˆ σ 2 = n−1 e0 e. ⎠ . Due to this equivalence. . σ 2 ) = − 2 + ¡ ¢2 e2 = 0. .8) .4 Model in Matrix Notation convenient to write the model and statistics in x0 1 x0 2 . ⎝ ⎝ . x0 n ⎞ ⎛ e1 e2 . We deﬁne ⎛ ⎞ ⎛ y1 ⎜ ⎜ y2 ⎟ ⎜ ⎟ ⎜ X=⎜ Y = ⎜ . The linear equation (3. residual vector. Thus the MLE (β. the ˆ OLS estimator β is frequently referred to as the Gaussian MLE. ˆ ˆˆ ˆ A useful result is obtained by inserting Y = Xβ + e into the formula for β to obtain ¢ ¡ 0 ¢−1 ¡ 0 ˆ X (Xβ + e) β = XX ¡ ¢−1 ¡ 0 ¢ ¡ 0 ¢−1 0 Xe X Xβ + X 0 X = XX ¡ 0 ¢−1 0 X e. . . and X is an n × k matrix. . including computation. σ 2 ) for the normal ˆ regression model are identical to the method of moment estimators. n X i=1 Thus the estimator (4. 4. For example n X i=1 For many purposes. it is matrix notation. ˆi 2 ∂σ 2 2ˆ σ 2 σ ˆ i=1 ˆ ˆ Solving for σ 2 yields the method of moments estimator (4. yn = x0 β + en . one for each observation. yn ⎟ ⎟ ⎟.4). We can stack these n equations together as y1 = x0 β + e1 1 y2 = x0 β + e2 2 . ⎟. Sample sums can also be written in matrix notation.12) is a system of n equations. = β+ XX 26 (4. ⎠ ⎜ ⎜ e=⎜ ⎝ ⎟ ⎟ ⎟. .Maximization with respect to σ 2 yields the ﬁrst-order condition n X ∂ n 1 ˆ ˆ Ln (β.

4.5

Projection Matrices

¢−1 0 ¡ X P = X X 0X M ¡ ¢−1 0 = In − X X 0 X X = In − P

Deﬁne the matrices and

where In is the n × n identity matrix. They are called projection matrices due to the property that for any matrix Z which can be written as Z = XΓ for some matrix Γ, (we say that Z lies in the range space of X) then ¢−1 0 ¡ X XΓ = XΓ = Z P Z = P XΓ = X X 0 X M Z = (In − P ) Z = Z − P Z = Z − Z = 0. As an important example of this property, partition the matrix X into two matrices X1 and X2 , so that X = [X1 X2 ] . Then P X1 = X1 and M X1 = 0. It follows that M X = 0 and M P = 0, so M and P are orthogonal. The matrices P and M are symmetric and idempotent1 . To see that P is symmetric, ³ ¡ ¢−1 0 ´0 P 0 = X X 0X X ³¡ ¡ ¢0 ¢−1 ´0 X 0X = X0 (X)0 ³¡ ¢0 ´−1 0 = X X 0X X ³ ¡ ¢0 ´−1 0 = X (X)0 X 0 X = P. To establish that it is idempotent, ³ ¡ ¢−1 0 ´ ³ ¡ 0 ¢−1 0 ´ X X 0X X XX X X ¡ 0 ¢−1 0 ¡ 0 ¢−1 0 = X XX XX XX X ¡ 0 ¢−1 0 = X XX X = P, and MM = M (In − P )

and

PP

=

= M − MP = M

1

A matrix A is idempotent if AA = A.

27

**since M P = 0. Another useful property is that tr P tr M where the trace operator tr A = = k = n−k
**

r X j=1

(4.9) (4.10)

ajj

is the sum of the diagonal elements of the matrix A. To show (4.9) and (4.10), ³ ¡ ¢−1 0 ´ X tr P = tr X X 0 X ³¡ ¢−1 0 ´ = tr X 0 X XX = tr (Ik ) = k, and tr M = tr (In − P ) = tr (In ) − tr (P ) = n − k. Given the deﬁnitions of P and M, observe that ¢−1 0 ¡ ˆ ˆ X Y = PY Y = X β = X X 0X ˆ e = Y − X β = Y − P Y = M Y. ˆ Furthermore, since Y = Xβ + e and M X = 0, then e = M (Xβ + e) = M e. ˆ Another way of writing (4.11) is ˆ ˆ Y = (P + M ) Y = P Y + M Y = Y + e. This decomposition is orthogonal, that is ˆ ˆ Y 0 e = (P Y )0 (M Y ) = Y 0 P M Y = 0. (4.12) (4.11)

and

4.6

Residual Regression

X = [X1 X2 ] ¶ .

Partition and β= µ

β1 β2

Then the regression model can be rewritten as Y = X1 β 1 + X2 β 2 + e. 28 (4.13)

Observe that the OLS estimator of β = (β 0 , β 0 )0 can be obtained by regression of Y on X = [X1 1 2 X2 ]. OLS estimation can be written as ˆ ˆ Y = X1 β 1 + X2 β 2 + e. ˆ (4.14)

Suppose that we are primarily interested in β 2 , not in β 1 , so are only interested in obtaining ˆ ˆ the OLS sub-component β 2 . In this section we derive an alternative expression for β 2 which does not involve estimation of the full model. Deﬁne ¡ 0 ¢−1 0 X1 . M1 = In − X1 X1 X1

0 Recalling the deﬁnition M = I − X (X 0 X)−1 X 0 , observe that X1 M1 = 0 and thus ¡ 0 ¢−1 0 M1 M = M − X1 X1 X1 X1 M = M

It follows that

**M1 e = M1 M e = M e = e. ˆ ˆ Using this result, if we premultiply (4.14) by M1 we obtain M1 Y ˆ ˆ = M1 X1 β 1 + M1 X2 β 2 + M1 e ˆ ˆ = M1 X2 β + e ˆ
**

2

(4.15)

0 0ˆ the second equality since M1 X1 = 0. Premultiplying by X2 and recalling that X2 e = 0, we obtain 0 0 0 0 ˆ ˆ X2 M1 Y = X2 M1 X2 β 2 + X2 e = X2 M1 X2 β 2 . ˆ

Solving, ˆ an alternative expression for β 2 . Now, deﬁne

¡ 0 ¢ ¢−1 ¡ 0 ˆ β 2 = X2 M1 X2 X2 M1 Y ˜ X2 = M1 X2 ˜ Y = M1 Y, (4.16) (4.17)

ˆ ˜ ˜ This shows that β 2 can be calculated by the OLS regression of Y on X2 . This technique is called residual regression. Furthermore, using the deﬁnitions (4.16) and (4.17), expression (4.15) can be equivalently written as ˜ ˜ ˆ ˆ Y = X2 β 2 + e. ˆ ˜ ˜ Since β 2 is precisely the OLS coeﬃcient from a regression of Y on X2 , this shows that the residual from this regression is e, numerically the same residual as from the joint regression (4.14). We ˆ have proven the following theorem. 29

the least-squares residuals from the regression of X2 and Y, respectively, on the matrix X1 only. Since the matrix M1 is idempotent, M1 = M1 M1 and thus ¢−1 ¡ 0 ¡ 0 ¢ ˆ X2 M1 Y β 2 = X2 M1 X2 ¢ ¢−1 ¡ 0 ¡ 0 X2 M1 M1 Y = X2 M1 M1 X2 ´−1 ³ ³ ´ ˜0 ˜ ˜0 ˜ = X2 X2 X2 Y

¡ ¢−1 0 ι. . Rather. and X2 is the vector of observed regressors. . but in most cases there is little computational advantage to using the two-step algorithm. .13).8)(3. . the FWL theorem can be used to speed computation. In the model (4. ! Thus the OLS estimator for the slope coeﬃcients is a regression with demeaned data.18) E (e | X) = ⎜ E (ei | X) ⎟ = ⎜ E (ei | xi ) ⎟ = 0. . 30 . . A common application of the FWL theorem.6. observe that ⎞ ⎛ ⎞ ⎛ . The FWL theorem says that β 2 is the OLS estimate from a regression of ˜ ˜ Y on X2 . . ⎟ ⎜ ⎟ ⎜ (4. . is the demeaning formula for regression. ⎠ ⎝ ⎠ ⎝ . Regress X2 on X1 . . Regress Y on X1 . which you may have seen in an introductory econometrics course.1 (Frisch-Waugh-Lovell). In this section we derive the small sample conditional mean and variance of the OLS estimator. or yi − y on x2i − x2 : ˆ β2 = Ã n X i=1 (x2i − x2 ) (x2i − x2 ) 0 !−1 Ã n X i=1 (x2i − x2 ) (yi − y) . 4. . ˆ ˜ ˜ ˆ 3. M1 = I − ι ι0 ι Observe that ¡ ¢−1 0 ˜ ι X2 X2 = M1 X2 = X2 − ι ι0 ι = X2 − X 2 and ˜ Y ¡ ¢−1 0 = M1 Y = Y − ι ι0 ι ιY = Y −Y. In some contexts. the primary use is theoretical.Theorem 4. By the independence of the observations and (3. obtain residuals X2 . Partition X = [X1 X2 ] where X1 = ι is a vector of ones. the OLS estimator of β 2 and the OLS residuals e may be equivalently computed by either the OLS regression (4. ˜ 2. ˆ which are “demeaned”. obtain OLS estimates β 2 and residuals e.9).7 Bias and Variance In this and the following section we consider the special case of the linear regression model (3. In this case.14) or via the ˆ following algorithm: ˜ 1. obtain residuals Y . Regress Y on X2 .9).

19) ˆ and thus the OLS estimator β is unbiased for β. and ³ ´ ¡ ¢−1 2 ˆ σ . . the properties of conditional expectations. Thus D is a diagonal matrix with i’th diagonal element σ 2 : i ⎛ 2 σ1 0 · · · ⎜ 0 σ2 · · · 2 ⎜ D = diag{σ 2 . . Then given (4. we can calculate ³ ´ ³¡ ´ ¢−1 0 ˆ E β−β |X = E X 0X Xe|X ¢−1 0 ¡ X E (e | X) = X 0X = 0. 1 n . σ2 n ⎞ We now calculate the ﬁnite sample bias of the method ¡ moments estimator σ 2 for σ 2 .19) we see that ¶ µ³ ³ ´ ´³ ´0 ˆ ˆ ˆ V ar β | X = E β−β β−β |X = where The i’th diagonal element of D is ¡ ¢ D = E ee0 | X . Next.Using (4.. X 0 DX = X 0 Xσ 2 . . ¡ X 0X ¢−1 ¡ ¢−1 X 0 DX X 0 X while the ij 0 th oﬀ-diagonal element of D is ¢ ¢ ¡ ¡ E e2 | X = E e2 | xi = σ 2 i i i E (ei ej | X) = E (ei | xi ) E (ej | xj ) = 0.20) . σ2 = ˆ ¢ 1 ¡ ¢ 1 0 1 1 1 ¡ e e = e0 M M e = e0 M e = tr e0 M e = tr M ee0 ˆˆ n n n n n 31 In the special case of the linear homoskedastic regression model...8). the properties additional assumption of conditional homoskedasticity E ei i of projection matrices.18). . σ 2 = σ 2 and we have the simplii ﬁcations D = In σ 2 . and the trace operator observe that.. σ 2 } = ⎜ . . We have shown that which implies ³ ´ ˆ E β|X =β ³ ´ ˆ E β =β (4. . ⎠ (4. ⎝ . 0 0 ··· 0 0 . . V ar β | X = X 0 X ⎟ ⎟ ⎟. under the of ˆ ¢ 2 | x = σ 2 . From (4. and (4. .12). for any random vector Z deﬁne the covariance materix V ar(Z) = E (Z − EZ) (Z − EZ)0 = EZZ 0 − (EZ) (EZ)0 .

˜ β L = A0 Y = A (Xβ + e) = β + Ae. which is (3.8 Gauss-Markov Theorem In this section we restrict attention to the homoskedastic linear regression model. By a calculation similar to those of the previous section. we can write The “best” linear estimator is obtained by ﬁnding the matrix A for which this variance is the smallest in the positive deﬁnite sense.10). As an alternative.9) plus E e2 | xi = σ 2 .7) is unbiased for σ 2 by this calculation.8. is a famous statement of the solution. however. says that the least-squares estimator is the best choice. ³ ´ ˜ E β | X = A0 Xβ. and thus can be written as ˜ β = A0 Y where A is an n × k function of X. What is the best choice of A? The Gauss-Markov theorem. which we now present. Thus σ 2 is biased towards zero. that this estimator is only unbiased in the special case of the homoskedastic linear regression model. Now consider the class of estimators of β which are linear functions of i the vector Y.8)¢ ¡ (3. This is the justiﬁcation for the common s ˆ preference of s2 over σ 2 in empirical practice. the best (minimumvariance) unbiased linear estimator is OLS.1 Gauss-Markov. The following result. The least-squares estimator is the special case obtained by setting A = X(X 0 X)−1 . 32 .Then ¢ ¡ E σ2 | X = ˆ ¢¤ 1 £ ¡ tr E M ee0 | X n ¡ ¢¤ 1 £ = tr M E ee0 | X n ¤ 1 £ tr M σ 2 = n n−k . ˜ so β is unbiased if (and only if) A0 X = Ik . In this case. It is not unbiased in the absence of homoskedasticity. Theorem 4. It is important to remember. the estimator ˆ 2 deﬁned (4. or in the projection model. as it yields the smallest variance among all unbiased linear estimators. known as the Gauss-Markov theorem. ³ ´ ˜ V ar β | X = A0 V ar (e | X) A = A0 Aσ 2 . In the homoskedastic linear regression model. 4. Thus since V ar (e | X) = In σ 2 under homoskedasticity. = σ2 n the ﬁnal equality by (4.

where 1 (·) is the indicator function.5) as required.. A broader justiﬁcation is provided in Chamberlain (1987). r.. Not only has the class of models has been restricted to homoskedastic linear regressions. That is. ξ j . xi = ξ j = π j ... . xi ) is discrete. Assume that the τ j and ξ j are known. ˆ β mle = ⎝ j j=1 j=1 33 . the deﬁnition (4. It is suﬃcient to show that the diﬀerence A0 A − (X 0 X)−1 is positive semi-deﬁnite.) In this discrete setting.Proof. This property is called semiparametric eﬃciency. That is. Let A be any n×k function of X such that A0 X = Ik . (We know the values yi and xi can take. The variance of the least-squares estimator is (X 0 X)−1 σ 2 and that of A0 Y is A0 Aσ 2 . the class of potential estimators has been restricted to linear unbiased estimators. We discuss the intuition behind his result in this section.21) j j=1 j=1 Thus β is a function of (π 1 . . but it is quite limited in scope. This latter restriction is particularly unsatisfactory. π r ) . . but the π j are unknown.. The MLE β mle for β is then the analog of (4. Suppose that the joint distribution of (yi . who established that in the projection model the OLS estimator has the smallest asymptotic meansquared error among feasible estimators.. 4. but we don’t know the probabilities. ¡ ¢ P yi = τ j . j = 1. π j is the percentage of the observations ˆ ˆ which fall in each category.. the maximum likelihood estimator (MLE) is ¡ ¢ 1X 1 (yi = τ j ) 1 xi = ξ j πj = ˆ n i=1 n for j = 1. and π j ... 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. Note that X 0 C = 0.9 Semiparametric Eﬃciency In the previous section we presented the Gauss-Markov theorem as a limited eﬃciency justiﬁcation for the least-squares estimator.21) with the parameters ˆ π j replaced by the estimates π j : ⎞−1 ⎛ ⎞ ⎛ r r X X ˆ πj ξj ξ0 ⎠ ⎝ ˆ πj ξ j τ j ⎠ . As the data are multinomial. Then we calculate that ³ ¡ ¡ ¢−1 ¡ ¢−1 ´0 ³ ¢−1∗ ´ ¡ 0 ¢−1 C + X X 0X − XX = C + X X 0X A0 A − X 0 X ¡ 0 ¢−1 ¡ 0 ¢−1 0 ¡ ¢−1 0 ¡ 0 ¢−1 ¡ 0 ¢−1 = C 0C + C 0X X X + XX X C + X 0X XX XX − XX = C 0C The matrix C 0 C is positive semi-deﬁnite (see Appendix 2. ¥ The Gauss-Markov theorem is an eﬃciency justiﬁcation for the least-squares estimator. and is a strong justiﬁcation for the least-squares estimator.3) can be rewritten as ⎞−1 ⎛ ⎞ ⎛ r r X X β=⎝ πj ξ j ξ 0 ⎠ ⎝ πj ξ j τ j ⎠ (4. for ﬁnite r. r for some constant vectors τ j . Set C = A − X (X 0 X)−1 .

ˆ r X j=1 πj ξ j ξ 0 ˆ j r n X1X ¡ ¢ = 1 (yi = τ j ) 1 xi = ξ j ξ j ξ 0 j n j=1 = 1 n 1 n i=1 n X X r i=1 j=1 n X i=1 ¡ ¢ 1 (yi = τ j ) 1 xi = ξ j xi x0 i = and r X j=1 xi x0 i πj ξ j τ j ˆ r n X1X ¡ ¢ = 1 (yi = τ j ) 1 xi = ξ j ξ j τ j n j=1 i=1 = r n ¡ ¢ 1 XX 1 (yi = τ j ) 1 xi = ξ j xi yi n = Thus ˆ β mle = Ã 1 n n i=1 j=1 n X i=1 xi yi !−1 Ã ! In other words.Substituting in the expressions for π j . Chamberlain (1987) extends this argument to the case of continuously-distributed data.10 Omitted Variables xi = µ x1i x2i ¶ .22) 34 .9). He proves that generically the OLS estimator (4. Since this is a regular parametric model the MLE is asymptotically eﬃcient (see Appendix A.10) for the class of models satisfying Assumption 3. We can write the model as yi = x0 β 1 + x0 β 2 + ei 1i 2i E (xi ei ) = 0 where the parameter of interest is β 1 . the maximum likelihood estimator is identical to the OLS estimator.5. if the data have a discrete distribution. 1X xi x0 i n i=1 1X xi yi n i=1 n ˆ = β ols 4. and any continuous distribution can be arbitrarily well approximated by a multinomial distribution.4) is an asymptotically eﬃcient estimator for the parameter β deﬁned in (3. He observes that the above argument holds for all multinomial distributions.1. (4. and thus so is the OLS estimator. Let the regressors be partitioned as Suppose we are interested in the coeﬃcient on x1i alone in the regression of yi on the full set xi .

when the columns of X are close to linearly dependent. except in special cases. log(p2 ) and log(p1 /p2 ). This is the situation when the X 0 X matrix is near singular.23) is an estimate of γ 1 = β 1 + Γβ 2 rather than β 1 . It is the consequence of omission of a relevant correlated variable. i.11 Multicollinearity ˆ If rank(X 0 X) < k + 1. this is rarely a problem for applied econometric practice.22) by least-squares. Goldberger (1991) calls (4. if X includes both the logs of two prices and the log of the relative prices. This happens when the columns of X are linearly dependent. First. 35 . the regression of x2i on x1i yields a set of zero coeﬃcients (they are uncorrelated). Typically. or second.22) is zero. 4. This is called strict multicollinearity. Most commonly. This is estimation of the equation yi = x0 γ 1 + ui 1i E (x1i ui ) = 0 Notice that we have written the coeﬃcient on x1i as γ 1 rather than β 1 . When this happens. In general. This is because the model being estimated is diﬀerent than (4. Observe that γ 1 6= β 1 unless Γ = 0 or β 2 = 0. Typically there are limits. and the error as ui rather than ei .23) the short regression to emphasize the distinction. we calculate γ1 = ¡ ¡ ¢¢−1 E x1i x0 E (x1i yi ) 1i ¡ ¡ ¢¢−1 ¡ ¡ 0 ¢¢ 0 E x1i x1i β 1 + x0 β 2 + ei = E x1i x1i 2i ¢¢−1 ¡ ¢ ¡ ¡ E x1i x0 β 2 = β 1 + E x1i x0 1i 2i = β 1 + Γβ 2 ¢¢−1 ¡ ¢ ¡ ¡ E x1i x0 Γ = E x1i x0 1i 2i (4. In this case. This is one diﬃculty with the deﬁnition and interpretation of multicollinearity.22) the long regression and (4. least-squares estimation of (4. By construction. This deﬁnition is not precise. Since the error is discovered quickly..22). the general model will be free of the omitted variables problem. which is often called “multicollinearity” for brevity. The more relevant issue is near multicollinearity. Thus the short and long regressions have the same coeﬃcient on x1i only under one of two conditions. To avoid omitted variables bias the standard advice is to include potentially relevant variables in the estimated model. the applied researcher quickly discovers the error as the statistical software will be unable to construct (X 0 X)−1 . the coeﬃcient on x2i in (4.23) where is the coeﬃcient from a regression of x2i on x1i . as many desired variables are not available in a given dataset. this arises when sets of regressors are included which are identically related. log(p1 ).Now suppose that instead of estimating equation (4.e. we regress yi on x1i only. then β is not deﬁned. there is some α such that Xα = 0. β 1 6= γ 1 . because we have not said what it means for a matrix to be “near singular”. To see this. For example. The diﬀerence Γβ 2 is known as omitted variable bias. the possibility of omitted variables bias should be acknowledged and discussed in the course of an empirical investigation.

A more relevant implication of near multicollinearity is that individual coeﬃcient estimates will be imprecise. We can see this most simply in a homoskedastic linear regression model with two regressors yi = x1i β 1 + x2i β 2 + ei . and 1 0 XX= n µ 1 ρ ρ 1 ¶ In this case The correlation ρ indexes collinearity.24) β (−i) = X(−i) X(−i) is the i’th diagonal element of the projection matrix X (X 0 X)−1 X 0 . it is statistically diﬃcult to disentangle the impact of β 1 from that of β 2 . We can see the eﬀect of collinearity on precision by observing that the asymptotic variance of a coeﬃcient ¡ ¢−1 approaches inﬁnity as ρ approaches 1. 1] and sum to k. To investigate the possibility of inﬂuential observations.27) (4. In extreme cases it is possible that the reported calculations will be in error due to ﬂoating-point calculation diﬃculties. = V ar β ρ 1 n n (1 − ρ2 ) −ρ 1 4. the precision of individual estimates are reduced. A convenient alternative expression is ¡ ¢−1 ˆ ˆ xi ei ˆ (4. This equals ´−1 ³ 0 ˆ X(−i) Y(−i) (4. We can also deﬁne the leave-one-out residual ˆ ˆ ei. the worse the precision of the individual coeﬃcient estimates.−i = yi − x0 β (−i) = (1 − hi )−1 ei . As a consequence. Thus the more “collinear” are the estimate σ 2 1 − ρ2 regressors. since as ρ approaches 1 the matrix becomes singular.One implication of near singularity of matrices is that the numerical reliability of the calculations is reduced. We derive expression (4. µ ¶−1 µ ¶ ³ ´ 2 σ2 1 ρ 1 −ρ ˆ|X = σ . the OLS estimator based on the sample excluding the i’th observation.12 Inﬂuential Observations where X(−i) and Y(−i) are the data matrices omitting the i’th row. ˆ i A simple comparison yields that ˆ ei − ei. the change in the coeﬃcient estimate by deletion of the i’th observation depends critically on the magnitude of hi .−i = (1 − hi )−1 hi ei . The hi take values in [0.26) As we can see.25) β (−i) = β − (1 − hi )−1 X 0 X where ¢−1 ¡ xi hi = x0 X 0 X i The i’th observation is inﬂuential on the least-squares estimate if the deletion of the observation ˆ from the sample results in a meaningful change in β.25) below. that is. What is happening is that when the regressors are highly dependent. ˆ ˆ (4. deﬁne the leave-one-out least-squares estimator of β. If the i’th observation 36 .

This implies has a large value of hi . ˆ = β − (1 − hi )−1 X 0 X and thus ˆ β (−i) = 37 . ˆ We now derive equation (4. ¢−1 ¡ 0 ¢−1 ¡ 0 ¢−1 ¡ 0 ¢−1 ¡ = XX + XX xi (1 − hi )−1 x0 X 0 X X X − xi x0 i i ¡ 0 ¢−1 ¡ 0 ¢ X X − xi x0 X Y − xi yi i ¢ ¡ ¡ ¢ ¡ ¢−1 ¡ 0 ¢−1 ¢−1 ¡ 0 X Y − xi yi + (1 − hi )−1 X 0 X X Y − xi yi = X 0X xi x0 X 0 X i ¢−1 ¡ ˆ xi ei .27). Investigations into the presence of inﬂuential observations can plot the values of (4. which is considerably more informative than plots of the uncorrected residuals ei . The key is equation (2.4 which states that ¡ ¢ (A + BCD)−1 = A−1 − A−1 BC C + CDA−1 BC CDA−1 . then this observation is a leverage point and has the potential to be an inﬂuential observation.25).1) in Section 2.

4.13

Exercises

1. Let X be a random variable with µ = EX and σ 2 = V ar(X). Deﬁne µ ¶ ¢ ¡ x−µ 2 . g x, µ, σ = (x − µ)2 − σ 2

µ ˆ Let (ˆ , σ 2 ) be the values such that g n (ˆ , σ 2 ) = 0 where g n (m, s) = n−1 µ ˆ 2 Show that µ and σ are the sample mean and variance. ˆ ˆ

Pn

i=1 g

¡ ¢ Xi , µ, σ 2 .

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.

0ˆ 3. Let e be the OLS residual from a regression of Y on X = [X1 X2 ]. Find X2 e. ˆ

4. Let e be the OLS residual from a regression of Y on X. Find the OLS coeﬃcient estimate ˆ from a regression of e on X. ˆ ˆ 5. Let y = X(X 0 X)−1 X 0 y. Find the OLS coeﬃcient estimate from a regression of y on X. ˆ ˆ 6. Prove that R2 is the square of the simple correlation between y and y. Pn 1 7. Explain the diﬀerence between n i=1 xi x0 and E (xi x0 ) . i i

0 0 ˆ 8. Let β n = (Xn Xn )−1 Xn Yn denote the OLS estimate when Yn is n × 1 and Xn is n × k. A new observation (yn+1 , xn+1 ) becomes available. Prove that the OLS estimate computed using this additional observation is ³ ´ ¡ 0 ¢−1 1 ˆ ˆ ˆ xn+1 yn+1 − x0 β n . β n+1 = β n + Xn Xn n+1 0 1 + x0 (Xn Xn )−1 xn+1 n+1

9. True or False. If yi = xi β + ei , xi ∈ R, E(ei | xi ) = 0, and ei is the OLS residual from the ˆ P ˆ regression of yi on xi , then n x2 ei = 0. i=1 i 10. 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 the three regressions Y Y Y = µ + D1 α1 + D2 α2 + e = D1 α1 + D2 α2 + e = µ + D1 φ + e (4.28) (4.29) (4.30)

(a) Can all three regressions (4.28), (4.29), and (4.30) be estimated by OLS? Explain if not. (b) Compare regressions (4.29) and (4.30). Is one more general than the other? Explain the relationship between the parameters in (4.29) and (4.30). 38

(d) Letting α = (α0 α0 )0 , write equation (4.29) as Y = Xα + e. Consider the assumption 1 2 E(xi ei ) = 0. Is there any content to this assumption in this setting? 11. Let D1 and D2 be deﬁned as in the previous exercise. (a) In the OLS regression Y = D1 γ 1 + D2 γ 2 + u, ˆ ˆ ˆ show that γ 1 is sample mean of the dependent variable among the men of the sample ˆ (Y 1 ), and that γ 2 is the sample mean among the women (Y 2 ). ˆ (b) Describe in words the transformations Y ∗ = Y − D1 Y 1 + D2 Y 2

(c) Compute ι0 D1 and ι0 D2 , where ι is an n × 1 is a vector of ones.

X ∗ = X − D1 X 1 + D2 X 2 . ˜ (c) Compare β from the OLS regresion ˜ ˜ Y ∗ = X ∗β + e ˆ with β from the OLS regression ˆ ˆ Y = D1 α1 + D2 α2 + X β + e. ˆ ˆ 12. The data ﬁle cps85.dat contains a random sample of 528 individuals from the 1985 Current Population Survey by the U.S. Census Bureau. The ﬁle contains observations on nine variables, listed in the ﬁle cps85.pdf. V1 V2 V3 V4 V5 V6 V7 V8 V9 = = = = = = = = = education (in years) region of residence (coded 1 if South, 0 otherwise) (coded 1 if nonwhite and non-Hispanic, 0 otherwise) (coded 1 if Hispanic, 0 otherwise) gender (coded 1 if female, 0 otherwise) marital status (coded 1 if married, 0 otherwise) potential labor market experience (in years) union status (coded 1 if in union job, 0 otherwise) hourly wage (in dollars)

Estimate a regression of wage yi on education x1i , experience x2i , and experienced-squared x3i = x2 (and a constant). Report the OLS estimates. 2i ˆ Let ei be the OLS residual and yi the predicted value from the regression. Numerically ˆ calculate the following: (a) (b) (c) (d) (e) Pn Pn Pn Pn ˆ i=1 ei ˆ i=1 x1i ei

ˆ i=1 x2i ei 2 ˆ i=1 x1i ei Pn 2 ˆ i=1 x2i ei

39

(f) (g) (h)

Pn

Pn

ˆˆ i=1 yi ei ˆ2 i=1 ei R2

Are the calculations (i)-(vi) consistent with the theoretical properties of OLS? Explain. 13. Use the data from the previous problem, restimate the slope on education using the residual regression approach. Regress yi on (1, x2i , x2 ), regress x1i on (1, x2i , x2 ), and regress the 2i 2i residuals on the residuals. Report the estimate from this regression. Does it equal the value from the ﬁrst OLS regression? Explain. In the second-stage residual regression, (the regression of the residuals on the residuals), calculate the equation R2 and sum of squared errors. Do they equal the values from the initial OLS regression? Explain.

40

1 Inequalities Asymptotic theory is based on a set of approximations. Cauchy-Schwarz Inequality. We list here some of the most critical deﬁnitions and inequalities.1) (5. These approximations are bounded through the use of mathematical inequalities. |a| = a a i i=1 If A is a m × n matrix. then its Euclidean norm is The following are an important set of inequalities which are used in asymptotic distribution theory. Triangle inequality |X + Y | ≤ |X| + |Y | . The Euclidean norm of an m × 1 vector a is Ã m !1/2 X ¡ 0 ¢1/2 = a2 . 41 . ij i=1 j=1 (5. If p > 1 and q > 1 and 1 p ⎞1/2 ⎛ m n XX ¡ 0 ¢1/2 |A| = tr A A =⎝ a2 ⎠ .3) E |XY | ≤ (E |X|p )1/p (E |Y |q )1/q . 5. If g(·) : R → R is convex. then g(E(X)) ≤ E(g(X)). then (5.2) + 1 q = 1.Chapter 5 Asymptotic Theory This chapter reviews the essential components of asymptotic theory. ´1/2 ³ ´1/2 ³ E |Y |2 E |XY | ≤ E |X|2 Holder’s Inequality. Jensen’s Inequality.

This is a probabilistic way of generalizing the mathematical deﬁnition of a limit. Applying expectations. =E U V p q p q Proof of Markov’s Inequality. If Xi ∈ Rk is iid and E |Xi | < ∞. if for all δ > 0. denoted Zn →p Z as n → ∞. ¥ Proof of Holder’s Inequality.2 Weak Law of Large Numbers Let Zn ∈ Rk be a random vector.2.4) Proof of Jensen’s Inequality.Markov’s Inequality. The WLLN shows that sample averages converge in probability to the population average. We say that Zn converges in probability to Z as n → ∞. g(x) ≥ a + bx yet g(EX) = a + bEX since the curve is tangent at EX.1 Weak Law of Large Numbers (WLLN). (E |X|p )1/p (E |Y |q )1/q ¥ µ ¶ ´ ³ U V 1 1 1/p 1/q ≤E + = + = 1. as stated. Since 1 + 1 = 1 an application of Jensen’s inequality shows that p q U Then 1/p V 1/q 1 1 ln U + ln V = exp p q E |XY | ∙ ¸ ≤ 1 1 U V exp (ln U ) + exp (ln V ) = + . Eg(X) ≥ a + bEX = g(EX). tangent lines lie below it. (5. ¥ 5. Let a + bx be the tangent line to g(x) at x = EX. Note EU = EV = 1. For any strictly increasing function g(X) ≥ 0. Let U = |X|p /E |X|p and V = |Y |q /E |Y |q . Since g(x) is convex.3). p p p q which is (5. Theorem 5. n→∞ lim P (|Zn − Z| > δ) = 0. n i=1 42 . P (g(X) > α) ≤ α−1 Eg(X). So for all x. Then Z ∞ −1 P (Y > α) = α αf (y)dy Zα∞ ≤ α−1 yf (y)dy α Z ∞ ≤ α−1 yf (y)dy = α−1 E(Y ) −∞ the second-to-last inequality using the region of integration {y > α}. Set Y = g(X) and let f denote the density function of Y. then as n → ∞ n 1X Xn = Xi →p E(X).

6) and (5. Finally.3 Convergence in Distribution Let Zn be a random variable with distribution Fn (x) = P (Zn ≤ x) .7).5) (where 1 (·) is the indicator function) which is possible since E |X| < ∞.4). the fact that X n = W n + Z n . Pick C < ∞ large enough so that E (|X| 1 (|X| > C)) ≤ ε (5. We say that Zn converges in distribution to Z as n → ∞. where Z has distribution F (x) = P (Z ≤ x) . V ) . P ¯X n ¯ > δ ≤ δ δ δ the equality by the deﬁnition of ε. Fn (x) → F (x) as n → ∞. ¯ ¯ E ¯Z n ¯ ≤ E |Zi | ≤ 2ε.Proof: Without loss of generality. denoted Zn →d Z. ¯¢2 ¡ ¯ 2 ≤ EW n E ¯W n ¯ = EWi2 n 4C 2 ≤ n ≤ ε2 (5. (5.1 Central Limit Theorem (CLT). Jensen’s inequality (5. We need to ¢show that for all δ > 0 and η > 0 there is some N < ∞ so that for all n ≥ N. P ¯X n ¯ > δ ≤ η. by Markov’s inequality (5. If Xi ∈ Rk is iid and E |Xi |2 < ∞. the triangle inequality.1) and (5. then as n→∞ n ¢ √ ¡ 1 X n Xn − µ = √ (Xi − µ) →d N (0. n i=1 0 where µ = EX and V = E (X − µ) (X − µ) .3. as needed. if for all x at which F (x) is continuous.5). ≤ E |Xi | 1 (|Xi | > C) + |E (Xi 1 (|Xi | > C))| (5. the fact that the Wi are iid and mean zero. ¯ ¯ ¯ ¯ ¯ ¯ ¢ E ¯X n ¯ ¡¯ ¯ E ¯W n ¯ + E ¯Z n ¯ 3ε ≤ ≤ = η. we can set E(X) = 0 (by recentering Xi on its expectation). ¡¯ ¯ P ¯X n ¯ > δ ≤ η. and the bound |Wi | ≤ 2C. Fix δ and η. Set ε = δη/3. Deﬁne the random vectors Wi = Xi 1 (|Xi | ≤ C) − E (Xi 1 (|Xi | ≤ C)) By the triangle inequality. 43 . We have shown that for any δ > 0 and η > 0 then for all ¢ ¡¯ ¯ ¥ n ≥ 36C 2 /δ 2 η 2 .7) the ﬁnal inequality holding for n ≥ 4C 2 /ε2 = 36C 2 /δ 2 η 2 . ≤ 2E |Xi | 1 (|Xi | > C) Zi = Xi 1 (|Xi | > C) − E (Xi 1 (|Xi | > C)) .6) By Jensen’s inequality (5. Theorem 5.1). 5.

1 1 c(λ) = c(0) + cλ (0)0 λ + λ0 cλλ (λ∗ )λ = λ0 cλλ (λ∗ )λ 2 2 (5. the deﬁnition of c(λ) and (5. it is suﬃcient to consider the case µ = 0 and V = Ik .8) ⎞ ⎛ nλ0 X 1 Xj ⎠ ln Cn (λ) = log E exp ⎝i √ n j=1 µ ¶ n Y 1 exp i √ λ0 Xj = log E n j=1 µ ¶ n Y 1 0 E exp i √ λ Xj = log n i=1 ¶ µ λ = nc √ n 1 0 λ cλλ (λn )λ = 2 44 . ¢ ¡ √ √ We now compute Cn (λ) = E exp iλ0 nX n the characteristic function of nX n . ∂λ∂λ cλ (λ) = cλλ (λ) = ∂ ∂ c(λ) = C(λ)−1 C(λ) ∂λ ∂λ ∂ ∂2 ∂2 −2 ∂ C(λ) 0 C(λ) c(λ) = C(λ)−1 0 0 C(λ) − C(λ) ∂λ ∂λ∂λ ∂λ∂λ ∂λ so when evaluated at λ = 0 Furthermore. For ¡ ¢ λ ∈ Rk . the independence of the Xi .8) where λ∗ lies on the line segment joining 0 and λ. so when evaluated at λ = 0 c(0) = 0 cλ (0) = 0 cλλ (0) = −Ik . By the properties of the exponential function. By a second-order Taylor series expansion of c(λ) about λ = 0. Then observe ¡ ¡ ¢¢ ∂ C(λ) = iE X exp iλ0 X ∂λ ¡ ¡ 0 ¢¢ ∂2 2 0 0 C(λ) = i E XX exp iλ X ∂λ∂λ C(0) = 1 ∂ C(0) = iE (X) = 0 ∂λ ¡ ¢ ∂2 0 0 C(0) = −E XX = −Ik .Proof: Without loss of generality. let C(λ) = E exp iλ0 X denote the characteristic function of X and set c(λ) = ln C(λ).

we ﬁnd jn ¢ ¡ √ √ 0 n (g (θn ) − g(θ0 )) = (gθ + an ) n (θn − θ0 ) →d gθ N (0. Recall that A ⊂ B implies P (A) ≤ P (B). Hence g(Zn ) →p g(c) as n → ∞. for each element of g. Proof: Since g is continuous at c. then ¡ ¢ √ 0 n (g (θn ) − g(θ0 )) →d N 0.4.1 Continuous Mapping Theorem 1 (CMT). Stacking across elements of g. ¥ 5. gθ Σgθ where gθ (θ) = ∂ g(θ) ∂θ0 and gθ = gθ (θ0 ). k ≤ m.4 Asymptotic Transformations Theorem 5. 45 . Theorem 5. gj (θn ) = gj (θ0 ) + gjθ (θ∗ ) (θn − θ0 ) jn where θnj lies on the line segment between θn and θ0 and therefore converges in probability to θ0 . It follows that ajn = gjθ (θ∗ ) − gjθ →p 0. Thus P (|g (Zn ) − g (c)| ≤ ε) ≥ P (|Zn − c| < δ) → 1 as n → ∞ by the assumption that Zn →p c. µ ¶ 1 Cn (λ) → exp − λ0 λ 2 the characteristic function of the N (0. where θ is m × 1 and Σ is m × m. we see that as n → ∞. and g(θ) : Rm → Rk . If Zn →p c as n → ∞ and g (·) is continuous at c. Proof : By a vector Taylor series expansion.√ where λn → 0 lies on the line segment joining 0 and λ/ n. for all ε > 0 we can ﬁnd a δ > 0 such that if |Zn − c| < δ then |g (Zn ) − g (c)| ≤ ε.3 Delta Method: If n (θn − θ0 ) →d N (0. If Zn →d Z as n → ∞ and g (·) is continuous. √ Theorem 5. then g(Zn ) →d g(Z) as n → ∞.2 Continuous Mapping Theorem 2. Since cλλ (λn ) → cλλ (0) = −Ik .4. Σ) .4. gθ Σgθ . then g(Zn ) →p g(c) as n → ∞. Σ) = N 0. Ik ) distribution. This is suﬃcient to establish the theorem.

In general.Chapter 6 Inference 6.1: Sampling Density of β 2 To illustrate the possibilities in one example. xi ) and the sample size n. The verticle line marks the true value of the projection coeﬃcient. and therefore has a sampling distribution. Using ˆ simulation methods. y) = 2πxy 2 2 ˆ and let β 2 be the slope coeﬃcient estimate computed on observations from this joint density.1 Sampling Distribution The least-squares estimator is a random vector.1 for sample sizes of n = 25. the density function of β 2 was computed and plotted in Figure 6. its distribution is a complicated function of the joint distribution of (yi . 46 . n = 100 and n = 800. ˆ Figure 6. since it is a function of the random data. let yi and xi be drawn from the joint density µ ¶ µ ¶ 1 1 1 2 2 exp − (ln y − ln x) exp − (ln x) f (x.

A preliminary step in this approach is the demonstration that estimators are consistent — that they converge in probability to the true parameters as the sample size gets large. 0). we can conclude ˆ that β →p β. (6. xi ei →p g (Q.1). 6.1).1) and ˆ We now deduce the consistency of β.1). (6.5. we will use the methods of asymptotic approximation.1) imply that n ¡ ¢ 1X xi x0 →p E xi x0 = Q (6. the OLS estimator β is has a statistical distribution which is unknown. Hence by the continuous mapping theorem (Theorem 5.3) Ã where g(A. Proof. This is illustrated in Figure 6. (6.8) implies that ˆ β−β = Ã n X i=1 xi x0 i !−1 n X i=1 xi ei .1. Assumption 3.3).1 Under Assumption 3.5.2. b) = A−1 b is a continuous function of A and b at all values of the arguments such that A−1 exist.4 implies that Q−1 exists and thus g(·. As the sample size increases the density becomes more concentrated about the population coeﬃcient. Assumption 3. Ã n ! n X 1X 0 1 ˆ xi xi .1 and the WLLN (Theorem 5.2 Consistency ˆ As discussed in Section 6.5. using (6. β →p β as n → ∞.1 by the fact that the sampling densities become more concentrated as n gets larger. n i=1 (6.4. ˆ Theorem 6. For a complete argument. To characterize the sampling distribution more fully.4. ¥ We can similarly show that the estimators σ 2 and s2 are consistent for σ 2 .1. xi ei = g i n n i=1 i=1 1X xi ei →p E (xi ei ) = 0. 0) = Q−1 0 = 0 β−β =g n n i=1 i=1 ˆ which implies β →p β as stated. First. ·) is continuous at (Q. Asymptotic (large sample) methods approximate sampling distributions based on the limiting experiment that the sample size n tends to inﬁnity. Equation (4.2).2.2) i i n i=1 n From (6. and the continuous mapping theorem (Theorem 5.From the ﬁgure we can see that the density functions are dispersed and highly non-normal.1.1). we can write !−1 Ã n ! n 1X 1X 0 ˆ β−β = xi xi xi ei n n i=1 i=1 Ã n ! n X X 1 1 xi x0 . ˆ 47 .

σ 2 →p σ 2 and s2 →p σ 2 as n → ∞.5. Ee4 < ∞ and E |xi |4 < ∞.4) ³ ´ ³ ´0 ³ ´ ˆ ˆ ˆ e2 = e2 − 2ei x0 β − β + β − β xi x0 β − β ˆi i i i σ2 = ˆ 1X 2 ei ˆ n i=1 i=1 →p σ 2 n 1X 2 = ei − 2 n Ã ! Ã n ! n ´ ³ ´0 1 X ³ ´ 1 X 0 ³ˆ 0 ˆ ˆ β−β + β−β β−β ei xi xi xi n n i=1 i=1 the last line using the WLLN. (6. it follows that s2 = ¥ n σ 2 →p σ 2 . (6. i Now deﬁne ¡ ¢ Ω = E xi x0 e2 . Assumption 6.1).3.2).3 Asymptotic Normality ˆ We now establish the asymptotic distribution of β after normalization. (6.2 Under Assumption 3. We need a strengthening of the moment conditions.2). i i Assumption 6.2. (6. To see this.5) Thus xi ei is iid with mean zero and has covariance matrix Ω.1). n 1 X √ xi ei →d N (0.Theorem 6.6) n i=1 48 . ´1/2 ¡ ¯ ¯¢ ¯ ³ ¯ ¯2 ´1/2 ¡ ¯ 4 ¯¢1/2 ³ ¯ 1/2 E ¯xi x0 e2 ¯ ≤ E ¯xi x0 ¯ = E |xi |4 < ∞.3.1 guarantees that the elements of Ω are ﬁnite.2. ˆ n−k 6.3) and Theorem (6.3. Ω) . By the central limit theorem (Theorem 5. by the Cauchy-Schwarz inequality (5. ˆ Proof.1 In addition to Assumption 3.1.5. Note that ˆ ei = yi − x0 β ˆ i i Thus and n ˆ = ei + x0 β − x0 β i i ³ ´ ˆ = ei − x0 β − β . E ¯ei ¯ E ¯e4 ¯ i i i i (6.1. ˆ Finally. since n/(n − k) → 1 as n → ∞. Thus σ 2 is consistent for σ 2 .

otherwise V 6= V 0 . When (6. 1) asymptotic distibution.1 Under Assumption 6. If α > 2 the 2 error εi has zero mean and variance α/(α − 2).1.7) holds. ´ √ ³ ˆ n β−β = Ã 1X xi x0 i n i=1 n !−1 Ã Theorem 6. and (6.2). However. e2 ) = 0.Then using (6.3.7) the asymptotic variance formulas simplify as ¡ ¢ ¡ ¢ (6. there is no simple answer to this reasonable question. ´ √ ³ˆ As V is the variance of the asymptotic distribution of n β − β . 1). for example. The expression V = Q−1 ΩQ−1 is called a sandwich form. The trouble is that no matter how large is the sample size. after rescaling. Q−1 ΩQ−1 .7) Cov(xi x0 . We say that ei is a Homoskedastic Projection Error when (6. How large should n be in order for the approximation to be useful? Unfortunately.8) Ω = E xi x0 E e2 = Qσ 2 i i V = Q−1 ΩQ−1 = Q−1 σ 2 ≡ V 0 (6.9) we deﬁne V 0 = Q−1 σ 2 whether (6. V is often referred to as ˆ the asymptotic covariance matrix of β. is approximately normal when the sample size n is suﬃciently large. xi ) which satisfy the conditions of Assumption 6.3. There is a special case where Ω and V simplify.7) is true then V = V 0 . In this context the normalized least-squares slope estimator n α−2 β 2 − β 2 has α →d Q−1 N (0. however. We illustrate this problem using a simulation. i i Condition (6.3. Let yi = β 0 + β 1 xi + εi where xi is N (0.1 we have already seen a simple example where the least-squares estimate is quite asymmetric and non-normal even for reasonably large sample sizes. as n → ∞ ´ √ ³ ˆ n β − β →d N (0. and εi is independent of xi with the Double Pareto density f (ε) = α |ε|−α−1 .3. its variance diverges q ³ ´ ˆ to inﬁnity. V ) where V = Q−1 ΩQ−1 .1.7) is true or false. The asymptotic distribution ´ Theorem 6. for any ˆ ﬁxed n the sampling distribution of β can be arbitrarily far from the normal distribution. setting n = 100 and varying the parameter α. As α approaches 2. Ω) ¡ ¢ = N 0. Theorem 6.3. We call V 0 the homoskedastic covariance matrix. when xi and ei are independent. the normal approximation is arbitrarily poor for some data distribution satisfying the assumptions.1 is commonly used to approximate the ﬁnite of √ ³ˆ sample distribution of n β − β . This holds true for all joint distibutions of (yi . (6. For α 49 . but this is not a necessary condition. 1 X √ xi ei n i=1 n ! the N (0. In Figure 6.2 we display the ﬁnite sample densities of the q ³ ˆ normalized estimator n α−2 β 2 − β 2 . Under (6.6). The approximation may be poor when n is small. |ε| ≥ 1. In´Figure 6.1 states that the sampling distribution of the least-squares estimator.9) In (6.1).

4 Let Covariance Matrix Estimation 1X ˆ xi x0 Q= i n i=1 n be the method of moments estimator for Q. the sampling distribution becomes highly skewed and non-normal. The homoskedastic covariance matrix V 0 = Q−1 σ 2 is typically estimated by ˆ ˆ (6. 4. ˆ ˆ Since Q →p Q and s2 →p σ 2 (see (6.2.3 is that the N (0. concentrating most of the probability mass around zero.11) 50 .10) without changing this result. As k increases. 6 and 8. The lesson from Figures 6.0 the density is very close to the N (0.10) V 0 = Q−1 s2 . σ may also be substituted for s ˆ ¡ ¢ To estimate V = Q−1 ΩQ−1 .Figure 6. for k = 1. The MME estimator is i i 1X ˆ xi x0 e2 Ω= i ˆi n i=1 n (6. Another example is shown in Figure 6.1) it is clear that V 0 →p V 0 . Here the model is yi = β 1 + εi where uk − Euk i i εi = ³ ¡ k ¢2 ´1/2 Eu2k − Eui i ´ √ ³ˆ and ui ∼ N (0.2 and 6.2) and Theorem 6.2: Density of Normalized OLS estimator α = 3. The estimator 2 2 in (6.3. 6. We show the sampling distribution of n β 1 − β 1 setting n = 100. As α diminishes the density changes signiﬁcantly. we need an estimate of Ω = E xi x0 e2 . 1) density. 1) asymptotic approximation is never guaranteed to be accurate. 1).

as it is not always clear which has been computed.. standard errors are not unique. the “Huber formula”. we focus on individual elements of β one-at-a-time. Generically.4.. ˆ ˆ The variance estimator V is an estimate of the variance of the asymptotic distribution of β. vis.1 A standard error s(β) for an estimator β is an estimate of the standard ˆ deviation of the distribution of β. β j . and V is an estimator of the variance of n β − β . ˆ ˆ Deﬁnition 6. A more easily interpretable measure of spread is its square root — the standard deviation. it is important to pay ˆ ˆ attention to the distinction between V 0 and V . as there may be more than one estimator of the variance of the estimator. ˆ ˆ When V is used rather than the traditional choice V 0 . This motivates the deﬁnition of a standard error. ˆ The estimator V 0 was the dominate covariance estimator used before 1980.Figure 6. the “Huber-White formula” or the “GMM covariance matrix”. j = 0.3: Sampling distribution where ei are the OLS residuals. In most cases. . the “Eicker-White formula”. or that they use the “White formula”. Thus q ˆ ˆ s(β j ) = n−1/2 Vjj . When reading and reporting applied work. these all mean the same thing. k. ´ p √ ³ˆ ˆ ˆ ˆ When β is scalar. The estimator of V is then ˆ ˆ ˆ ˆˆ V = Q−1 ΩQ−1 This estimator was introduced to the econometrics literature by White (1980).. and was still the standard choice for much empirical work done in the early 1980s. The methods switched during ˆ the late 1980s and early 1990s.. When β is a vector. 1. so that by the late 1990s White estimate V emerged as the standard covariance matrix estimator. we set s(β) = n−1/2 V . or that they use a “heteroskedasticity-robust covariance matrix estimator”. It is therefore important to understand what formula and method is 51 . many authors will state that their “standard errors have been corrected for heteroskedasticity”.

085 p ˆ and that for β 1 is . (. n n i=1 52 .199 2.35/988 = . and then the square roots.20 −0.035 ¶ .5) and the WLLN (Theorem 5.80 40.1) show that n ¡ ¢ 1X xi x0 e2 →p E xi x0 e2 = Ω. First. The standard errors for β 0 are 7.20 = V 14.80 2.2.83 −0. the standard method to calculate the standard errors is to ˆ ﬁrst calculate n−1 V .199 2.3) so by the WLLN ³ ´ ³ ´3/4 ¡ ¯ ¯¢ 1/4 E |xi |3 |ei | ≤ E |xi |4 < ∞.6 ¶µ 1 14.80 40. E ¯e4 ¯ i ³ ´ 1X |xi |3 |ei | →p E |xi |3 |ei | . by Holder’s inequality (5.14 14.98 −0.480 .6 Therefore the two covariance matrix estimates are µ ¶−1 µ ¶ 1 14. p ˆ In this case the two estimates are quite similar.14 205.493 0.117 Educationi .83 ¶−1 = µ 7.12) in turn.4) ˆ Ω = = n n n ³ ´0 ³ ´0 ³ ´ 1X 2X 1X 0 2 0 ˆ ˆ ˆ xi xi ei − xi xi β − β xi ei + xi x0 β − β xi x0 β − β .12) i i n n n i=1 i=1 We now examine each sum on the right-hand-side of (6. From a computational standpoint. It is also important to understand that a particular standard error may be relevant under one set of model assumptions.1.493 −0.020.used by an author when studying their work. Using (6.14 14.83 ¶−1 µ . then take the diagonal elements. We calculate that s2 = 0.14 6.480 ˆ0 = 0. Ω 2. we return to the log wage regression of Section 4.085) (. To illustrate. (6.020) 6. just as any other estimator.30 + 0. but not under another set of assumptions.14 205. from which it follows that V →p V as n → ∞. i i i i n i=1 Second. We can write the estimated equation with standards errors using the format \ log(W agei ) = 1. (6.039 and ˆ V = µ 1 14.14 205.5 Consistency of the White Covariance Matrix Estimate 1X xi x0 e2 i ˆi n i=1 i=1 n ˆ ˆ We now show Ω →p Ω.2/988 = .80 .20 and µ ¶ ˆ = 0.

11) with the leave-one-out estimator ei. Using this substitution.13) of Efron (1982). ¯ ¯n ¯ n i=1 so Together.25).26). Andrews (1991) suggested an similar estimator based on cross-validation. MacKinnon and White (1985) present numerical (simulation) evidence i ˆ ˆ that V ∗ works better than V as an estimator of V . these establish consistency.5. the jackknife estimator of the ˆ variance matrix for β is ˆ V ∗ = (n − 1) where n X³ ˆ β i=1 ¯ (−i) − β n ´³ ˆ β ¯ (−i) − β ´0 (6. ¯ ¯ ¯ ¯ ¯n ¯ n i=1 i=1 1X |xi |4 →p E |xi |4 . n i=1 n ˆ ˆ Theorem 6.13) Using formula (4. i=1 Third. you can show that 1 Xˆ ¯ β= β (−i) . They also suggest that the scaling factor (n − 1)/n in (??) can be omitted.14) Ã 1X (1 − hi )−1 xi ei ˆ n i=1 n !Ã 1X (1 − hi )−1 xi ei ˆ n i=1 n !0 and hi = x0 (X 0 X)−1 xi .6 Alternative Covariance Matrix Estimators ˆ MacKinnon and White (1985) suggested a small-sample corrected version of V based on the jackˆ knife principle.12 the deﬁnition of β (−i) as the least-squares estimator with the i’th observation deleted. Recall from Section 4.1 As n → ∞. From equation (3. n i=1 n − 1 ˆ −1 ˆ ∗ ˆ −1 ˆ V∗ = Q Ω Q n where 1X ˆ (1 − hi )−2 xi x0 e2 − Ω∗ = i ˆi n i=1 n (6. is deﬁned ˆ ˆ by replacing the OLS residual ei in (6. Andrews’ proposed estimator is ˆ ˆ ˆ ˆ V ∗∗ = Q−1 Ω∗∗ Q−1 53 .−i = (1 − hi )−1 ei ˆ presented in (4. ¯ n ¯ n ¯2 X ¯1 X ³ ´0 ³ ´¯ ¯ ¯ 0 ˆ ˆ ˆ − β ¯ ≤ ¯β − β ¯ 1 xi xi β − β xi xi β |xi |4 →p 0. which.¯ ¯ ¯ ¯ˆ and thus since ¯β − β ¯ →p 0. 6. by the WLLN ¯ n ¯ Ã n ! ¯ 1X ¯1 X ¯ ¯ ³ ´0 ¯ ¯ ¯ ¯ˆ 3 0 ˆ xi xi β − β xi ei ¯ ≤ ¯β − β ¯ |xi | |ei | →p 0. Ω →p Ω and V →p V.

.7 Functions of Parameters Sometimes we are interested in some lower-dimensional function of the parameter vector β = (β 1 . but omits the mean correction. (6. By a ﬁrst-order Taylor series approximation: ³ ´ 0 ˆ ˆ h(β) ' h(β) + Hβ β − β . we may be interested in a single coeﬃcient β j or a ratio β j /β l . In this case. = N (0.. 54 . then the natural estimate for the variance of ˆ is θ ˆ0 ˆ ˆ ˆ Vθ = Hβ V Hβ where ∂ ˆ ˆ h(β). . The estimate of θ is ˆ = h(β). For example.15) where 0 Vθ = Hβ V Hβ . Hβ = ∂β In many cases. Hβ = Thus ´ √ ³ ´ √ ³ ˆ n ˆ − θ = n h(β) − h(β) θ ³ ´ 0√ ˆ ' Hβ n β − β 0 →d Hβ N (0. Hβ = R and Hβ = R.where ˆ It is similar to the MacKinnon-White estimator V ∗ . Let h : Rk → Rq denote this function and let θ = h(β) denote the parameter of interest. Vθ ). V ) where ∂ h(β) ∂β (k + 1) × q. Ω∗∗ = i ˆi n i=1 n 6. In these cases we can write the parameter of interest as a function of β. the function h(β) is linear: h(β) = R0 β ˆ ˆ ˆ for some k × q matrix R. ˆ θ What is an appropriate standard error for ˆ Assume that h(β) is diﬀerentiable at the true θ? value of β. ˆ ˆ If V is the estimated covariance matrix for β. Andrews ˆ ˆ (1991) argues that simulation evidence indicates that V ∗∗ is an improvement on V ∗ . 1X ˆ (1 − hi )−2 xi x0 e2 . β k+1 ).. so Vθ = R0 V R.

For example. however. ˆ its estimate and s(ˆ its asymptotic θ θ) standard error. 55 . θ could be a single element of β). we say that tn is an exactly pivotal statistic. µ I 0 ¶ ˆ the upper-left block of V . In general. ˆ When q = 1q h(β) is real-valued). A simple null and composite hypothesis takes the form H0 : θ = θ0 H1 : θ 6= θ0 where θ0 is some pre-speciﬁed value. Consider the studentized statistic tn (θ) = Theorem 6. and θ = h(β) is some function of the parameter vector. 1) Proof. θ) β 6.8. if R is a “selector matrix” R= so that if β = (β 1 . see Section X). the statistic tn has an exact t distribution.1 tn (θ) →d N (0. In this case. β 2 ). Vθ ) √ Vθ = N (0.15) tn (θ) = ˆ−θ θ s(ˆ θ) ´ √ ³ˆ n θ−θ q = ˆ Vθ N (0.8 t tests Let θ = h(β) : Rk → R be any parameter of interest. (For example. the standard error for ˆ is the square root of n−1 Vθ . In special cases (such as the normal regression model. we say that tn (θ) is asymptotically pivotal. and is therefore exactly free of unknowns. s(ˆ = n−1/2 H 0 V Hβ . that (so θ ˆ ˆ ˆ is. then θ = R0 β = β 1 and ˆ Vθ = ¡ I 0 ¢ ˆ V µ I 0 ¶ ˆ = V11 . s(ˆ θ) (6. By (6. pivotal statistics are unavailable and so we must rely on asymptotically pivotal statistics. Since the standard normal distribution does not depend on the parameters.16) ¥ Thus the asymptotic distribution of the t-ratio tn (θ) is the standard normal. 1) →d ˆ−θ θ .

6. in that the reader is allowed to pick the level of signiﬁcance α. p-values and hypothesis tests are equivalent since pn < α if and only if |tn | > zα/2 . 1]. 1].96 and z.645. 1). pn →d U [0. tn →d N (0. Thus an equivalent statement of a Neyman-Pearson test is to reject at the α% level if and only if pn < α. or asymptotic p-value function p(t) = P (|Z| > |t|) = 2 (1 − Φ(|t|)) . P (1 − p(tn ) ≤ u) establishing that 1 − pn →d U [0. An alternative approach. is to report an asymptotic p-value. Let zα/2 is the upper α/2 quantile of the standard normal distribution. in contrast to Neyman-Pearson rejection/acceptance reporting where the researcher picks the level. To see this fact. or “accepts” H0 . 1]. then P (Z > zα/2 ) = α/2 and P (|Z| > zα/2 ) = α. Otherwise the test does not reject. from which it follows that pn →d U [0. associated with Fisher. and is a function of the data and hence is / random. The rejection/acceptance dichotomy is associated with the Neyman-Pearson approach to hypothesis testing. note that the asymptotic distribution of |tn | is F (x) = 1 − p(x). 1). That is. It is designed to cover θ with high probability. however. Deﬁne the tail probability. The asymptotic p-value for the above statistic is constructed as follows.025 = 1. 56 .The standard test for H0 against H1 is the t-statistic (or studentized statistic) tn = tn (θ0 ) = ˆ − θ0 θ . A test of asymptotic signiﬁcance α rejects H0 if |tn | > zα/2 . In a sense. Thus P (1 − pn ≤ u) = = P (F (tn ) ≤ u) ¡ ¢ = P |tn | ≤ F −1 (u) ¡ ¢ → F F −1 (u) = u. under H0 . This is because ¡ ¢ P (reject H0 | H0 true) = P |tn | > zα/2 | θ = θ0 ¢ ¡ → P |Z| > zα/2 = α.05 = 1. Either θ ∈ Cn or θ ∈ Cn . Another helpful observation is that the p-value function has simply made a unit-free transformation of the test statistic. regardless of the complication of the distribution of the original test statistic. The p-value is more general. That is.9 Conﬁdence Intervals A conﬁdence interval Cn is an interval estimate of θ. s(ˆ θ) Under H0 . Then the asymptotic p-value of the statistic tn is pn = p(tn ). The coverage probability is P (θ ∈ Cn ). If the p-value pn is small (close to zero) then the evidence against H0 is strong. For example. z. so the “unusualness” of the test statistic can be compared to the easy-to-understand uniform distribution. if Z ∼ N (0.

¢ ¡ ¢ ¡ P (θ ∈ Cn ) = P |tn (θ)| ≤ zα/2 → P |Z| ≤ zα/2 = 1 − α. θ θ) (6. and it is desired to test the joint restrictions simultaneously. the most common professional choice isi95%. The t-test of the previous setion rejects H0 : θ0 = θ if |tn (θ)| > zα/2 where tn (θ) is the t-statistic (6. While there is no hard-and-fast guideline for choosing the coverage probability 1 − α. However we often can calculate the asymptotic coverage probability limn→∞ P (θ ∈ Cn ). We say that Cn has asymptotic (1 − α)% coverage for θ if P (θ ∈ Cn ) → 1 − α as n → ∞. θ The interval has been constructed so that as n → ∞. ˆ + zα/2 s(ˆ .05. We have the null and alternative H0 : θ = θ0 H1 : θ 6= θ0 . This corresponds to selecting the conﬁdence h i h ˆ ± 1. A good method for construction of a conﬁdence interval is the collection of parameter values which are not rejected by a statistical test. Hβ = ∂β The Wald statistic for H0 against H1 is ´0 ³ ´ ³ ˆ θ θ Wn = n ˆ − θ0 Vθ−1 ˆ − θ0 ´0 ³ ´−1 ³ ´ ³ ˆ ˆ ˆ0 ˆ ˆ h(β) − θ0 .96s(ˆ ≈ ˆ ± 2s(ˆ .We typically cannot calculate the exact coverage probability P (θ ∈ Cn ). 6. ˆ The natural estimate of θ is ˆ = h(β) and has asymptotic covariance matrix estimate θ ˆ0 ˆ ˆ ˆ Vθ = Hβ V Hβ where ∂ ˆ ˆ h(β).17) = ˆ − zα/2 s(ˆ θ and Cn is an asymptotic (1 − α)% conﬁdence interval.18) .10 Wald Tests Sometimes θ = h(β) is a q × 1 vector. Thus values of θ within two standard errors of the estimated interval θ θ) θ θ) ˆ are considered “reasonable” candidates for the true value θ. and values of θ outside two standard θ errors of the estimated ˆ are considered unlikely or unreasonable candidates for the true value.16) and zα/2 is the upper α/2 quantile of the standard normal distribution. In this case the t-statistic approach does not work. = n h(β) − θ0 Hβ V Hβ 57 (6. A conﬁdence interval is then constructed as the values of θ for which this test does not reject: © ª Cn = θ : |tn (θ)| ≤ zα/2 ( ) ˆ−θ θ = θ : −zα/2 ≤ ≤ zα/2 s(ˆ θ) i h θ). or α = .

8.84 = z. We have established: ³ ´0 ³ ´ ˆ − θ0 V −1 ˆ − θ0 →d Z 0 V −1 Z = χ2 .3. We know that the Wald statistic takes the form 0 Wn = nˆ Vθ−1 ˆ θ ˆ θ ³ ´−1 ˆ ˆ0 ˆ β 2.5. h(β) = R0 β. where p(x) = P χ2 ≥ x is the tail q q probability function of the χ2 distribution.15) showed that n ˆ − θ →d Z ∼ N (0. ˜ 58 ¡ 0 ¢−1 0 ˜ β 1 = X1 X1 X1 Y .10.05) = 3.1.When h is a linear function of β. then the Wald statistic can be written in the form ¶ µ 2 ˆ σ − σ2 ˜ Wn = n (6. = nβ 2 R0 V R R0 β is linear with R = µ 0 I ¶ ¢−1 ¡ ˆ ˆ ˆ . ˜˜ n ˜ e = Y − X1 β 1 . Hence β β by Theorem A. then Wn →d χ2 .11 F Tests Y = X1 β 1 + X2 β 2 + e Take the linear model where X1 is n × k1 and X2 is n × k2 and k + 1 = k1 + k2 . 6.2.19) σ2 ˆ where σ2 = ˜ 1 0 e e. and there are q = k2 restrictions. Thus Vθ = H 0 V Hβ →p H 0 V Hβ = Vθ > 0 if Hβ has full rank q. Wn = n R0 β − θ0 ´ √ ³ The delta method (6. An asymptotic Wald test rejects H0 in favor of H1 if Wn exceeds χ2 (α). In this case. the covariance What we will show in this section is that if V is replaced with V 0 = σ 2 n−1 X 0 X matrix estimator valid under homoskedasticity. so by the continuous mapping ˆ ˆ ˆ ˆ ˆ theorem. then the Wald statistic takes the form ´0 ³ ´ ³ ´−1 ³ ˆ ˆ ˆ R0 V R R0 β − θ 0 . the upper-α quantile q 2 of the χ2 distribution. Vθ ). The Wald test fails to reject if Wn is q 1 ¡ ¢ less than χ2 (α). The asymptotic p-value for Wn is pn = p(Wn ). Hβ (β) is a continuous function of β.1 Under H0 and Assumption 6.025 . Furthermore. For example. 1] under H0 . and Theorem 6.1 showed θ ˆ that V →p V. Also h(β) = a selector matrix. q and pn is asymptotically U[0. a chiq square random variable with q degrees of freedom. Hβ (β) →p Hβ . The null hypothesis is H0 : β 2 = 0. the test rejects at the α% level iﬀ pn < α. As before. if rank(Hβ ) = q. χ2 (. θ = β 2 . ˆ Wn = n θ θ q θ θ Theorem 6.

Now consider the residual regression of e on X2 = M1 X2 . ˜ ˆ ˜ 2 β 2 + e and X 0 e = 0. Because of this connection we call (6. = F = n k2 σ 2 /(n − k) ˆ ¢−1 ¡ ˆ While it should be emphasized that equality (6. Thus ¡ 0 ¢−1 R = X2 M1 X2 and ³ ¡ ´−1 ³ ¢−1 ´−1 ¡ 0 ¢ ˆ = σ −2 n−1 R0 X 0 X ˆ R = σ −2 n−1 X2 M1 X2 ˆ R0 V 0 R ³ ´−1 ˆ0 ˆ ˆ Wn = nβ 2 R0 V 0 R β2 = 0 ˆ ˆ0 β 2 (X2 M1 X2 ) β 2 . The elegant feature about (6. and σ2 = ˆ 1 0 e e.19). Thus ˜ ˆ e=X ˜ ˆ 2 ³ ´0 ³ ´ ˜ ˆ ˜ ˆ e0 e = X2 β 2 + e ˜˜ ˆ ˆ X2 β 2 + e ˆ0 ˜ 0 ˜ ˆ = β 2 X2 X2 β 2 + e0 e ˆˆ 0 0 ˆ ˆ ˆˆ = β X M1 X2 β + e0 e.19) is that it is directly computable from the standard output from two simple OLS regressions. 59 . ˆˆ n ˆ e = Y − X β. 2 2 2 or alternatively. note that partitioned matrix inversion (2. since ˆ0 0 ˆ β 2 X2 M1 X2 β 2 = e0 e − e0 e.are from OLS of Y on X1 .19) the F form of the Wald statistic. still this formula often ﬁnds good use in reading applied papers.19) only holds if V 0 = σ 2 n−1 X 0 X ˆ . By the FWL theorem. the residual is ˜ ˜ e = M1 Y. note that if we regress Y on X1 alone. First. as the sum of squared errors is a typical output from statistical packages. µ e0 e − e0 e ˜˜ ˆˆ e0 e ˆˆ ¶ =n µ σ2 − σ2 ˜ ˆ 2 σ ˆ ¶ . ˜˜ ˆˆ ˆˆ σ 2 = n−1 e0 e ˆ we conclude that Wn = n as claimed. X2 ). Also.2) ¢−1 ¡ R = R0 R X 0X 0 µ 0 0 X1 X1 X1 X2 0X 0 X2 1 X2 X2 ¶−1 0 0 where M1 = I − X1 (X1 X1 )−1 X1 . ˆ ¡ ¢−1 0 ˆ β = X 0X XY are from OLS of Y on X = (X1 . 2 σ ˆ To simplify this expression further. This statistic is typically reported as an “F-statistic” which is deﬁned as ¡ 2 ¢ σ − σ 2 /k2 ˜ ˆ n − k Wn . We now derive expression (6.

The modelling assumption that the error ei is independent of xi and N (0. this implies that conditional on X µ ¶ µ ¶ µ µ 2 0 −1 ¶¶ ˆ β−β (X 0 X)−1 X 0 σ (X X) 0 = e ∼ N 0. This is ¢ ¡ 2 ˆ σ y − σ 2 / (k − 1) ˜ F = . σ 2 /(n − k) ˆ σ2 = ˜y 1 (y − y)0 (y − y) n is the sample variance of yi . Since linear functions of normals are also normal. In σ 2 . σ 2 ) can be be used to calculate a set of exact distribution results. Then (n − k) s2 σ2 1 0 ee ˆˆ σ2 1 0 e Me = σ2 ∙ ¸ 1 0 In−k 0 eH = H 0e 0 0 σ2 ∙ ¸ In−k 0 0 = u u 0 0 = ∼ χ2 . when sample sizes were very small and researchers wanted to know if there was “any explanatory power” to their regression. when an OLS regression is reported. introduced in Section 4.12 Normal Regression Model As an alternative to asymptotic distribution theory. This is rarely an issue today. n−k 60 . there is an exact distribution theory available for the normal linear regression model.where In many statistical packages. under the normality assumption the error vector e is independent of X and ¢ ¡ has distribution N 0. In particular. H 0 H) ∼ N (0. these cases are atypical.3. s The spectral decomposition of M yields ¸ ∙ In−k−1 0 H0 M =H 0 0 (see equation (2. e ˆ M 0 σ2 M where M = I − X (X 0 X)−1 X 0 . an “F statistic” is reported. equivalently the residual variance from an intercept-only model. including the estimated error variance 2. Let u = σ −1 H 0 e ∼ N (0. Since uncorrelated normal variables are independent. as sample sizes are typically suﬃciently large that this F statistic is highly “signiﬁcant”.4)) where H 0 H = In . This was a popular statistic in the early days of econometric reporting. it follows ˆ that β is independent of any function of the OLS residuals. While there are special cases where this F statistic is useful. 6. In ) . This special F statistic is testing the hypothesis that all slope coeﬃcients (other than the intercept) are zero.

1 shows that under the strong assumption of ˆ normality. σ 2 ) where β 1 is the OLS estimate from ˜ 2 ˜ a regression of yi on x1i only.1 we showed that in large samples. Furthermore.3. 0.1 and Theorem 6. In contrast. β 2 . ˆ 2 2 2 ˜ ˜ The MLE of the model under the null hypothesis is (β 1 . σ2 ˆ 61 . σ 2 ) = − log 2πˆ 2 − 2 e0 e ˆˆ 2 2ˆ σ ¡ ¢ n n n = − log σ 2 − log (2π) − .a chi-square distribution with n − k degrees of freedom.10) then ´ ³ ˆ • β ∼ N 0. β has an exact normal distribution and t has an exact t distribution. so as a practical matter it does not matter which distribution is used. 1) and tn−k distributions. with residual variance σ . The Gaussian log-likelihood at these estimates is ¡ ¢ n 1 ˆ ˆ ˆ σ Ln (β 1 . 0. The log-likelihood of this model is ¡ 2¢ n n n ˜ ˜ ˜ Ln (β 1 . The critical values are quite close if n − k ≥ 30. σ 2 ) − Ln (β 1 . σ 2 (X 0 X)−1 • (n−k)s2 σ2 ˆ β j −β j ˆ s(β ) j ∼ χ2 . n−k • ∼ tn−k ˆ In Theorem 6.12. which is twice the diﬀerence in the log-likelihood function evaluated under the null and alternative hypotheses. if standard errors are calculated using the homoskedastic formula (6. β and t are approximately normally distributed. σ 2 ) = − log σ − log (2π) − . Theorem 6. As inference (conﬁdence intervals) are based on the t-ratio. β 2 . σ 2 ) ¡ ¢¢ ¡ ¡ ¢ ˆ = n log σ 2 − log σ 2 ˜ µ 2¶ σ ˜ = n log .8. β 2 .) Now let us partition β = (β 1 . 0. 1) jj rh ∼ tn−k = rh i ∼q i = q χ2 ˆ s(β j ) n−k σ2 2 0 X)−1 0 X)−1 (X s (X n−k n−k χn−k jj jj a t distribution with n − k degrees of freedom. σ ˆ ˆ βj − βj βj − βj N (0. We summarize these ﬁndings Theorem 6. σ 2 ) discussed above.12. the notable distinction is between the N (0. 2 2 2 The LR statistic for H0 is ´ ³ ˆ ˆ ˆ ˜ ˜ LR = 2 Ln (β 1 . (Unless the sample size is unreasonably small. a good testing procedure is based on the likelihood ratio statistic. σ 2 ). and standard errors are calculated using the homoskedastic formula (6.10) µ h i ¶ 2 (X 0 X)−1 N 0. β 2 ) and consider tests of the linear restriction H0 : β 2 = 0 H1 : β 2 6= 0 In the context of parametric models. The estimator under the alternative is the unrestricted estimator ˆ ˆ ˆ (β 1 .1 If ei is independent of xi and distributed N(0.

σ 2 s2 β ˆ ˆ While the hypothesis β s = 1 is unaﬀected by the choice of s. they can work quite poorly when the restrictions are nonlinear. It is easy to see that in each case there are values of s for which the line is for the case β test statistic is signiﬁcant relative to asymptotic critical values. as Wn (s) →d χ2 under H0 for 1 any s.By a ﬁrst-order Taylor series approximation ¶ µ 2 ¶ µ σ ˜ σ2 ˜ − 1 = Wn . features of this distribution can be calculated. Our ﬁrst-order asymptotic theory is not useful to help pick s. we ﬁnd that the standard Wald test for H0 (s) is ´2 ³ s ˆ β −1 Wn (s) = n 2s−2 . 62 . 6.4 the Wald statistic Wn (s) as a function ˆ of s. 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. setting n/σ 2 = 10. To demonstrate this eﬀect. while there are other values of s for which test test statistic is insigniﬁcant. The increasing solid line is for the case β = 0. This is an unfortunate feature of the Wald statistic. Then the standard Wald test for H0 is ˆ ³ ´2 ˆ β−1 Wn = n . The decreasing dashed ˆ = 1. Through repetition. LR = n log 1 + 2 − 1 ' n σ ˆ σ2 ˆ the F statistic. σ 2 ) and consider the hypothesis H0 : β = 1. This is a context where Monte Carlo simulation can be quite useful as a tool to study and compare the exact distributions statistical procedures in ﬁnite samples. ˆ Let β and σ 2 be the sample mean and variance of yi . we have plotted in Figure 6. σ2 ˆ Now notice that H0 is equivalent to the hypothesis H0 (s) : β s = 1 for any positive integer s. Letting h(β) = β s .8. the statistic Wn (s) varies with s. and noting Hβ = sβ s−1 .13 Problems with Tests of NonLinear Hypotheses While the t and Wald tests work well when the hypothesis is a linear restriction on β.7. In the present context of the Wald statistic. This can be seen by a simple example introduced by Lafontaine and White (1986). Take the model yi = β + ei ei ∼ N (0. This is distressing since the choice of s seems arbitrary and irrelevant to the actual hypothesis. This produces random draws from the sampling distribution of interest. The method uses random simulation to create an artiﬁcial dataset to apply the statistical tools of interest.

These probabilities are calculated as the percentage of the 50.4. which is not surprising given the dependence of Wn (s) on s as shown in Figure 6.1 you can also see the impact of variation in sample size. Test performance deteriorates as s increases.Figure 6. and σ is varied among 1 and 3.1 we report the results of a Monte Carlo simulation where we vary these three parameters. and rarely fall outside of the 3%-8% range. remember that the ideal Type I error probability is 5% (. To interpret the table. and σ 2 .000 random samples. Given the simplicity of the model. n. the only test which meets this criterion is the conventional Wn = Wn (1) test. looking for tests for which rate rejection rates are close to 5%. In Table 2. so these probabilities are Type I error. the Type I error probability improves towards 5% as the sample size n increases.05) with deviations indicating+ distortion. no magic choice of n for which all tests perform uniformly well. Type I error rates between 3% and 8% are considered reasonable. Any other choice of s leads to a test with unacceptable Type I error probabilities. The second through seventh columns contain the Type I error probabilities for diﬀerent combinations of n and σ. There is.4: Wald Statistic as a function of s P (Wn (s) > 3. n is varied among 20. we compare the rates row by row. Table 4. For this particular example. Typically. Error rates avove 10% are considered excessive.000 simulated Wald statistics Wn (s) which are larger than 3. In Table 4. In each case.84. Each row of the table corresponds to a diﬀerent value of s — and thus corresponds to a particular choice of test statistic. The value of s is varied from 1 to 10. Table 4. When comparing statistical procedures. The null hypothesis β s = 1 is true.84 | β = 1) .1 Type I error Probability of Asymptotic 5% Wn (s) Test 63 .1 reports the simulation estimate of the Type I error probability from 50. 100 and 500. however. Rates above 20% are unexceptable. this probability depends only on s.

17 .06 .05 .07 .15 .30 .14 . While this is clear in this particular example.33 .06 . Deﬁne θ ˆ ˆ ⎞ ⎛ 0 ⎟ ⎜ ⎟ ⎜ ⎜ 1 ⎟ ⎟ ⎜ ⎜ ˆ ⎟ ˆ H1 = ⎜ β 2 ⎟ ⎟ ⎜ ⎟ ⎜ ⎜ β ⎟ ˆ ⎠ ⎝ − 1 ˆ2 β2 ´ s(ˆ θ) −r .19 .08 .05 .13 .24 .12 .06 .06 .28 .05 . Other choices are arbitrary and would not be used in practice. In this case a t-statistic for H0 is t1n = ³ˆ β1 ˆ β2 where r is a known constant.23 .16 Rejection frequencies from 50. ˆ ˆ ˆ ˆ ˆ Let β = (β 0 .10 . in other examples natural choices are not always obvious and the best choices may in fact appear counter-intuitive at ﬁrst.13 .06 .20 .10 . This point can be illustrated through another example.08 . β 2 ) be the least-squares estimates of (6.06 . so the hypothesis can be stated as H0 : θ = r.25 .s 1 2 3 4 5 6 7 8 9 10 Note: σ=1 σ=3 n = 20 n = 100 n = 500 n = 20 n = 100 n = 500 .12 .15 .09 .06 . 64 .15 .21 .26 .14 .000 simulated random samples In this example it is not surprising that the choice s = 1 yields the best test statistic. Take the model yi = β 0 + x1i β 1 + x2i β 2 + ei E (xi ei ) = 0 and the hypothesis H0 : β1 =r β2 (6.08 .05 . let V be an estimate of the ˆ asymptotic variance matrix for β and set ˆ = β 1 /β 2 .15 .08 .13 .22 . deﬁne θ = β 1 /β 2 .34 .25 .08 . β 1 .35 .07 .10 .06 .22 .07 .07 .20 .31 . Equivalently.20) ³ ´1/2 ˆ0 ˆ ˆ so that the standard error for ˆ is s(ˆ = n−1 H1 V H1 θ θ) .18 .11 .05 . An alternative statistic can be constructed through reformulating the null hypothesis as H0 : β 1 − rβ 2 = 0.20).05 .07 .

the rejection rates for the t1n statistic diverge greatly from this value.06 . The implication of Table 4.645) are calculated from 50.26 .00 .15 . However.645) greatly exceed 5%.05 .09 .05.06 . In all cases.00 .00 .00 .25 . −1 H V H ˆ 2 n 2 ⎞ 0 H2 = ⎝ 1 ⎠ .0 and n among 100 and 500. This leaves β 2 as a free parameter. such as the GMM distance statistic which will be presented in Section 9. especially for small values of β 2 .28 . and alternatives to asymptotic critical values should be considered.14 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.50. 1) variables.A t-statistic based on this formulation of the hypothesis is ´2 ³ ˆ 1 − rβ 2 ˆ β =³ ´1/2 . .05 .06 . this formulation should be used.12 . The common message from both examples is that Wald statistics are sensitive to the algebraic formulation of the null hypothesis. If no linear formulation is feasible.000 simulated samples. σ 2 ) draw with σ = 3.00 .06 . . We let x1i and x2i be mutually independent N (0.00 The one-sided Type I error probabilities P (tn < −1.2 Type I error Probability of Asymptotic 5% t-tests n = 100 n = 500 P (tn < −1. In this section we describe the method in more detail.06 .50 .00 .05 .00 . and are close to the ideal 5% rate for both sample sizes.10 . . the rejection rates for the linear t2n statistic are invariant to the value of β 2 . Table 4.645) are close to zero in most cases.05 . while the right tail probabilities P (t1n > 1. the entries in the table should be 0. ei be an independent N (0. if the hypothesis can be expressed as a linear restriction on the model parameters.06 . It is also prudent to consider alternative tests to the Wald statistic.75.15 .2.05 . We vary β 2 among .1. −r ⎛ t2n where To compare t1n and t2n we perform another simple Monte Carlo simulation.06 .06 .10 .00 . along with sample size n.2 is that the two t-ratios have dramatically diﬀerent sampling behavior. 6.75 1.7.05 .07 . and normalize β 0 = 0 and β 1 = 1. Ideally.02 .645) P (tn > 1.25.06 . The left tail probabilities P (t1n < −1.645) and P (tn > 1. In contrast.645) P (tn > 1. The results are presented in Table 4.645) t1n t2n t1n t2n t1n t2n t1n t2n . 65 .05 β2 .645) P (tn < −1.05 .47 . then the “most linear” formulation should be selected.00 . and 1.06 .05 .05 .10 .

We will discuss this choice later. The basic idea is that for any given F. Typically.. x∗ ... or variance of the distribution ˆ − θ. ∗ ∗ • The statistic Tn = Tn (y1 . They constitute a random sample of size B from the distribution of Gn (x. Monte Carlo simulation uses numerical simulation to compute Gn (x. the exact (ﬁnite sample) distribution Gn is generally unknown. for example an estimator ˆ or a t-statistic (ˆ − θ)/s(ˆ The exact distribution of Tn is θ θ θ). 1) random numbers.. F ) = P (Tnb ≤ x) = P (Tn ≤ x | F ) . a chi-square can be generated by sums of squares of normals. F ) for selected choices of F. This is useful to investigate the performance of the statistic Tn in reasonable situations and sample sizes. F ) = P (Tn ≤ x | F ) . θ) be a statistic of interest. and from these most random variables can be constructed. where games of chance are played.) For step 2. . x∗ . F ) can be calculated numerically through simulation. i = 1. x1 . the distribution function Gn (x. The above experiment creates one random draw from the distribution Gn (x. n... n n For step 1. which implies restrictions on F . it is important that the statistic be evaluated at the “true” value of θ corresponding to the choice of F. b = 1. B. F ). This is one observation from an unknown distribution. let the b0 th experiment result in the draw Tnb . Notationally. or equivalently the value θ is selected directly by the researcher. we can estimate any feature of interest using (typically) a method of moments estimator. and calculate θ of θ \ ˆ Bias(θ) = B B 1 X 1 Xˆ Tnb = θb − θ B B b=1 b=1 B 1 X \θ) (Tnb )2 M SE(ˆ = B b=1 µ ¶2 \ = M SE(ˆ − Bias(ˆ \θ) \ ˆ V ar(θ) θ) 66 . We then set Tn = ˆ − θ. where B is a large number. The researcher chooses F (the distribution of the data) and the sample size n. While the asymptotic distribution of Tn might be known. xi ) which are random draws from a population distribution F. from one observation very little can be said.. (For example. run the above experiment. These results are stored. So the researcher repeats the experiment B times.. yn . are drawn from the distribution F using i the computer’s random number generator.Recall.. Let θ be a parameter and let Tn = Tn (y1 . For example: Suppose we are interested in the bias. yn . our data consist of observations (yi . x∗ ) . Then the following experiment is conducted ∗ • n independent random pairs (yi . θ) is calculated on this pseudo data. Clearly. we set B = 1000 or B = 5000. The name Monte Carlo derives from the famous Mediterranean gambling resort. A “true” value of θ is implied by this choice. 1] and N (0. From a random sample.. mean-squared error (MSE).. most computer packages have built-in procedures for generating U [0. The method of Monte Carlo is quite simple to describe.. xn . Gn (x. .

Again our dependent variable is the natural log of wages. and 5000.21) is therefore an ³ ´ p ˆ unbiased estimator of P with standard error s P = P (1 − P ) /B. and our regressors include years of education. Generally. The random variable 1 (Tnb ≥ 1. Suppose we are interested in the 5% and 95% quantile of Tn = ˆ We then compute the 10% θ. For the dependent variable we use the natural log of wages. immigrant. a larger B results in more precise estimates of the features of interest of Gn . In practice. experience squared. if we set B = 999. For example.22/ B. and poorly for others. As P is unknown. Hence the ³ ´ ˆ standard errors for B = 100. then we can set s P = (. . therefore. the performance will depend on n and F. For regressors we include years of education. female. For example.007. s P = . it is simple to make inferences about rejection probabilities from statistical tests. and .15 Estimating a Wage Equation We again return to our wage equation. equalling 1 with probability P = E1 (Tnb ≥ 1. union member. then the 5% sample quantile is 50’th sorted value and the 95% sample quantile is the 950’th sorted value. respectively. are. B. 1000. potential work experience. and hispanic. if we are ³ ´ p √ ˆ assessing an asymptotic 5% test. As discussed above.003. an estimator or test may perform wonderfully for some values.022. where N = (B +1)α. the researcher must select the number of experiments. If the standard error is too large to make a reliable inference. In particular. so that coeﬃcients may be interpreted as semi-elasticities. Quite simply. potential work experience. We separately estimate equations for white and non-whites. and dummy variable indicators for the following: married. this may ˆ be approximated by replacing P with P or with an hypothesized value. we included a dummy 67 . such as the percentage estimate reported in (6.96) is iid Bernoulli. The typical purpose of a Monte Carlo simulation is to investigate the performance of a statistical procedure (estimator or test) in realistic settings. the choice of B is often guided by the computational demands of the statistical procedure. We us the sample of wage earners from the March 2004 Current Population Survey. excluding military. but requires more computational time. B b=1 (6. It is therefore convenient to pick B so that N is an integer. In many cases. experience squared. for a selection of choices of n and F.95) /B ' . and therefore it is straightforward to calculate standard errors for any quantity of interest. Since the results of a Monte Carlo experiment are estimates computed from a random sample of size B.05) (.Suppose we are interested in the ¯Type I error associated with an asymptotic 5% two-sided ¯ ¯ ¯ θ) θ t-test.21) the percentage of the simulated t-ratios which exceed the asymptotic 5% critical value. Often this is called the number of replications. The average (6. hispanic. female. We now expand the sample all non-military wage earners.96) . Then qα is the N ’th number in this ordered sequence. Furthermore. and non-white. then B will have to be increased. immigrant. The α% sample quantile is a number qα such that α% of the sample are less than qα . and estimate a multivariate regression. It is therefore useful to conduct a variety of experiments. A simple way to compute sample quantiles is to sort the sample {Tnb } from low to high. union member.21).96) . 6. and 90% sample quantiles of the sample {Tnb }. We would then set Tn = ¯ˆ − θ¯ /s(ˆ and calculate B X ˆ= 1 P 1 (Tnb ≥ 1. and dummy variable indicators for the following: married.

00057 . excluding the coeﬃcients on the state dummy variables.49452 σ ˜ Notice that the two statistics are close. Alternatively. as the 1% critical value for the χ2 distribution is 76.008 . the restricted standard deviation estimate is σ = .232 .032 . if β 2 > 0 and β 3 < 0 the solution is θ=− From Table 4. Computing the Wald statistic (6.69.48772 = 515.010 Intercept Education Experience Experience2 Married Female Union Member Immigrant Hispanic Non-White σ ˆ Sample Size R2 One question is whether or not the state dummy variables are relevant.1 we ﬁnd the point estimate ˆ ˆ = − β 2 = 28. Wn = n 1 − 2 = 18. reestimating the model with the 50 state dummies excluded.027 .097 −. but not equal.808 . Ignoring the other coeﬃcients.007 . 2β 3 68 . Table 4. 808 1 − .070 .1 OLS Estimates of Linear Equation for Log(Wage) ˆ β 1.101 .34 ˆ s(β) .033 −.014 .18) that the state coeﬃcients are jointly zero.121 −.00002 . The F form of the Wald statistic (6. we ﬁnd Wn = 550.1.19) is ˜ µ ¶ µ ¶ σ2 ˆ .001 . θ ˆ 2β 3 β2 .4877 18.808 so the parameter estimates are quite precise and reported in Table 4.102 −. Using either statistic the hypothesis is easily rejected. we can write this eﬀect as log(W age) = β 2 Experience + β 3 Experience2 + · · · Our question is: At which level of experience θ do workers achieve the highest wage? In this quadratic model.002 .010 . this adds 50 regressors).013 . 50 Another interesting question which can be addressed from these estimates is the maximal impact of experience on mean wages.102 −.variable for state of residence (including the District of Columbia. The available sample is 18.4945.

we found better Type I error rates by reformulating the hypothesis as a linear restriction. In the present context we are interested in forming a conﬁdence interval. there is not a simple expression for this set.0. Notice that the upper end of the conﬁdence interval is the same as that from the delta method.Using the Delta Method. this is a poor choice.96.5]. but the lower end is substantially lower. This set is [27.5]. Since tn (θ) is a non-linear function of θ. This is the set of θ such that |tn (θ)| ≤ 1. However. implying a 95% conﬁdence θ) interval of [27. as the coverage probability of this conﬁdence interval is one minus the Type I error of the hypothesis test based on the t-test. These t-statistics take the form ˆ ˆ β + 2β 3 θ tn (θ) = ³ 2 ´1/2 ˆ h0 V hθ θ hθ = µ −1 2θ ¶ where ˆ ˆ ˆ and V is the covariance matrix for (β 2 β 3 ). so we have to go one step further. 29. 29. Instead. but it can be found numerically quite easily. we can calculate a standard error of s(ˆ = . 69 . Our desired conﬁdence interval will be the set of parameter values θ which are not rejected by the hypothesis test.40. In the previous section we discovered that such t-tests had very poor Type I error rates. not testing a hypothesis.9.

β − β = Ik − X 0 X 70 . ˜ (e) Find an appropriate formula to calculate standard errors for the elements of β. β 2 ). In the model Y = Xβ + e. the least-squares estimate of β subject to the restriction h(β) = 0 is ˜ β = argmin Sn (β) h(β)=0 Sn (β) = (Y − Xβ)0 (Y − Xβ) . show that the least-squares estimate of β = (β 1 . (d) Under the ´ standard assumptions plus R0 β = r. ˜ (b) Verify that R0 β = r. 1. In the model Y = X1 β 1 + X2 β 2 + e. Take the model Y = Xβ + e with the restriction R0 β = r where R is a known k × s matrix. ﬁnd the least-squares estimate of β = (β 1 .6. In the model Y = X1 β 1 + X2 β 2 + e. 2. In the model Y = X1 β 1 + X2 β 2 + e. then ¶ µ ¡ ¢−1 h 0 ¡ 0 ¢−1 i−1 0 ¡ 0 ¢−1 0 ˜ XX R R XX R R X e.16 Exercises For exercises 1-4. β 2 ). ﬁnd the asymptotic distribution of √ ³˜ n β − β as n → ∞. λ) = Sn (β) + λ0 R0 β − r 2 where λ is s × 1. and rank(R) = s. subject to the constraint that β 1 = c (where c is some given vector) is simply the OLS regression of Y − X1 c on X2 . (a) Show that the solution is ´ ¢−1 h 0 ¡ 0 ¢−1 i−1 ³ 0 ¡ ˆ ˜ ˆ Rβ−r R R XX R β = β − X 0X ´ h ¡ ¢−1 i−1 ³ 0 ˆ ˆ Rβ−r R λ = R0 X 0 X ¢−1 0 ¡ ˆ XY β = X 0X where is the unconstrained OLS estimator. subject to the constraint that β 1 = −β 2 . r ˜ is a known s × 1 vector. β 2 ) subject to the constraint that β 2 = 0 is the OLS regression of Y on X1 . 3. 0 < s < k. Explain why β solves the minimization of the Lagrangian ¢ ¡ 1 L(β. show that the least-squares estimate of β = (β 1 . 4. β minimizes the sum of squared errors Sn (β) over all β such that the restriction holds. (c) Show that if R0 β = r is true. the following deﬁnition is used. ˜ That is.

5. You have two independent samples (Y1 , X1 ) and (Y2 , X2 ) which satisfy Y1 = X1 β 1 + e1 and Y2 = X2 β 2 + e2 , where E (x1i ei1 ) = 0 and E (x2i e2i ) = 0, and both X1 and X2 have k ˆ ˆ columns. Let β 1 and β 2 be the OLS estimates of β 1 and β 2 . For simplicity, you may assume that both samples have the same number of observations n. ´ ´ √ ³³ ˆ ˆ (a) Find the asymptotic distribution of n β 2 − β 1 − (β 2 − β 1 ) as n → ∞. (b) Find an appropriate test statistic for H0 : β 2 = β 1 . (c) Find the asymptotic distribution of this statistic under H0 . 6. The model is yi = x0 β + ei i E (xi ei ) = 0 ¡ ¢ Ω = E xi x0 e2 . i i

ˆ ˆ (a) Find the method of moments estimators (β, Ω) for (β, Ω). ˆ ˆ (b) In this model, are (β, Ω) eﬃcient estimators of (β, Ω)? (c) If so, in what sense are they eﬃcient? 7. Take the model yi = x0 β 1 + x0 β 2 + ei with Exi ei = 0. Suppose that β 1 is estimated by 1i 2i regressing yi on x1i only. Find the probability limit of this estimator. In general, is it consistent for β 1 ? If not, under what conditions is this estimator consistent for β 1 ? 8. Verify that equation (6.13) equals (6.14) as claimed in Section 6.6. 9. Prove that if an additional regressor Xk+1 is added to X, Theil’s adjusted R increases if ˆ ˆ ˆ and only if |tk+1 | > 1, where tk+1 = β k+1 /s(β k+1 ) is the t-ratio for β k+1 and ¡ ¢1/2 ˆ s(β k+1 ) = s2 [(X 0 X)−1 ]k+1,k+1

2

is the homoskedasticity-formula standard error.

**10. Let Y be n × 1, X be n × k (rank k). Y = Xβ + e with E(xi ei ) = 0. Deﬁne the ridge regression estimator Ã n !−1 Ã n ! X X 0 ˆ xi xi + λIk xi yi β=
**

i=1 i=1

ˆ ˆ where λ > 0 is a ﬁxed constant. Find the probability limit of β as n → ∞. Is β consistent for β?

∗ ∗ 11. 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 = x0 β + ei i

E (xi ei ) = 0

∗ yi = yi + ui

71

**where ui is a measurement error satisfying E (xi ui ) = 0
**

∗ E (yi ui ) = 0

ˆ Let β denote the OLS coeﬃcient from the regression of yi on xi . (a) Is β the coeﬃcient from the linear projection of yi on xi ? ˆ (b) Is β consistent for β as n → ∞? ´ √ ³ˆ (c) Find the asymptotic distribution of n β − β as n → ∞.

12. The data set invest.dat contains data on 565 U.S. ﬁrms extracted from Compustat for the year 1987. The variables, in order, are • Ii Investment to Capital Ratio (multiplied by 100). Total Market Value to Asset Ratio (Tobin’s Q). Cash Flow to Asset Ratio. Long Term Debt to Asset Ratio.

• Qi • Di • Ci

The ﬂow 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 conﬁdence intervals for the coeﬃcients. (c) This regression is related to Tobin’s q theory of investment, which suggests that investment should be predicted solely by Qi . Thus the coeﬃcient on Qi should be positive and the others should be zero. Test the joint hypothesis that the coeﬃcients on Ci and Di are zero. Test the hypothesis that the coeﬃcient on Qi is zero. Are the results consistent with the predictions of the theory?

2 (d) Now try a non-linear (quadratic) speciﬁcation. Regress Ii on Qi , Ci , Di , Q2 , Ci2 , Di , i Qi Ci , Qi Di , Ci Di . Test the joint hypothesis that the six interaction and quadratic coeﬃcients are zero.

13. 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 ﬁle 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 speciﬁcation ln T Ci = β 1 + β 2 ln Qi + β 3 ln P Li + β 4 ln P Ki + β 5 ln P Fi + ei . (6.22)

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 diﬀer. (b) Using a Wald statistic, test the hypothesis H0 : β 3 + β 4 + β 5 = 1. 72

(c) Estimate (6.22) by least-squares imposing this restriction by substitution. Report your parameter estimates and standard errors. (d) Estimate (6.22) subject to β 3 + β 4 + β 5 = 1 using the restricted least-squares estimator from problem 4. Do you obtain the same estimates as in part (c)? .

73

. Then i 0 E (η i | xi ) = α0 + z1i α1 0 η i = α0 + z1i α1 + ξ i and we have the regression equation (7. Typically. σ 2 } and σ 2 = σ 2 (x ) = E e2 | x . z1i are squares (and perhaps levels) of some (or all) elements of xi . σ1 We now discuss this estimation problem. σ 2 }. . A ˆ ˆn feasible GLS (FGLS) estimator replaces the unknown D with an estimate D = diag{ˆ 2 .1) XD Y β = X 0 D−1 X ¡ ¢ 2 . The GLS estimator (7. we know that the least-squares estimator is semi-parametrically eﬃcient for the projection coeﬃcient. where z1i is some q × 1 function of xi .. ..1) infeasible since the matrix D is unknown.2) E (ξ i | xi ) = 0.. The theory of Chamberlain (1987) can be used to show that in this model the semiparametric eﬃciency bound is obtained by the Generalized Least Squares (GLS) estimator ¢−1 ¡ 0 −1 ¢ ¡ ˜ (7. The GLS estimator is sometimes called where D = diag{σ 1 i i n i i the Aitken estimator.. The error ξ i in this regression error ξ i is generally heteroskedastic and has the conditional variance ¡ ¢ V ar (ξ i | xi ) = V ar e2 | xi i ´ ³¡ ¢¢2 ¡ = E e2 − E e2 | xi | xi i i ¡ 4 ¢ ¡ ¡ 2 ¢¢2 = E ei | xi − E ei | xi .. the least-squares estimator is ineﬃcient. in the linear regression model yi = x0 β + ei i E (ei | xi ) = 0. 74 .Chapter 7 Additional Regression Topics 7.1 Generalized Least Squares In the projection model. Let η i = e2 . Suppose that we model the conditional variance using the parametric form 0 σ 2 = α0 + z1i α1 i = α0 zi . Often the functional form is kept simple for parsimony. However.

. Since there is not a unique speciﬁcation for the conditional variance the FGLS estimator is not unique. Then set ˜i ˜ D = diag{˜ 2 . and hence we can estimate 0 σ 2 = zi α by i ˜ (7. Then we could estimate α by OLS: ¢−1 0 ¡ Z η →p α α = Z 0Z ˆ and where √ n (ˆ − α) →d N (0. Then ˆ ¢−1 0 ¡ Zη ˆ α = Z 0Z ˜ (7.5) ˆ Thus the fact that η i is replaced with η i is asymptotically irrelevant. xi ). Vα = E zi zi (7. we have the OLS residual ei = yi − x0 β = ei − x0 (β − β). This is the feasible GLS. We may call (7. Thus ˆ i i η − η i = e2 − e2 ˆ ˆi i = φi . and will depend on the model (and estimation method) for the skedastic regression. If σ 2 < 0 for some i. ³ ´ ˆ ˆ ˆ = −2ei x0 β − β + (β − β)0 xi x0 (β − β) i i say. Furthermore. This suggests 75 . estimator of β. as it is estimating the conditional variance of the regression of yi on xi . Vα ) α ¡ ¡ 0 ¢¢−1 ¡ 0 2 ¢ ¡ ¡ 0 ¢¢−1 E zi zi ξ i E zi zi . then the FGLS ˜i estimator is not well deﬁned.Suppose ei (and thus η i ) were observed.4) the skedastic regression. We have shown that α is consistently estimated by a simple procedure. if σ 2 ≈ 0 for some i. σ 2 } σ1 ˜n and ³ ´−1 ˜ ˜ ˜ β = X 0 D−1 X X 0 D−1 Y. ˜i Suppose that σ 2 > 0 for all i. Note that n n n ´ 1X √ ³ √ −2 X 1 X ˆ ˆ ˆ √ zi φi = zi ei x0 n β − β + zi (β − β)0 xi x0 (β − β) n i i n n n i=1 i=1 i=1 →p 0 Let be from OLS regression of η i on zi . Vα ) (7. there is no guarantee that σ 2 > 0 for all i.6) σ 2 = α0 zi .. or FGLS..3) ˆ ˆ While ei is not observed. then the FGLS estimator will force ˜i the regression equation through the point (yi . which is typically undesirable.4) ¡ ¢−1 −1/2 0 √ √ n (˜ − α) = n (ˆ − α) + n−1 Z 0 Z α α n Zφ →d N (0. One typical problem with implementation of FGLS estimation is that in a linear regression ˜i speciﬁcation..

that there is a need to bound the estimated variances away from zero. First. In the linear regression model. It may be the case that α0 zi is only an approximation to the true conditional variance σ 2 = i ˜ E(e2 | xi ). e2 }. Since the form of the skedastic regression is unknown.1 If the skedastic regression is correctly speciﬁed. V = E α0 zi i i Examining the asymptotic distribution of Theorem 7.1. Instead. This may be written as !−1 Ã n !Ã n !−1 Ã n 1 X −4 2 0 1 X −2 0 1 X −2 0 ˜ σ i xi xi ˜ σ i ei xi xi ˜ ˆ σ i xi xi ˜ V = n n n i=1 i=1 i=1 ³ ´−1 ³ ´³ ´−1 ˜ ˜ ˆ˜ ˜ X 0 D−1 DD−1 X X 0 D−1 X = n X 0 D−1 X ˜ which is consistent for V as n → ∞. V ˜ An appropriate solution is to use a White-type estimator in place of V 0 . Why then do we not exclusively estimate regression models by FGLS? This is a good question. 1992).1. FGLS is asymptotically superior to OLS. We just state the result without proof.1. FGLS estimation depends on speciﬁcation and estimation of the skedastic regression. ´ √ ³ ˜ ˜ n β GLS − β F GLS →p 0. β should perhaps be i i called a quasi-FGLS estimator of β. V ). Unless σ 2 = α0 zi . . the natural estimator of the asympˆ totic variance of β is Ã n !−1 µ ¶−1 1 X −2 0 1 0 ˜ −1 ˜0 = XD X σ i xi xi ˜ = . σ 2 ] σi i for some σ 2 > 0. This is an estimator which is robust to misspeciﬁcation of the condie1 ˆn tional variance. In this case we interpret α0 zi as a linear projection of e2 on zi . ´´−1 ³ ³¡ ´´−1 ³ ³¡ ¢−1 ¢−2 2 0 ´´ ³ ³¡ 0 ¢−1 E α0 zi E α zi xi x0 σ i xi xi xi x0 ..2) is correctly speciﬁed. i ˜ 0 is inconsistent for V . but the proof of this can be tricky.. Theorem 7. This estimator V 0 is appropriate when the skedastic regression (7. and it may be estimated with considerable 76 V takes a sandwich form√. Its asymptotic variance is not that given in Theorem 7. V = E σ −2 xi x0 i i ˆ where D = diag{ˆ2 . similar to the covariance matrix of the OLS estimator.1. A trimming rule might make sense: σ 2 = max[˜ 2 . and was proposed by Cragg (Journal of Econometrics. It is possible to show that if the skedastic regression is correctly speciﬁed. and thus where ´ √ ³ ˜ n β F GLS − β →d N (0. ¢¢−1 ¡ ¡ .. There are three reasons.1. then FGLS is asymptotically equivalent to GLS. V n n i=1 .

This hypothesis does not imply that ξ i is independent of xi . Vα = E zi zi (7. and not a conditional mean.1 Under H0 and ei independent of xi . OLS is a more robust estimator of the parameter vector. and the cost is a reduction of robustness to misspeciﬁcatio 7.4) and constructing a Wald statistic. and this requires trimming to solve.3) for α simpliﬁes to ¡ ¡ 0 ¢¢−1 ¡ 2 ¢ E ξi . The GLS and FGLS estimators.2 Testing for Heteroskedasticity ¢ ¡ The hypothesis of homoskedasticity is that E e2 | xi = σ 2 . The main diﬀerences between popular “tests” is which transformations of xi enter zi . Second. the two tests coincide. q Most tests for heteroskedasticity take this basic form.2). It is consistent not only in the regression model. then the OLS and FGLS estimators will converge in probability to diﬀerent limits. If i this simpliﬁcation is replaced by the standard formula (under independence of the error). In the linear regression model the conditional mean of yi given xi = x is 77 .5) and the asymptotic variance (7. White (1980) proposed that the test for heteroskedasticity be based on setting zi to equal all non-redundant elements of xi . If the equation of interest is a linear projection. its squares. Theorem 7. We may therefore test this hypothesis by the estimation (7. require the assumption of a correct conditional mean. and replaced E ξ 2 with 2σ 4 which holds when ei is N(0. σ 2 ). FGLS will do worse than OLS in practice. on the other hand. 7. The point is that the eﬃciency gains from FGLS are built on the stronger assumption of a correct conditional mean.3 Forecast Intervals m(x) = E (yi | xi = x) = x0 β. The asymptotic distribution (7. the estimated conditional variances may contain more noise than information about the true conditional variances. Third. And the FGLS probability limit will depend on the particular function selected for the skedastic regression. but also under the assumptions of linear projection. as they will be estimating two diﬀerent projections.error. but the only diﬀerence is that they a ¢ allowed for general choice of zi . BreuschPagan (1979) proposed what might appear to be ¡ distinct test. This introduces an element of arbitrariness which is unsettling to empirical researchers. Motivated by the form of the asymptotic variance ˆ of the OLS estimator β.7) under independence show that this test has an asymptotic chi-square distribution. and all cross-products. in which case ξ i is ˜ independent of xi and the asymptotic variance (7. the Wald test of H0 is asymptotically χ2 . Typically. In this case. or equivalently that i H0 : α1 = 0 in the regression (7. individual estimated conditional variances may be negative.2.7) Hence the standard test of H0 is a classic F (or Wald) test for exclusion of all regressors from the skedastic regression (7.4). however. we impose the stronger hypothesis and test the hypothesis that ei is independent of xi .

We would also like a measure of uncertainty for ˆ the forecast. Letting h(β) = x0 β and θ = h(β). In general. A reasonable rule is the conditional mean m(x) as it is the mean-square-minimizing forecast. Notice that this is a (linear) ˆ ˆ θ function of β. s 7. We describe this setting as non-linear regression. but this turns out to be incorrect. ³ ´ ˆ ˆ The forecast error is ei = yi − m(x) = ei − x0 β − β . θ) = E (yi | xi = x) which is a non-linear function of the parameters θ. s(x) ˆ But no such asymptotic approximation can be made. the width of the conﬁdence set is dependent on x. p h i ˆ ˆ x0 β ± 2 n−1 x0 V x . Given the diﬃculty. ¡ ¢ ˆ Assuming E e2 | xi = σ 2 . As the out-of-sample error ei is ˆ ˆ independent of the in-sample estimate β. 78 . this has variance ³ ´³ ´0 ¡ ¢ ˆ ˆ Eˆ2 = E e2 | xi = x + x0 E β − β β − β x ei i = σ 2 (x) + n−1 x0 V x. e. σ 2 ). If we have an estimate of the conditional variance function. which is generally invalid. we want to estimate m(x) at a particular point x. To get an accurate forecast interval. Notice that this is diﬀerent from the standard ˆ ˆ error for the conditional mean. so p ˆ s(ˆ = n−1 x0 V x.4 NonLinear Least Squares In some cases we might use a parametric regression function m(x. this would require the asymptotic normality of the ratio ³ ´ ˆ ei − x0 β − β . In the present case. the validity of an asymptotic conﬁdence interval is based on the asymptotic normality of the studentized ratio. we need to estimate the conditional distribution of ei given xi = x. A point forecast ˆ is the estimated conditional mean m(x) = x0 β. so a standard ˆ i p ˆ error for the forecast is s(x) = σ 2 + n−1 x0 V x. For a given value of xi = x. then the forecast standard error is s(x) = σ 2 (x) + n−1 x0 V x ˜ It would appear natural to conclude that an asymptotic 95% forecast interval for yi is i h ˆ s x0 β ± 2ˆ(x) . we may want to forecast (guess) yi out-of-sample. we see that m(x) = ˆ = x0 β and Hβ = x. q 2 0 0 ˆ ˜ ˆ ˜ σ (x) = α z from (7. the natural estimate of this variance is σ 2 + n−1 x0 V x.6).g.In some cases. Thus an asymptotic 95% conﬁdence interval for m(x) is θ) It is interesting to observe that if this is viewed as a function of x. The only special exception is the case where ei has the exact distribution N (0. which is a much more diﬃcult task. many applied forecasters focus h i ˆ on the simple approximate interval x0 β ± 2ˆ(x) .

See Appendix E.8) Theorem 7. ¢¢−1 ¡ ¡ ¢¢ ¡ ¡ ¢¢−1 ¡ ¡ E mθi m0 e2 E mθi m0 V = E mθi m0 θi θi i θi Sketch of Proof. This can be done using arguments θ for optimization estimators. In other cases. ˆ must be found by numerical θ methods.Examples of nonlinear regression functions include x 1 + θ3 x m(x. θ) = θ1 + θ2 m(x.1 If the model is identiﬁed and m(x. ˆ ei . When it exists. θi 79 . θ))2 . but we won’t cover that argument here. θ) = θ1 + θ2 xθ3 m(x. ˆ ˆ One motivation for the choice of NLLS as the estimation method is that the parameter θ is the solution to the population problem minθ E (yi − m(xi . θ) is diﬀerentiable with respect to θ. then the FOC for minimization are 0= n X i=1 mθ (xi . let ∂ m(x.4. it is adopted as a ﬂexible approximation to an unknown regression function. The NLLS residuals are ei = yi − m(xi . mθ (x. θ))2 Since sum-of-squared-errors function Sn (θ) is not quadratic. Since ˆ →p θ0 . which alters some of the analysis. θ) is diﬀerentiable. we call this the nonlinear least squares (NLLS) θ). V ) θ where mθi = mθ (xi . G known x2 − θ5 m(x. so the minimization of Sn (θ) only needs to be examined for θ close to θ0 . Let 0 yi = ei + m0 θ0 . θ) = (θ1 + θ2 x1 ) 1 (x2 < θ3 ) + (θ4 + θ5 x1 ) 1 (x2 > θ3 ) In the ﬁrst ﬁve examples. m(x. estimator. ˆ is close to θ θ θ0 for n large. When the regression function is nonlinear. In the ﬁnal two examples. θ) = θ1 + θ2 exp(θ3 x) m(x. θ) = θ1 + θ2 x1 + (θ3 + θ4 x1 ) Φ θ6 m(x. θ) = G(x0 θ). θ) = ∂θ Nonlinear regression is frequently adopted because the functional form m(x. θ)ˆ (7. θ) is (generically) diﬀerentiable in the parameters θ. θ). θ0 ). θ) = θ1 + θ2 x + θ4 (x − θ3 ) 1 (x > θ3 ) µ ¶ m(x. θ) is suggested by an economic model. it must be shown that ˆ →p θ0 . First. When m(x. m is not diﬀerentiable with respect to θ3 . ´ √ ³ n ˆ − θ0 →d N (0. The least squares estimator ˆ minimizes the sum-of-squared-errors θ Sn (θ) = n X i=1 (yi − m(xi .

Hansen shows how to use simulation (similar to the bootstrap) to construct the asymptotic critical values (or p-values) in a given application. Hansen (1996).1. θ which is that stated in the theorem.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 . tests of H0 do not have asymptotic normal or chi-square distributions. ˆ and ei = yi − m(xi . m(xi . but these are not the only measures of central tendency. Thus the NLLS 0 estimator ˆ has the same asymptotic distribution as the (infeasible) OLS regression of yi on mθi . ˆ ˆ θ) ˆ θ). Furthermore. γ is not identiﬁed. θ0 ) + m0 (θ − θ0 ) θi = ei − m0 (θ − θ0 ) θi 0 = yi − m0 θ. θi Hence the sum of squared errors function is Sn (θ) = n X i=1 (yi − m(xi . The asymptotic theory of such tests have been worked out by Andrews and Ploberger (1994) and B. the parameter estimates are not asymptotically normally distributed. ¥ Based on Theorem 7.For θ close to the true value θ0 . We have discussed projections and conditional means. Thus when the truth is that β 2 = 0. θ0 )) − m(xi . 7. This renders the distribution theory presented in the previous section invalid. θ) ' (ei + m(xi . Suppose that m(xi . Thus we want to test H0 : β 2 = 0. An alternative good measure is the conditional median. θ) ' m(xi . θ))2 ' n X¡ ¢2 0 yi − m0 θ θi i=1 0 and the right-hand-side is the SSE function for a linear regression of yi on mθi . by a ﬁrst-order Taylor series approximation. and this is often a useful hypothesis (sub-model) to consider. In particular.4. θ) = β 0 zi + β 0 xi (γ). 80 . This means that under H0 . Identiﬁcation is often tricky in nonlinear regression models. θ0 ) + m0 (θ − θ0 ) . the model is yi = β 0 zi + εi 1 and both β 2 and γ have dropped out. θi Thus ¡ ¢ yi − m(xi . However. 1 2 The model is linear when β 2 = 0. an estimate of the asymptotic variance V is ˆ V = Ã 1X mθi m0 ˆ ˆ θi n i=1 n !−1 Ã 1X mθi m0 e2 ˆ ˆ θi ˆi n i=1 n !Ã 1X mθi m0 ˆ ˆ θi n i=1 n !−1 where mθi = mθ (xi . under H0 .

Two useful facts about the median are that (7. These facts deﬁnitions motivate three estimators of θ. The ﬁrst deﬁnition is the 50th empirical 1 P quantile. and let M ed (Y | X) = m(X) be this function evaluated at the random variable X. These distinctions are illusory. however. let Y be a continuous random variable. The median θ0 = M ed(Y ) is the value such that P (Y ≤ θ0 ) = P (Y ≥ θ0 ) = .10) The LAD estimator has the asymptotic distribution 81 . Let m(x) = M ed (Y | X = x) denote the conditional median of Y given X = x. The least absolute deviations (LAD) estimator of β minimizes this function ˆ β = argmin Ln (β) β Equivalently.5 • E (sgn (ei ) | xi ) = 0 • E (xi sgn (ei )) = 0 • β 0 = minβ E |yi − x0 β| i These facts motivate the following estimator.5. and the substantive assumption is that the median function is linear in x. Let ¯ 1 X¯ ¯yi − x0 β ¯ Ln (β) = i n i=1 n be the average of absolute deviations. The second is the value which minimizes n n |yi − θ| . and the third deﬁnition is the i=1 1 P solution to the moment equation n n sgn (yi − θ) . as i=1 these estimators are indeed identical.9) θ0 = argmin E |Y − θ| θ and E sgn (Y − θ0 ) = 0 where sgn (u) = ½ 1 −1 if u ≥ 0 if u < 0 is the sign function. Now let’s consider the conditional median of Y given a random variable X. the linear function M ed (yi | xi = x) = x0 β is the conditional median function.To recall the deﬁnition and properties of the median. i n n i=1 (7. The linear median regression model takes the form yi = x0 β + ei i M ed (ei | xi ) = 0 In this model. Conditional analogs of the facts about the median are • P (yi ≤ x0 β 0 | xi = x) = P (yi > x0 β | xi = x) = . it is a solution to the moment condition ³ ´ 1X ˆ xi sgn yi − x0 β = 0.

then there are many innovations near to the median. ¢¢−1 ¡ ¢¢−1 ¢¡ ¡ 1¡ ¡ 0 Exi x0 E xi x0 f (0 | xi ) E xi xi f (0 | xi ) i i 4 The variance of the asymptotic distribution inversely depends on f (0 | x) .5. When f (0 | x) is large.5. and this improves estimation of the median. ∂ g(β) = ∂β 0 ¡ ¡ ¢¢ ∂ 0 0 Exi 1 − 2 · 1 yi ≤ xi β ∂β £ ¡ ¡ ¢ ¢¤ ∂ = −2 0 E xi E 1 ei ≤ x0 β − x0 β 0 | xi i i ∂β " Z 0 # xi β−x0 β 0 i ∂ = −2 0 E xi f (e | xi ) de ∂β −∞ £ ¢¤ ¡ = −2E xi x0 f x0 β − x0 β 0 | xi i i i £ 0 ¤ ∂ 0 g(β 0 ) = −2E xi xi f (0 | xi ) . ∂β ˆ g(β) ' ´ ³ ∂ ˆ g(β 0 ) β − β 0 . Computation of standard error for LAD estimates typically is based on equation (7.1 ´ √ ³ˆ n β − β 0 →d N (0.5.1) n X √ −1/2 n (g n (β 0 ) − g(β 0 )) = −n gi (β 0 ) →d N (0. This can be done with kernel estimation techniques.11) This simpliﬁcation is similar to the simpliﬁcation of the asymptotic covariance of the OLS estimator under homoskedasticity. by a Taylor series expansion and the fact g(β 0 ) = 0 82 . (7. Let g(β) = Eg (β).10) is equivalent to g n (β) = 0. ˆ Proof of Theorem 7.1 is advanced. we provide a sketch here for completeness. by the central limit theorem (Theorm 5. the conditional density of the error at its median. In the special case where the error is independent of xi . ∂β 0 so Third. then f (0 | x) = f (0) and the asymptotic variance simpliﬁes V = (Exi x0 )−1 i 4f (0)2 (7. V ). Pn −1 0 β)) .Theorem 7. First. where V = and f (e | x) is the conditional density of ei given xi = x. Second using the law of iterated expectations and the chain rule of i diﬀerentiation.11).3. Exi x0 ) i i=1 since Egi (β 0 )gi (β 0 )0 = Exi x0 . the height of the error density at its median. While a complete proof of Theorem 7. See Chapter 16. The main diﬃculty is the estimation of f (0). We need where g n (β) = n i i=1 gi (β) and gi (β) = xi (1 − 2 · 1 (yi ≤ xi three preliminary result.1: Since sgn (a) = 1 − 2 · 1 (a ≤ 0) .

The quantile Qτ is the value such that τ (percent) of the mass of the distribution is less than Qτ . 1] the τ ’th quantile Qτ of a random variable with distribution function F (u) is deﬁned as Qτ = inf {u : F (u) ≥ τ } When F (u) is continuous and strictly monotonic. Exi x0 ) →d i 2 = N (0. ¥ 7. For the random variables (yi . The third line follows from an asymptotic empirical process argument. when F (y | x) is continuous and strictly monotonic in y. (7. then (7. xi ) with conditional distribution function F (y | x) the conditional quantile function qτ (x) is Qτ (x) = inf {y : F (y | x) ≥ τ } .Together ´ √ ³ ˆ n β − β0 ' ' ¶−1 √ ∂ ˆ ng(β) 0 g(β 0 ) ∂β ´ ¡ £ ¤¢−1 √ ³ ˆ ˆ = −2E xi x0 f (0 | xi ) n g(β) − g n (β) i µ ¤¢−1 √ 1¡ £ 0 n (g n (β 0 ) − g(β 0 )) E xi xi f (0 | xi ) 2 ¤¢−1 1¡ £ 0 E xi xi f (0 | xi ) N (0. so you can think of the quantile as the inverse of the distribution function. the quantile regression function qτ (x) describes how the τ ’th quantile of the conditional distribution varies with the regressors. This linear speciﬁcation assumes that Qτ (x) = β 0 x where the coeﬃcients β τ τ vary across the quantiles τ .6 Quantile Regression The method of quantile regression has become quite popular in recent econometric practice. As functions of x. V ). then F (Qτ (x) | x) = τ . For ﬁxed τ .12) Qτ = argmin Eρτ (U − θ) . For τ ∈ [0.9) for the median to all quantiles. The median is the special case τ = . θ where ρτ (q) is the piecewise linear function ½ −q (1 − τ ) q<0 ρτ (q) = qτ q≥0 = q (τ − 1 (q < 0)) .13) This generalizes representation (7. then F (Qτ ) = τ . If the random variable U has τ ’th quantile Qτ .5. The following alternative representation is useful. However for computational convenience it is typical to assume that they are (approximately) linear in x (after suitable transformations). the quantile regression functions can take any shape. Again. We then have the linear quantile regression model yi = x0 β τ + ei i 83 .

The null model is yi = x0 β + εi i 84 . the τ ’th conditional quantile of ei is zero. the quantile regression estimator β τ for β τ solves the minimization problem ˆ β τ = argmin Lτ (β) n β∈Rk n where Lτ (β) = n and ρτ (q) is deﬁned in (7. otherwise its properties are unspeciﬁed without further restrictions.5. the unconditional density of ei at 0.1 n β τ − β τ →d N (0. the asymptotic variance depends on the conditional density of the quantile regression error. ¡ ¡ ¢¢−1 ¡ ¢¢−1 ¢¡ ¡ Exi x0 E xi x0 f (0 | xi ) Vτ = τ (1 − τ ) E xi x0 f (0 | xi ) i i i ¢ 1X ¡ ρτ yi − x0 β i n i=1 In general. i By construction. if the model yi = x0 β + ei has i been ﬁt by OLS. Furthermore. let zi = h(xi ) denote functions of xi which are not linear functions of xi (perhaps 0˜ ˜ ˜ squares of non-binary regressors) and then ﬁt yi = x0 β +zi γ + ei by OLS. conventional Newton-type optimization methods are inappropriate. A asymptotic distribution theory for the quantile regression estimator can be derived using similar arguments as those for the LAD estimator in Theorem 7.12).where ei is the error deﬁned to be the diﬀerence between yi and its τ ’th conditional quantile x0 β τ . One simple test for neglected nonlinearity is to add nonlinear functions of the regressors to ˆ ˆ the regression.13). where and f (e | x) is the conditional density of ei given xi = x. since it has discontinuous derivatives. n numerical methods are necessary for its minimization.1. Vτ ). and form a Wald statistic i for γ = 0. 7. it is useful to have a general test of the adequacy of the speciﬁcation.6. Since the quanitle regression criterion function Lτ (β) does not have an algebraic solution. and we have the simpliﬁcation Vτ = τ (1 − τ ) ¡ ¡ 0 ¢¢−1 . then f (0 | xi ) = f (0) . E xi xi f (0)2 A recent monograph on the details of quantile regression is Koenker (2005). ´ √ ³ˆ Theorem 7. Thus. Another popular approach is the RESET test proposed by Ramsey (1969). and are widely available. fast linear programming methods have been developed for this problem. and test their signiﬁcance using a Wald test. When the error ei is independent of xi .7 Testing for Omitted NonLinearity If the goal is to estimate the conditional expectation E (yi | xi ) . ˆ Given the representation (7. Fortunately.

⎠ . In this case i ¡ ¢−1 2 ˜ σ = Q−1 σ 2 . Another is to regress yi on x1i and x2i jointly. Q−1 σ 2 < Q11 − Q12 Q−1 Q21 11 22 85 ¡ ¡ ¢ ¢−1 2 ¡ ¢−1 2 ˆ lim nV ar(β 1 ) = Ex1i x0 − Ex1i x0 Ex2i x0 Ex2i x0 σ = Q11 − Q12 Q−1 Q21 σ . then 22 Q11 > Q11 − Q12 Q−1 Q21 . since Q12 Q−1 Q21 > 0. The RESET test appears to work well as a test of functional form against a wide range of smooth alternatives. and n→∞ say. and consequently 22 ¡ ¢−1 2 σ . One estimator of β 1 is to regress 0 0 ˜ yi on x1i alone. β 2 ). and form the Wald statistic Wn for γ = 0. yielding predicted values yi = x0 β.14) may be written as ³ ´3 ³ ´m ³ ´2 ˜ ˆ ˜ ˆ ˜ ˆ yi = x0 β + x0 β γ 1 + x0 β γ 2 + · · · x0 β γ m−1 + ei ˜ ˜ i i i i which has essentially approximated G(·) by a m’th order polynomial.8 Irrelevant Variables yi = x0 β 1 + x0 β 2 + ei 1i 2i E (xi ei ) = 0. However. they will (typically) have diﬀerence asymptotic variances. yielding ˆ ˜ ˆ ˆ (β 1 .be an (m − 1)-vector of powers of yi . 1i 2i 2i 1i 22 . If Q12 = 0 (so the variables are orthogonal) then these two variance matrices equal. Under which conditions is β 1 or β 1 superior? It is easy to see that both estimators are consistent for β 1 . and the two estimators have equal asymptotic eﬃciency. Typically. To see why this is the case. or 4 seem to work best. note that (7. 7. 3. ⎟ zi = ⎝ . The comparison between the two estimators is straightforward when the error is conditionally ¢ ¡ homoskedastic E e2 | xi = σ 2 . Now let ˆ i ⎛ 2 ⎞ yi ˆ ⎜ . Thus the null is rejected at the α% level if Wn m−1 exceeds the upper α% tail critical value of the χ2 m−1 distribution.14) by OLS. Wn →d χ2 . Otherwise. Then run the auxiliary regression ˆ 0 ˜ yi = x0 β + zi γ + ei ˜ ˜ i ˆ which is estimated by OLS. To implement the test. It is particularly powerful at detecting single-index models of the form yi = G(x0 β) + εi i where G(·) is a smooth “link” function. m must be selected in advance. yi ˆm (7. β 1 = (X1 X1 )−1 (X1 Y ) . In the model x2i is “irrelevant” if β 1 is the parameter of interest and β 2 = 0. small values such as m = 2. It is easy (although somewhat tedious) to show that under the null hypothesis. lim nV ar( β 1 ) = Ex1i x0 1i 11 n→∞ say.

We thus consider the question of the inclusion of X2 in the linear regression Y = X1 β 1 + X2 β 2 + ε. n→∞ σx ˜ When µ 6= 0. 7. To simplify some of ¢ statistical discussion. we see that β 1 has a lower asymptotic variance. xKi = xK )?” is well posed. estimated ˆ ˆ ˆ2 the variances σ 2 and σ 2 . there is no clear ranking of the eﬃciency of the ˜ restricted estimator β 1 versus the unrestricted estimator. respectively. In contrast.. Let Exi = µ. It is important that the model selection question be well-posed. Let β 1 be the ˜ be the estimate under the constraint β = 0. take the model yi = β 0 + β 1 xi + ei and suppose that β 0 = 0. E (ε | X1 . . There are many possible desirable properties for a model selection procedure. “Which subset of (x1 . that it selects the true model with probability one if the sample is suﬃciently large.. For example. To ﬁx notation. X2 ) = 0 Note that M1 ⊂ M2 . x2i = σ 2 .. A model selection procedure is consistent if ´ ³ c → 1 P M = M1 | M1 ³ ´ c P M = M2 | M2 → 1 86 . σ2 ˜ lim nV ar( β 1 ) = 2 n→∞ σ x + µ2 while σ2 ˆ lim nV ar( β 1 )−1 = 2 . In many cases the problem of model selection can be reduced to the comparison of two nested models.. How does a researcher go about selecting an econometric speciﬁcation. models 1 and 2 are estimated by OLS. we will on ˆ1 ¡ occasion use the homoskedasticity assumption E e2 | x1i . when economic theory does not provide complete guidance? This is the question of model selection. the question: “What is the right model for y?” is not well posed.7). To be concrete. x Then under (6. and E (xi − µ)2 = σ 2 . we say that M2 is true if β 2 6= 0. i A model selection procedure is a data-dependent rule which selects one of the two models.9 Model Selection In earlier sections we discussed the costs and beneﬁts of inclusion/exclusion of variables. In the absence of the homoskedasticity assumption. as the larger problem can be written as a sequence of such comparisons.). E (ε | X1 . where X1 is n × k1 and X2 is n × k2 . etc. We c can write this as M. We conclude that the inclusion of irrelevant variable reduces estimation eﬃciency if these variables are correlated with the relevant variables. this result can be reversed when the error is conditionally heteroskedastic.. with residual vectors e1 and e2 . because it does not make clear the conditioning set. This is equivalent to the comparison of the two models M1 : M2 : Y = X1 β 1 + ε. One useful property is consistency. and β 1 0 (The least-squares estimate with the intercept omitted. . estimate of β 1 from the unconstrained model... However. the question. ˆ For example. Y = X1 β 1 + X2 β 2 + ε. X2 ) = 0.˜ ˆ This means that β 1 has a lower asymptotic variance matrix than β 1 . xK ) enters the regression function E(yi | x1i = x1 .

In contrast to the AIC.15) then where σ 2 is the variance estimate for model m. based on Bayesian arguments. One popular choice is the Akaike Information Criterion (AIC). the BIC places a larger penalty than the AIC on the number of estimated parameters and is more parsimonious. LRn →p 0. log(n) 87 . then P M ´ ³ c 1 − α but P M = M2 | M2 could vary dramatically. this rule only makes sense when the true model is ﬁnite dimensional. etc. as the rule tends to overﬁt. ¢ ¡ ˆ1 ˆ2 (7. His criterion. since (7. For some critical level α. k = While many modiﬁcations of the AIC have been proposed.16) holds under M1 . A common approach to model selection is to base the decision on a statistical test such as the Wald Wn . That is. since under M1 . k ´ ³ c P M = M1 | M1 P (AIC1 < AIC2 | M1 ) µ ¶ k1 k1 + k2 2 2 σ | M1 = P log(ˆ 1 ) + 2 < log(ˆ 2 ) + 2 σ n n = P (LRn < 2k2 | M1 ) ¡ ¢ → P χ22 < 2k2 < 1. If the truth is inﬁnite dimensional. M)) between the true density and the model density. Another common approach to model selection is to to a selection criterion. k A major problem with this approach is that the critical level α is indeterminate. the most popular to be one proposed by Schwarz. Indeed. Indeed. as ³ ´ c P M = M1 | M1 → 1 − α < 1. The AIC for model m is ¡ ¢ km ˆm AICm = log σ 2 + 2 . The reasoning which helps guide the choice of α in hypothesis testing (controlling Type I error) is ´ not ³ c = M1 | M1 ≈ relevant for model selection. ˆm The AIC can be derived as an estimate of the KullbackLeibler information distance K(M) = E (log f (Y | X) − log f (Y | X. let cα satisfy ¢ ¡ P χ22 > cα . and km is the number of coeﬃcients in the model.16) LRn = n log σ 2 − log σ 2 ' Wn →d χ22 . AIC selection is inconsistent. n (7. Then select M1 if Wn ≤ cα . else select M2 .However. depending on the sample size. The model selection rule is as follows. BIC model selection is consistent. it is more appropriate to view model selection as determining the best ﬁnite sample approximation. if α is set to be a small number.17) Since log(n) > 2 (if n > 8). The rule is to select M1 if AIC1 < AIC2 . Another problem is that if α is held ﬁxed. is ¡ ¢ km ˆm BICm = log σ 2 + log(n) . known as the BIC. else select M2 . then this model selection procedure is inconsistent. n (7.

which regressors enter the regression). The general problem is the model yi = β 1 x1i + β 2 x1i + · · · + β K xKi + εi . and the AIC or BIC in principle can be implemented by estimating all possible subset models. In the ordered case. 048. 88 . . 576. We have discussed model selection between two models. For example. and 220 = 1. However. 210 = 1024. .. In the unordered case. . xKi }. Also under M2 . a full-blown implementation of the BIC selection criterion would seem computationally prohibitive.15). Similarly for ˆ the BIC. selecting based on (7. β 2 = β 3 = · · · = β K = 0 M2 : β 1 6= 0. there are 2K such models. one can show that thus LRn →p ∞. . a model consists of any possible subset of the regressors {x1i . The AIC selection criteria estimates the K models by OLS. stores the residual variance σ 2 for each model. which can be a very large number. the models are M1 : β 1 6= 0. and then selects the model with the lowest AIC (7. . There are two leading cases: ordered regressors and unordered.17). E (εi | xi ) = 0 and the question is which subset of the coeﬃcients are non-zero (equivalently.. β 2 6= 0.so ´ ³ c P M = M1 | M1 = = P (BIC1 < BIC2 | M1 ) P (LRn < log(n)k2 | M1 ) µ ¶ LRn = P < k2 | M1 log(n) → P (0 < k2 ) = 1. β K 6= 0. MK : β 1 6= 0. In the latter case. β 2 6= 0. The methods extend readily to the issue of selection among multiple regressors.. β 3 = · · · = β K = 0 . . which are nested. log(n) µ LRn > k2 | M2 log(n) ¶ ´ ³ c P M = M2 | M2 = P → 1. .

where M1 = I − X X 0 Ω−1 X ˆ ¡ 0 −1 ¢−1 0 −1 0 (c) Prove that M1 Ω−1 M1 = Ω−1 − Ω−1 X X Ω X XΩ . the estimates (β. V . where xji is one of the xi .. ˜ the residual vector is e = Y − X β. V ar(e | X) = σ 2 Ω with Ω known. i ˆ ˆ suppose β is the OLS estimate of β with covariance matrix V . and the out-of-sample value of the regressors. . Let θ satisfy Eg(Yi − θ) = 0. in which situations would you expect that β would perform better than OLS? 89 (a) Why is this a reasonable estimator for σ 2 ? ¡ ¢−1 0 −1 XΩ . else equals zero). For any predictor g(xi ) for y. You wish to forecast an out-of-sample value of y given that ˆ ˆ ˆ ˆ X = x. σ 2 ). 4. Deﬁne g(u) = τ − 1 (u < 0) where 1 (·) is the indicator function (takes the value 1 if the argument is true. Is θ a quantile of the distribution of Yi ? 3. Show that the function g(x) which minimizes the MAE is the conditional median M (x). ji ˜ (a) In which contexts would β be a good estimator? ˜ (b) Using your intuition.7. based on a sample of size n. (b) Find an estimate of the variance of this forecast. the residuals e. (b) Prove that e = M1 e. 2.12). ˆ ˆ n−k 6. Verify equation (7. ˆ (a) Find a point forecast of y. . E(e | X) = 0. Let σ 2 be the estimate of σ 2 . xi ) be a random sample with E(Y | X) = Xβ. Let (yi .. 5. Thus the available information is the sample (Y.10 Exercises 1. In the homoskedastic regression model Y = Xβ + e with E(ei | xi ) = 0 and E(e2 | xi ) = σ 2 . X). wn ) and wi = x−2 . x. the GLS estimator is ¡ ¢−1 ¡ 0 −1 ¢ ˜ β = X 0 Ω−1 X XΩ Y . the mean absolute error (MAE) is E |yi − g(xi )| . Consider the Weighted Least Squares (WLS) estimator of β ¢−1 ¡ 0 ¢ ¡ ˜ X WY β = X 0W X where W = diag(w1 . In a linear model Y = Xβ + e.. and an estimate of σ 2 is ˆ s2 = 1 e0 Ω−1 e.

the variable ln Qi has a regression slope of a2 . Find an asymptotic 95% conﬁdence interval for g(x). the regression slope is a2 + a6 . Do you agree with this modiﬁcation? (b) Now try a non-linear speciﬁcation. . and V is the = g(x estimate of V ar ˆ . The model is non-linear because of the parameter a7 . The model is yi = xi β + ei E (ei | xi ) = 0 where xi ∈ R.. Do so. (ii) AIC criterion. This model is called a smooth threshold model. The model works best when a7 is selected so that several values (in this example. I recommend the concentration method: Pick 10 (or more or you like) values of a7 in this range. Examine the data and pick an appropriate range for a7 . Assess the merits of this new speciﬁcation using (i) a hypothesis test. 8. add the variable (ln Qi )2 to the regression. you estimated a cost function on a cross-section of electric companies. (iii) BIC criterion. θ) + ei with E (ei | xi ) = 0. For values of ln Qi much below a7 . In Chapter 6. (d) Calculate standard errors for all the parameters (a1 . Consider the two estimators ˆ β = ˜ β = Pn i=1 x y Pn i 2 i i=1 xi n 1 X yi . For values much above a7 . The equation you estimated was ln T Ci = β 1 + β 2 ln Qi + β 3 ln P Li + β 4 ln P Ki + β 5 ln P Fi + ei . θ is the NLLS estimator.. a7 ).. (7. at least 10 to 15) of ln Qi are both below and above a7 . impose the restriction a3 + a4 + a5 = 1. Exercise 13. 90 . are both estimators consistent for β? (b) Are there conditions under which either estimator is eﬃcient? 9.ˆ ˆ 7. and ﬁnd the value of a7 for which the sum of squared errors is minimized. You are interested in the conditional mean function E (yi | xi = x) = θ g(x) at some x. where zi = ln Qi (1 + exp (− (ln Qi − a7 )))−1 .18) (a) Following Nerlove. Consider model (7. Record the sum of squared errors. For each value of a7 . calculate zi and estimate the model by OLS. In addition. and the model imposes a smooth transition between these regimes. n xi i=1 (a) Under the stated assumptions. (c) Estimate the model by non-linear least squares.18) plus the extra term a6 zi . Suppose that yi ³ ´ i .

(h) Do the OLS and FGLS estimates diﬀer greatly? Note any interesting diﬀerences.dat contains 550 observations on 20 variables taken from the May 1978 current population survey. if desired. (d) Test whether the error variance is diﬀerent for men and women. (g) Using this model for the conditional variance.10. (e) Test whether the error variance is diﬀerent for whites and nonwhites. Estimate such a model. Report the results. (a) Start by an OLS regression of LNWAGE on the other variables. 91 . (b) Consider augmenting the model by squares and/or cross-products of the conditioning variables. test for general heteroskedasticity and report the results. (f) Construct a model for the conditional variance. Interpret. Variables are listed in the ﬁle cps78. The data ﬁle cps78. re-estimate the model from part (c) using FGLS. Report coeﬃcient estimates and standard errors. Interpret. Estimate your selected model and report the results. Note any interesting diﬀerences. (i) Compare the estimated standard errors. The goal of the exercise is to estimate a model for the log of earnings (variable LNWAGE) as a function of the conditioning variables.pdf. (c) Are there any variables which seem to be unimportant as a determinant of wages? You may re-estimate the model without these variables.

F ). For example. xi ) . Efron (1979) proposed the bootstrap.1 Deﬁnition of the Bootstrap Let F denote a distribution function for the population of observations (yi . The bootstrap distribution is itself random. for example an estimator ˆ or a t-statistic ˆ − θ /s(ˆ Note that we θ θ θ). x∗ . x1. Asymptotic inference is based on approximating Gn (x. Ideally. In the next sections we describe computation of the bootstrap distribution. x∗ ) denote random variables with the distribution Fn . write Tn as possibly a function of F . yn . ∗ . This is generally impossible since F is unknown.. yn . as it depends on the sample through the estimator Fn .. xn . Fn ) constructed on n 1. The exact CDF of Tn when the data are sampled from the distribution F is Gn (x. F ). F ) ³ ´ be a statistic of interest. n n ∗ Let (yi . the t-statistic is a function of the parameter θ which itself is a function of F.. Let Tn = Tn (y1 . we say that Tn is asymptotically pivotal and use the distribution function G(x) for inferential purposes. . n (8. inference would be based on Gn (x. Bootstrap inference is based on G∗ (x). In a seminal contribution. That is... A random sample from this i ∗ ∗ ∗ distribution is call the bootstrap data.Chapter 8 The Bootstrap 8.1) We call G∗ the bootstrap distribution. Gn (x. The statistic Tn = Tn (y1 . F ) = G(x) does not depend on F. F ) = limn→∞ Gn (x. The unknown F is replaced by a consistent estimate Fn (one choice is discussed in the next section). meaning that G changes as F changes. x∗ . Fn ). F ) depends on F. We call T ∗ this sample is a random variable with distribution Gn n n n ∗ the bootstrap statistic. The distribution of Tn is identical to that of Tn when the true CDF of Fn rather than F. When G(x.. P (T ∗ ≤ x) = G∗ (x). which makes a diﬀerent approximation. F ) we obtain G∗ (x) = Gn (x. F ) with G(x. Plugged into Gn (x. 92 . F ) = P (Tn ≤ x | F ) In general.

. The EDF is a consistent estimator of the CDF. In general. ∗ P (yi ≤ y. as the sample size gets larger. The method of moments estimator is the corresponding sample moment: n 1X 1 (yi ≤ y) 1 (xi ≤ x) . Furthermore. x) is called the empirical distribution function (EDF). For n = 25. x∗ ≤ x) = Fn (y. Thus by the WLLN (Theorem 5. That is. n. x) = n i=1 Figure 8. where 1(·) is the indicator function.8. F (y. x) (1 − F (y.2 The Empirical Distribution Function Fn (y. I have plotted the EDF and true CDF for three random samples of size n = 25. Notationally.. xi ). 1 (yi ≤ y) 1 (xi ≤ x) is an iid random variable with expectation F (y. √ n (Fn (y. To see this. in the Figure below. i 93 . x)) →d N (0. x) →p F (y. i = 1. x) .3. x) = P (yi ≤ y.1). note that for any (y.1). x). To see the eﬀect of sample size on the EDF..2) Fn (y. This is a population moment. x) − F (y. The random draws are from the N (0. x). by the CLT (Theorem 5. Fn is a nonparametric estimate of F. it is helpful to think of a random pair (yi . Note that while F may be either discrete or continuous. x∗ ) with the i distribution Fn .2. 50. Fn (y. (8. 1) distribution. x))) .1: Empirical Distribution Functions The EDF is a valid discrete probability distribution which puts probability mass 1/n at each ∗ pair (yi . x). the EDF step function gets uniformly close to the true CDF. Recall that F (y. Fn is by construction a step function. but the approximation appears to improve for the large n. . and 100. the EDF is only a crude approximation to the CDF. xi ≤ x) = E (1 (yi ≤ y) 1 (xi ≤ x)) .

´ For example. (yn . However. the t-ratio ˆ − θ /s(ˆ depends on θ. x∗ )) = h(y.2) as the estimate Fn of F. In consequence. When the statistic Tn³is a function of F. x∗ . As the bootstrap statistic replaces F with θ θ) ˆ Fn . in principle the distribution G∗ n ∗ ∗ could be calculated by direct methods. a bootstrap ∗ ∗ sample {y1 .. xi ) P (yi = yi . yn . . B is known as the number of bootstrap replications. x) i = = the empirical sample average.. Sampling from the EDF is particularly easy. x∗ } will necessarily have some ties and multiple values. as there are 2n possible samples {(y1 . x)dFn (y.. The popular alternative is to use simulation to approximate the distribution. xi ) from the observed data with replacement. . n i=1 8. Since the EDF Fn is a multinomial (with n support points). so long as the computational costs are reasonable. x∗ )}. 94 The bias of ˆ is θ . the value of θ implied by Fn . 8. xi ) .3 Nonparametric Bootstrap The nonparametric bootstrap is obtained when the bootstrap distribution (8. x∗ = xi ) i n 1X h (yi . x∗ . B = 1000 typically suﬃces. Typically θn = θ. (When in doubt use θ.. n X i=1 ∗ h (yi . x∗ . Fn ) is calculated for each bootstrap sample.∗ We can easily calculate the moments of functions of (yi . it is typically through dependence on a parameter. which is generally n 1 not a problem. n i This is repeated B times.. x∗ ) .) at the sample estimate ˆ θ. It is desireable for B to be large. yn . it similarly replaces θ with θn . ∗ • The random vectors (yi . . A theory for the determination of the number of bootstrap replications B has been developed by Andrews and Buchinsky (2000). The algorithm is identical to our discussion of Monte Carlo simulation.. Since Fn is a discrete probability distribution putting probability mass 1/n at each sample point. x∗ ) : i Z ∗ Eh (yi .. ∗ ∗ ∗ The bootstrap statistic Tn = Tn (y1 . the parameter ˆ estimate. with the following points of clariﬁcation: • The sample size n used for the simulation is the same as the sample size.4 Bootstrap Estimation of Bias and Variance θ τ n = E(ˆ − θ0 ). n 1 such a calculation is computationally infeasible..1) is deﬁned using the EDF (8. xi ) randomly from the sample. This is i equivalent to sampling a pair (yi . x∗ ) are drawn randomly from the empirical distribution.. sampling from the EDF is equivalent to random sampling a pair (yi .

so a biased-corrected estimator of θ should be larger than ˆ and θ.Let Tn (θ) = ˆ − θ. n If this is calculated by the simulation described in the previous section.. It appears superior to calculate bootstrap conﬁdence intervals. it is not clear if it is useful. in particular. in which case the normal approximation is suspected. Intuitively. The (estimated) bootstrap biased-corrected estimator is ˜∗ = ˆ − τ ∗ θ ˆn θ ∗ = ˆ − (ˆ − ˆ θ θ) θ ∗ = 2ˆ − ˆ . θ Let Tn = ˆ The variance of ˆ is ∗ θ Let Tn = ˆ . which are based on the asymptotic normal approximation to the t-ratio. θ θ If ˆ is biased. B ´2 X³ ∗ ∗ ˆ − ˆ∗ .. θ θ is upward-biased and the biased-corrected estimator should be lower than ˆ θ. n estimate of τ n is ∗ τ ∗ = E(Tn ). θ q ˆ ˆ∗ s(β) = Vn . 95 . the bootstrap makes θ the following experiment. θ θ θ. Similarly if ˆ is higher than ˆ then the estimator θ θ. x∗ . The primary use of asymptotic standard errors is to construct asymptotic conﬁdence intervals. θ. and we turn to this next. x∗ ) and Tn = ˆ − θn = ˆ − ˆ The bootstrap θ θ(y ∗ 1 θ θ θ. If this is lower than ˆ this suggests that the calculated from such samples? The answer is θ θ. θ = ˆ −ˆ B 1 X ˆ∗ ˆ θb − θ B b=1 ∗ ∗ Note. It has variance ∗ Vn = E(Tn − ETn )2 . Then θ ∗ τ n = E(Tn (θ0 )). that the biased-corrected estimator is not ˆ . Suppose that ˆ is the truth. ∗ ∗ the best guess is the diﬀerence between ˆ and ˆ . the estimate of τ ∗ is n τ∗ = ˆn = B 1 X ∗ Tnb B b=1 but τ n is unknown. Ideally. ∗ ∗ ∗ ∗ The bootstrap counterparts are ˆ = ˆ 1 . this would be ˜ = ˆ − τ n. Then what is the average value of ˆ θ θ ˆ∗ . ∗ ∗ ∗ Vn = E(Tn − ETn )2 . yn . ˆn = 1 θb θ V B b=1 The simulation estimate is A bootstrap standard error for ˆ is the square root of the bootstrap estimate of variance. the use of the bootstrap presumes that such asymptotic approximations might be poor. However.. it might be desirable to construct a biased-corrected estimator (one with reduced θ bias). estimator is downward-biased. While this standard error may be calculated and reported.

was popularized by Efron early in the history of the bootstrap.) Then C1 can alternatively be written as ∗ θ C1 = [ˆ + qn (α/2). if we deﬁne φ = f (θ) as the parameter of interest for a monotonic function f. ˆ lies in the region [qn (α/2). That is. Let Tn (θ) = ˆ − θ and let qn (α) be the quantile function of its distribution. qn (1 − α/2)]. θ−θ then Gn (−x. θ. F0 )) ³ ³ α ´´ α´ ³ − 1− 1− = 1− 2 2 = 1−α 96 . F0 ).. F0 ) which generally is not 1−α! There is one important exception. the quantile qn (x) is estimated by qn (x). qn (α. then percentile method ∗ ∗ applied to this problem will produce the conﬁdence interval [f (qn (α/2)). F ) = α. F ). ∗ qn (1 − α/2)]. Fn ) denote the quantile function of the bootstrap distribution. qn (1 − α/2)]. F0 ) denote the quantile function of the true sampling distribution. F ) may be non-unique. If ˆ 0 has a symmetric distribution. T ∗ }. This is because it is easy to compute. (These are the original quantiles. F ) denote its quantile function. . so ¡ ¢ 0 P θ0 ∈ C1 = Gn (−qn (α/2). ∗ ˆ∗ Computationally. with θ subtracted. Note that this function will change depending on the underlying statistic Tn whose distribution is Gn . θ It will be useful if we introduce an alternative deﬁnition C1 . as we show now. F0 ) − Gn (−qn (1 − α/2).5 Percentile Intervals For a distribution function Gn (x. The latter has coverage probability ³ ´ ¢ ¡ 0 θ P θ0 ∈ C1 = P ˆ + qn (α/2) ≤ θ0 ≤ ˆ + qn (1 − α/2) θ ´ ³ θ = P −qn (1 − α/2) ≤ ˆ − θ0 ≤ −qn (α/2) = Gn (−qn (α/2). ˆ + q ∗ (1 − α/2)]. This is the function which solves Gn (qn (α.8. This motivates a conﬁdence interval proposed by Efron: ∗ C1 = [qn (α/2). and also has the feature that it is translation invariant. let qn (α.. However. the x’th sample quantile of the ∗ . F ). F0 ) = 1 − Gn (x. which is a naturally good property. simple to motivate. θ n ˆ + qn (1 − α/2)]. F0 ) − Gn (−qn (1 − α/2). This is often called the percentile conﬁdence interval. C1 is in a deep sense very poorly motivated. F0 ) = (1 − Gn (qn (α/2).. q∗ The interval C1 is a popular bootstrap conﬁdence interval often used in empirical practice.] ∗ Let qn (α) = qn (α. θ This is a bootstrap estimate of the “ideal” conﬁdence interval 0 θ C1 = [ˆ + qn (α/2). F ) is discrete. f (qn (1 − α/2))]. and qn (α) = qn (α. In (1 − α)% of samples. [When Gn (x. but we will ignore such complications. as discussed in the section on Monte Carlo simulation. The simulated statistics {Tn1 nB ˆ∗ (1 − α)% Efron percentile interval is then [ˆn (α/2). F0 )) − (1 − Gn (qn (1 − α/2). θ Let Tn = ˆ an estimate of a parameter of interest.

025)].975). if the alternative is H1 : θ > θ0 . These t-statistics are sorted to ﬁnd the estimated quantiles qn (α) and/or qn (1 − α). these critical values can be estimated from a bootstrap simulation by sorting ³ ∗ ´ ∗ ∗ the bootstrap t-statistics Tn = ˆ − ˆ /s(ˆ ). it also follows that if there exists some monotonic transformation f (·) such that f (ˆ is symmetrically distributed about f (θ0 ). ˆ − q ∗ (.6 Percentile-t Equal-Tailed Interval Suppose we want to test H0 : θ = θ0 against H1 : θ < θ0 at size α.975). and this is important. but is not widely used in practice. many argue that the percentile interval should not be used unless the sampling distribution is close to unbiased and symmetric. θ) then the idealized percentile bootstrap method will be accurate. by the translation invariance argument presented above. which are centered at the sample estimate ˆ These are sorted θ θ θ. The 95% conﬁdence interval is then [θ − ˆ to yield the quantile estimates qn ˆ ˆn ∗ (. θ. n Computationally. that the bootstrap test θ θ θ ∗ statistic is centered at the estimate ˆ and the standard error s(ˆ ) is calculated on the bootstrap θ. θ Let Tn (θ) = ˆ − θ. 8. this³interval can be estimated from a bootstrap simulation by sorting the ´ ∗ ∗ bootstrap statistics Tn = ˆ − ˆ . θ ˆ − qn (α/2)]. ˆ − q ∗ (α/2)]. Therefore. where c would be selected so that θ θ) P (Tn (θ0 ) < c) = α. θ n Notice that generally this is very diﬀerent from the Efron interval C1 ! They coincide in the special case that G∗ (x) is symmetric about ˆ but otherwise they diﬀer. Note. Thus c = qn (α). Since this is unknown. C1 may perform quite poorly. The problems with the percentile method can be circumvented by an alternative method. and the test rejects if Tn (θ0 ) < qn (α).025) and q ∗ (. θ When ˆ does not have a symmetric distribution. θ ˆn qn ˆ This conﬁdence interval is discussed in most theoretical treatments of the bootstrap. ∗ (. θ This motivates a bootstrap analog ∗ θ C2 = [ˆ − qn (1 − α/2). Based on these arguments. We would set Tn (θ) = ´ ³ ˆ − θ /s(ˆ and reject H0 in favor of H1 if Tn (θ0 ) < c. ˆ∗ ˆ∗ 97 . Then so an exact (1 − α)% conﬁdence interval for θ0 would be 0 C2 = [ˆ − qn (1 − α/2). θ However.0 and this idealized conﬁdence interval is accurate. a bootstrap test replaces qn (α) with the bootstrap estimate ∗ ∗ qn (α). the bootstrap test rejects if Tn (θ0 ) > qn (1 − α). ∗ Similarly. Computationally. θ sample. C1 and C1 are designed for the case that ˆ has a symmetric distribution about θ0 . θ 1 − α = P (qn (α/2) ≤ Tn (θ0 ) ≤ qn (1 − α/2)) ´ ³ θ = P ˆ − qn (1 − α/2) ≤ θ0 ≤ ˆ − qn (α/2) .

discussed above. 1 − α = P (qn (α/2) ≤ Tn (θ0 ) ≤ qn (1 − α/2)) ´ ³ ´ ³ θ) θ = P qn (α/2) ≤ ˆ − θ0 /s(ˆ ≤ qn (1 − α/2) ³ ´ = P ˆ − s(ˆ n (1 − α/2) ≤ θ0 ≤ ˆ − s(ˆ n (α/2) . θ θ)qn This motivates a bootstrap analog θ θ)q ∗ C3 = [ˆ − s(ˆ n (1 − α/2). ∗ ∗ The bootstrap estimate is qn (α). θ θ)q θ θ)q ˆ − s(ˆ n (α/2)]. It is equal-tailed or central since the probability that θ0 is below the left endpoint approximately equals the probability that θ0 is above the right endpoint. qn (α) is the 1 − α quantile of the distribution of |Tn (θ0 )| . Equivalently. The ideal critical value c = qn (α) solves the equation Gn (qn (α)) = 1 − α. qn¯ ¯ ∗ ∗ ¯ ¯ ∗ θ θ θ statistics |Tn | = ¯ˆ − ˆ¯ /s(ˆ ). n n ∗ (α) is estimated from a bootstrap simulation by sorting the bootstrap tComputationally. The bootstrap test rejects if ∗ |Tn (θ0 )| > qn (α). the 1 − α quantile of the distribution of |Tn | . Let θ θ)q ∗ θ θ)q ∗ C4 = [ˆ − s(ˆ n (α).7 Symmetric Percentile-t Intervals Suppose we want to test H0 : θ = θ0 against H1 : θ 6= θ0 at size α. 8. and taking the upper α% quantile. θ θ)q ˆ − s(ˆ ∗ (α/2)]. or the number which solves the equation ∗ ∗ ∗ Gn (qn (α)) = G∗ (qn (α)) − G∗ (−qn (α)) = 1 − α. ˆ + s(ˆ n (α)]. ∗ 98 . each α/2.´ ³ θ θ). this is based on the critical values from the one-sided hypothesis tests. We would set Tn (θ) = ³ ´ ˆ − θ /s(ˆ and reject H0 in favor of H1 if |Tn (θ0 )| > c. Let Tn (θ) = ˆ − θ /s(ˆ Then so an exact (1 − α)% conﬁdence interval for θ0 would be 0 θ θ)q C3 = [ˆ − s(ˆ n (1 − α/2). which is a symmetric distribution function. where c would be selected so that θ θ) P (|Tn (θ0 )| > c) = α. This is often called a percentile-t conﬁdence interval. Note that P (|Tn (θ0 )| < c) = P (−c < Tn (θ0 ) < c) = Gn (c) − Gn (−c) ≡ Gn (c). Computationally.

where qn (α) is the (1 − α)% quantile of the distribution of ³ ∗ ³ ∗ ´0 ´ ∗ Wn = n ˆ − ˆ Vθ∗−1 ˆ − ˆ . or the size of the divergence for any particular sample size n. F ) = Φ n→∞ v or ³x´ Gn (x. it says nothing. Deﬁnition 8. F ) = Φ + o (1) . θ [This is a typical mistake made by practitioners. It is designed to work well since ³ ´ P (θ0 ∈ C4 ) = P ˆ − s(ˆ n (α) ≤ θ0 ≤ ˆ + s(ˆ n (α) θ θ)q ∗ θ θ)q ∗ ∗ = P (|Tn (θ0 )| < qn (α)) ' P (|Tn (θ0 )| < qn (α)) = 1 − α. the critical value qn (α) is found as the quantile from simulated values of W´ . The bootstrap test ∗ ∗ rejects if Wn ≥ qn (α). The following notation will be helpful. where qn (α) is the (1 − α)% quantile of the distribution of Wn .8. Basically. writing Tn ∼ Gn (x. Deﬁnition 8. If θ is a vector. Let an be a sequence. F ) then ³x´ . The ideal test rejects if Wn ≥ qn (α).8. and a function h(x) is odd if h(−x) = −h(x). We say that a function g(x) is even if g(−x) = g(x). and vice-versa. θ θ θ ˆ θ ∗ ∗ Computationally.∗ where qn (α) is the bootstrap critical value for a two-sided hypothesis test. Equivalently. an = O(n−r ) if it declines to zero like n−r .3 an = o(n−r ) if nr |an | → 0 as n → ∞. then to test H0 : θ = θ0 against H1 : θ 6= θ0 at size α. Let Tn be a statistic such that (8.8 Asymptotic Expansions Tn →d N (0.8. we would use a Wald statistic ³ ´ ³ ´0 ˆ θ Wn (θ) = n ˆ − θ Vθ−1 ˆ − θ θ or some other asymptotically chi-square statistic. A better asymptotic approximation may be obtained through an asymptotic expansion. about the rate v of convergence.] 8. C4 is called the symmetric percentile-t interval. Thus here Tn (θ) = Wn (θ). v 2 ). n ´ ³ ∗ ³ ∗ θ θ Note in the simulation that the Wald statistic is a quadratic form in ˆ − ˆ . The derivative of an even function is odd. (8. not ˆ − θ0 .2 an = O(1) if |an | is uniformly bounded. however.1 an = o(1) if an → 0 as n → ∞ Deﬁnition 8.4) v ¡ ¢ While (8. lim Gn (x.4) says that Gn converges to Φ x as n → ∞. 99 .3) ´ √ ³ˆ If Tn = n θ − θ0 then v = V while if Tn is a t-ratio then v = 1.

Theorem 8. A bootstrap test is based on G∗ (x).3). Fn ) − g1 (x. F ) + n−1 g2 (x. which from Theorem 8. F0 ) = n 1 n1/2 (g1 (x. To the second order. Moreover.1 Under regularity conditions and (8. F ) + O(n−3/2 ) Gn (x. The diﬀerence is Φ(x) − Gn (x. Fn ) − g1 (x. F ) converges to the normal limit at rate n1/2 . F ) ≈ Φ + n−1/2 g1 (x. F0 ) = Φ(x) + since v = 1. F0 ) + O(n−1 ) 1/2 1 n 1 g (x. First. F0 ) = 1 g (x. and g1 is continuous with respect to F. Fn ) = Φ(x) + n 1 g (x. √ Since Fn converges to F at rate n.1. To a second order of approximation. Gn (x. To a third order of approximation. Fn ) + O(n−1 ). F0 ) (Fn − F0 ) ∂F = O(n−1/2 ). An asymptotic test is based on Φ(x). ³x´ 1 1 + 1/2 g1 (x. the density function is skewed.1 as follows. so the order of the error is O(n−1/2 ). F0 )) + O(n−1 ). the diﬀerence between the bootstrap distribution and the true distribution is G∗ (x) − Gn (x. F ).8. 1/2 1 n Because Φ(x) appears in both expansions.1 has the expansion n G∗ (x) = Gn (x. the exact distribution is P (Tn < x) = Gn (x. ³x´ Gn (x. We can interpret Theorem 8.uniformly over x. F0 )) converges to 0 at rate n. F ) = Φ v n n 8.8. where g1 is an even function of x. we can assess the accuracy of one-sided hypothesis tests and conﬁdence regions based on an asymptotically normal t-ratio Tn . the diﬀerence √ (g1 (x. g1 (x. ³x´ Gn (x. g1 and g2 are diﬀerentiable functions of x and continuous in F relative to the supremum norm on the space of distribution functions.8. F0 ) + O(n−1 ) 1/2 1 n = O(n−1/2 ).8. F ) v which adds a symmetric non-normal component to the approximate density (for example. Fn ) − g1 (x. adding leptokurtosis). F0 ) ≈ ∂ g1 (x. F ) + g2 (x.9 One-Sided Tests Using the expansion of Theorem 8. and g2 is an odd function of x. F ) ≈ Φ + n−1/2 g1 (x. 100 . Heuristically. v Since the derivative of g1 is odd.

F0 ) + O(n−3/2 ) = O(n−1 ). F ) − Gn (−x. if Tn has exact distribution Gn (x. 101 2 g2 (x. Hence the diﬀerence between the asymptotic distribution and the exact distribution is Φ(x) − Gn (x. F ) ¶ µ 1 1 = Φ(x) + 1/2 g1 (x. Thus asymptotic critical values are taken from the Φ distribution. From Theorem 8.1. Similarly. as F is a function. F ) + g2 (x. then the random variable |X| has distribution function H(x) = H(x) − H(−x) since P (|X| ≤ x) = P (−x ≤ X ≤ x) = H(x) − H(−x). n or ∗ P (Tn ≤ x) = P (Tn ≤ x) + O(n−1 ). F0 ) = The order of the error is O(n−1 ). A two-sided hypothesis test rejects H0 for large values of |Tn | . F ) + O(n−3/2 ). F ) is only heuristic. 1). F ) − Gn (−x. = P (X ≤ x) − P (X ≤ −x) For example. F ). then |Z| has distribution function Φ(x) = Φ(x) − Φ(−x) = 2Φ(x) − 1. F ) n n µ ¶ 1 1 − Φ(−x) + 1/2 g1 (−x. This rate can be used to show that one-tailed bootstrap inference based on the tratio achieves a so-called asymptotic reﬁnement — the Type I error of the test converges at a faster rate than an analogous asymptotic test. Since Tn →d Z. which is an improved rate of convergence over the asymptotic test (which converged at rate O(n−1/2 )). if Z ∼ N (0. F ).The “derivative” ∂ ∂F g1 (x. We conclude that G∗ (x) − Gn (x. F0 ) distribution. then |Tn | →d |Z| ∼ Φ. we can calculate that Gn (x. n (8. and exact critical values are taken from the Gn (x.5) where the simpliﬁcations are because g1 is even and g2 is odd.10 Symmetric Two-Sided Tests If a random variable X has distribution function H(x) = P (X ≤ x). F ) = Gn (x. F ) = Gn (x. F ) + g2 (−x. 8. then |Tn | has the distribution function Gn (x. F ) + O(n−3/2 ) n n 2 = Φ(x) + g2 (x.8. n . F0 ) = O(n−1 ).

similar to a one-sided test. the asymptotic two-sided test has a better coverage rate than the asymptotic one-sided test. as it depends on the unknown V. which is not pivotal. but one-sided tests might have more power against alternatives of interest.5) to the bootstrap distribution.8. Gn (x. the choice between symmetric and equal-tailed conﬁdence intervals is unclear. This is in contrast to the use of the asymptotic distribution.Interestingly. F0 ) = Φ n v ´ ˆ v ³x x (ˆ − v) + O(n−1/2 ) v = −φ v v ˆ ˆ = O(n−1/2 ) 102 . whose error is O(n−1 ). Twosided tests will have more accurate size (Reported Type I error).11 Percentile Conﬁdence Intervals ´ √ ³ To evaluate the coverage rate of the percentile interval. and for the bootstrap ³x´ + O(n−1/2 ). A reader might get confused between the two simultaneous eﬀects. is an even function. g1 . Theorem 8. and bootstrap tests have better rates of convergence than asymptotic tests. and needs to be determined on a case-by-case basis. F0 ) = ∗ 2 (g2 (x. n Thus the diﬀerence between the bootstrap and exact distributions is Gn (x) − Gn (x. Another way of writing this is ∗ P (|Tn | < x) = P (|Tn | < x) + O(n−3/2 ) so the error from using the bootstrap distribution (relative to the true unknown distribution) is O(n−3/2 ).1 shows that a ﬁrst-order approximation ³x´ + O(n−1/2 ). meaning that the errors in the two directions exactly cancel out. we ﬁnd Gn (x) = Gn (x. Fn ) = Φ n v ˆ ˆ where V = V (Fn ) is the bootstrap estimate of V. V ). set Tn = n ˆ − θ0 . G∗ (x) = Gn (x. The analysis shows that there may be a trade-oﬀ between one-sided and two-sided tests. Two-sided tests have better rates of convergence than the one-sided tests. F0 )) + O(n−3/2 ) n = O(n−3/2 ). The diﬀerence is ³x´ ³x´ −Φ + O(n−1/2 ) G∗ (x) − Gn (x. Fn ) − g2 (x. 8. and g2 is continuous in F. This is because the ﬁrst term in the asymptotic expansion. Therefore. √ the last equality because Fn converges to F0 at rate n. We know that θ Tn →d N (0. Thus a two-sided bootstrap test also achieves an asymptotic reﬁnement. Fn ) + O(n−3/2 ). F ) = Φ v √ where v = V . Conﬁdence 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. Fn ) = Φ(x) + ∗ 2 g2 (x. Applying (8.

where Fn is the EDF. ∗ The non-parametric bootstrap distribution resamples the observations (yi . Based on these arguments. 2002) tends to advocate the use of the percentile-t bootstrap methods rather than percentile methods. The disadvantage is that the bootstrap distribution may be a poor approximation when the error is not independent of the regressors.Hence the order of the error is O(n−1/2 ). Therefore if standard errors are available.. and is thus an ineﬃcient estimator of the true distribution (when in fact the regression assumption is true... x∗ ) from the EDF. The advantage of the EDF is that it is fully nonparametric. 8. i i The the bootstrap distribution does not impose the regression assumption. 1992. . then all inferential statements are made conditionally on the 103 . √ The good news is that the percentile-type methods (if appropriately used) can yield nconvergent asymptotic inference. The advantage is that in this case it is straightforward to construct bootstrap distributions. If this is done. Fn ). the theoretical literature (e. A parametric method is to sample x∗ from an estimated parametric i distribution. n Any other consistent estimate of F0 may be used to deﬁne a feasible bootstrap estimator. i For the regressors x∗ ..) One approach to this problem is to impose the very strong assumption that the error εi is independent of the regressor xi . This is equivalent to treating the regressors as ﬁxed i in repeated samples. Yet these methods do not require the calculation of standard errors! This means that in contexts where standard errors are not available or are diﬃcult to calculate. it may be ineﬃcient in contexts where more is known about F. σ 2 ). it is unclear if there are any beneﬁts from using the percentile bootstrap over simple asymptotic methods. a nonparametric method is to sample the x∗ randomly from the EDF i i or sample values {x1 .g.. It is no better than the rate obtained from an asymptotic one-sided conﬁdence region. To impose independence. i which implies ∗ ˆ yi = x∗0 β + ε∗ i i E (x∗ ε∗ ) = 0 i i but generally E (ε∗ | x∗ ) 6= 0. A third approach sets x∗ = xi . A parametric i method is to generate the bootstrap errors ε∗ from a parametric distribution. it is suﬃcient to sample the x∗ and ε∗ independently. We discuss some bootstrap methods appropriate for the case of a regression model where yi = x0 β + εi i E (ei | xi ) = 0. such as the normal i ˆ ε∗ ∼ N (0. en }. There are diﬀerent ways to impose independence. The bad news is that the rate of convergence is disappointing. and then create i i ∗ = x∗0 β + ε∗ . But since it is fully nonparametric.12 Bootstrap Methods for Regression Models The bootstrap methods we have discussed have set G∗ (x) = Gn (x. the percentile bootstrap methods provide an attractive inference method. Horowitz. Hall. . A non-parametric method is ˆ yi i i e ˆ to sample the bootstrap errors ε∗ randomly from the OLS residuals {ˆ1 . and works in nearly any context.. it imposes no conditions. xn }.

.. Using the non-parametric bootstrap. Consider the following bootstrap procedure for a regression of yi on xi . yn } be a random sample from the empirical distribution function and let Tn = y ∗ be n ∗ ∗ its sample mean. One proposal which imposes the regression condition without independence is the Wild Bootstrap. ˆ 3.. 4. Unfortunately this is a harder problem. Take a random sample {y1 . OLS estimator from the regression of Y on X. ˆ Draw (with replacement) n such vectors.. Let ∗ ∗ ∗ {y1 . creating a random bootstrap data set (Y ∗ . ˆ∗ (b) Regress Y ∗ on X ∗ . . yn } with µ = Eyi and σ 2 = V ar(yi ). Find the population moments ETn and V ar(Tn ). n]. Let β denote the ˆ the OLS residuals.. which is a valid statistical approach. The idea is to construct a conditional distribution for ε∗ so that i E (ε∗ | xi ) = 0 i ¡ ∗2 ¢ E εi | xi = e2 ˆi ¡ ∗3 ¢ E εi | xi = e3 .. It does not really matter. and e = Y − X β ˆ (a) Draw a random vector (x∗ . draw a ˆ ˆ random integer i0 from [1. The methods discussed above are unattractive for most applications in econometrics because they impose the stringent assumption that xi and εi are independent. . generate ∗ bootstrap samples. you sample ε∗ using this two-point distribution. n} .. calculate the estimate ˆ on these samples and then calculate θ ∗ ∗ θ θ)/s(ˆ θ).13 Exercises 1. Consider the following bootstrap procedure. Find the bootstrap moments ETn and V ar(Tn ). X ∗ ). ei ) : i = 1. 2. . i 8. however. whether or not the xi are really “ﬁxed” or random. yielding OLS estimates β and any other statistic of interest. Set y∗ = x∗0 β + e∗ .. Show that this bootstrap procedure is (numerically) identical to the non-parametric bootstrap.observed values of the regressors. F0 (x) (1 − F0 (x))) .. Let the statistic of interest be the sample mean Tn = y n . Let Fn (x) denote the EDF of a random sample.. Tn = (ˆ − ˆ 104 . Show that √ n (Fn (x) − F0 (x)) →d N (0.. Typically what is desirable is to impose only the regression condition E (εi | xi ) = 0. .. e∗ ) from the pair {(xi . ˆi A conditional distribution with these features will preserve the main important features of the data. That is. 2. This can be achieved using a two-point distribution of the form Ã Ã √ ! ! √ 1+ 5 5−1 ∗ √ ei ˆ = P εi = 2 2 5 Ã Ã √ ! ! √ 1− 5 5+1 ∗ √ P εi = = ei ˆ 2 2 5 For each xi . and set x∗ = xi0 and e∗ = ei0 .

What is wrong with this procedure? Tn = θ/s(θ) n 6. and the colonial dummy. with variables listed in the ﬁle hprice1. θ respectively. Suppose that in an application. Using the non-parametric ∗ bootstrap. and thus reject H0 if ˆ ˆ > q ∗ (. ∗ ∗ ∗ Let qn (. You do this as follows. (a) Report the 95% Efron Percentile interval for θ. θ θ) ∗ 1000 samples are generated from the bootstrap distribution.95). Calculate 95% conﬁdence intervals for the regression coeﬃcients using both the asymptotic normal approximation and the percentile-t bootstrap. (b) Report the 95% Alternative Percentile interval for θ. You replace c with qn (.2 and s(ˆ = . 105 .3. You want to test H0 : θ = 0 against H1 : θ > 0. The dataﬁle hprice1.5% and 97. lot size.95) denote the 95% quantile of Tn . ˆ − s(ˆ n (. can you report the 95% Percentile-t interval for θ? 7. The ˆ are sorted. calculate the estimates ˆ on these samples and θ then calculate ∗ ∗ ∗ θ θ Tn = ˆ /s(ˆ ). θ θ)q ∗ θ θ)q ∗ θ/s(θ) 5.95).2.pdf. (c) With the given information.dat contains data on house prices (sales). and deﬁne the bootstrap conﬁdence interval h i C = ˆ − s(ˆ n (. you generate bootstrap samples.05) and qn (.Show that C exactly equals the Alternative percentile interval (not the percentile-t interval).05) .75 and 1.95) denote the 5% θ) ∗ and 95% quantiles of Tn . Estimate a linear regression of price on the number of bedrooms. size of house.95).5% quantiles of the ˆ are . Using the non-parametric bootstrap. ˆ = 1. and the 2. and ˆ is calculated on each θ ∗ ∗ θ sample. ∗ ∗ where s(ˆ is the standard error in the original data. Let qn (. The test for H0 is to reject if Tn = ˆ ˆ > c where c is picked so that Type I error is α.

β 0 ) = x(y − z 0 β) 106 (9. This situation is called overidentiﬁed. g(y. We know that without further restrictions. This model falls in the class (9. is not necessarily eﬃcient. where the dimensions of zi and xi are k × 1 and × 1 . an asymptotically eﬃcient estimator of β is the OLS estimator. x.1) by setting g(y. while β 1 is of dimension k < . 1 this class of models are called moment condition models.2) .1 Overidentiﬁed Linear Model yi = x0 β + ei i = x0 β 1 + x0 β 2 + ei 1i 2i E (xi ei ) = 0 where x1i is k × 1 and x2 is r × 1 with = k + r. The variables zi may be components and functions of xi . z. otherwise it is overidentiﬁed.Chapter 9 Generalized Method of Moments 9. β) be an × 1 function of a k × 1 parameter β with ≥ k such that Eg(yi . β) = x(y − x0 β). however. In our previous example. In econometrics. If k = the model is just identiﬁed. This method. β 0 ) = 0 (9. x. Let g(y.1) where β 0 is the true value of β. Now we can write the model as yi = x0 β 1 + ei 1i E (xi ei ) = 0. which can be written as 0 yi = zi β + ei Consider the linear model E (xi ei ) = 0. these are known as estimating equations. zi . Now suppose that we are given the information that β 2 = 0. but this is not required. In this case. as their are restrictions in E (xi ei ) = 0. how should β 1 be estimated? One method is OLS regression of yi on x1i alone. This is a special case of a more general class of moment condition models. with ≥ k. We call r the number of overidentifying restrictions. x. There are − k = r more moment restrictions than free parameters. xi . In the statistics literature. As an important special case we will devote special attention to linear moment condition models. z.

the square of the Euclidean length. then. This is a non-negative measure of the “length” of the vector g n (β). The GMM estimator minimizes Jn (β). and the GMM estimator is the MME.3 Distribution of GMM Estimator ¡ 0¢ Q = E xi zi Assume that Wn →p W > 0. then g n (β) = 0. Let and where gi = xi ei . β GMM = Z 0 XWn X 0 Z 9. β GMM does not change.1 β GMM = argmin Jn (β). The idea of the generalized method of moments (GMM) is to deﬁne an estimator which sets g n (β) “close” to zero. The ﬁrst order conditions for the GMM estimator are 0 = ∂ ˆ Jn (β) ∂β ∂ ˆ ˆ = 2 g n (β)0 Wn g n (β) ∂β ´ 1 1 ³ ˆ = −2 Z 0 XWn X 0 Y − Z β n n so which establishes the following.9. let Jn (β) = n · g n (β)0 Wn g n (β). if Wn = I.2. Jn (β) = n · g n (β)0 g n (β) = n · |g n (β)|2 . For example. for if Wn is replaced ˆ by cWn for some c > 0.2. ˆ Deﬁnition 9. gi (β) = xi yi − zi β = n n n i=1 i=1 Deﬁne the sample analog of (9. i i .3) The method of moments estimator for β is deﬁned as the parameter value which sets g n (β) = 0. ¡ ¢−1 0 ˆ Z XWn X 0 Y.1 ˆ 2Z 0 XWn X 0 Z β = 2Z 0 XWn X 0 Y While the estimator depends on Wn . but this is generally not possible when > k.2) g n (β) = (9.2 GMM Estimator n n ¢ 1¡ 0 ¢ 1X 1X ¡ 0 X Y − X 0 Zβ . For some × weight matrix Wn > 0. β ˆ Note that if k = . Proposition 9. the dependence is only up to scale. Then µ ¶ ¶ µ 1 0 1 0 Z X Wn X Z →p Q0 W Q n n 107 ¡ ¢ ¡ 0¢ Ω = E xi x0 e2 = E gi gi .

as we are only considering alternative weight matrices Wn . However. a better choice is Wn = (X 0 X)−1 . where In general. ˆ we still call β the eﬃcient GMM estimator. n β − β →d N 0. If it is known that E (gi (β)) = 0. and this is all that is known. The optimal weight matrix W0 is one which minimizes V. This yields the eﬃcient GMM estimator: ¢−1 0 ¡ ˆ Z XΩ−1 X 0 Y. as shown by Gary Chamberlain (1987). Q0 Ω−1 Q ¢−1 ¡ 0 ¢¡ ¢−1 ¡ . ˆ n i=1 gi = gi − g n . without imposing additional assumptions. no estimator has greater asymptotic eﬃciency than the eﬃcient linear GMM estimator. This is a weak concept of optimality. For example.1 ´ √ ³ˆ n β − β →d N (0. it turns out that the GMM estimator is semiparametrically eﬃcient.3. The proof is left as an exercise.3.2 For the eﬃcient GMM estimator. we can set Wn = I . W0 = Ω−1 is not known in practice. Ω) . Given any such ﬁrst0ˆ ˆ ˆ ˆ step estimator. no semiparametric estimator (one which is consistent globally for the class of models considered) ¢−1 ¡ .and µ µ ¶ ¶ 1 0 1 0 Z X Wn X e →d Q0 W N (0. ˆ∗ ˆ 108 . GMM estimators are asymptotically normal with “sandwich form” asymptotic variances. but it can be estimated consistently. Since the GMM estimator has this can have a smaller asymptotic variance than G0 Ω−1 G asymptotic variance.4 Estimation of the Eﬃcient Weight Matrix ˆ Given any weight matrix Wn > 0. In the linear model. as the distribution of the data is unknown. 9. This estimator is eﬃcient only in the sense that it is the best (asymptotically) in the class of GMM estimators with this set of moment conditions. This turns out to be W0 = Ω−1 . V ) . the GMM estimator β is consistent yet ineﬃcient. this is a semi-parametric problem. we can deﬁne the residuals ei = yi −zi β and moment equations gi = xi ei = g(wi . No estimator can do better (in this ﬁrst-order asymptotic sense). For any Wn →p W0 . This results shows that in the linear model. Q W ΩW Q Q0 W Q V = Q0 W Q Thus we have Theorem 9. β). β = Z 0 XΩ−1 X 0 Z ´ ³ ¡ ¢−1 ´ √ ³ˆ . ˆ Construct n 1X ˆ g n = g n (β) = gi . as it has the same asymptotic distribution. it is semiparametrically eﬃcient. ˆ We have described the estimator β as “eﬃcient GMM” if the optimal (variance minimizing) weight matrix is selected. n n We conclude: Theorem 9. Chamberlain showed that in this context.

A common alternative choice is to set Ã n !−1 1X 0 gi gi ˆˆ Wn = n i=1 1 X ∗ ∗0 gi gi ˆ ˆ n i=1 n !−1 = Ã 1X 0 gi gi − g n g 0 ˆˆ n n i=1 n !−1 . V = n Z 0 X g 0 g − ng n g 0 n where ˆ The β which minimizes this function is called the continuously-updated GMM estimator. First. i. Then the eﬃcient GMM estimator is ˆ ³ ¡ ¢−1 0 ´−1 0 ¡ 0 ¢−1 0 ˆ ˆˆ XZ Z X g g − ng n g 0 ˆˆ X Y. and GMM using Wn as the weight matrix is asymptotically eﬃcient. A simple solution is to use the centered moment conditions to construct the weight matrix. Heaton and Yaron (1996). and construct the residual ei = yi − zi β. There is little point in using an ineﬃcient GMM estimator as it is easy to compute. Here is a simple way to compute the eﬃcient GMM estimator. However. but can be numerically tricky to obtain in some cases. we actually mean “eﬃcient GMM”. we can let the weight matrix be considered as a function of β. J(β) = n · g n (β) n i=1 In most cases.4) above. Alastair Hall (2000) has shown that the uncentered estimator is a poor choice. (9. under the alternative hypothesis the moment conditions are violated. ˆ An estimator of the asymptotic variance of β can be seen from the above formula. 109 . these two estimators are asymptotically equivalent under the hypothesis of correct speciﬁcation. There is an important alternative to the two-step GMM estimator just described.5 GMM: The General Case In its most general form. ∗ gi (β) = gi (β) − g n (β) 9. When constructing hypothesis tests. as in (9. Set ³ ¡ ¢−1 0 ´−1 ˆ ˆˆ XZ . when we say “GMM”. This is a current area of research in econometrics.e. The criterion function is then Ã n !−1 1X ∗ 0 ∗ 0 gi (β)gi (β) g n (β). β = Z 0 X g 0 g − ng n g 0 n n ˆ Asymptotic standard errors are given by the square roots of the diagonal elements of V . ˆ and let g be the associated n × matrix. so the uncentered estimator will contain an undesirable bias term and the power of the test will be adversely aﬀected. GMM applies whenever an economic or statistical model implies the × 1 moment condition E (gi (β)) = 0. set Wn = (X 0 X)−1 . Egi 6= 0. Instead.4) which uses the uncentered moment conditions. and was introduced by L. 0ˆ ˆ ˆ ˆ estimate β using this weight matrix. The estimator appears to have some better properties than traditional GMM. Since Egi = 0. Then set gi = xi ei . Hansen.and deﬁne Wn = Ã Then Wn →p Ω−1 = W0 .

and is thus a natural test statistic for H0 : Egi = 0. However. and in this case it would be impossible to ﬁnd a value of β 1 so that both E (x1i (yi − x0 β 1 )) = 0 and E (x2i (yi − x0 β 1 )) = 0 hold simultaneously.6. and if the weight matrix is asymptotically eﬃcient. so that the linear equation may be written as yi = β 0 x1i + ei . ˆˆ n is a quadratic form in g n . β) = 0 holds. and thus g n can be used to assess whether or not the hypothesis that Egi = 0 is true or not. Since the GMM estimator depends upon the ﬁrst-stage estimator. 9. Under the hypothesis of correct speciﬁcation. The general theory of GMM estimation and testing was exposited by L. This estimator can be iterated if needed. Theorem 9. 1 2 It is possible that β 2 = 0. In this sense an 1i 1i exclusion restriction can be seen as an overidentifying restriction.6 Over-Identiﬁcation Test Overidentiﬁed models ( > k) are special in the sense that there may not be a parameter value β such that the moment condition Eg(wi . Thus the model — the overidentifying restrictions — are testable. Hansen (1982). Note that g n →p Egi . Identiﬁcation requires l ≥ k = dim(β). 1 it is possible that β 2 6= 0. G0 Ω−1 G .1 (Sargan-Hansen).Often.5. ´ ³ ¡ ¢−1 ´ √ ³ˆ Theorem 9. perhaps obtained by ﬁrst ˆ setting Wn = I. this is all that is known. where ³ ´−1 ∂ 0 ˆ ˆ ˆ ˆ Ω = (E (gi gi ))−1 and G = E ∂β 0 gi (β). The variance of β may be estimated by G0 Ω−1 G where P ∗ ∗0 P ∂ ˆ ˆ ˆ g g and G = n−1 ˆ ˆ Ω = n−1 0 gi (β). For example. The GMM estimator minimizes J(β) = n · g n (β)0 Wn g n (β) where g n (β) = and Wn = Ã 1X gi (β) n i=1 n ˜ ˜ with gi = gi (β) constructed using a preliminary consistent estimator β. The criterion function at the parameter estimates is J = n g 0 Wn g n n ¡ ¢−1 2 0 = n g n g 0 g − ng n g 0 gn . 110 . and then β recomputed. often the weight ˆ matrix Wn is updated. ˆ J = J(β) →d χ2−k .1 Under general regularity conditions. take the linear model yi = β 0 x1i + β 0 x2i + ei with E (x1i ei ) = 0 and E (x2i ei ) = 0. i i i i ∂β 1X 0 gi gi − g n g 0 ˆˆ n n i=1 n !−1 . n β − β →d N 0.

This reasoning is not compelling. h(β)=0 ˆ ˜ The two minimizing criterion functions are J(β) and J(β).7 Hypothesis Testing: The Distance Statistic We described before how to construct estimates of the asymptotic covariance matrix of the GMM estimates. current evidence suggests that the D statistic appears to have quite good sampling properties. the Wald statistic can work quite poorly. Based on this information alone. Hansen (1982) for the general case. D ≥ 0 2. but it is typically cause for concern. the hypothesis is H0 : h(β) = 0. This result was established by Sargan (1958) for a specialized case. This test shares the useful feature of LR tests in that it is a natural by-product of the computation of alternative models. In contrast. it is unclear what is wrong. and is the preferred test statistic. Proposition 9. however. 111 .The proof of the theorem is left as an exercise. If h is linear in β. 1. and some current research suggests that this restriction is not necessary for good performance of the test. the GMM criterion function is J(β) = n · g n (β)0 Wn g n (β) For h : Rk → Rr . a better approach is to directly use the GMM criterion function. When over-identiﬁed models are estimated by GMM. as this ensures that D ≥ 0. If the hypothesis is non-linear. and sometimes called a LR-like statistic (the LR is for likelihood-ratio). For a given weight matrix Wn . The estimates under H1 are ˆ β = argmin J(β) β and those under H0 are ˜ β = argmin J(β). then D equals the Wald statistic. it is customary to report the J statistic as a general test of model adequacy. D →d χ2 r 3. The GMM distance statistic is the diﬀerence ˜ ˆ D = J(β) − J(β). The idea was ﬁrst put forward by Newey and West (1987). 9. Newey and West (1987) suggested to use the same weight matrix Wn for both null and alternative. If h is non-linear. If the statistic J exceeds the chi-square critical value. and by L.7. The GMM overidentiﬁcation test is a very useful by-product of the GMM methodology. The degrees of freedom of the asymptotic distribution are the number of overidentifying restrictions. These may be used to construct Wald tests of statistical hypotheses. This is sometimes called the GMM Distance statistic.1 If the same weight matrix Wn is used for both null and alternative. and it is advisable to report the statistic J whenever GMM is the estimation method. we can reject the model.

β). How is k to be determined? This is an area of theory still under development. in linear regression. etc. an unconditional moment restriction can always be constructed. except that in this case zi = xi . 0 Our linear model yi = zi β + ei with instruments xi falls into this class under the stronger 0 assumption E (ei | xi ) = 0. It implies a conditional moment restriction of the form E (ei (β) | xi ) = 0 where ei (β) is some s × 1 function of the observation and the parameters. and restrictive. so is just-identiﬁed) yields the best GMM estimator possible. ei (β.. note that Ri = −E (zi | xi ) 112 ¡ ¢ σ 2 = E ei (β)2 | xi . For example. β)ei (β) which satisﬁes Egi (β) = 0 and hence deﬁnes a GMM estimator. the model implies more than an unconditional moment restriction of the form Egi (β) = 0. In a joint model of the conditional mean and variance ⎧ yi − x0 β ⎨ i . Which should be selected? This is equivalent to the problem of selection of the best instruments. but its form does help us think about construction of optimal instruments. In practice. In many cases. That is for any × 1 function φ(xi . γ) = ⎩ 0 β)2 − f (x )0 γ (yi − xi i Here s = 2. i Ai = −σ −2 Ri i . ei (β) = yi − x0 β. The obvious problem is that the class of functions φ is inﬁnite. The form was uncovered by Chamberlain (1987). Ai is unknown. β). while in a nonlinear i regression model ei (β) = yi − g(xi . .9. If xi is a valid instrument satisfying E (ei | xi ) = 0. Another approach is to construct the optimal instrument. are all valid instruments. 0 In the linear model ei (β) = yi − zi β. s = 1.. then xi . It turns out that this conditional moment restriction is much more powerful.8 Conditional Moment Restrictions In many contexts. and then use the ﬁrst k instruments. A recent study of this problem is Donald and Newey (2001). It is also helpful to realize that conventional regression models also fall into this class. x3 . Given a conditional moment restriction. Then ei (β) = yi − zi β. Let ¶ µ ∂ ei (β) | xi Ri = E ∂β and Then the “optimal instrument” is so the optimal moment is gi (β) = Ai ei (β). Setting gi (β) to be this choice (which is k × 1. x2 . we can set gi (β) = φ(xi . Take the case s = 1.. Which should be used? i i One solution is to construct an inﬁnite list of potent instruments.. than the unconditional moment restriction discussed above.

so ˆ Since i=1 pi gi β no recentering or adjustments are needed. the bootstrap applied J test will yield the wrong answer. it fails to achieve an asymptotic reﬁnement when the model is over-identiﬁed. not from the bootstrap data. caution should be applied when interpreting such results. z ∗ ) drawn from the EDF F . Z ∗ ). we need the best conditional mean model for zi given xi . as this is a very new area of research. A straightforward application of the nonparametric bootstrap works in the sense of consistently achieving the ﬁrst-order asymptotic distribution. However. deﬁne the bootstrap GMM criterion ³ ³ ´0 ´ ∗ ˆ ˆ J ∗ (β) = n · g ∗ (β) − g n (β) Wn g ∗ (β) − g n (β) n n ˆ ˆ∗ where e = Y − Z β are the in-sample residuals. this sampling scheme preserves the moment conditions of the model. so Ai = σ −2 xi . Furthermore. .. Thus according to ∗ . ∗ ˆ Let β minimize J ∗ (β). there are very few empirical applications of bootstrap GMM. i i 9. namely that improved eﬃciency can be obtained if the observations are weighted inversely to the conditional variance of the errors.. not just an arbitrary linear projection. this implies that the bootstrap estimator is ¢−1 ¡ ∗0 ∗ ∗ ¡ ∗0 ∗ ¡ ¢¢ ˆ∗ Z X W X Y − X 0e . and deﬁne all statistics and tests accordingly. Hence eﬃcient GMM is GLS. X ∗ . However. ˆ where g n (β) is from the in-sample data. note that the eﬃcient instrument Ai involves the estimation of the conditional mean of zi given xi . zi ). to get the best instrument for zi . 113 ..and so In the case of linear regression. Brown and Newey argue that this bootstrap procedure will be more eﬃcient than the Hall-Horowitz GMM bootstrap. and construct conﬁdence intervals for β. random variables (yi i i n ³ ³ ´´ ˆ ˆ E gi β = g n (β) 6= 0. identically as in the regression model. x∗ . ˆ Brown and Newey (2002) have an alternative solution. The eﬃcient instrument is also inversely proportional to the conditional variance of ei . In the linear model. Using the EDF of (yi . This is the same as the GLS estimator. as we i discussed earlier in the course. wn } with the empirical likelihood probabilities {ˆi } described in Chapter X. They note that we can sample from the p observations {w³. Ai = σ −2 E (zi | xi ) .9 Bootstrap GMM Inference ˆ Let β be the 2SLS or GMM estimator of β. ˆ ˆ The problem is that in the sample. To date. The bootstrap J statistic is J ∗ (β ). we can apply the ˆ bootstrap methods discussed in Chapter 8 to compute estimates of the bias and variance of β. zi = xi . 1 ´ Pn ˆ = 0. This has been shown by Hahn (1996). ˆ β = Z ∗0 X ∗ W ∗ X ∗0 Z ∗ n n ∗ This means that wi do not satisfy the same moment conditions as the population distribution. β is the “true” value and yet g n (β) 6= 0. jeopardizing the theoretical justiﬁcation for percentile-t methods. A correction suggested by Hall and Horowitz (1996) can solve the problem. In other words. i ¡ ¢ σ 2 = E e2 | xi . Given the bootstrap sample (Y ∗ . In the case of endogenous variables. xi .

a GMM estimator with arbitrary weight matrix). Take the model yi = zi β + ei with E (xi ei ) = 0. V − V0 is positive semi-deﬁnite (for then V0 is the smaller possible covariance matrix and W = Ω−1 is the eﬃcient weight matrix). we know that V − V0 is positive semi-deﬁnite if and only if V0−1 − V −1 = A is positive semi-deﬁnite. From matrix algebra. (c) Since Ω is positive deﬁnite. i 0 0ˆ ˆ 3. 114 . Deﬁne the estimate of the optimal GMM weight matrix Ã n !−1 1X 0 2 xi xi ei ˆ .g. verify that A can be written as ³ ¡ ¢−1 0 ´ G H. 2. and therefore positive semideﬁnite. Take the single equation Y E(e = Zβ + e | X) = 0 ˆ Assume E(e2 | xi ) = σ 2 . Show that Wn →p Ω i i i 4. the asymptotic variˆ ance of β GMM is ¡ ¢−1 0 ¡ ¢−1 V = Q0 W Q Q W ΩW Q Q0 W Q ¡ ¢−1 (a) Let V0 be this matrix when W = Ω−1 .10 Exercises yi = x0 β + ei i E (xi ei ) = 0 0 e2 = zi γ + η i i 1. Wn = n i=1 ¡ ¢ −1 where Ω = E x x0 e2 . Take the model E (zi ηi ) = 0. (b) We want to show that for any W.9. Let ei = yi − zi β where β is consistent for ˆ β (e. Letting H = CQ and G = C 0−1 W Q. Hint: The matrix I −G (G0 G)−1 G0 is symmetric and idempotent. In the linear model estimated by GMM with general weight matrix W. ˆ ˆ Find the method of moments estimators (β. then (X ´ ¢−1 ¡ √ ³ ˆ n β − β →d N (0. Show that V0 = Q0 Ω−1 Q . γ ) for (β. Show that if β is estimated by GMM with weight matrix Wn = i 0 X)−1 . A = H 0 I − G G0 G (d) Show that A is positive semideﬁnite. γ). Write out the matrix A. σ 2 Q0 M −1 Q ) 0 where Q = E(xi zi ) and M = E(xi x0 ). there exists a nonsingular matrix C such that C 0 C = Ω−1 .

h(β)=0 (9. λ) = 1 Jn (β) + λ0 (R0 β − r) where λ is 2 s × 1. In the linear model Y = Xβ + e with E(xi ei ) = 0. 115 . is deﬁned as Jn (β) = ˜ β = argmin Jn (β).6) equals the Wald statistic. The equation of interest is yi = g(xi . zi is l × 1 and β is k × 1.6) (a) Show that you can rewrite Jn (β) in (9. β) + ei E (zi ei ) = 0. The GMM estimator of β. 6. λ = R0 Vn R (c) Show that if R0 β = r then ´ √ ³˜ n β − β →d N (0. Deﬁne the Lagrangian L(β. VR ) where ¡ ¢−1 0 R V. the distance statistic D in (9.5) n 1 P ˆ where ei are the OLS residuals and Ωn = n n xi x0 e2 .5) as ³ ´ ³ ´0 ˆ ˆ ˆ −1 β − β Jn (β) = β − β Vn where (b) Now focus on linear restrictions: h(β) = R0 β − r. Show that the minimizer is ³ ´−1 ³ ´ ˆ ˜ ˆ ˆ ˆ R0 β − r β = β − Vn R R0 Vn R ´ ´−1 ³ ³ ˆ ˆ ˆ R0 β − r . zi ). h(β)=0 The GMM test statistic (the distance statistic) of the hypothesis h(β) = 0 is ˜ D = Jn (β) = min Jn (β). l ≥ k. subject ˆ i ˆi i=1 to the restriction h(β) = 0. xi . the Generalized Method of Moments (GMM) criterion function for β is deﬁned as 1 ˆn (Y − Xβ)0 X Ω−1 X (Y − Xβ) (9. VR = V − V R R0 V R (d) Show that in this setting.5. Show how to construct an eﬃcient GMM estimator for β. The observed data is (yi . Vn = X X i=1 ˜ and hence R0 β = r. Thus ˜ β = argmin Jn (β) R0 β−r Ã n ! ¡ 0 ¢−1 ¡ 0 ¢−1 X 0 2 ˆ XX xi xi ei ˆ .

7. Take the linear model 0 yi = zi β + ei E (xi ei ) = 0. Deﬁne ¶µ ¶−1 µ 1 0 ˆ −1 1 0 1 0 ˆ −1 0−1 0 1 0 XZ Z XΩ XZ Z XΩ C Dn = Il − C n n n n 1 0 Xe g n (β 0 ) = n ¡ ¢ 0 R = C 0 E xi zi Show that Jn →d χ2 as n → ∞ by demonstrating each of the following: l−k (a) Since Ω > 0. we can write Ω−1 = CC 0 and Ω = C 0−1 C −1 ³ ´0 ³ ´−1 ˆ ˆ ˆ C 0 ΩC (b) Jn = n C 0 g n (β) C 0 g n (β) ˆ (c) C 0 g n (β) = Dn C 0 g n (β 0 ) (d) Dn →p Il − R (R0 R)−1 R0 (e) n1/2 C 0 g n (β 0 ) →d N ∼ N (0. l−k Hint: Il − R (R0 R)−1 R0 is a projection matrix. Il ) ³ ´ (f) Jn →d N 0 Il − R (R0 R)−1 R0 N ³ ´ (g) N 0 Il − R (R0 R)−1 R0 N ∼ χ2 . 116 .. Let ˆ ˆ ˆ Jn = ng n (β)0 Ω−1 g n (β) denote the test of overidentifying restrictions. ˆ and consider the GMM estimator β of β.

. (10. In this case the moment condition places no additional restrictions on the multinomial distribution. the log-likelihood function for this multinomial distribution is n X ln(pi ).1) we ﬁnd that the MLE is pi = n−1 yielding the log-likelihood −n log(n). The n ﬁrst order conditions are 0 = p−1 − µ. To be a valid multinomial distribution. Now consider the case of an overidentiﬁed model with moment condition Egi (β 0 ) = 0 where g is × 1 and β is k × 1 and for simplicity we write gi (β) = g(yi . (10. pn ) are those which maximize the log-likelihood subject to the constraint (10.. This is equivalent to maximizing Ã n ! n X X log(pi ) − µ pi − 1 i=1 i=1 where µ is a Lagrange multiplier.3) i=1 117 . these probabilities must satisfy the requirements that pi ≥ 0 and n X pi = 1.1) i=1 First let us consider a just-identiﬁed model.. The multinomial distribution which places probability pi at each observation (yi . 2001) and has been extended to moment condition models by Qin and Lawless (1994). xi ... β). pn ) which places probability pi at each observation. .1 Non-Parametric Likelihood Since each observation is observed once in the sample.. The maximum likelihood estimator of the probabilities (p1 . The idea is to construct a multinomial distribution F (p1 . pn ) = i=1 An alternative to GMM is empirical likelihood. The idea is due to Art Owen (1988.. ...2) Ln (p1 . zi .Chapter 10 Empirical Likelihood 10. Combined with i the constraint (10. . xi . zi ) will satisfy this condition if and only if n X pi gi (β) = 0 (10.1). It is a non-parametric analog of likelihood estimation.

λ(β)) = −n ln (n) − n X i=1 ˆ The EL estimate β is the value which maximizes Ln (β). or equivalently minimizes its negative ˆ (10.. λ) is a convex function of λ. the Lagrange multiplier λ(β) minimizes Rn (β. λ.7) 118 . but must be obtained numerically (see section 6. The solution (when it exists) is well deﬁned since Rn (β. λ). The ﬁrst-order-conditions of and λ are 1 = µ + nλ0 gi (β) pi n X pi = 1 i=1 n X i=1 The empirical likelihood estimator is the value of β which maximizes the multinomial loglikelihood (10. we also obtain the Lagrange multiplier λ = λ(β). pn . The Lagrangian for this maximization problem is Ã n ! n n X X X ln(pi ) − µ pi − 1 − nλ0 pi gi (β) L∗ (β. (see section 6. p1 . we ﬁnd µ = n and 1 ¢.6) β = argmin [−Ln (β)] β ¡ ¢ ln 1 + λ(β)0 gi (β) and maximized empirical likelihood ˆ Numerical methods are required for calculation of β.where λ and µ are Lagrange multipliers.4) (10. probabilities 1 ³ ´´ . Ln (β) = Rn (β.2) subject to the restrictions (10. p (10. Multiplying the ﬁrst equation by pi . summing over i. and using the second and third equations.. µ.1) and (10.3). This yields the (proﬁle) empirical log-likelihood function for β. . pi gi (β) = 0.5). The solution cannot be obtained explicitly. µ) = n i=1 i=1 i=1 L∗ n with respect to pi . λ ¡ ¢ ln 1 + λ0 gi (β) . (10.5) This minimization problem is the dual of the constrained maximization problem. pi = ¡ n 1 + λ0 gi (β) Substituting into L∗ we ﬁnd n Rn (β. λ) = −n ln (n) − n X i=1 For given β. λ) : λ(β) = argmin Rn (β.. pi = ³ ˆ ˆ ˆ0 n 1 + λ gi β ˆ Ln = n X i=1 ln (ˆi ) .5) ˆ ˆ As a by-product of estimation.

β) = −xi zi .1 Under regularity conditions. i=1 i=1 i=1i Expanding (10. (β. β λ ∂ ³ ´. ˆ The asymptotic variance V for β is the same as for eﬃcient GMM. and n β − β 0 and nλ are asymptotically independent. G = −E (xi zi ). g n = n n gi (β 0 ) and Ωn = n n g (β 0 ) gi (β 0 )0 . ˆ ˆ Proof.9) and (10. in the linear model.13) yields n n Expanding (10. λ) = − (10.11) around β = β 0 and λ = λ0 = 0 yields ³ ´ ˆ ˆ 0 ' −g n − Gn β − β 0 + Ωn λ (10.11) 0 = ∂λ ˆ0 ˆ i=1 1 + λ gi β ³ ´0 ˆ n X Gi . Thus the EL estimator is asymptotically eﬃcient.14) ³ ´ ˆ ˆ 0 ' −G0 Ω−1 g n − G0 Ω−1 Gn β − β 0 + G0 Ω−1 Ωn λ n n n n n n ³ ´ ˆ = −G0 Ω−1 g n − G0 Ω−1 Gn β − β 0 n n n n 119 . ´ √ ³ ˆ n β − β 0 →d √ ˆ nλ →d N (0. and Ω = E xi x0 e2 .8) and ¢ 0 0 For example.12) around β = β 0 and λ = λ0 = 0 yields ³ ´ ˆ 0 ' G0 λ − λ0 . Gi (.12) ∂β ˆ ˆ0 i=1 1 + λ gi β 1 P 1 P 1 P Let Gn = n n Gi (β 0 ) . 0 = Rn (β. V ) ¢−1 ¡ V = G0 Ω−1 G ¡ ¢−1 0 Vλ = Ω − G G0 Ω−1 G G G = EGi (β 0 ) ¡ ¢ Ω = E gi (β 0 ) gi (β 0 )0 (10.10.13) Premultiplying by G0 Ω−1 and using (10. Vλ ) ´ √ ˆ √ ³ˆ where V and Vλ are deﬁned in (10.10). λ) jointly solve ³ ´ ˆ n gi β X ∂ ³ ³ ´´ Rn (β. λ) = − (10.2.9) (10.2 Deﬁne Asymptotic Distribution of EL Estimator Gi (β) = ∂ gi (β) ∂β 0 (10. n (10. i i Theorem 10.10) ¡ Ω−1 N (0.

V ) ˆ Solving (10.3. and it is advisable to report the statistic LRn whenever EL is the estimation method. by a Taylor expansion.15).17) 2 ln 1 + λ gi β . They are asymptotically ﬁrst-order equivalent. Vλ ) Furthermore.ˆ Solving for β and using the WLLN and CLT yields ´ ¢−1 0 −1 √ ¡ √ ³ ˆ ' − G0 Ω−1 Gn n β − β0 Gn Ωn ng n n n ¡ 0 −1 ¢−1 0 −1 G Ω N (0. Proof. n ³ ´´ ´ √ ³ 1 X ³ ˆ ˆ √ g wi . it is an asymptotically eﬃcient estimator for β. The overidentiﬁcation test is a very useful by-product of EL estimation. Since the EL estimator achieves this bound. Twice the diﬀerence between the unrestricted empirical likelihood −n log (n) and the maximized empirical likelihood for the model (10. 120 . This means that no consistent estimator for this class of models can have a lower asymptotic variance than V .16). The same statistic can be constructed for empirical likelihood.15) (10. tests are based on the diﬀerence in the log likelihood functions. (10.14) for λ and using (10. First. The EL overidentiﬁcation test is similar to the GMM overidentiﬁcation test.15) yields ³ ¡ ¢−1 0 −1 ´ √ √ ˆ nλ ' Ω−1 I − Gn G0 Ω−1 Gn Gn Ωn ng n n n n ³ ¡ 0 −1 ¢−1 0 −1 ´ N (0.7) is n ³ ´´ ³ X ˆ ˆ0 (10. and (10. and have the same interpretation. Ω) →d Ω−1 I − G G Ω G GΩ = Ω−1 N (0.3 Overidentifying Restrictions In a parametric likelihood context. 10. LRn = i=1 Theorem 10. since ³ ¡ ¢−1 0 ´ G0 I − Ω−1 G G0 Ω−1 G G =0 (10. ¥ Chamberlain (1987) showed that V is the semiparametric eﬃciency bound for β in the overidentiﬁed moment condition model. β 0 ) = 0 then LRn →d χ2−k .1 If Eg(wi . β ' n g n + Gn β − β 0 n i=1 ³ ¡ ¢−1 0 −1 ´ √ Gn Ωn ng n ' I − Gn G0 Ω−1 Gn n n √ ˆ ' Ωn nλ.16) ´ √ ˆ √ ³ˆ n β − β 0 and nλ are asymptotically uncorrelated and hence independent. Ω) GΩ G →d = N (0.

Instead we use the diﬀerence in log-likelihoods: ³ ´ ˆ ˜ LRn = 2 Ln − Ln . the simple overidentifying restrictions test (10.wisc.17) is not appropriate. To test the hypothesis h(β) while maintaining (10. since ln(1 + x) ' x − x2 /2 for x small. h(β)=0 ˜ Fundamentally. the restricted EL estimator β is simply an EL estimator with − k + a overidenti˜ ˆ fying restrictions.18) are assumed to hold and are not being challenged (at least for the test discussed in this section).1 Under (10. By “maintained” we mean that the overidentfying restrictions contained in (10. Theorem 10. 10.18) and H0 : h(β) = 0. so there is no fundamental change in the distribution theory for β relative to β.18).4.prc 121 . 10. Vλ ) ¥ where the proof of the ﬁnal equality is left as an exercise. LRn →d χ2 . a The proof of this result is more challenging and is omitted.4 Testing Egi (β) = 0 (10. Vλ )0 Ω−1 N (0.18) Let the maintained model be where g is × 1 and β is k × 1. The hypothesis of interest is h(β) = 0. where h : Rk → Ra .5 Numerical Computation Gauss code which implements the methods discussed below can be found at http://www.ssc. LRn = n X i=1 ˆ0 ' 2λ →d ˆ0 ˆ ' nλ Ωn λ = χ2−k n X i=1 ³ ´´ ³ ˆ ˆ0 2 ln 1 + λ gi β i=1 n ³ ´ X ³ ´ ³ ´0 ˆ − λ0 ˆ ˆ ˆ ˆ gi β gi β gi β λ N (0.edu/~bhansen/progs/elike. The restricted EL estimator and likelihood are the values which solve ˜ β = argmax Ln (β) h(β)=0 ˜ ˜ Ln = Ln (β) = max Ln (β). This test statistic is a natural analog of the GMM distance statistic.Second.

λ) G∗ (β. λ)0 − ∂β∂β Rn (β. λ) = i i ∂β∂β 0 1 + λ0 gi (β) i=1 n where δ > 0 is a scalar steplength (to be discussed next). λ) = i The ﬁrst derivatives of (10. λ) = ∂λ∂λ i=1 ¶ n 2 Xµ ∂ Gi (β) ∗ gi (β. λ)0 0 Rn (β. Eﬃcient convergence requires a good choice of steplength δ. For p = 0. λp ) A quadratic function can be ﬁt exactly through these three points. A complication is that λ must be constrained so that 0 ≤ pi ≤ 1 which holds if ¢ ¡ n 1 + λ0 gi (β) ≥ 1 122 (10. The starting value λ1 can be set to the zero vector. λ) gi (β. δ 4R2 + 4R0 − 8R1 yielding the steplength to be plugged into (10. Set δ 0 = 0. One method uses the following quadratic approximation. λ) = − gi (β. Deﬁne ∗ gi (β. δ 1 = 1 and δ 2 = 1.19) X ∂2 ∗ ∗ gi (β. λ) yielding the iteration rule λj+1 = λj − δ (Rλλ (β. λj ))−1 Rλ (β.Derivatives The numerical calculations depend on derivatives of the dual likelihood function (10. λj )) Rp = Rn (β. λ) = − ∂β n X i=1 G∗ (β. 1. λ) G∗ (β. λj ) is smaller than some prespeciﬁed tolerance. λ) .. i Rββ = Inner Loop The so-called “inner loop” solves (10. 2.19) is continued until the gradient Rλ (β. λj ))−1 Rλ (β.19). λ) ∂λ i=1 n ∂ Rn (β. The value of δ which minimizes this quadratic is ˆ = R2 + 3R0 − 4R1 . λ)0 − Rn (β.5) for given β.4) are Rλ = Rβ = The second derivatives are Rλλ = Rλβ = ¡ Gi (β)0 λ 1 + λ0 gi (β) gi (β) ¢ 1 + λ0 gi (β) X ∂ ∗ Rn (β. The modiﬁed Newton method takes a quadratic approximation to Rn (β. λ) = G∗ (β.20) . The iteration (10. λj ) . set 2 λp = λj − δ p (Rλλ (β. (10.4). λ) = i ∂λ∂β 0 1 + λ0 gi (β) i=1 ⎞ ⎛ ¡ 0 ¢ ∂2 n 2 X 0 gi (β) λ ∂ ⎠ ⎝G∗ (β.

123 . where λβ = ∂ −1 λ(β) = −Rλλ Rλβ . λ(β)) + λ0 Rλ (β. λ(β)) = 0. λ) = 0 at λ = λ(β).6). and the rule is iterated until convergence. The gradient for (10. It is not guaranteed that Lββ > 0.20) fails. the stepsize δ needs to be decreased. λ(β)) + Rλβ λβ + λ0 Rλβ + λ0 Rλλ λβ β β −1 0 = Rλβ Rλλ Rλβ − Rββ . λ(β)) β ∂β ¡ ¢ 0 = − Rββ (β. If not. The Newton iteration rule is β j+1 = β j − δL−1 Lβ ββ where δ is a scalar stepsize.6) is Lβ = ∂ ∂ Ln (β) = Rn (β. λ) = Rβ + λ0 Rλ = Rβ β ∂β ∂β since Rλ (β.for all i. The Hessian for (10. the eigenvalues of Lββ should be adjusted so that all are positive. If (10. ∂β 0 the second equality following from the implicit function theorem applied to Rλ (β. Outer Loop The outer loop is the minimization (10. This can be done by the modiﬁed Newton method described in the previous section.6) is Lββ = − ∂ Ln (β) ∂β∂β 0 ¤ ∂ £ = − 0 Rβ (β.

what happens? It is helpful to solve for qi and pi in terms of the errors. ¸µ ¶ µ ¶ ∙ qi e1i 1 β1 = 1 −β 2 pi e2i 124 Assume that ei = . and x∗ is not observed. Eei = 0. Instead i i i we observe xi = x∗ + ui where ui is an k × 1 measurement error. This cannot happen if β is deﬁned by linear projection. i e2i The question is. if we regress qi on pi . so requires a structural interpretation. Suppose that (yi . then i ¡ ¡ ¢¢−1 ¡ 0 ¢ ˆ E ui ui β 6= β. E(yi | x∗ ) = x∗0 β is linear. x∗ ) are joint random i variables. Example: Measurement error in the regressor. β is the parameter of interest. independent of yi and x∗ . The variables qi and pi (quantity and price) are determined jointly by the demand equation qi = −β 1 pi + e1i µ ¶ qi = β 2 pi + e2i . and the supply equation e1i is iid. β 1 + β 2 = 1 and Eei e0 = I2 (the latter for simplicity). β →p β ∗ = β − E xi x0 i This is called measurement error bias. It follows that if β is the OLS estimator. Example: Supply and Demand. Then i i yi = x∗0 β + ei i = (xi − ui )0 β + ei = x0 β + vi i where vi = ei − u0 β. i The problem is that £ ¡ ¢¤ ¡ ¢ E (xi vi ) = E (x∗ + ui ) ei − u0 β = −E ui u0 β 6= 0 i i i ˆ if β 6= 0 and E (ui u0 ) 6= 0.Chapter 11 Endogeneity 0 We say that there is endogeneity in the linear model y = zi β + ei if β is the parameter of interest and E(zi ei ) 6= 0. The coeﬃcient β must have meaning separately from the deﬁnition of a conditional mean or linear projection. In matrix notation.

which does not equal either β 1 or β 2 .1) if E (xi ei ) = 0. Deﬁnition 11. That is x2i are variables which could be included in the equation for yi (in the sense 125 .2) Any solution to the problem of endogeneity requires additional information which we call instruments. at least the intercept). and x2i are the excluded exogenous variables.1) where zi is k × 1. In a typical set-up. so should be included in xi . and assume that E(zi ei ) 6= 0 so there is endogeneity. We call z1i exogenous and z2i endogenous.1 Instrumental Variables 0 yi = zi β + ei Let the equation of interest be (11. So we have the partition µ ¶ z1i k1 (11. Thus we make the partition ¶ µ k1 z1i (11. ¶ µ ∙ β2 β1 1 −1 β 2 e1i + β 1 e2i (e1i − e2i ) ¸µ e1i e2i ¶ ˆ Hence if it is estimated by OLS. z1i is an instrumental variable for (11.1). 11.1) the structural equation.4) xi = x2i 2 where z1i = x1i are the included exogenous variables. β →p β ∗ .3) zi = z2i k2 where E(z1i ei ) = 0 yet E(z2i ei ) 6= 0. In matrix notation. this can be written as Y = Zβ + e. This is called simultaneous equations bias. By the above deﬁnition.so µ qi pi ¶ = = = The projection of qi on pi yields qi = β ∗ pi + εi E (pi εi ) = 0 where β∗ = β − β1 E (pi qi ) ¡ 2¢ = 2 2 E pi ∙ 1 β1 1 −β 2 ¸−1 µ e1i e2i ¶ .1. some regressors in zi will be uncorrelated with ei (for example.1 The × 1 random vector xi is an instrumental variable for (11. (11. We call (11.

that they are uncorrelated with ei ) yet can be excluded, as they would have true zero coeﬃcients in the equation. The model is just-identiﬁed if = k (i.e., if 2 = k2 ) and over-identiﬁed 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.

11.2

Reduced Form

The reduced form relationship between the variables or “regressors” zi and the instruments xi is found by linear projection. Let ¡ ¢−1 ¡ 0 ¢ E xi zi Γ = E xi x0 i be the × k matrix of coeﬃcients from a projection of zi on xi , and deﬁne ui = zi − x0 Γ i as the projection error. Then the reduced form linear relationship between zi and xi is zi = Γ0 xi + ui . In matrix notation, we can write (11.5) as Z = XΓ + u where u is n × k. By construction, E(xi u0 ) = 0, i so (11.5) is a projection and can be estimated by OLS: ˆ ˆ Z = XΓ + u ¡ 0 ¢−1 ¡ 0 ¢ ˆ XZ . Γ = XX Y = (XΓ + u) β + e = Xλ + v, where λ = Γβ and v = uβ + e. Observe that ¡ ¢ E (xi vi ) = E xi u0 β + E (xi ei ) = 0. i ˆ ˆ Y = X λ + v, ¡ 0 ¢−1 ¡ 0 ¢ ˆ λ = XX XY 126 (11.8) (11.7) (11.6) (11.5)

Substituting (11.6) into (11.2), we ﬁnd

Thus (11.7) is a projection equation and may be estimated by OLS. This is

The equation (11.7) is the reduced form for Y. (11.6) and (11.7) together are the reduced form equations for the system Y = Xλ + v Z = XΓ + u. ˆ ˆ As we showed above, OLS yields the reduced-form estimates (λ, Γ)

11.3

Identiﬁcation

The structural parameter β relates to (λ, Γ) through (11.8). The parameter β is identiﬁed, meaning that it can be recovered from the reduced form, if rank(Γ) = k. (11.9) Assume that (11.9) holds. If = k, then β = Γ−1 λ. If > k, then for any W > 0, β = (Γ0 W Γ)−1 Γ0 W λ. If (11.9) is not satisﬁed, then β cannot be recovered from (λ, Γ). Note that a necessary (although not suﬃcient) condition for (11.9) is ≥ k. Since X and Z have the common variables X1 , we can rewrite some of the expressions. Using (11.3) and (11.4) to make the matrix partitions X = [X1 , X2 ] and Z = [X1 , Z2 ], we can partition Γ as ∙ ¸ Γ11 Γ12 Γ = Γ21 Γ22 ∙ ¸ I Γ12 = 0 Γ22 (11.6) can be rewritten as Z1 = X1 Z2 = X1 Γ12 + X2 Γ22 + u2 . (11.10) β is identiﬁed if rank(Γ) = k, which is true if and only if rank(Γ22 ) = k2 (by the upperdiagonal structure of Γ). Thus the key to identiﬁcation of the model rests on the 2 × k2 matrix Γ22 in (11.10).

11.4

Estimation

0 yi = zi β + ei

The model can be written as E (xi ei ) = 0 or Eg (wi , β) = 0 ¢ ¡ 0 g (wi , β) = xi yi − zi β .

This 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 Wn , is ¡ ¢−1 0 ˆ β = Z 0 XWn X 0 Z Z XWn X 0 Y. 127

11.5

Special Cases: IV and 2SLS

This estimator is often called the instrumental variables estimator (IV) of β, where X is used as an instrument for Z. Observe that the weight matrix Wn has disappeared. In the just-identiﬁed case, the weight matrix places no role. This is also the MME estimator of β, and the EL estimator. Another interpretation stems from the fact that since β = Γ−1 λ, we can construct the Indirect Least Squares (ILS) estimator: ˆ ˆ ˆ β = Γ−1 λ ³¡ ¢−1 ¡ 0 ¢´−1 ³¡ 0 ¢−1 ¡ 0 ¢´ XX XZ XY = X 0X ¡ 0 ¢−1 ¡ 0 ¢ ¡ 0 ¢−1 ¡ 0 ¢ = XZ XX XX XY ¡ 0 ¢−1 ¡ 0 ¢ = XZ XY .

If the model is just-identiﬁed, so that k = , then the formula for GMM simpliﬁes. We ﬁnd that ¡ ¢−1 0 ˆ β = Z 0 XWn X 0 Z Z XWn X 0 Y ¡ ¢−1 −1 ¡ 0 ¢−1 0 = X 0Z Wn Z X Z XWn X 0 Y ¡ ¢−1 0 = X 0Z XY

which again is the IV estimator.

¡ ¢ Recall that the optimal weight matrix is an estimate of the inverse of Ω = E xi x0 e2 . In the i i ¢ ¡ 2 special case that E ei | xi = σ 2 (homoskedasticity), then Ω = E (xi x0 ) σ 2 ∝ E (xi x0 ) suggesting i i the weight matrix Wn = (X 0 X)−1 . Using this choice, the GMM estimator equals ³ ¡ ¢−1 0 ´−1 0 ¡ 0 ¢−1 0 ˆ β 2SLS = Z 0 X X 0 X XZ ZX XX XY

then

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 eﬃcient GMM, but otherwise it is ineﬃcient. It is useful to observe that writing ¡ ¢−1 0 X, PX = X X 0 X ¡ 0 ¢−1 0 ˆ X Z, Z = PX Z = X X X ˆ β = ¢−1 0 Z PX Y Z 0 PX Z ´ ³ ˆ ˆ ˆ = Z 0 Z Z 0 Y. ¡

The source of the “two-stage” name is since it can be computed as follows ˆ ˆ ˆ • First regress Z on X, vis., Γ = (X 0 X)−1 (X 0 Z) and Z = X Γ = PX Z. ³ ´−1 ˆ ˆ ˆ ˆ ˆ • Second, regress Y on Z, vis., β = Z 0 Z Z 0 Y. 128

¶ µ 1 0 Z e = E (zi ei ) = Γ0 E (xi ei ) + E (ui ei ) = S21 E n and E µ 1 0 ZZ n ¶ ¡ 0¢ = E zi zi ¢ ¡ ¢ ¢ ¡ ¢ ¡ ¡ = Γ0 E xi x0 Γ + E ui x0 Γ + Γ0 E xi u0 + E ui u0 i i i i = Γ0 QΓ + S22 where Q = E (xi x0 ) . Z2 = [PX Z1 . First. Using (11.11). In our notation. Here we present a simpliﬁed version of one of his results.13) which is zero only when S21 = 0 (when Z is exogenous). Recall. X2 ].11) (11.12) Z = XΓ + u ξ = (e. PX Z2 ] = [Z1 . X is n × l so there are l instruments. We now derive a similar result for the 2SLS estimator. ¡ ¢−1 ¡ 0 ¢ ˆ β 2SLS = Z 0 PX Z Z PX Y . the model is Y = Zβ + e (11.12). ˆ since Z1 lies in the span of X. we regress Y on Z1 and Z2 . Thus in the second stage. So only the endogenous variables Z2 are replaced by their ﬁtted values: ˆ ˆ ˆ Z2 = X1 Γ12 + X2 Γ22 .ˆ It is useful to scrutinize the projection Z. Then i h ˆ ˆ ˆ Z = Z1 . Z2 ] and X = [Z1 . 11. let’s analyze the approximate bias of OLS applied to (11. PX Z2 ] h i ˆ = Z1 . 129 .6 Bekker Asymptotics Bekker (1994) used an alternative asymptotic framework to analyze the ﬁnite-sample bias in the 2SLS estimator. Hence by a ﬁrst-order approximation i µ µ ¶¶−1 µ ¶ ³ ´ 1 0 1 0 ˆ OLS − β E β ZZ Ze ≈ E E n n ¡ ¢−1 = Γ0 QΓ + S22 S21 (11. u) E (ξ | X) = 0 ¢ ¡ E ξξ 0 | X = S As before. Z = [Z1 . Z2 .

The consequences of identiﬁcation failure for inference are quite severe.Let PX = X (X 0 X)−1 X 0 . except when S21 = 0 (when Z is exogenous). n n n ¢ ¡ ¡ ¢ ¢ ¢ 1 ¡ 0 1 ¡ E Z PX Z = Γ0 E xi x0 Γ + Γ0 E (xi ui ) + E ui x0 Γ + E u0 PX u i i n n = Γ0 QΓ + αS22 .14) In general this is non-zero. l . The parameter β fails to be identiﬁed if Γ22 has deﬁcient rank.12) and this result. 130 .14) approaches that in (11. Bekker (1994) showed further that under the alternative asymptotic approximation that α is ﬁxed as n → ∞ (so that the number of instruments goes to inﬁnity proportionately with sample ˆ size) then the expression in (11. By the spectral decomposition of an idempotent matrix. the expression in (11. Together µ µ ¶¶−1 µ ¶ ³ ´ 1 0 1 0 ˆ 2SLS − β E β Z PX Z Z PX e ≈ E E n n ¡ ¢−1 = α Γ0 QΓ + αS22 S21 . Let q = H 0 ξS −1/2 which satisﬁes Eqq 0 = In and partition q = (q1 q2 ) where q1 is l × 1. It is also close to zero when α = 0. Hence µ ¶ 1 0 1 1/20 ¡ 0 ¢ 1/2 ξ PX ξ S E q Λq S E = n n µ ¶ 1 0 1 1/20 S E q q1 S 1/2 = n n 1 l 1/20 1/2 S S = n = αS where α= Using (11. n and (11.13).14) is the probability limit of β 2SLS − β 11. 0). Indeed as α → 1. ¢ 1 ¡ ¢ 1 ¡ ¢ 1 ¡ 0 E Z PX e = E Γ0 X 0 e + E u0 PX e = αS21 .7 Identiﬁcation Failure Recall the reduced form equation Z2 = X1 Γ12 + X2 Γ22 + u2 . the bias in the 2SLS estimator will be large. Bekker (1994) pointed out that it also has the reverse implication — that when α = l/n is large. indicating that the bias in 2SLS approaches that of OLS as the number of instruments increases. P = HΛH 0 0 0 where Λ = diag(Il .

15) is unaﬀected.16) ˆ β−β = 1 √ n 1 √ n Pn Pn i=1 xi ei i=1 xi zi →d N1 ∼ Cauchy. once again take the simple case k = l = 1. Qc + N2 ˆ As in the case of complete identiﬁcation failure.15) therefore ¢¢ ¡ ¡ 1 X 1 X √ xi zi = √ xi ui →d N2 ∼ N 0. E x2 e2 i i n i=1 n (11.16) instead takes the form 1 X √ xi zi n i=1 n = = 1 X 2 1 X √ xi γ + √ xi ui n n 1 n i=1 n X x2 c + i i=1 n n 1 √ n →d Qc + N2 therefore ˆ β − β →d i=1 n X i=1 xi ui N1 . where C is a full rank matrix. N2 since the ratio of two normals is Cauchy. The n−1/2 is picked because it provides just the right balancing to allow a rich distribution theory. Then by the CLT ¡ ¡ ¢¢ 1 X √ xi ei →d N1 ∼ N 0. To see the consequences. We see that β is identiﬁed if and only if Γ22 = γ 6= 0. but is weak. meaning that inferences about parameters of interest can be misleading. This occurs when Γ22 is full rank. This result carries over to more general settings. the instrument xi is weak for zi if γ = n−1/2 c. as the Cauchy distribution does not have a ﬁnite mean. we ﬁnd that β is inconsistent for β and the ˆ is non-normal. In addition. Here.Take the simplest case where k = l = 1 (so there is no X1 ). 131 . and was examined by Phillips (1989) and Choi and Phillips (1992). E x2 u2 i i n n i=1 i=1 n n (11. so E (zi xi ) = 0. standard test statistics have non-standard asymptotic distribution of β distributions. but (11. viz Γ22 = n−1/2 C. Then the model may be written as yi = zi β + ei zi = xi γ + ui and Γ22 = γ = E (xi zi ) /Ex2 . Thus identiﬁcation hinges on the existence of correlation between the excluded exogenous variable and the included endogenous variable. but small. This can be handled in an asymptotic analysis by modeling it as local-to-zero. Suppose that identiﬁcation does not complete fail. which occurs i when E (zi xi ) 6= 0. Suppose this condition fails. This is particularly nasty. Then (11.

132 . one constructive recommendation is to explicitly estimate the reduced form equation for Z2 . and attempt to assess the likelihood that Γ22 has deﬁcient rank. tests of deﬁcient rank are diﬃcult to implement. It is not even suﬃcient that each column of Γ22 is non-zero (each column corresponds to a distinct endogenous variable in X2 ). construct the test of Γ22 = 0. Γ22 6= 0 is not suﬃcient for identiﬁcation. Unfortunately. The bottom line is that it is highly desirable to avoid identiﬁcation failure. Therefore in the case of k2 = 1 (one RHS endogenous variable). If k2 = 1. and at a minimum check that the test rejects H0 : Γ22 = 0. the equation to focus on is the reduced form Z2 = X1 Γ12 + X2 Γ22 + u2 and identiﬁcation requires rank(Γ22 ) = k2 . this requires Γ22 6= 0. this cannot be interpreted as deﬁnitive proof that Γ22 has full rank. So while a minimal check is to test that each columns of Γ22 is non-zero. When k2 > 1. which is straightforward to assess using a hypothesis test on the reduced form. Once again. In any event.The distribution theory for this model was developed by Staiger and Stock (1997) and extended to nonlinear GMM estimation by Stock and Wright (2000). it appears reasonable to explicitly estimate and report the reduced form equations for X2 . Further results on testing were obtained by Wang and Zivot (1998).

and Γ is l × k. 5. Show that the GLS estimator of β can be written as an i IV estimator using some instrument xi . Let β denote the IV estimator of β using as an instrument a dummy variable di (takes only the values 0 and 1). l ≥ k.) 3. at ˆˆ If X is indeed endogeneous. 133 . ˆ ˆ Let the OLS estimator for β be β and the OLS residual be e = Y − Z β. 6. xi is l × 1. Z is n × k. u is n × k. show that if rank(Γ) < k then β cannot be identiﬁed. The parameter Γ is deﬁned by the population moment condition ¡ ¢ E xi u0 = 0 i ˆ Show that the method of moments estimator for Γ is Γ = (X 0 X)−1 (X 0 Z) . Consider the single equation model ˆ where yi and zi are both real-valued (1 × 1). Explain why this is true.11. ˜ 0 e < e0 e. ˆ ˜ ˜ Let the IV estimator for β using some instrument X be β and the IV residual be e = Y − Z β.8 Exercises yi = zi β + ei . in the sense that e ˜ ˜ least in large samples? 4. 1. That is. Take the linear model Y = Zβ + e. we claim that β is identiﬁed (can be recovered from the reduced form) if rank(Γ) = k. Take the linear model yi = xi β + ei E (ei | xi ) = 0. where xi and β are 1 × 1. 2. The reduced form between the regressors zi and instruments xi takes the form zi = x0 Γ + ui i or Z = XΓ + u where zi is k × 1. In the structural model Y = Zβ + e Z = XΓ + u with Γ l × k. In the linear model 0 yi = zi β + ei E (ei | zi ) = 0 ¡ 2 ¢ suppose σ 2 = E ei | zi is known. will IV “ﬁt” better than OLS. (Find an expression for xi . X is n × l. Find a simple expression for the IV estimator in this context.

Report the estimates and standard errors. of the mother).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). In this model. The data ﬁle card. There are 2215 observations with 19 variables. I suggest that you use the 2SLS estimates as the ﬁrst-step to get the weight matrix. and then calculate the GMM estimator from this weight matrix without further iteration. Suppose that price and quantity are determined by the intersection of the linear demand and supply curves Demand : Supply : Q = a0 + a1 P + a2 Y + e1 Q = b0 + b1 P + b2 W + e2 ¡ ¢ (a) Show that E (xi ei ) = 0 and E x2 ei = 0. Report estimates and standard errors. Report estimates and standard errors. using the instrument near4.(b) Deﬁne the 2SLS estimator of β. in years. Does this diﬀer from 2SLS and/or OLS? 7. a dummy indicating that the observation lives near a 4-year college. 134 .pdf. (d) Re-estimate the model by eﬃcient GMM. (a) Estimate the model by OLS. f atheduc (the education. and the remaining variables as exogenous.. How does this diﬀer from OLS? (c) Find the eﬃcient GMM estimator of β based on the moment condition E (zi (yi − xi β)) = 0. are the parameters identiﬁed? 8. We want to estimate a wage equation log(W age) = β 0 + β 1 Educ + β 2 Exper + β 3 Exper2 + β 4 South + β 5 Black + e where Educ = Eduation (Years) Exper = Experience (Years). in years. (e) Calculate and report the J statistic for overidentiﬁcation. Is zi = (xi i for estimation of β? x2 ) a valid instrumental variable i where income (Y ) and wage (W ) are determined outside the market. (b) Now treat Education as endogenous. Estimate the model by 2SLS. listed in card. (f) Discuss your ﬁndings. using zi as an instrument for xi . (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). of the father) and motheduc (the education. and South and Black are regional and racial dummy variables.

Yt−k ) is independent of t for all k.. The following two theorems are essential to the analysis of stationary time series.3 ρ(k) = γ(k)/γ(0) = Corr(Yt . Yt−k ) = γ(k) is independent of t for all k.1 {Yt } is covariance (weakly) stationary if E(Yt ) = µ is independent of t.. . Theorem 12. A stationary time series is ergodic if γ(k) → 0 as k → ∞.Chapter 12 Univariate Time Series A time series yt is a process observed in sequence over time. There proofs are rather diﬃcult. Deﬁnition 12. Yt−1 . and T to denote the number of observations. To indicate the dependence on time. . so classical assumptions are not valid.1 Stationarity and Ergodicity Deﬁnition 12. and multivariate (yt ∈ Rm is vector-valued). 135 . we adopt new notation.1.. and Cov (Yt . however.2 {Yt } is strictly stationary if the joint distribution of (Yt ..4 (loose).) is a random variable..1. Deﬁnition 12. γ(k) is called the autocovariance function. .. Deﬁnition 12.1 If Yt is strictly stationary and ergodic and Xt = f (Yt . The primary model for univariate time series is autoregressions (ARs). we expect that Yt and Yt−1 are not independent. 12.. Because of the sequential nature of time series. We can separate time series into two categories: univariate (yt ∈ R is scalar).1. Yt−k ) is the autocorrelation function. The primary model for multivariate time series is vector autoregressions (VARs). and use the subscript t to denote the individual observation. then Xt is strictly stationary and ergodic..1. t = 1. T .1.

then as T → ∞. If Xt is strictly stationary and ergodic and E |Xt | < ∞. ˆ 2. then as T → ∞. Part (1) is a direct consequence of the Ergodic theorem. ˆ ˆ T t=1 T t=1 T t=1 By Theorem 12. µ →p E(Yt ). For Part (2). T 1X Xt →p E(Xt ). γ (k) →p γ(k). Thus an application of the Ergodic Theorem yields T 1X Yt Yt−k →p E(Yt Yt−k ).1. ˆ ˆ T t=1 γ (k) = ˆ The sample autocorrelation ρ(k) = ˆ γ (k) ˆ .2 (Ergodic Theorem).1 above. T t=1 Thus γ (k) →p E(Yt Yt−k ) − µ2 − µ2 + µ2 = E(Yt Yt−k ) − µ2 = γ(k). and it has a ﬁnite mean by the assumption that EYt2 < ∞. note that γ (k) = ˆ T 1X (Yt − µ) (Yt−k − µ) ˆ ˆ T t=1 = T T T 1X 1X 1X Yt Yt−k − Yt µ − ˆ Yt−k µ + µ2 . the sequence Yt Yt−k is strictly stationary and ergodic.Theorem 12. ˆ 3. ˆ ˆ γ ¥ 136 . T t=1 This allows us to consistently estimate parameters using time-series moments: The sample mean: T 1X µ= ˆ Yt T t=1 The sample autocovariance T 1X (Yt − µ) (Yt−k − µ) .1.1. ρ(k) →p ρ(k). ˆ Proof.3 If Yt is strictly stationary and ergodic and EYt2 < ∞. γ (0) ˆ Theorem 12. 1. ˆ Part (3) follows by the continuous mapping theorem: ρ(k) = γ (k)/ˆ (0) →p γ(k)/γ(0) = ρ(k).

. should be a MDS. . E(Yt−k et ) = 0. 12. Letting et = Yt − E (Yt | It−1 ) .. then we have the autoregressive model Yt = α + ρ1 Yt−1 + ρ2 Yt−1 + · · · + ρk Yt−k + et E (et | It−1 ) = 0.1 et is a martingale diﬀerence sequence (MDS) if E (et | It−1 ) = 0.} is the past history of the series... k=0 Loosely speaking.. Y1 . some versions of the life-cycle hypothesis imply that either changes in consumption...2. Regression errors are naturally a MDS.. it is even more important in the time-series context.} are jointly random. A linear AR model (the most common type used in practice) speciﬁes linearity: E (Yt | It−1 ) = α + ρ1 Yt−1 + ρ2 Yt−1 + · · · + ρk Yt−k .3 Stationarity of AR(1) Process Yt = ρYt−1 + et . . Most asset pricing models imply that asset returns should be the sum of a constant plus a MDS.2 Autoregressions In time-series. t By back-substitution. Some time-series processes may be a MDS as a consequence of optimizing behavior. In fact. this series converges if the sequence ρk et−k gets small as k → ∞. Thus for k > 0. The last property deﬁnes a special time-series process. as it is diﬃcult to derive distribution theories without this property. 137 . ∞ X = ρk et−k . We consider the conditional expectation E (Yt | It−1 ) where It−1 = {Yt−1 . E(et ) = 0 and Ee2 = σ 2 < ∞. YT .. Y2 .... This occurs when |ρ| < 1.. we ﬁnd Yt = et + ρet−1 + ρ2 et−2 + . . For example. A mean-zero AR(1) is Assume that et is iid.. the series {..12. Yt−k ) . An autoregressive (AR) model speciﬁes that only a ﬁnite number of past lags matter: E (Yt | It−1 ) = E (Yt | Yt−1 . . Yt−2 . Deﬁnition 12. 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. A useful property of a MDS is that et is uncorrelated with any function of the lagged information It−1 . or consumption growth rates.

3. Yt − ρ1 LYt − ρ2 L2 Yt − · · · − ρk Lk Yt = et . 138 . The AR(k) model is Using the lag operator.4 Lag Operator An algebraic construct which is useful for the analysis of autoregressive models is the lag operator. In general. ∞ X k=0 ρ2k V ar (et−k ) = 12. We call ρ(L) the autoregressive polynomial of Yt . Deﬁning L2 = LL. we see that L2 Yt = LYt−1 = Yt−2 . 1 − ρ2 V ar(Yt ) = If the equation for Yt has an intercept. Lk Yt = Yt−k .1 The lag operator L satisﬁes LYt = Yt−1 .3.1.1 If |ρ| < 1 then Yt is strictly stationary and ergodic. We say that the root of the polynomial is 1/ρ. except that the mean of Yt can be computed from the relationship EYt = α + ρEYt−1 . We call ρ(L) the autoregressive polynomial of Yt . The AR(1) model can be written in the format Yt − ρYt−1 + et or (1 − ρL) Yt−1 = et .5 Stationarity of AR(k) Yt = ρ1 Yt−1 + ρ2 Yt−1 + · · · + ρk Yt−k + et . or ρ(L)Yt = et where ρ(L) = 1 − ρ1 L − ρ2 L2 − · · · − ρk Lk .4. and solving for EYt = EYt−1 we ﬁnd EYt = α/(1 − ρ). We can compute the moments of Yt using the inﬁnite sum: EYt = ∞ X k=0 ρk E (et−k ) = 0 σ2 . the above results are unchanged.Theorem 12. 12. an AR(1) is stationary iﬀ |ρ| < 1. Note that an equivalent way to say this is that an AR(1) is stationary iﬀ the root of the autoregressive polynomial is larger than one (in absolute value). Deﬁnition 12. The operator ρ(L) = (1 − ρL) is a polynomial in the operator L. since ρ(z) = 0 when z = 1/ρ. From Theorem 12.

the requirement is that all roots are larger than one. it is helpful to deﬁne the process ut = xt et . This is a special case of non-stationarity. t t T t=1 ¥ Combined with (12.5. it is also strictly stationary and ergodic. and by Theorem 12. we see that Ã !−1 Ã ! T T 1X 1X ˆ β=β+ xt x0 xt et →p Q−1 0 = 0. β = X 0X (12. The Fundamental Theorem of Algebra says that any polynomial can be factored as ¢¡ ¢ ¡ ¢ ¡ ρ(z) = 1 − λ−1 z 1 − λ−1 z · · · 1 − λ−1 z 1 2 k 12. “has a unit root”.” If one of the roots equals 1. and is of great interest in applied time series. since E (ut | It−1 ) = E (xt et | It−1 ) = xt E (et | It−1 ) = 0. T t=1 T t=1 ¢−1 0 ¡ ˆ X Y. so is xt x0 .1 The AR(k) is strictly stationary and ergodic if and only if |λj | > 1 for all j..6. Thus T T 1X 1X xt et = ut →p E (ut ) = 0. Note that ut is a MDS. Theorem 12.1. .1 If the AR(k) process Yt is strictly stationary and ergodic and EYt2 < ∞.1. We know that an AR(1) is stationary iﬀ the absolute value of the root of its autoregressive polynomial is larger than one. t T t=1 T t=1 139 . then ˆ β →p β as T → ∞.. Thus t by the Ergodic Theorem. and hence Yt . Proof. The vector xt is strictly stationary and ergodic. For an AR(k). By Theorem 12.1.where the λ1 .6 Let Estimation xt = β = ¡ 1 Yt−1 Yt−2 · · · α ρ1 ρ2 · · · yt = x0 β + et . t Yt−k ¢0 ρk . ¢0 Then the model can be written as The OLS estimator is ¡ ˆ To study β.1) Theorem 12. which satisfy ρ(λj ) = 0. we say that ρ(L).1) and the continuous mapping theorem. T ¢ ¡ 1X xt x0 →p E xt x0 = Q. λk are the complex roots of ρ(z).. One way of stating this is that “All roots lie outside the unit circle.1. Let |λ| denote the modulus of a complex number λ.

Y0 ). This creates an iid bootstrap sample.8 Bootstrap for Autoregressions In the non-parametric bootstrap. This is identical in form to the asymptotic distribution of OLS in cross-section regression. then as T → ∞. t t Theorem 12. which may be diﬀerent than the i properties of the actual ei . The implication is that asymptotic inference is the same.2 If the AR(k) process Yt is strictly stationary and ergodic and EYt4 < ∞. i 4.7. we constructed the bootstrap sample by randomly resampling from the data values {yt . Ω) . ´ √ ³ ¡ ¢ ˆ T β − β →d N 0. as this imposes inappropriate independence. Y−k+2 .. the asymptotic covariance matrix is estimated just as in the cross-section case. Q−1 ΩQ−1 . Divide the sample into T /m blocks of length m. e ˆ 3. Estimate β and residuals et . Method 1: Model-Based (Parametric) Bootstrap. t then as T → ∞. ˆ 1.7 Asymptotic Distribution Theorem 12. Simulate iid draws e∗ from the empirical distribution of the residuals {ˆ1 . xt }. T t=1 where Ω = E(xt x0 e2 ). Fix an initial condition (Y−k+1 . T 1 X √ ut →d N (0.12. this cannot work in a time-series application. T t=1 Since xt et is a MDS... ˆ 2. It also presumes that the AR(k) structure is the truth.. .7. . If ut is a strictly stationary and ergodic MDS and E(ut u0 ) = Ω < ∞.1 to see that T 1 X √ xt et →d N (0.. Method 2: Block Resampling 1. In particular. Clearly. there are two popular methods to implement bootstrap resampling for time-series data.1 MDS CLT. 140 . eT }.7.. we can apply Theorem 12. Brieﬂy. 12. Ω) . t This construction imposes homoskedasticity on the errors e∗ . Create the bootstrap series Yt∗ by the recursive formula ˆ ˆ ∗ ˆ ∗ ˆ ∗ Yt∗ = α + ρ1 Yt−1 + ρ2 Yt−2 + · · · + ρk Yt−k + e∗ .

The results may be sensitive to the block length.4) by OLS.3) sequentially by OLS. • You can estimate (12. (12. ﬁrst estimate (12. This allows for arbitrary stationary serial correlation. This is a ﬂexible approach which can extract a wide range of seasonal factors.. however. ∆s Yt = Yt − Yt−s . Resample complete blocks. • You can estimate (12. • Include dummy variables for each season. ρk ).. and then perform regression (12. But the long-run trend is also eliminated.3) or Yt = α0 + α1 t + ρ1 Yt−1 + ρ2 Yt−1 + · · · + ρk Yt−k + et . If s is the number of seasons (typically s = 4 or s = 12). • Use “seasonally adjusted” data. 5. . or the season-to-season change.2)-(12. Seasonal Eﬀects There are three popular methods to deal with seasonal data. The seasonal factor is typically estimated by a two-sided weighted average of the data for that season in neighboring years.2) (12. Thus the seasonally adjusted data is a “ﬁltered” series. May not work well in small samples. The seasonal adjustment. 3. draw T /m blocks.2. 141 . and for modelmisspeciﬁcation. The series ∆s Yt is clearly free of seasonality. That is.3) replacing St with St . 4. also alters the time-series correlations of the data.2). (12. and the way that the data are partitioned into blocks. This presumes that “seasonality” does not change over the sample. • First apply a seasonal diﬀerencing operator. For each simulated sample.. Paste the blocks together to create the bootstrap time-series Yt∗ .9 Trend Stationarity Yt = µ0 + µ1 t + St St = ρ1 St−1 + ρ2 St−2 + · · · + ρk St−l + et .4) There are two essentially equivalent ways to estimate the autoregressive parameters (ρ1 . 12. The reason why these two procedures are (essentially) the same is the Frisch-Waugh-Lovell theorem. heteroskedasticity. get the ˆ ˆ residual St . and perhaps this was of relevance. This procedure is sometimes called Detrending. 6.

We model this as an AR(1): ut = θut−1 + et with et a MDS. The hypothesis of no omitted serial correlation is H0 : θ = 0 H1 : θ 6= 0.5) and (12. and under H1 it is an AR(2). One approach to model selection is to choose k based on a Wald tests... We want to test H0 against H1 . 142 . AIC(k) = log σ 2 (k) + ˆ 2k . you need more initial conditions. An appropriate test is the Wald test of this restriction. (12. or Yt = α(1 − θ) + (ρ + θ) Yt−1 − θρYt−2 + et = AR(2). T where σ 2 (k) is the estimated residual variance from an AR(k) ˆ One ambiguity in deﬁning the AIC criterion is that the sample available for estimation changes as k changes.6) 12. we take (12.. (A simple exclusion test).. to ﬁnd Yt − θYt−1 = α − θα + ρYt−1 − θρYt−1 + ut − θut−1 .5) and lag the equation once: Yt−1 = α + ρYt−2 + ut−1 . and this is equivalent to the restriction that the coeﬃcients on Yt−k−1 .10 Testing for Omitted Serial Correlation For simplicity. (If you increase k.g.6). Another is to minimize the AIC or BIC information criterion. The best remedy is to ﬁx a upper value k. Thus under H0 . Yt−k−m are jointly zero. and then reserve the ﬁrst k as initial conditions. e.5) We are interested in the question if the error ut is serially correlated. Yt is an AR(1). AR(k) on this (uniﬁed) sample. AR(2). this is the same as saying that under the alternative Yt is an AR(k+m). H0 may be expressed as the restriction that the coeﬃcient on Yt−2 is zero. . We then multiply this by θ and subtract from (12. An appropriate test of H0 against H1 is therefore a Wald test that the coeﬃcient on Yt−2 is zero..12.11 Model Selection What is the appropriate choice of k in practice? This is a problem of model selection.) This can induce strange behavior in the AIC. (12. if the null hypothesis is that Yt is an AR(k).. and the alternative is that the error is an AR(m).5). . and then estimate the models AR(1). In general. To combine (12. let the null hypothesis be an AR(1): Yt = α + ρYt−1 + ut .

However. as the models are equivalent. Yt has a unit root when ρ(1) = 0. (12. Thus if Yt has a (single) unit root. we say that if Yt is non-stationary but ∆d Yt is stationary. An alternative asymptotic framework has been developed to deal with non-stationary data. In this case. the natural test statistic is the t-statistic for H0 from OLS estimation of (12. We do not have the time to develop this theory in detail.7) is (1 − L) − α0 L − α1 (L − L2 ) − · · · − αk−1 (Lk−1 − Lk ) But this must equal ρ(L). Thus a time series with unit root is I(1). Indeed. then ∆Yt is a stationary AR process.12 Autoregressive Unit Roots The AR(k) model is ρ(L)Yt = µ + et ρ(L) = 1 − ρ1 L − · · · − ρk Lk . and test statistics are asymptotically non-normal. Hence. this is the most popular unit root test. The DF parameterization is convenient because the parameter α0 summarizes the information about the unit root. or ρ1 + ρ2 + · · · + ρk = 1. It would seem natural to assess the signiﬁcance of the ADF statistic using the normal table. The ergodic theorem and MDS CLT do not apply. Because of this property. Yt is non-stationary. then Yt is “integrated of order d”. Under H0 . since ρ(1) = −α0 . the model for Yt is ∆Yt = µ + α1 ∆Yt−1 + · · · + αk−1 ∆Yt−(k−1) + et . Yt is non-stationary. Note that the model is stationary if α0 < 0. Thus ρ(1) = (1 − 1) − α0 − (1 − 1) − · · · − (1 − 1) = −α0 .7) These models are equivalent linear transformations of one another. under H0 .7).12. A helpful way to write the equation is the so-called Dickey-Fuller reparameterization: ∆Yt = µ + α0 Yt−1 + α1 ∆Yt−1 + · · · + αk−1 ∆Yt−(k−1) + et . To see this. which is an AR(k-1) in the ﬁrst-diﬀerence ∆Yt . or I(d). As we discussed before. but simply assert the main results. So the natural alternative is H1 : α0 < 0. so conventional normal asymptotics are invalid. and is called the Augmented Dickey-Fuller (ADF) test for a unit root. 143 . Since α0 is the parameter of a linear regression. observe that the lag polynomial for the Yt computed from (12. the hypothesis of a unit root in Yt can be stated as H0 : α0 = 0.

Most texts include as well the critical values for the extreme polar case where the intercept has been omitted from the model. If the series has a linear trend (e. The natural solution is to include a time trend in the ﬁtted OLS equation. and may be used for testing the hypothesis H0 . Stock Prices). however. then the series itself is nonstationary. In this context. it is a silly waste of time to ﬁt an AR model to the level of the series without a time trend. but it may be stationary around the linear time trend. If a time trend is included.g. This is not really a correct conclusion. As T → ∞. Another popular setting includes as well a linear time trend. but does not depend on the parameters α. These are included for completeness (from a pedagogical perspective) but have no relevance for empirical practice where intercepts are always included. conventional rate of T The second result states that the t-statistic for α0 converges to a limit distribution which is ˆ non-normal. The ﬁrst result states that α0 converges to its true value (of zero) at rate T. the ADF test rejects H0 in favor of H1 when ADF < c.1 (Dickey-Fuller Theorem). GDP. When conducting the ADF test. this means that it is computed as the t-ratio for α0 from OLS estimation of (12. All we can say is that there is insuﬃcient evidence to conclude whether the data are stationary or not. We have described the test for the setting of with an intercept. Note: The standard error s(ˆ 0 ) is the conventional (“homoskedastic”) standard error. Since the alternative hypothesis is one-sided. the test procedure is the same.Theorem 12. as the AR model cannot conceivably describe this data. This model is ∆Yt = µ1 + µ2 t + α0 Yt−1 + α1 ∆Yt−1 + · · · + αk−1 ∆Yt−(k−1) + et . (12.8). and a diﬀerent table should be consulted. If the test does not reject H0 . rather than the ˆ 1/2 . this means that the evidence points to Yt being stationary. Assume α0 = 0. 144 . T α0 →d (1 − α1 − α2 − · · · − αk−1 ) DFα ˆ ADF = α0 ˆ → DFt . Thus the Dickey-Fuller test is robust to heteroskedasticity. a common conclusion is that the data suggests that Yt is non-stationary.12.8) This is natural when the alternative hypothesis is that the series is stationary about a linear time trend. where c is the critical value from the ADF table. If the test rejects H0 . They are skewed to the left. This is called a “super-consistent” rate of convergence. s(ˆ 0 ) α The limit distributions DFα and DFt are non-normal. This distribution has been extensively α tabulated. but diﬀerent critical values are required. The ADF test has a diﬀerent distribution when the time trend has been included. But the theorem does not require an assumption of homoskedasticity. and have negative means.

.. Let It−1 = (Yt−1 . deﬁning the mk + 1 vector ⎜ Yt−1 ⎜ ⎜ xt = ⎜ Yt−2 ⎜ . Alternatively.Chapter 13 Multivariate Time Series A multivariate time series Yt is a vector process m × 1. .1 Vector Autoregressions (VARs) A VAR model speciﬁes that the conditional mean is a function of only a ﬁnite number of lags: E (Yt | It−1 ) = E (Yt | Yt−1 . .and each of A1 through Ak are m × m matrices. Yt−k ) = A0 + A1 Yt−1 + A2 Yt−2 + · · · Ak Yt−k . we have the VAR model Yt = A0 + A1 Yt−1 + A2 Yt−2 + · · · Ak Yt−k + et E (et | It−1 ) = 0.. Yt−2 . 13... ⎝ . A linear VAR speciﬁes that this conditional mean is linear in the arguments: E (Yt | Yt−1 . Deﬁning the m × 1 regression error et = Yt − E (Yt | It−1 ) . and the m × (mk + 1) matrix A0 A1 A2 · · · 145 . .. The typical goal is to ﬁnd the conditional expectation E (Yt | It−1 ) . Observe that A0 is m × 1. .. Note that since Yt is a vector. this conditional expectation is also a vector.. Yt−k ) . Yt−k A= ¡ ⎛ 1 ⎞ ⎟ ⎟ ⎟ ⎟ ⎟ ⎠ Ak ¢ .) be all lagged information at time t.

For example. j Unrestricted VARs were introduced to econometrics by Sims (1980).then Yt = Axt + et . m. The VAR model is a system of m equations.. ¡ Y1 Y2 · · · where Y = Ym 0 the T × m matrix of the stacked yt . Restrictions may be imposed on individual equations. Then the VAR system can be written as the equations Yjt = a0 xt + ejt .. 146 . ˆj ˆ And if we stack these to create the estimate A. Consider the moment conditions j = 1. The GMM framework gives a convenient method to impose such restrictions on estimation. each with the same set of regressors. or as a linear projection. ˆ An alternative way to compute this is as follows.3 Restricted VARs The unrestricted VAR is a system of m equations. either as a regression. These are implied by the VAR model. . ⎠ 0 Ym+1 = Y 0 X(X 0 X)−1 . A restricted VAR imposes restrictions on the system. we ﬁnd ⎞ ⎛ Y10 ⎜ Y0 ⎟ ˆ = ⎜ . Note that a0 = Yj0 X(X 0 X)−1 .2 Estimation E (xt ejt ) = 0. 13.2 ⎟ X(X 0 X)−1 A ⎜ . One way to write this is to let a0 be the jth row j of A. This (system) estimator is known as the SUR (Seemingly Unrelated Regressions) estimator. The GMM estimator corresponding to these moment conditions is equation-by-equation OLS aj = (X 0 X)−1 X 0 Yj . some regressors may be excluded from some of the equations. or across equations. ⎟ ⎝ .. and was originally derived by Zellner (1962) ¢ 13.

(A simple version of this estimator is called Cochrane-Orcutt.3) which is a valid regression model. (13. which once was very popular.2) may be estimated by iterated GLS.5 Testing for Omitted Serial Correlation Consider the problem of testing for omitted serial correlation in equation (13. t and xt = This is just a conventional regression. and is typically rejected empirically. and is still routinely reported by conventional regression packages. Then the null and alternative are H0 : θ = 0 H1 : θ 6= 0. h¡ 1 Yt−1 · · · 0 Yt−k Zt−1 · · · 0 Zt−k ¢0 i . we are only interested in a single equation out of a VAR system. xt−1 ) to the regression. Another interesting fact is that (13.) Since the common factor restriction appears arbitrary. Let yt = Yjt .1). direct estimation of (13.2) is a special case of (13. under the restriction γ = −βθ. it is convenient to re-deﬁne the variables.1) yt = x0 β + et . j Often. t or yt = θyt−1 + x0 β + x0 γ + ut . the model (13. So testing H0 versus H1 is equivalent to testing for the signiﬁcance of adding (yt−1 . Suppose that et is an AR(1). general. Let et = ejt and β = aj . to get t ¡ ¢ ¡ ¢ yt − θyt−1 = x0 β + et − θ x0 β + et−1 t t−1 = x0 β − θxt−1 β + et − θet−1 .2) is uncommon in recent applications.4 Single Equation from a VAR Yjt = a0 xt + et . If valid. t t−1 (13. We see that an appropriate.3). Then yt = x0 β + et t et = θet−1 + ut E (ut | It−1 ) = 0.2) Take the equation yt = x0 β + et . This restriction. which is called a common factor restriction.13. and subtract oﬀ the equation once lagged multiplied by θ. This can be done by a Wald test. 1). and Zt be the other variables. Then the single equation takes the form (13. In this case. t and et is iid N (0. You may have heard of the Durbin-Watson test for omitted serial correlation. This takes the form 0 and xt consists of lagged values of Yjt and the other Ylt s. with time series data. Otherwise it is invalid. 147 . may be tested if desired. and simple way to test for omitted serial correlation is to test the signiﬁcance of extra lagged values of the dependent variable and regressors. 13. The DW test is appropriate only when regression yt = x0 β + et is not dynamic (has no lagged values on the RHS).

you may either use a testingbased approach (using. Note. Zt . The latter has more information. Zt ).t−1 ) . then we say that Zt Granger-causes Yt .. If Zt is some sort of forecast of the future. conditional on information in lagged Yt . such as a futures price.13. Yt−2 . If this condition does not hold.. The reason why we call this “Granger Causality” rather than “causality” is because this is not a physical or structure deﬁnition of causality. That is. In a linear VAR.) I2t = (Yt . however. Zt−1 . . lagged Zt does not help to forecast Yt . the equation for Yt is 0 0 Yt = α + ρ1 Yt−1 + · · · + ρk Yt−k + Zt−1 γ 1 + · · · + Zt−k γ k + et . If it is found that Zt does not Granger-cause Yt . then we deduce that our time-series model of E (Yt | It−1 ) does not require the use of Zt . This deﬁnition of causality was developed by Granger (1969) and Sims (1972). Zt−2 . We say that Zt does not Granger-cause Yt if E (Yt | I1. In this equation. . or an information criterion approach. The hypothesis can be tested by using the appropriate multivariate Wald test. That is. This idea can be applied to blocks of variables. Yt−1 . that Zt may still be useful to explain other features of Yt . 148 . and et (k) is the OLS residual vector from the ˆ model with k lags. and the information set I2t is generated by both Yt and Zt . The formula for the AIC and BIC are ³ ´ p ˆ AIC(k) = log det Ω(k) + 2 T ³ ´ p log(T ) ˆ BIC(k) = log det Ω(k) + T T X 1 ˆ et (k)ˆt (k)0 ˆ e Ω(k) = T t=1 p = m(km + 1) where p is the number of parameters in the model. the Wald statistic). Yt and/or Zt can be vectors. Zt does not Granger-cause Yt if and only if H0 : γ 1 = γ 2 = · · · = γ k = 0.7 Granger Causality Partition the data vector into (Yt . .6 Selection of Lag Length in an VAR If you want a data-dependent rule to pick the lag length k in a VAR.. Yt−1 . This may be tested using an exclusion (Wald) test. then Zt may help to forecast Yt even though it does not “cause” Yt . Yt−2 . Deﬁne the two information sets I1t = (Yt . such as the conditional variance. 13.t−1 ) = E (Yt | I2.. for example. The log determinant is the criterion from the multivariate normal likelihood.) The information set I1t is generated only by the history of Yt .

then β = (1 − 1)0 speciﬁes that the spread (the diﬀerence in returns) is stationary. For example. 13.9 Cointegrated VARs A(L)Yt = et A(L) = I − A1 L − A2 L2 − · · · − Ak Lk We can write a VAR as or alternatively as ∆Yt = ΠYt−1 + D(L)∆Yt−1 + et where Π = −A(1) = −I + A1 + A2 + · · · + Ak . Often. and was articulated in detail by Engle and Granger (1987). If the series Yt is not cointegrated. If Y = (log(Consumption) log(Income))0 . If Yt is cointegrated with a single cointegrating vector (r = 1). When the data have time trends.8 Cointegration The idea of cointegration is due to Granger (1981). then Yt is I(0). m × r. however. In some cases. but it is useful in the construction of alternative estimators and tests. so zt = β 0 Yt is known. the asymptotic distribution of the test is aﬀected by the presence of the time trend. The asymptotic distribution was worked out in B. yet under H1 zt is I(0). Yt is I(1) and cointegrated. If r = m. of rank r.1 The m × 1 series Yt is cointegrated if Yt is I(1) yet there exists β. Under H0 . it may be believed that β is known a priori. In other cases. When β is unknown. then β = (1 − 1)0 speciﬁes that log(Consumption/Income) is stationary. so the ADF critical values are not appropriate. if Yt is a pair of interest rates. Y2t ) and β = (β 1 β 2 ) and normalize β 1 = 1.13. β may not be known. in general.8. a good method to estimate β. then it turns out that β can be consistently estimated by an OLS regression of one component of Yt on the others. β is not consistent. Hansen (1992). Thus H0 can be tested using a univariate ADF test on zt . Furthermore ˆ this estimation is super-consistent: T (β 2 − β 2 ) →d Limit. This is not. Thus Yt = ˆ (Y1t . as ﬁrst shown by Stock (1987). then r = 0. from OLS of Y1t on Y2t . The r vectors in β are called the cointegrating vectors. β = (1 − 1)0 . such that zt = β 0 Yt is I(0). Engle and Granger (1987) suggested using an ADF test on the estimated ˆ0 ˆ residual zt = β Yt . Whether or not the time trend is included. For 0 < r < m. Then β 2 = (Y20 Y2 )−1 Y2 Y1 →p β 2 . it may be necessary to include a time trend in the estimated cointegrating regression. The asymptotic distribution was worked out by Phillips and Ouliaris (1990). Deﬁnition 13. We are often interested in testing the hypothesis of no cointegration: H0 : r = 0 H1 : r > 0. 149 . Suppose that β is known. Their justiﬁcation was Stock’s result that β is superˆ ˆ consistent under H1 . Then under H0 zt is I(1).

and diﬀerent authors have adopted diﬀerent identiﬁcation schemes. they are typically called the “Johansen Max and Trace” tests. Their asymptotic distributions are non-standard. The restricted model can be written as ∆Yt = αβ 0 Yt−1 + D(L)∆Yt−1 + et ∆Yt = αzt−1 + D(L)∆Yt−1 + et . Ω). and are similar to the Dickey-Fuller distributions. One diﬃculty is that β is not identiﬁed without normalization. When r > 1. this is the MLE of the restricted parameters under the assumption that et is iid N (0. which is least-squares subject to the stated restriction. Thus cointegration imposes a restriction upon the parameters of a VAR.9.1 (Granger Representation Theorem). this does not work. 1991.Theorem 13. 1995). In the context of a cointegrated VAR estimated by reduced rank regression. Yt is cointegrated with m × r β if and only if rank(Π) = r and Π = αβ 0 where α is m × r. rank(α) = r. If β is unknown. As they were discovered by Johansen (1988. we typically just normalize one element to equal unity. this can be estimated by OLS of ∆Yt on zt−1 and the lags of ∆Yt . These tests are constructed as likelihood ratio (LR) tests. If β is known. Equivalently. then estimation is done by “reduced rank regression”. it is simple to test for cointegration by testing the rank of Π. When r = 1. 150 .

the goal is to describe P (Yi = 1 | xi ) . 1. allowing the construction of the likelihood function. However. 2.. For the parametric approach. The two standard choices for F are 151 . estimation is by MLE. . so that F (z) = 1 − F (−z).. i As P (Yi = 1 | xi ) = E (Yi | xi ) . you were to write a thesis involving LDV estimation. They still constitute the majority of LDV applications.. The linear probability model speciﬁes that P (Yi = 1 | xi ) = x0 β.} • Censored: Y = {x : x ≥ 0} The traditional approach to the estimation of limited dependent variable (LDV) models is parametric maximum likelihood. as this is the full conditional distribution.1 Binary Choice The dependent variable Yi = {0. A major practical issue is construction of the likelihood function. 2. Given some regressors xi . 1}. If. this yields the regression: Yi = x0 β + ei which can be estimated by i OLS. you would be advised to consider employing a semi-parametric estimation approach. however. This represents a Yes/No outcome. . typically assumed to be symmetric about zero. 1} • Multinomial: Y = {0. The standard alternative is to use a function of the form ¡ ¢ P (Yi = 1 | xi ) = F x0 β i where F (·) is a known CDF. eliminating the dependence on a parametric distributional assumption.Chapter 14 Limited Dependent Variables A “limited dependent variable” Y is one which takes a “limited” set of values. A more modern approach is semi-parametric.. 1. 14. k} • Integer: Y = {0. Even so estimation of a linear probability model is a useful starting point for subsequent analysis. The most common cases are • Binary: Y = {0. A parametric model is constructed. We will discuss only the ﬁrst (parametric) approach. due to time constraints.. the linear probability model does not impose the restriction that 0 ≤ P (Yi | xi ) ≤ 1.

we need the conditional distribution of an individual observation. 1. This model is identical to the latent variable model Yi∗ = x0 β + ei i ei ∼ F (·) ½ 1 Yi = 0 For then P (Yi = 1 | xi ) = P (Yi∗ > 0 | xi ) ¡ ¢ = P x0 β + ei > 0 | xi i ¢ ¡ = P ei > −x0 β | xi i ¢ ¡ = 1 − F −x0 β i ¡ ¢ = F x0 β . If F is logistic. i | Hence the log-likelihood function is ln (β) = = = = n X i=1 n X i=1 n X i=1 log f (yi | xi ) ¡ ¡ ¢y ¡ ¢ ¢ log F x0 β i (1 − F x0 β )1−yi i i yi =1 X £ ¡ ¢ ¡ ¢¤ yi log F x0 β + (1 − yi ) log(1 − F x0 β ) i i ¡ ¢ X ¡ ¢ log F x0 β + log(1 − F x0 β ). Standard errors and test statistics are computed by asymptotic approximations. Thus i the conditional density is f (yi xi ) = pyi (1 − pi )1−yi i ¡ 0 ¢yi ¡ ¢ = F xi β (1 − F x0 β )1−yi . we call this the logit model. i i yi =0 ˆ The MLE β is the value of β which maximizes ln (β). −1 . such that P (Y = 1) = p and P (Y = 0) = 1 − p.• Logistic: F (u) = (1 + e−u ) • Normal: F (u) = Φ(u). 152 . To construct the likelihood. then we can write the density of Y as f (y) = py (1 − p)1−y . In the Binary choice model. and if F is normal. y = 0. Recall that if Y is Bernoulli. i if Yi∗ > 0 . Details of such calculations are left to more advanced courses. Yi is conditionally Bernoulli with P (Yi = 1 | xi ) = pi = F (x0 β) . we call this the probit model. otherwise Estimation is by maximum likelihood.

He proposed the latent variable model ∗ yi = x0 β + ei i ei ∼ iid N (0. incomes above a threshold are typically reported at the threshold. 2. i If Y = {0.2 Count Data exp (−λi ) λk i . Thus the model is that there is some latent process yi with unbounded support. . this motivates the label Poisson “regression.. The functional form for λi has been picked to ensure that λi > 0. In surveys. i so the model imposes the restriction that the conditional mean and variance of Yi are the same.3 Censored Data The idea of “censoring” is that some data above or below a threshold are mis-reported at the ∗ threshold. The conditional density is the Poisson with parameter λi . 14. Since E (Yi | xi ) = λi = exp(x0 β) i is the conditional mean.. or it may be a by-product of economic constraints. the consumption level (purchases) in a particular period was zero..” Also observe that the model implies that V ar (Yi | xi ) = λi = exp(x0 β). 2.1) yi = ∗ <0 . but we observe only ½ ∗ ∗ yi if yi ≥ 0 (14. The censoring process may be explicit in data collection. This model (now called the Tobit) speciﬁes that the latent (or ideal) value of consumption may be negative (the household 153 . . 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. This model speciﬁes that P (Yi = k | xi ) = λi k = 0.. This may be considered restrictive. σ 2 ) with the observed variable yi generated by the censoring equation (14. The ﬁrst censored regression model was developed by Tobin (1958) to explain consumption of durable goods. An example of a data collection censoring is top-coding of income.) The observed data yi therefore come from a mixed continuous/discrete distribution. k! = exp(x0 β). A generalization is the negative binomial. i i i=1 ˆ The MLE is the value β which maximizes ln (β). The log-likelihood function is ln (β) = n X i=1 log f (yi | xi ) = n X¡ ¢ − exp(x0 β) + yi x0 β − log(yi !) . Tobin observed that for many households. 1.1). a typical approach is to employ Poisson regression. 0 if yi (This is written for the case of the threshold being zero. 1. any known value can substitute.}.14.

¶¸ µ 0 ¶1(y=0) ∙ µ z − x0 β 1(y>0) xi β −1 i f (y | xi ) = Φ − . where the goal is to assess the eﬀect of “training” on the general population. The naive approach to estimate β is to regress yi on xi .4 Sample Selection The problem of sample selection arises when the sample is a non-random selection of potential observations.would prefer to sell than buy). not just on the volunteers. σ φ σ σ where 1 (·) is the indicator function. not E (Yi∗ | xi ) = x0 β. and they wish to extrapolate the eﬀects of the experiment on a general population. [Note: it is still possible to estimate E (Yi | xi ) by LS techniques. and the latter is of interest. if you ask for volunteers for an experiment. Hence the log-likelihood is a mixture of the probit and the normal: ln (β) = n X i=1 ¶¸ µ 0 ¶ X ∙ µ xi β yi − x0 β −1 i + . log Φ − log σ φ = σ σ y =0 y >0 X i i log f (yi | xi ) ˆ The MLE is the value β which maximizes ln (β). P (yi = y | xi ) = σ φ σ 0 Therefore. This occurs when the observed data is systematically diﬀerent from the population of interest. Thus OLS will be biased i for the parameter of interest β. The Tobit framework postulates that this is not inherently interesting. 154 . observe that the probability of being censored is ∗ P (yi = 0 | xi ) = P (yi < 0 | xi ) ¢ ¡ 0 = P xi β + ei < 0 | xi ¶ µ x0 β ei i <− | xi = P σ σ µ 0 ¶ xβ = Φ − i . All that is reported is that the household purchased zero units of the good. σ The conditional distribution function above zero is Gaussian: ¶ µ Z y z − x0 β −1 i dz.] Consistent estimation will be achieved by the MLE. This does not work because regression estimates E (Yi | xi ) . the density function can be written as y > 0. you should worry that the people who volunteer may be systematically diﬀerent from the general population. 14. For example. that the parameter of β is deﬁned by an alternative statistical structure. To construct the likelihood. This has great relevance for the evaluation of anti-poverty and job-training programs.

For example. ρ from OLS of yi on xi and λ(z 0 γ ).2) is a valid regression equation for the observations for which Ti = 1. Heckman (1979) observed that we could consistently estimate β and ρ from this equation. ˆ ³ ´ ˆ ˆ ˆ • Estimate β. as this is a two-step estimator. using regressors zi . where vi is independent of e0i ∼ N (0. It is unknown. zi . ¡ 0 ¢ yi = x0 β + ρλ zi γ + ui i (14. e1i = ρe0i + vi . if γ were known. 155 . Zi ) = E e1i | {e0i > −zi γ}. which is non-zero. e1i ρ σ2 It is presumed that we observe {xi . The problem is that this is equivalent to conditioning on the event {Ti = 1}. alternatively. The dependent variable yi is observed if (and only if) Ti = 1. The model is often completed by specifying that the errors are jointly normal ¶ µ µ ¶¶ µ 1 ρ e0i ∼ N 0. The binary dependent variable is Ti . 1). Thus E (ui | Ti = 1.A simple sample selection model can be written as the latent model yi = x0 β + e1i i ¡ 0 ¢ Ti = 1 zi γ + e0i > 0 where 1 (·) is the indicator function. A useful fact about the standard normal distribution is that φ(x) . . The equation for Ti is an equation specifying the probability that the person is employed. Zi ¡ 0 ¢ = ρλ zi γ . Ti } for all observations. The “Heckit” estimator is thus calculated as follows • Estimate γ from a Probit. Under the normality assumption. i • The OLS standard errors will be incorrect. Else it is unobserved. They can be corrected using a more complicated formula. The naive estimator of β is OLS regression of yi on xi for those observations for which yi is available. Zi ) = 0. but also can be consistently estimated by a Probit model for selection. However. by viewing the Probit/OLS estimation equations as a large joint GMM problem. where Hence ¡ 0 ¢ e1i = ρλ zi γ + ui . E (e0i | e0i > −x) = λ(x) = Φ(x) and the function λ(x) is called the inverse Mills ratio. yi could be a wage. ¡ ¢ 0 E (e1i | Ti = 1. Or. which can be observed only if a person is employed. Zi ¢ ¡ ¢ ¡ 0 0 = ρE e0i | {e0i > −zi γ}. Zi + E vi | {e0i > −zi γ}.

and the estimator can be quite sensitive to this assumption. 0ˆ The estimator can also work quite poorly if λ (zi γ ) does not have much in-sample variation. and hence improve the second stage estimator’s precision. 156 . Another 0ˆ potential problem is that if zi = xi .The Heckit estimator is frequently used to deal with problems of sample selection. This can happen if the Probit equation does not “explain” much about the selection choice. Some modern econometric research is exploring how to relax the normality assumption. However. so the second step OLS estimator will not be able to precisely estimate β. then λ (zi γ ) can be highly collinear with xi . If 0ˆ this is valid. it will ensure that λ (zi γ ) is not collinear with xi . the estimator is built on the assumption of normality. Based this observation. it is typically recommended to ﬁnd a valid exclusion restriction: a variable should be in zi which is not in xi .

. i = 1. then OLS is inconsistent.. .. Condition (15. It is a strong assumption.1) fails if the individual-speciﬁc unobserved eﬀect ui is correlated with the observed explanatory variables xit . xit } : For i = 1. The motivation is to allow ui to be arbitrary. .1) If this condition fails. OLS can be improved upon via a GLS technique.. collected over time. 157 . it The typical maintained assumptions are that the individuals i are mutually independent. If (15. and most applied researchers try to avoid its use. and thus achieve invariance. and the t subscript denotes time.. where the i subscript denotes the individual... n. 15. and that eit is uncorrelated with xit . OLS of yit on xit is called pooled estimation.2 Fixed Eﬀects This is the most common technique for estimation of non-dynamic linear panel regressions. ti . . T .1 Individual-Eﬀects Model The standard panel data speciﬁcation is that there is an individual-speciﬁc eﬀect which enters linearly in the regression yit = x0 β + ui + eit . There are several derivations of the estimator.1) is called the random eﬀects hypothesis. It is consistent if E (xit ui ) = 0 (15. or unbalanced : {yit . In either event. however. This is often believed to be plausible if ui is an omitted variable. and have arbitrary correlated with xi . A panel may be balanced : {yit . (15.. xit }... OLS appears a poor estimation choice.. that eit is iid across individuals and time. 15.1) is true.. xit } : t = 1. An observation is the pair {yit . t = ti ..Chapter 15 Panel Data A panel is a set of observations on individuals. n. that ui and eit are independent. The goal is to eliminate ui from the estimator.

⎟ di = ⎝ . it i (15. u). ⎟ u = ⎝ . else Then note that and an n × 1 dummy vector with a “1” in the i0 th place. with di as a regressor along with xi .g. Observe that • This is generally consistent. as the parameter space is too large. . their gender) will be collinear with di . ⎠ . i yit = x0 β + d0 u + eit . • If xit contains an intercept. ˆ ˆ OLS on (15.2) is a valid regression. which is quite large as typically n is very large.First.2) ⎞ di1 ⎜ . let dij = and ⎧ ⎨ 1 ⎩ 0 ⎛ if i = j . you do not want to actually implement conventional OLS estimation. OLS estimation of β proceeds by the FWL theorem. 158 . so the intercept is typically omitted from xit . Let ⎛ ⎞ u1 ⎜ . so (15. so will have to be omitted. it will be collinear with di . Conventional inference applies. ⎠. din Observe that E (eit | xit . un ui = d0 u. • Any regressor in xit which is constant over time for all individuals (e. di ) = 0. Stacking the observations together: Y = Xβ + Du + e. Computationally. . Y ∗ = Y − D(D0 D)−1 D0 Y where X ∗ = X − D(D0 D)−1 D0 X.. then by the FWL theorem. ˆ β = ¢−1 ¡ 0 ¢ ¡ 0 X (1 − PD ) Y X (1 − PD ) X ¡ ¢−1 ¡ ∗0 ∗ ¢ = X ∗0 X ∗ X Y .2) yields estimator (β. • There are n + k regression parameters.

3) A dynamic panel regression has a lagged dependent variable This is a model suitable for studying dynamic behavior of individual agents. i ∗ yit = x∗0 β + e∗ . if eit is iid. 15. Taking ﬁrst-diﬀerences of (15. it and then take individual-speciﬁc means by taking the average for the i0 th individual: ti ti ti 1 X 1 X 0 1 X yit = x β + ui + eit Ti t=t Ti t=t it Ti t=t i i i or Subtracting.4) by IV using yt−2 as an instrument for ∆yt−1 (which is just identiﬁed). Another derivation of the estimator is to take the equation yit = x0 β + ui + eit . there are more instruments available. Unfortunately.4) . 159 (15. two periods back. k ≥ 2. Alternatively. This is conveniently organized by the GMM principle. A more sophisticated GMM estimator recognizes that for time-periods later in the sample. But if there are valid instruments. but loses a time-series observation). A simple application of GMM yields the parameter estimates and standard errors.Since the regression of yit on di is a regression onto individual-speciﬁc dummies. Hence a valid estimator of α and β is to estimate (15. e So OLS on (15. the predicted value from these regressions is the individual speciﬁc mean y i . as this enables the moments from the diﬀerent timeperiods to be stacked together to create a list of all the moment conditions.3 Dynamic Panel Regression yit = αyit−1 + x0 β + ui + eit . at least if T is held ﬁnite as n → ∞. GMM using yt−2 and yt−3 as instruments (which is overidentiﬁed. and the residual is the demean value ∗ yit = yit − yi .4) will be inconsistent. the dependent variable and regressors in deviationit from-mean form. are valid instruments. as yt−2 is uncorrelated with ∆eit .3) eliminates the individual-speciﬁc eﬀect: ∆yit = α∆yit−1 + ∆x0 β + ∆eit . so the instrument list should be diﬀerent for each equation. it However. Typically. it it which is free of the individual-eﬀect ui . This is because the sample mean of yit−1 is correlated with that of eit . we ﬁnd y i = x0 β + ui + ei . then IV or GMM can be used to estimate the equation. The standard approach to estimate a dynamic panel is to combine ﬁrst-diﬀerencing with IV or GMM. then it will be correlated with ∆yit−1 : E (∆yit−1 ∆eit ) = E ((yit−1 − yit−2 ) (eit − eit−1 )) = −E (yit−1 eit−1 ) = −σ 2 . ∗ ˆ The ﬁxed eﬀects estimator β is OLS of yit on x∗ . it (15. Thus values of yit−k . the ﬁxed eﬀects estimator is inconsistent. we use lags of the dependent variable.

The goal is to estimateP(x) from a random sample (X1 .. The kernel density estimator of f (x) is 1X ˆ f (x) = Kh (Xi − x) . we cannot deﬁne d F (x) since F (x) is a step function. The kernel functions are used to smooth the data.. Let 1 ³u´ . the choice between these three rarely makes a meaningful diﬀerence in the estimates. and then see how the method generalizes to estimating the entire function.1 Kernel Density Estimation d Let X be a random variable with continuous distribution F (x) and density f (x) = dx F (x). While we are typically interested in estimating the entire function f (x). Three common choices for kernels include the Gaussian µ 2¶ x 1 K(x) = √ exp − 2 2π the Epanechnikov K(x) = and the Biweight or Quartic K(x) = 15 16 3 4 ¡ ¢ 1 − x2 .. |x| ≤ 1 0 |x| > 1 ¡ ¢2 1 − x2 .Chapter 16 Nonparametrics 16. n i=1 n 160 . Deﬁnition 1 K(u) is a second-order kernel function if it is a symmetric zero-mean density function. The ˆ ˆ EDF F i=1 dx standard nonparametric method to estimate f (x) is based on smoothing using a kernel. |x| ≤ 1 0 |x| > 1 In practice. Kh (u) = K h h be the kernel K rescaled by the bandwidth h. Xn } While F (x) can be estimated by the f ˆ (x) = n−1 n 1 (Xi ≤ x) . we can simply focus on the problem where x is a speciﬁc ﬁxed number. . The amount of smoothing is controlled by the bandwidth h > 0.

The mean is Z ∞ ˆ xf (x)dx = = = = −∞ 1X n 1 n 1 n 1 n n i=1 −∞ n XZ ∞ Z ∞ xKh (Xi − x) dx (Xi + uh) K (u) du n i=1 −∞ n X Z ∞ Xi Xi i=1 n X i=1 1X K (u) du + h n −∞ i=1 Z ∞ uK (u) du −∞ the sample mean of the Xi .This estimator is the average of a set of weights. where the second-to-last equality used the change-of-variables u = (Xi − x)/h which has Jacobian h. Conversely. f (x) is a valid density. If a large number of the observations Xi are near ˆ x. The bandwidth h controls the meaning of “near”. The second moment of the estimated density is Z ∞ n Z 1X ∞ 2 2ˆ x f (x)dx = x Kh (Xi − x) dx n −∞ i=1 −∞ n Z 1X ∞ = (Xi + uh)2 K (u) du n −∞ i=1 Z ∞ Z n n n X 2X 1X 2 ∞ 2 1 2 Xi + Xi h K(u)du + h u K (u) du = n n n −∞ −∞ = where σ2 = K 1 n i=1 n X i=1 i=1 i=1 Xi2 + h2 σ 2 K Z ∞ x2 K (x) dx ˆ is the variance of the kernel. and Z ∞ ˆ f (x)dx = −∞ Z ∞ −∞ 1X 1X Kh (Xi − x) dx = n n i=1 i=1 n n Z ∞ −∞ Kh (Xi − x) dx = 1X n i=1 n Z ∞ K (u) du = 1 −∞ where the second-to-last equality makes the change-of-variables u = (Xi − x)/h. That is. ˆ We can also calculate the moments of the density f (x). f (x) ≥ 0 for all x. then the weights are relatively large and f (x) is larger. if only a few Xi are near x. 161 . It follows that the variance of the density f (x) is Ã n !2 ¶2 µZ ∞ Z ∞ n 1X 1X 2 ˆ ˆ x2 f (x)dx − xf (x)dx = Xi + h2 σ 2 − Xi K n n −∞ −∞ i=1 i=1 −∞ = σ ˆ 2 + h2 σ 2 K Thus the variance of the estimated density is inﬂated by the factor h2 σ 2 relative to the sample K moment. ˆ(x) is small. then the weights are small and f ˆ ˆ Interestingly.

16. 2 Z ∞ ˆ The bias of f (x) is then 1 1X ˆ EKh (Xi − x) − f (x) = f 00 (x)h2 σ 2 . The sharper the derivative. Bias(x) = E f (x) − f (x) = K n 2 i=1 n ˆ We see that the bias of f (x) at x depends on the second derivative f 00 (x). 162 . so we approximate it using a second order Taylor expansion of f (x + hu) in the argument hu about hu = 0. ˆ the greater the bias. Thus 1 f (x + hu) ' f (x) + f 0 (x)hu + f 00 (x)h2 u2 2 and therefore µ ¶ 1 00 0 2 2 K (u) f (x) + f (x)hu + f (x)h u du EKh (X − x) ' 2 −∞ Z ∞ Z ∞ Z ∞ 1 = f (x) K (u) du + f 0 (x)h K (u) udu + f 00 (x)h2 K (u) u2 du 2 −∞ −∞ −∞ 1 00 2 2 = f (x) + f (x)h σ K . ˆ We now examine the variance of f (x). and is larger the greater the curvature in f (x). so is estimating a smoothed version of f (x). nh (x)2 dx is called the roughness of K. The bias results from this smoothing. Intuitively. The last expression shows that the expected value is an average of f (z) locally about x. using ﬁrst-order Taylor approximations and the fact that n−1 is of smaller order than (nh)−1 V ar (x) = = ' = ' = where R(K) = R∞ −∞ K 1 V ar (Kh (Xi − x)) n 1 1 EKh (Xi − x)2 − (EKh (Xi − x))2 n n ¶ Z ∞ µ z−x 2 1 1 K f (z)dz − f (x)2 2 nh −∞ h n Z ∞ 1 K (u)2 f (x + hu) du nh −∞ Z f (x) ∞ K (u)2 du nh −∞ f (x) R(K) .2 Asymptotic MSE for Kernel Estimates ∞ For ﬁxed x and bandwidth h observe that Z ∞ Z Kh (z − x) f (z)dz = EKh (X − x) = −∞ Kh (uh) f (x + hu)hdu = −∞ Z ∞ K (u) f (x + hu)du −∞ The second equality uses the change-of variables u = (z − x)/h. This integral (typically) is not analytically solvable. which is valid as h → 0. Since it is an average of iid random variables. the estimator f (x) smooths data local to Xi = x.

h must tend to zero. We thus say that the optimal bandwidth is of order O(n−1/5 ). h0 = argmin AM ISEh h It can be found by solving the ﬁrst order condition d R(K) AM ISEh = h3 σ 4 R(f 00 ) − =0 K dh nh2 yielding h0 = µ R(K) 4 R(f 00 ) σK ¶1/5 n−1/2 . The ﬁrst two are determined by the kernel function. That is. Equation (16. and gives a nearly optimal rule when f (x) is close to normal. how should the bandwidth be selected? This is a diﬃcult problem. A classic simple solution proposed by Silverman (1986)has come to be known as the reference bandwidth or Silverman’s Rule-of-Thumb. That is. R∞ R∞ K σ 2 = −∞ x2 K (x) dx R(K) = −∞ K (x)2 dx K √ Gaussian 1 1/(2 π) Epanechnikov 1/5 1/5 Biweight 1/7 5/7 An obvious diﬃculty is that R(f 00 ) is unknown. We deﬁne the asymptotically optimal bandwidth h0 as the value which minimizes this approximate MSE. The asymptotically optimal choice given in (16. 1) distribution and σ 2 is an estimate of σ 2 = V ar(X).1) AM ISEh = AM SEh (x)dx = 4 nh R where R(f 00 ) = (f 00 (x))2 dx is the roughness of f 00 . Notice that the ﬁrst term (the squared bias) is increasing in h and the second term (the variance) is decreasing in nh.2) depends on R(K). ˆ It uses formula (16.06n−1/5 163 . but at a slower rate than n−1 . (16. The downside is that if the density is very far from normal. This choice for h gives an optimal rule when f (x) is ˆ normal. where φ is the N (0. In practice. we need h → 0 but nh → ∞.2) but replaces R(f 00 ) with σ −5 R(φ00 ). and there is a large and continuing literature on the subject. and R(f 00 ).Together.2) This solution takes the form h0 = cn−1/5 where c is a function of K and f. AM SEh (x) = f 00 (x)2 h4 σ 4 + K 4 nh A global measure of precision is the asymptotic mean integrated squared error (AMISE) Z h4 σ 4 R(f 00 ) R(K) K + . We can calculate that √ R(φ00 ) = 3/ (8 π) . Their K values for the three functions introduced in the previous section are given here. σ 2 . Thus for the AMISE to decline with n. the rule-of-thumb h can be quite ineﬃcient. the asymptotic mean-squared error (AMSE) for ﬁxed x is the sum of the approximate squared bias and approximate variance 1 f (x) R(K) . Together with the above table. we ﬁnd the reference rules for the three kernel functions introduced earlier. (16. Gaussian Kernel: hrule = 1.1) is an asymptotic approximation to the MSE. but at a very slow rate. Note that this h declines to zero. but not of n.

there are modern versions of this estimator work well.78n−1/5 Unless you delve more deeply into kernel estimation methods the rule-of-thumb bandwidth is a good practical bandwidth choice. This is more treacherous than may ﬁrst appear. in particular the iterative method of Sheather and Jones (1991). Another popular choice for selection of h is cross-validation. which are known as smoothed cross-validation which is a close cousin of the bootstrap. I now discuss some of these choices. and then plug this estimate into the formula (16.34n−1/5 Biweight (Quartic) Kernel: hrule = 2. There are other approaches. This works by constructing an estimate of the MISE using leave-one-out estimators. but implementation can be delicate. There are some desirable properties of cross-validation bandwidths.Epanechnikov Kernel: hrule = 2. perhaps adjusted by visual inspection of the resulting estimate ˆ f (x). 164 . but they are also known to converge very slowly to the optimal values. as the optimal h for estimation of the roughness R(f 00 ) is quite diﬀerent than the optimal h for estimation of f (x).2). However. They are also quite ill-behaved when the data has some discretization (as is common in economics). The plug-in approach is to estimate R(f 00 ) in a ﬁrst step. in which case the cross-validation rule can sometimes select very small bandwidths leading to dramatically undersmoothed estimates. Fortunately there are remedies.

. An event A is any collection of possible outcomes of an experiment. the sets {HH. For example.. / Complement: Ac = {x : x ∈ A}. the event that the ﬁrst coin is heads. When S is countable. A2 .. . Associativity: A ∪ (B ∪ C) = (A ∪ B) ∪ C. including S itself and the null set ∅. An event is a subset of S. When S is the real line. A2 . We only deﬁne probabilities for events contained in B. let B be the smallest sigma algebra which contains all of the open sets in S. A probability function assigns probabilities (numbers between 0 and 1) to events A in S. There are two outcomes. and A1 . i=1 i=1 Some important properties of the probability function include the following • P (∅) = 0 165 . T T }. is called a partition i=1 of S. (A ∩ B)c = Ac ∪ B c . then the collection A1 . if the sets A1 . T H. a probability function P satisﬁes P (S) = 1. then P P (∪∞ Ai ) = ∞ P (Ai ). We say that A and B are disjoint or mutually exclusive if A ∩ B = ∅. one event is A = {HH. We now can give the axiomatic deﬁnition of probability. including ∅ and S. P (A) ≥ 0 for all A ∈ B. ∈ B implies ∪∞ Ai ∈ B. Furthermore. HT } and {T H} are disjoint. A2 . The following are elementary set operations: Union: A ∪ B = {x : x ∈ A or x ∈ B}. Given S and B. Take the simple example of tossing a coin. A ∩ B = B ∩ A. The following are useful properties of set operations. so we can write S = {H.A set B is called a sigma algebra (or Borel ﬁeld) if ∅ ∈ B .. HT. A2 . Distributive Laws: A∩(B ∪ C) = (A ∩ B)∪(A ∩ C) .. HT }. A∪(B ∩ C) = (A ∪ B)∩(A ∪ C) . S} which is known as the trivial sigma i=1 algebra. If two coins are tossed in sequence. For any sample space S. This is straightforward when S is countable. DeMorgan’s Laws: (A ∪ B)c = Ac ∩ B c . A simple example is {∅. T }. ∈ B are pairwise disjoint.. We call B the sigma algebra associated with S. then B is the collection of all open and closed intervals.. are pairwise disjoint and ∪∞ Ai = S. B is simply the collection of all subsets of S. A ∈ B implies Ac ∈ B. . .Appendix A Probability A. Continuing the two coin example. . when S is uncountable we must be somewhat more careful. heads and tails.1 Foundations The set S of all possible outcomes of an experiment is called the sample space for the experiment. Intersection: A ∩ B = {x : x ∈ A and x ∈ B}. Communtatitivity: A ∪ B = B ∪ A. and if A1 . we can write the four outcomes as S = {HH.. A ∩ (B ∩ C) = (A ∩ B) ∩ C.

49. Furthermore. Depending on these two distinctions. i∈I i∈I 166 . and is with replacement if this is allowed. 983.. Ak are mutually independent when for any subset {Ai : i ∈ I}. Counting may be ordered or unordered. as µ ¶ n n! = . Rearranging the deﬁnition. Counting is ordered or not depending on whether the sequential order of the numbers is relevant to winning the lottery. This selection is without replacement if you are not allowed to select the same number twice.• P (A) ≤ 1 • P (Ac ) = 1 − P (A) • P (B ∩ Ac ) = P (B) − P (A ∩ B) • P (A ∪ B) = P (A) + P (B) − P (A ∩ B) • If A ⊂ B then P (A) ≤ P (B) • Bonferroni’s Inequality: P (A ∩ B) ≥ P (A) + P (B) − 1 • Boole’s Inequality: P (A ∪ B) ≤ P (A) + P (B) For some elementary probability models. it is useful to have simple rules to count the number of objects in a set. n ≥ r. These counting rules are facilitated by using the binomial coeﬃcients which are deﬁned for nonnegative integers n and r. the number of ¡49 if potential combinations is 6 = 13. the conditional probability function is a valid probability function where S has been replaced by B.1) We say that the events A and B are statistically independent when (A. Without Replacement Ordered Unordered n! (n−r)! ¡n¢ r In the lottery example. Ã ! \ Y P Ai = P (Ai ) . For example. .. ¢ counting is unordered and without replacement. r r! (n − r)! When counting the number of objects in a set. . in which case we ﬁnd P (A ∩ B) = P (A) P (B) (A.. 816. P (B) With Replacement nr ¡n+r−1¢ r For any B. 2.1) holds. we have four expressions for the number of objects (possible arrangements) of size r from n objects... there are two important distinctions. We can say that the occurrence of B has no information about the likelihood of event A when P (A | B) = P (A). we can write P (A ∩ B) = P (A | B) P (B) which is often quite useful. consider a lottery where you pick six numbers from the set 1.. we say that the collection of events A1 . If P (B) > 0 the conditional probability of the event A given the event B is P (A | B) = P (A ∩ B) . Counting may be with replacement or without replacement.

.. 167 . A2 . τ r ... and use lower case letters such as x for potential values and realized values.3) P (X = τ j ) = π j .. We say that the random variable X is continuous if F (x) is continuous in x.3) is unavailable. these cases are unusualRand are typically ignored. then for each i = 1. This induces a new sample space — the real line — and a new probability function on the real line. we represent the relative probabilities by the probability density function (PDF) f (x) = so that F (x) = and P (a ≤ X ≤ b) = Z d F (x) dx x f (u)du Z b −∞ f (x)dx.Theorem 2 (Bayes’ Rule). . For any set B and any partition A1 . P (B | Ai ) P (Ai ) P (Ai | B) = P∞ j=1 P (B | Aj ) P (Aj ) A. While there are examples of continuous random variables which do not possess a PDF. The probability function for X takes the form j = 1. F (x) is right-continuous We say that the random variable X is discrete if F (x) is a step function. Instead. For a random variable X we deﬁne its cumulative distribution function (CDF) as F (x) = P (X ≤ x) . A function F (x) is a CDF if and only if the following three properties hold: 1. 2.. we denote random variables by uppercase letters such as X. the range of X consists of a countable set of real numbers τ 1 ... In the latter case.2 Random Variables A random variable X is a function from a sample space S into the real line. (A. a These expressions only make sense if F (x) is diﬀerentiable. of the sample space. ∞ A function f (x) is a PDF if and only if f (x) ≥ 0 for all x ∈ R and −∞ f (x)dx. limx→−∞ F (x) = 0 and limx→∞ F (x) = 1 2.2) Sometimes we write this as FX (x) to denote that it is the CDF of X. In this case P (X = τ ) = 0 for all τ ∈ R so the representation (A. F (x) is nondecreasing in x 3. Pr where 0 ≤ π j ≤ 1 and j=1 π j = 1. Typically.. ... . r (A.

when it does and E |X|m < ∞ then ¯ ¯ dm ¯ = E (X m ) m M (λ)¯ dλ λ=0 which is why it is called the moment generating function. Since E (a + bX) = a + bEX. For example. The CF always exists.A. of the random variable X is kXkp = (E |X|p )1/p . and when E |X|m < ∞ ¯ ¯ dm ¯ = im E (X m ) . However.4) g(x)f (x)dx Z g(x)f (x)dx g(x)>0 I2 = − g(x)<0 If I1 = ∞ and I2 < ∞ then we deﬁne Eg(X) = ∞. 168 . If X is discrete. More generally. If I1 < ∞ and I2 = ∞ then we deﬁne Eg(X) = −∞. We can also write σ 2 = V ar(X). The MGF does not necessarily exist. √ where i = −1 is the imaginary unit. we deﬁne the mean or expectation Eg(X) as follows. Eg(X) = j=1 and if X is continuous Eg(X) = The latter is well deﬁned and ﬁnite if Z If (A. this allows the convenient expression V ar(a + bX) = b2 V ar(X). the characteristic function (CF) of X is C(λ) = E exp (iλX) . evaluate I1 = Z Z ∞ g(x)f (x)dx. m C(λ)¯ dλ λ=0 The Lp norm. Two special moments are the mean µ = EX and variance σ 2 = E (X − µ)2 = EX 2 − µ2 . We √ call σ = σ 2 the standard deviation of X.3 Expectation For any measurable real function g. −∞ ∞ −∞ |g(x)| f (x)dx < ∞. p ≥ 1. we say that expectation is a linear operator.4) does not hold. (A. The moment generating function (MGF) of X is M (λ) = E exp (λX) . If both I1 = ∞ and I2 = ∞ then Eg(X) is undeﬁned. r X g(τ j )π j . For m > 0. we deﬁne the m0 th moment of X as EX m and the m0 th central moment as E (X − EX)m .

1. n. Xm = xm ) = x1 + · · · + xm n! px1 px2 · · · pxm . X2 = x2 . 2. . we now list some important discrete distribution function. 0≤p≤1 x = 0. x EX = V ar(X) = Poisson P (X = x) = exp (−λ) λx .4 Common Distributions For reference. x = 1. 0≤p≤1 V ar(X) = λ We now list some important continuous distributions. 1.. 2.. x EX = np V ar(X) = np(1 − p) Geometric P (X = x) = p(1 − p)x−1 .. EX = p V ar(X) = p(1 − p) Binomial µ ¶ n x P (X = x) = p (1 − p)n−x . x! EX = λ x = 0.... . 2...... 0≤p≤1 x = 0..A. λ>0 x = 1... 1 EX = p 1−p V ar(X) = p2 Multinomial P (X1 = x1 .. . . 0≤p≤1 p1 + · · · + pm = 1 EX = V ar(X) = Negative Binomial µ ¶ r+x−1 P (X = x) = p(1 − p)x−1 . 169 . 1. Bernoulli P (X = x) = px (1 − p)1−x . . m x1 !x2 ! · · · xm ! 1 2 = n.

π (1 + x2 ) EX = ∞ −∞ < x < ∞ Γ(α + β) α−1 x (1 − x)β−1 . Γ(α)Γ(β) α α+β αβ (α + β + 1) (α + β)2 0 ≤ x ≤ 1. α > 0. β > 0 V ar(X) = ∞ ExponentiaI f (x) = ³x´ 1 . V ar(X) = R1 0 x−2 dx Z ∞ 0 exp (−x) dx (1 + exp (−x))2 Lognormal Ã ! (ln x − µ)2 1 exp − f (x) = √ . (1 + exp (−x))2 EX = 0 −∞ < x < ∞. σ>0 Pareto βαβ . α > 0. exp θ θ EX = θ 0 ≤ x < ∞. xβ+1 βα . β>2 (β − 1)2 (β − 2) 170 β>0 .Beta f (x) = µ = V ar(X) = Cauchy f (x) = 1 . α ≤ x < ∞. θ>0 V ar(X) = θ2 Logistic f (x) = exp (−x) . 2σ 2 2πσx ¢ ¡ EX = exp µ + σ 2 /2 ¡ ¢ ¡ ¢ V ar(X) = exp 2µ + 2σ 2 − exp 2µ + σ 2 f (x) = EX = V ar(X) = 0 ≤ x < ∞. β>1 β−1 βα2 .

Y ) is F (x. y)f (x. y). y)dy. y)dydx −∞ f (x. P ((X < Y ) ∈ A) = For any measurable function g(x. 0 ≤ x < ∞. y)dxdy. Y ) = The marginal distribution of X is FX (x) = P (X ≤ x) = = so the marginal density of X is fX (x) = d FX (x) = dx Z ∞ y→∞ Z x Z ∞ −∞ ∂2 F (x. b−a a+b 2 (b − a)2 12 a≤x≤b A. the joint probability density function is f (x. −∞ 171 .Uniform f (x) = EX = V ar(X) = Weibull µ γ¶ x γ γ−1 x . y) = P (X ≤ x. y) f (x. Eg(X. exp − f (x) = β β µ ¶ 1 EX = β 1/γ Γ 1 + γ ¶ ¶¶ µ µ µ 2 1 2/γ 2 V ar(X) = β −Γ 1+ Γ 1+ γ γ γ > 0. y)dxdy A Z ∞ −∞ Z ∞ g(x. Y ) is a function from the sample space into R2 . y). If F is continuous. Y ≤ y) . The joint CDF of (X. y) = For a Borel measurable set A ∈ R2 . β > 0 1 .5 Multivariate Random Variables A pair of bivariate random variables (X. ∂x∂y Z Z f (x. −∞ lim F (x.

The reverse. then V ar (X + Y ) = V ar(X) + V ar(Y ).6) The Cauchy-Schwarz Inequality implies that |ρXY | ≤ 1. then Cov(X. −∞ The random variables X and Y are deﬁned to be independent if f (x. 172 . Another implication of (A. If X and Y are independent. The correlation between X and Y is Corr(X. Y ) = σ XY = E ((X − EX) (Y − EY )) = EXY − EXEY. if X and Y are independent then E(XY ) = EXEY. If X and Y are independent. is not true. σxσY (A. Y ) = ρXY = σ XY .5) if the expectations exist. (A. y)dx. The covariance between X and Y is Cov(X. y) = g(x)h(y). A useful fact is that V ar (X + Y ) = V ar(X) + V ar(Y ) + 2Cov(X.The correlation is a measure of linear dependence. the variance of the sum is the sum of the variances. X 2 ) = 0. y) = fX (x)fY (y). the marginal density of Y is fY (y) = Z ∞ f (x. Furthermore.Similarly. 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 ) . then MZ (λ) = E exp (λ (X + Y )) = E (exp (λX) exp (λY )) ¡ ¢ ¡ ¢ = E exp λ0 X E exp λ0 Y = MX (λ)MY (λ). if EX = 0 and EX 3 = 0. however.5) is that if X and Y are independent and Z = X + Y. For example. then σ XY = 0 and ρXY = 0. free of units of measurement. then Z Z E (g(X)h(Y )) = g(x)h(y)f (y. Y ). For example. X and Y are independent if and only if there exist functions g(x) and h(y) such that f (x. An implication is that if X and Y are independent.

then Σ is a diagonal matrix and Ã k ! k X X V ar Xi = V ar (Xi ) i=1 i=1 A. One way to derive this expression from the deﬁnition of conditional probability is fY |X (y | x) = = = = = = ∂ lim P (Y ≤ y | x ≤ X ≤ x + ε) ∂y ε→0 ∂ P ({Y ≤ y} ∩ {x ≤ X ≤ x + ε}) lim ∂y ε→0 P (x ≤ X ≤ x + ε) F (x + ε.. we deﬁne the expectation Z g(x)f (x)dx Eg(X) = Rk where the symbol dx denotes dx1 · · · dxk . y) fX (x) The conditional density of Y given X = x is deﬁned as fY |X (y | x) = if fX (x) > 0. In particular. y) fX (x + ε) ∂2 ∂x∂y F (x. ∂x1 · · · ∂xk For a measurable function g : Rk → Rs .. y) − F (x. fX (x) 173 .6 Conditional Distributions and Expectation f (x.A k × 1 random vector X = (X1 . y) fX (x) f (x. y) . Letting x = (x1 . Xk )0 is a function from S to Rk . . we have the k × 1 multivariate mean µ = EX and k × k covariance matrix ¡ ¢ Σ = E (X − µ) (X − µ)0 = EXX 0 − µµ0 If the elements of X are mutually independent.. xk )0 .. . it has the distribution and density functions F (x) = P (X ≤ x) ∂k f (x) = F (x).. y) ∂ lim ∂y ε→0 FX (x + ε) − FX (x) ∂ lim ∂y ε→0 ∂ ∂x F (x + ε..

if E (Y | X = x) = α + βx.9) where m(x) = E (Y | X = x) . then E (Y | X) = α + βX. a transformation of X. functions of X.The conditional mean or conditional expectation is the function Z ∞ yfY |X (y | x) dy. Evaluated at x = X. Z) | X) = E (Y | X) Conditioning Theorem. For example. Thus h(X) = g(X)m(X). we deﬁne the conditional variance of Y given X = x as σ 2 (x) = V ar (Y | X = x) ´ ³ = E (Y − m(x))2 | X = x ¢ ¡ = E Y 2 | X = x − m(x)2 . Similarly. −∞ −∞ (A. meaning that when X equals x.8) (A. For any function g(x). x) dydx = E(Y ). m(x) = E (Y | X = x) = −∞ The conditional mean m(x) is a function. We write this as E(Y | X) = m(X) and V ar (Y | X) = σ 2 (X). Simple Law of Iterated Expectations: E (E (Y | X)) = E (Y ) Proof : E (E (Y | X)) = E (m(X)) Z ∞ m(x)fX (x)dx = Z−∞ Z ∞ ∞ = yfY |X (y | x) fX (x)dydx −∞ −∞ Z ∞Z ∞ = yf (y.7) Law of Iterated Expectations: E (E (Y | X. then the expected value of Y is m(x). E (g(X)Y | X) = g (X) E (Y | X) Proof : Let h(x) = E (g(X)Y | X = x) Z ∞ g(x)yfY |X (y | x) dy = −∞ Z ∞ = g(x) yfY |X (y | x) dy = g(x)m(x) −∞ (A. which is the same as E (g(X)Y | X) = g (X) E (Y | X) . The following are important facts about conditional expectations. 174 . the conditional mean m(X) and conditional variance σ 2 (x) are random variables.

1] and Y = − ln(X). Let Y = g(X) where g(x) : Rk → Rk is one-to-one. g(x) = − ln(x) and h(y) = exp(−y) so the Jacobian is J(y) = − exp(y). If g(x) is an increasing function. Let h(y) denote the inverse of g(x). an exponential density. take the case X ∼ U [0.10) shows that fY (y) = exp(−y). This same formula (A. J(y) = det ∂y 0 Consider the univariate case k = 1. but its justiﬁcation requires deeper results from analysis. and invertible. φ(x) = √ exp − 2 2π 175 The standard normal density is −∞ < x < ∞. then g(X) ≤ Y if and only if X ≥ h(Y ). diﬀerentiable. The Jacobian is ¶ µ ∂ h(y) . Here.7 Transformations Suppose that X ∈ Rk with continuous distribution function FX (x) and density fX (x). so FY (y) = P (g(X) ≤ y) = 1 − P (X ≥ h(Y )) = 1 − FX (h(Y )) and the density of Y is fY (y) = −fX (h(Y )) We can write these two cases jointly as fY (y) = fX (h(Y )) |J(y)| . (A. then g(X) ≤ Y if and only if X ≤ h(Y ). A. As the range of X is [0. 0 ≤ y ≤ ∞. . Since fX (x) = 1 for 0 ≤ x ≤ 1 (A. so the distribution function of Y is FY (y) = P (g(X) ≤ y) = FX (h(Y )) = P (X ≤ h(Y )) so the density of Y is fY (y) = d d FY (y) = fX (h(Y )) h(y).A. As one example. 1]. that for Y is [0. dy dy If g(x) is a decreasing function.∞).10) d h(y).8 Normal and Related Distributions µ 2¶ 1 x .10) holds for k > 1. dy This is known as the change-of-variables formula.

k/2 1/2 2 (2π) det (ΣY ) where µY = a+Bµ and ΣY = BΣB 0 . This shows that if X is multivariate normal with uncorrelated elements. In this case the density function can be written as j Ã Ã !! (x1 − µ1 )2 /σ 2 + · · · + (xk − µk )2 /σ 2 1 1 k f (x) = exp − 2 (2π)k/2 σ 1 · · · σ k Ã ¡ ¢2 ! k Y xj − µj 1 = exp − 2σ 2 (2π)1/2 σ j j j=1 which is the product of marginal univariate normal densities. Σ) and Y = a + BX with B an invertible matrix. the multivariate normal density is ¶ µ (x − µ)0 Σ−1 (x − µ) 1 . f (x) = √ 2σ 2 2πσ which is the univariate normal density. exp − f (x) = k/2 1/2 2 (2π) det (Σ) The mean and covariance matrix of the distribution are µ and Σ. x ∈ Rk . Theorem A. It is conventional to write X ∼ N (0. and it is conventional to write X ∼ N(µ. q × q. (A. Q has the density f (y) = Γ 1 ¡r¢ 2 2r/2 yr/2−1 exp (−y/2) . then they are mutually independent. The mean and variance of the distribution are µ and σ 2 . This shows that linear transformations of normals are also normal. then Z 0 A−1 Z ∼ χ2 . then using the change-of-variables formula. and to denote the standard normal density function by φ(x) and its distribution function by Φ(x). For x ∈ Rk . Σ). By iterated integration by parts. Its mean and r variance are µ = r and σ 2 = 2r.1 Let X ∼ N (0. denoted χ2 . and it is conventional to write X ∼ N (µ. A) with A > 0. X has density Ã ! (x − µ)2 1 exp − . then Σ = diag{σ 2 } is a diagonal matrix. Theorem A. ¢ ¡ It useful to observe ¢ that the MGF and CF of the multivariate normal are exp λ0 µ + λ0 Σλ/2 ¡ 0 and exp iλ µ − λ0 Σλ/2 .8. respectively.This density has all moments ﬁnite. y ≥ 0. where we used the fact that det (BΣB 0 )1/2 = det (Σ)1/2 det (B) . q 176 . x ∈ Rk . then by the change-of-variables formula. Another useful fact is that if X ∼ N (µ. σ 2 ). we can also show that EX 2 = 1 and EX 4 = 3. Ir ) and set Q = X 0 X. If Z is standard normal and X = µ + σZ. 1). −∞ < x < ∞. The latter has no closed-form solution.2 If Z ∼ N(0. the density of Y is Ã ! (y − µY )0 Σ−1 (y − µY ) 1 Y exp − f (y) = . If X ∈ Rk is multivariate normal and the elements of X are mutually uncorrelated. Since it is symmetric about zero all odd moments are zero.8.11) and is known as the chi-square density with r degrees of freedom.

C −1 AC −10 ) = N (0. 1) and Q ∼ χ2 be independent. tr = p Q/r The density of tr is and is known as the student’s t distribution with r degrees of freedom.3.8.8.11) 2 with r = 1.8.6) can be used to show that the MGF j=1 ¥ Proof of Theorem A. 1). (A. 1) the distribution of Z 2 is ¡ ¢ √ P Z 2 ≤ y = 2P (0 ≤ Z ≤ y) µ 2¶ Z √y 1 x √ exp − dx = 2 2 2π 0 Z y ³ s´ 1 ¡1¢ = s−1/2 exp − ds 1/2 2 0 Γ 2 2 1 ¡r¢ 2 2r/2 y r/2−1 exp (ty) exp (−y/2) dy of Q is (1 − 2t)−r/2 .1. which we showed in (A.12) E exp (tQ) = 0 Γ (A. For Z ∼ N(0. which can be found by applying change-of-variables to the gamma function. Then A−1 = C −10 C −1 and C −1 Z ∼ N (0.13).Theorem A.13) is the MGF of the density (A.13) = (1 − 2t)−r/2 R∞ where the second equality uses the fact that 0 y a−1 exp (−by) dy = b−a Γ(a). Using the simple law of iterated expectations. tr has distribution 177 .8.11) is Z ∞ ¢ ¡ Γ r+1 2 f (x) = ´ r+1 ¡r¢ ³ √ 2 2 πrΓ 2 1 + x r (A. Set r Z . C −1 CC 0 C −10 ) = N (0. q Proof of Theorem A.11). we see that the MGF of Z 2 is (1 − 2t)−1/2 . From (A.3 Let Z ∼ N (0. The MGF for the density (A. Proof of Theorem A. Iq ). Since we can write Q = P 2 X 0 X = r Zj where the Zj are independent N(0. Thus the density of Z 2 is (A. ¡ ¢ √ using the change—of-variables s = x2 and the fact Γ 1 = π. The fact that A > 0 means that we can write A = CC 0 where C is non-singular.2. Thus ¥ ¡ ¢0 ¡ ¢ Z 0 A−1 Z = Z 0 C −10 C −1 Z = C −1 Z C −1 Z ∼ χ2 .

1+ √ r r rπΓ 2 ! Z P p ≤x Q/r ( r ) Q E Z≤x r !# " Ã r Q |Q E P Z≤x r Ã r ! Q EΦ x r Ã = ¥ A.. θ) = P (Y = y.9 Maximum Likelihood If the distribution of Yi is F (y. θ) . Yn ) is n ³ ´ Y ˜. viewed as a function of θ. If the distribution F is continuous then the density of Yi can be written as f (y. the likelihood and log-likelihood are constructed by setting f (y. The space Θ is the set of permissible value for θ. 178 .. i=1 The likelihood of the sample is this joint density evaluated at the observed sample values. In this setting the method of maximum likelihood is the appropriate technique for estimation and inference on θ. θ) . The log-likelihood function is its natural log Ln (θ) = n X i=1 ln f (Yi . If the distribution F is discrete. we say that the distribution is parametric and that θ is the parameter of the distribution F.θ = fn Y f (Yi . θ) where F is a known distribution function and θ ∈ Θ is an unknown m × 1 vector. .function F (x) = = = = Thus its density is f (x) = = = Ã r ! Q d E Φ x dx r Ã Ã r !r ! Q Q E φ x r r ! µ ¶¶ r Ã Z ∞µ qx2 1 q 1 r/2−1 ¡r¢ √ exp − q exp (−q/2) dq 2r r Γ 2 2r/2 2π 0 ¢ µ ¡ ¶− r+1 Γ r+1 x2 ( 2 ) 2¡ ¢ . θ) and the joint ˜ density of a random sample Y = (Y1 .. θ) .

ˆ is consistent θ for this value. θ)¯ =0 ¯ ∂θ θ=θ0 H = Ω ≡ I0 (A. under conventional regularity conditions.14) ¶ ∂ ∂ 0 ln f (Yi .9. θ) ≡ L(θ). θ) The Cramer-Rao Theorem gives a lower bound for estimation.15) Two important features of the likelihood are Theorem A. the log-likelihood is a sample average 1X 1 Ln (θ) = ln f (Yi .3 Under regularity conditions. the restriction to unbiased estimators means that the theorem has little direct relevance for ﬁnite sample eﬃciency. θ0 ) . The ﬁrst-order condition for the latter problem is 0= ∂ ∂ L(θ) = E ln f (Yi . θ) ∂θ ∂θ which holds at θ = θ0 by (A. ˆ →p θ0 as n → ∞. The maximum likelihood estimator or MLE ˆ is the parameter value which maximizes θ the likelihood (equivalently.17) The matrix I0 is called the information. then θ V ar(˜ ≥ (nI0 )−1 .17) is often called the information matrix equality. ∂θ ∂θ (A. θ θ∈Θ ˆ In some simple cases.Deﬁne the Hessian H = −E and the outer product matrix Ω=E µ ∂2 ln f (Yi . However.16) (A. the MLE θ Why do we believe that the MLE ˆ is estimating the parameter θ? Observe that when stanθ dardized. we can see that it is an estimator of the maximizer θ of the right-hand-side.1 ¯ ¯ ∂ E ln f (Yi . Theorem A. More typically. 179 . ´ ¡ −1 ¢ √ ³ˆ n θ − θ0 →d N 0. but these ˆ must be found by numerical methods. we can ﬁnd an explicit expression for θ as a function of the data. We can write this as ˆ = argmax Ln (θ). θ) →p E ln f (Yi .9. which maximizes the log-likelihood). n n i=1 n As the MLE ˆ maximizes the left-hand-side.2 Cramer-Rao Lower Bound. θ0 ) ∂θ∂θ 0 (A. cases are rare. If ˜ is an unbiased estimator of θ ∈ R. and the equality (A. θ0 ) ln f (Yi .16). I0 . In fact.9. θ Theorem A.

Thus in large samples the MLE has approximately the best possible variance. the approximate variance of the MLE is (nI0 )−1 which is the CramerRao lower bound. θ) θ In this case.14) n ³ ´ 1 X ∂2 ˆ ˆ ln f Yi . θ) dy which is the same as the minimizer of KLIC = Z f (y) ln µ f (y) f (y. A minor adjustment to Theorem (A. ˆ ln f Yi . θ). ˆ .18) H =− n ∂θ∂θ0 i=1 the Kullback-Leibler information distance between the true density f (y) and the parametric density f (y. θ V = H −1 ΩH −1 The moment estimator for V is ˆ where H is given in (A. θ θ Ω= n ∂θ ∂θ n i=1 ˆ ˆ ˆˆ V = H −1 ΩH −1 Asymptotic standard errors (sometimes called qmle standard errors) are then the square roots of ˆ the diagonal elements of n−1 V . but it is not literally believed that the model f (y. since the information matrix equality (A. θ) is necessarily the true density. Thus the QMLE θ0 is the value which makes the parametric density “closest” to the true value according to this measure of distance.18) and ³ ³ ´ ∂ ´0 1X ∂ ˆ ln f Yi .9. we refer to Ln (ˆ as a quasi-likelihood and the its maximizer ˆ as a quasi-mle or QMLE. Asymptotic standard errors for ˆ are then the square roots of the diagonal −1 I −1 . In this case there is not a “true” value of the parameter θ. V ) . Typically. θ) ¶ dy Thus in large samples. but has a diﬀerent covariance matrix than in the pure MLE case. to estimate the asymptotic variance of the MLE we use an estimate based on the Hessian formula (A. θ) is used to approximate the true unknown density f (y).ˆ ˆ−1 θ We then set I0 = H −1 .17) does not hold. The QMLE is consistent for the pseudo-true value. ˆ elements of n 0 Sometimes a parametric density function f (y. 180 . θ (A. θ) = f (y) ln f (y. Therefore the MLE is called asymptotically eﬃcient. Instead we deﬁne the pseudo-true value θ0 as the maximizer of Z E ln f (Yi .3) yields the asymptotic distribution of the QMLE: ´ √ ³ n ˆ − θ0 →d N (0.

θ0 ) ∂ ∂θ f ∂ (Yi . (Yi . θ0 ) d˜ = E ˜ . θ) d˜ θ = Eθ θ(˜ y y where y = (y1 . θ0 ) ln f (Yi .16). θ0 ) ∂θ ∂θ ¥ ³ ´ X ∂ ∂ ˜ ln fn Y .. θ0 ) ∂ ∂ − ln f (Yi . function of the data.9. ∂θ ¯θ=θ0 Z ! = 0. θ0 )0 . θ0 = ln f (Yi . θ0 ) − f (Yi . θ0 ) f (Yi . yn ). θ0 ) ¯ ∂ f (y. y y θ(˜ y y y θS 1 = θ(˜ ∂θ ∂θ By the Cauchy-Schwarz inequality ¯ ³ ´¯2 ³ ´ ¯ ¯ θS θ 1 = ¯E ˜ ¯ ≤ V ar (S) V ar ˜ ³ ´ V ar ˜ ≥ θ 1 1 = . θ0 ) (Yi .17). θ)¯ = ¯ ∂θ θ=θ0 = = = Similarly. θ0 ) = ∂θ∂θ0 = Taking expectations yields (A. S= ∂2 f ∂θ∂θ0 ∂2 f ∂θ∂θ0 ∂ ∂θ Z Ã ∂2 f ∂θ∂θ0 ¯ ¯ ln f (y. θ0 ) ∂θ ∂θ n i=1 which by Theorem (A. θ) d˜ = ˜ y ) ∂ ln f (˜.. Write the estimator ˜ = ˜ Y ) as a θ θ( ˜ ˜ is unbiased for any θ. ¯ ¯ ∂ E ln f (Yi . Diﬀerentiating with respect to θ and evaluating at θ0 yields ˜ Z Z ³ ´ ˜ y ) ∂ f (˜.2.. θ) f (˜. ∂2 ln f (Yi . θ0 ) ∂θ f (Yi . f (Yi . θ0 ) dy ¯ ¯ ¯ θ=θ0 f (y. θ) dy ¯ ¯ ∂θ θ=θ0 ¯ ∂ ¯ 1¯ = 0. θ0 )2 (Yi . Proof of Theorem A.Proof of Theorem A. θ) dy ¯ ∂θ f (y. θ) f (y. V ar (S) nH so ¥ 181 . θ0 )0 f (Yi . To see (A.9.1) has mean zero and variance nH. . Since θ Z ˜ = ˜ y)f (˜.1.9. θ) ¯θ=θ0 ¯ Z ¯ ∂ f (y. we can show that E By direction computation.

θ0 ) + ' 0 ln f (Yi .Proof of Theorem A. H −1 ΩH −1 = N 0. an application of the CLT yields 1 X ∂ √ ln f (Yi .9. θn ) = ln f (Yi . θ0 ) .3 Taking the ﬁrst-order condition for maximization of Ln (θ). θn ) − H(θn ) + H(θn ) − n n ∂θ∂θ0 ∂θ∂θ0 i=1 i=1 →p H by an application of a uniform WLLN. θ0 ) →d N (0. ¥ 182 . Let ∂2 H(θ) = − ∂θ∂θ0 ln f (Yi . θ) . θ n ) θ − θ 0 .) Rewriting this equation. θ0 ) is mean-zero with covariance matrix Ω. the ﬁnal equality using Theorem A. ˆ = θ ∂θ i=1 n n ´ ³ X ∂ X ∂2 ˆ ln f (Yi . − ln f (Yi . Ω) = N 0. ¯ ¯ ∂ Ln (θ)¯ 0 = ¯ ˆ ∂θ θ=θ n ³ ´ X ∂ ln f Yi . and making a ﬁrst-order Taylor series expansion.9. Ω) . ∂θ n i=1 n The analysis of the sample Hessian is somewhat more complicated due to the presence of θn . θn ) ∂θ ∂θ∂θ0 i=1 i=1 ln f (Yi . H −1 . Together.1 . the speciﬁc value of θn varies by row in this expansion. If it is continuous in θ. ´ ¡ ¢ ¡ ¢ √ ³ˆ n θ − θ0 →d H −1 N (0. (Technically. ∂θ ∂θ∂θ i=1 i=1 θ where θn lies on a line segment joining ˆ and θ0 . we ﬁnd ´ ³ ˆ − θ0 = θ Since ∂ ∂θ Ã !−1 Ã n ! n X ∂2 X ∂ ln f (Yi . then since θn →p θ0 we ﬁnd H(θn ) →p H and so ¶ n n µ ∂2 1X 1 X ∂2 − ln f (Yi .

a line search). b] with G gridpoints. 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)/ε to 2 (b − a)/ε which can be substantial. In most cases. the approximation error is bounded by ε. At each point evaluate the criterion function and ﬁnd the gridpoint which yields the smallest value of the criterion.. MLE and GMM estimators take this form.. The estimate of ˆ is j 0 ˆ θ (k). where j = argmin0≤j≤G Q(θ(j)). Grid Search. . Construct an equally spaced grid on the region [a. numerical methods are required to θ obtain ˆ θ.Appendix B Numerical Optimization Many econometric estimators are deﬁned by an optimization problem of the form ˆ = argmin Q(θ) θ θ∈Θ (B. the ﬁrst-step grid may not bracket ˆ θ which thus would be missed. b] with G gridpoints. Q(θ) can be computed for given θ. which is θ(ˆ) j j θ. 183 . which is {θ(j) = a + j(b − a)/G : j = 0.. but ˆ is not available in closed form. G}. G}. Plots of Q(θ(j)) against θ(j) can help diagnose errors in grid selection. At each point evaluate the criterion function and let j = argmin0≤j≤G Q(θ(j)). The disadvantage is that if the function Q(θ) is irregular.. In this context grid search may be employed. For example NLLS. Two-Step Grid Search. In this case. GLS. Let k = argmin0≤k≤G Q(θ0 (k)). For an error bound of ε pick G so that G2 = (b − a)/ε For the ﬁrst step deﬁne an equally spaced grid on the region [a. b] be an interval. which is {θ0 (k) = j j ˆ θ θ(ˆ − 1) + 2k(b − a)/G2 : k = 0. Pick some ε > 0 and set G = (b − a)/ε to be the number of gridpoints. ¯θ(ˆ) − θ¯ method is quite robust but potentially costly. For the second ˆ step deﬁne an equally spaced grid on [θ(ˆ − 1).. θ(ˆ + 1)] with G gridpoints. . . B. The gridsearch method can be reﬁned by a two-step execution.1) where the parameter is θ ∈ Θ ⊂ Rm and the criterion function is Q(θ) : Θ → R.. which is {θ(j) = a + j(b − a)/G : j = 0. to ¯ ¯ ˆ¯ ≤ ε.1 Grid Search Many optimization problems are either one dimensional (m = 1) or involve one-dimensional optimization as a sub-problem (for example.. This value θ(ˆ) is the gridpoint estimate of ˆ If the grid is ˆ suﬃciently ﬁne ¯ capture small oscillations in Q(θ). G}... Let Θ = [a. This j that is.

they can be calculated numerically. To pick λi write the criterion function as a function of λ Q(λ) = Q(θi + Di gi λ) 184 .B. Then for ε small gj (θ) ' Similarly. however. numerical derivatives can be quite unstable. Let δ j be the j 0 th unit vector (zeros everywhere except for a one in the j 0 th row). Another productive modiﬁcation is to add a scalar steplength λi . a one-dimensional optimization. Allowing the steplength to be a free parameter allows for a line search. One modiﬁcation to prevent this possibility is quadratic hill-climbing which sets ˆi+1 = θi − (H(θi ) + αi Im )−1 g(θi ). ε where gi = g(θi ) and Di = H(θi )−1 for Newton’s method and Di = (H(θi ) + αi Im )−1 for quadratic hill-climbing. numerical derivatives can work well but can be computationally costly relative to analytic derivatives. Take the j 0 th element of g(θ). By a Taylor’s expansion for θ close to ˆ θ ³ ´ 0 = g(ˆ ' g(θ) + H(θ) ˆ − θ θ) θ gjk (θ) ' which implies ˆ = θ − H(θ)−1 g(θ). which are designed to converge to ˆ All require the choice of a starting value θ1 . θ where αi is set just above the smallest eigenvalue of H(θi ) if H(θ) is not positive deﬁnite. ∂θ∂θ 0 If the functions g(θ) and H(θ) are not analytically available. and all require the computation θ. Q(θ + δ j ε + δ k ε) − Q(θ + δ k ε) − Q(θ + δ j ε) + Q(θ) ε2 In many cases.2) θi+1 = θi − Di gi λi Q(θ + δ j ε) − Q(θ) .. of the gradient of Q(θ) ∂ Q(θ) g(θ) = ∂θ and some require the Hessian ∂2 H(θ) = Q(θ).. In some cases. Most gradient methods are a variant of Newton’s method which is based on a quadratic approximation. θ where One problem with Newton’s method is that it will send the iterations in the wrong direction if H(θi ) is not positive deﬁnite.2 Gradient Methods Gradient Methods are iterative methods which produce a sequence θi : i = 1. 2. . θ This suggests the iteration rule ˆi+1 = θi − H(θi )−1 g(θi ). In this case the iteration rule takes the form (B.

a random variable is drawn and added to the current value of the parameter. One example is the simplex method of Nelder-Mead (1965). The half-step method considers the sequence λ = 1. . and Rodgers (1994). These problems have motivated alternative choices for the weight matrix Di . If the criterion has improved over Q(θi ). If the resulting criterion is decreased. This can be deﬁned by examining whether |θi − θi−1 | . Ferrier. Newton’s method does not perform well if Q(θ) is irregular. this new value is accepted. it relies on an iterative sequence. use this value. A quadratic approximation evaluates the ﬁrst and second derivatives of Q(λ) with respect to λ. There are two common methods to perform a line search. 185 . As the iterations continue. In these cases. 0 0 ∆gi Di−1 ∆gi ∆θi ∆gi For any of the gradient methods. At each iteration. A more recent innovation is the method of simulated annealing (SA). .a one-dimensional optimization problem. and it can be quite computationally costly if H(θ) is not analytically available. |Q (θi ) − Q (θi−1 )| or |g(θi )| has become small. the variance of the random innovations is shrunk. it may still be accepted depending on the extent of the increase and another randomization. Like the gradient methods. The SA method has been found to be successful at locating global minima. For a review see Goﬀe. the iterations continue until the sequence has converged in some sense. The DFP method sets Di = Di−1 + The BFGS methods sets Di = Di−1 + 0 ∆θi ∆θ0 ∆θi ∆θ0 ∆θi ∆gi Di−1 Di−1 ∆gi ∆θ0 i i i 0 −¡ ∆gi Di−1 ∆gi + + . The SA algorithm stops when a large number of iterations is unable to improve the criterion... These methods are called Quasi-Newton methods. Let ∆gi = gi − gi−1 ∆θi = θi − θi−1 and . The latter property is needed to keep the algorithm from selecting a local minimum. otherwise move to the next element in the sequence. The SA method is a sophisticated random search.3 Derivative-Free Methods All gradient methods can be quite poor in locating the global minimum when Q(θ) has several local minima. If the criterion is increased. alternative optimization methods are required. ¢2 0 ∆θ0 ∆gi ∆θ0 ∆gi ∆θ0 ∆gi ∆θi ∆gi i i i 0 ∆θi ∆θ0 Di−1 ∆gi ∆gi Di−1 i + . The downside is that it can take considerable computer time to execute. 1/4. Each value in the sequence is considered and the criterion Q(θi + Di gi λ) evaluated. Two popular methods are do to Davidson-Fletcher-Powell (DFP) and Broyden-Fletcher-Goldfarb-Shanno (BFGS). 1/2. 1/8. B. Furthermore. the methods are not well deﬁned when Q(θ) is non-diﬀerentiable. and picks λi as the value minimizing this approximation.

T. 1383-1414. B. D. 68.S. 239-253. A. and improved inference . (1994): “Alternative approximations to the distributions of instrumental variable estimators.W. Petroc and F. (1991).” Annals of Statistics. H. (1993). [10] Bekker.W.” Review of Economic Studies.” Journal of Business and Economic Statistics. 16. New York: Wiley. L. D. [9] Basmann. [7] Andrews.K. (1968): Convergence of Probability Measures. eﬃcient bootstrapping. 23-51. and H. and W. 20.. 59. [6] Andrews. 62.” Proceedings of the Royal Statistical Society. 821-8516.” In B. Rubin (1949): “Estimation of the parameters of a single equation in a complete system of stochastic equations. (1988): “Laws of large numbers for dependent non-identically distributed random variables.W. [3] Anderson.W.W. 458-467. 62. [15] Brown. (1935): “On least squares and linear combinations of observations. “Tests for parameter instability and structural change with unknown change point. P. Econometrica.” The Annals of Mathematical Statistics.K.Bibliography [1] Aitken. 47. T. Csake. P. 1709-1722.” Econometrica. and M.W. 307-345. D.K.K. [4] Andrews. A.” Econometrica. 4.W. “Asymptotic normality of series estimators for nonparameric and semiparametric regression models.K. P. 61. [5] Andrews. 42-48. [13] Bose. [2] Akaike. 77-83. Buchinsky: (2000): “A three-step method for choosing the number of bootstrap replications. Newey (2002): “GMM.” Econometrica.” Econometrica.K. [11] Billingsley. [8] Andrews.R.C.” Econometrica. D. D. (1979): Probability and Measure. [12] Billingsley.’ Econometric Theory. (1973): “Information theory and an extension of the maximum likelihood principle. R. New York: Wiley. and A. [14] Breusch. Pagan (1979): “The Lagrange multiplier test and its application to model speciﬁcation in econometrics. Ploberger (1994): “Optimal tests when a nuisance parameter is present only under the alternative. eds. 46-63. 55. and W. Second International Symposium on Information Theory. (1988): “Edgeworth correction by bootstrap in autoregressions. 25. 657-681.A. (1957): “A generalized classical method of linear estimation of coeﬃcients in a structural equation. 186 .

363-390.W. G.[16] Carlstein. 74. 14.A. [33] Goldberger. 187 . Newey (2001): “Choosing the number of instruments.” Journal of Econometrics.” Journal of Econometrics.A.” Econometrica. 55.V. Tibshirani (1993): An Introduction to the Bootstrap. [17] Chamberlain.R. G.F.” Econometrica. 34. estimation and testing. 1-26. (1991): A Course in Econometrics. [28] Eicker. 305-334. 7. and D.” Econometrica.” Annals of Mathematical Statistics. 51.B. Hinkley (1997): Bootstrap Methods and their Application. D. [31] Gallant. [24] Dufour. and D. (1963): “Asymptotic normality and consistency of the least squares estimators for families of linear regressions.” Econometrica. New York: Chapman-Hall. 113-150. 251-276. 591-603. 427-431. Waugh (1933): “Partial time regressions as compared with individual trends.” Econometrica. Bradley (1979): “Bootstrap methods: Another look at the jackknife. (1986): “The use of subseries methods for estimating the variance of a general statistic from a stationary time series. Arthur S.M. and H.” Journal of the American Statistical Association.J. 1. F. Oxford: Oxford University Press. Cambridge: Harvard University Press. Bradley and R. J. J. 387-401.C. 1161-1191. 447-456. (1997): “Some impossibility theorems in econometrics with applications to structural and dynamic models.C. R. A.C. and P. [25] Efron. [30] Frisch. White (1988): A Uniﬁed Theory of Estimation and Inference for Nonlinear Dynamic Models.F. (1987): “Asymptotic eﬃciency in estimation with conditional moment restrictions. and W. [26] Efron. and F. Society for Industrial and Applied Mathematics. [27] Efron. [29] Engle. (1960): “Tests of equality between sets of coeﬃcients in two linear regressions. and C.W. Bradley (1982): The Jackknife. 34. [23] Donald S. A. E. I. [19] Chow. Cambridge University Press.” Annals of Statistics. and Other Resampling Plans. (1994): Stochastic Limit Theory: An Introduction for Econometricians.” Annals of Statistics. 65. [32] Gallant. A. [21] Davison.K.G. 55. the Bootstrap. [22] Dickey. 69. Nychka (1987): “Seminonparametric maximum likelihood estimation.J. Phillips (1992): “Asymptotic and ﬁnite sample distribution theory for IV estimators and tests in partially identiﬁed structural equations.” Econometrica. 28. R. Granger (1987): “Co-integration and error correction: Representation. New York: Basil Blackwell. Fuller (1979): “Distribution of the estimators for autoregressive time series with a unit root. and W. 1365-1387. [18] Choi. 1171-1179. [20] Davidson.

homoskedasticity and serial independence of regression residuals. and J. 413-430. 153161. 68.[34] Goﬀe. (1996): “A note on bootstrapping generalized method of moments estimators. J. Econometrica. Teräsvirta (1993): Modelling Nonlinear Economic Relationships. P. Vol. 64. [35] Gauss. G. IV. eds. G.. 64. (1996): “Inference when a nuisance parameter is not identiﬁed under the null hypothesis.” Econometrica. L.” Econometrica. and T.” Journal of Business and Economic Statistics. Economic Letters. G.L. 65-99. R. 262-280. (1982): “Large sample properties of generalized method of moments estimators. (1994): “Methodology and theory for the bootstrap. [42] Hall. [45] Hansen.” Econometrica.F. New York: Elsevier Science. A. [39] Hall. [49] Heckman.W.” Review of Economic Studies. 359-383. (1978): “Speciﬁcation tests in econometrics.L. Vol.P. Heaton.” in Werke. 50.J. [40] Hall.” Econometrica. Oxford. 187-197.” Econometrica.E.” Journal of Econometrics.M. [46] Hansen. (1969): “Investigating causal relations by econometric models and crossspectral methods. Bera (1980): “Eﬃcient tests for normality. P. P.J. 66. 240-254. R. [38] Granger. and A.F.W.J. 1517-1527. (1997): “One step estimators for over-identiﬁed generalized method of moments models. Ferrier and J.E.” Econometric Theory. 12. (1992): “Eﬃcient estimation and testing of cointegrating vectors in the presence of deterministic trends. [43] Hahn. L. [47] Hansen. [51] Imbens.” Handbook of Econometrics.H. 891-916.K. (1809): “Theoria motus corporum coelestium. VII. 14. B. J.” Journal of Econometrics. 188 . 255-259.D. K.W. and A. Rogers (1994): “Global optimization of statistical functions with simulated annealing. Yaron (1996): “Finite sample properties of some alternative GMM estimators. J.W. Engle and D.” Journal of Econometrics. 1251-1271. [36] Granger. New York: Springer-Verlag. [52] Jarque. Oxford University Press.A. [41] Hall.. 64.L. W. 16.. 47. Horowitz (1996): “Bootstrap critical values for tests based on GeneralizedMethod-of-Moments estimation. 53. 6. Spady and P. B. [37] Granger. (1979): “Sample selection bias as a speciﬁcation error. R. 424-438.” Econometrica. (1992): The Bootstrap and Edgeworth Expansion. [50] Imbens. 46. [48] Hausman. (1981): “Some properties of time series data and their use in econometric speciﬁcation. C.P.” Econometrica. C. (2000): “Covariance matrix estimation and the power of the overidentifying restrictions test. C. [44] Hansen. 121-130. 333-357. J. McFadden. 1029-1054. Johnson (1998): “Information theoretic approaches to inference in moment condition models. 37.W. C. 87-121. 60.

” Annals of Statistics. 58. 237-249. [68] Nelder. and K.W. 305-325. (1988): “Statistical analysis of cointegrating vectors. J. [67] Muirhead.. [66] Magnus. H. 1217-1241.J. P.” Journal of Economic Dynamics and Control. Oxford University Press. 21. [54] Johansen. 65.J.C. 308-313. M. [69] Newey. 1551-1580.” Journal of Econometrics. Juselius (1992): “Testing structural hypotheses in a multivariate cointegration analysis of the PPP and the UIP for the UK. and C.R. Y. (1963): “Seasonal adjustment of economic time series. 28. 12. 231-254. 17. Cambridge University Press.[53] Johansen. and H. Neudecker (1988): Matrix Diﬀerential Calculus with Applications in Statistics and Econometrics. New York: John Wiley and Sons. 211-244. J.” Journal of the American Statistical Association.” in Engle. S. R. Oxford.” Econometrica. 35-40.” Econometrica..G. S. P.” Journal of Econometrics. [64] MacKinnon. and H. and M.” Computer Journal. 59. (2001): “Asymptotic optimality and empirical likelihood for testing moment restrictions. [58] Kitamura. F.K. West (1987): “Hypothesis testing with eﬃcient method of moments estimation. (1989): “The jackknife and the bootstrap for general stationary observations. New York: Wiley. 861-874.” International Economic Review.D. S. (1991): “Estimation and hypothesis testing of cointegration vectors in the presence of linear trend.) Long-Run Economic Relationships: Readings in Cointegration. Schmidt. [57] Kitamura. [63] Lovell. and Y. Roger (2005): Quantile Regression. Y. 1661-1672. Oxford University Press. [59] Koenker. Phillips. (1982): Aspects of Multivariate Statistical Theory.” Econometrica. W.” Economics Letters. [61] Kwiatkowski. 993-1010. [65] MacKinnon. R. J. [55] Johansen.B. White (1985): “Some heteroskedasticity-consistent covariance matrix estimators with improved ﬁnite sample properties. Art B. [62] Lafontaine. 75. (1988): “Empirical likelihood ratio conﬁdence intervals for a single functional. 7. [56] Johansen. Granger (eds. (1990): “Critical values for cointegration. and K.G. and K. [70] Owen.. and R.” Biometrika. Stutzer (1997): “An information-theoretic alternative to generalized method of moments. 54. [60] Kunsch.F. 159-178. Shin (1992): “Testing the null hypothesis of stationarity against the alternative of a unit root: How sure are we that economic time series have a unit root?” Journal of Econometrics.C. (1995): Likelihood-Based Inference in Cointegrated Vector Auto-Regressive Models. Mead (1965): “A simplex method for function minimization. R. 69. 53. 29. S. J. 189 . 777-787. White (1986): “Obtaining any Wald statistic you want. D.

62. 71. B. S.H. [76] Qin. (1986): Density Estimation for Statistics and Data Analysis. C. 1303-1313. P.C. J. 163-185. New York: McGraw-Hill.” Econometrica. (1994): “A residual-based test of the null of cointegration against the alternative of no cointegration.C.” Journal of the American Statistical Association.H. (1980): “Macroeconomics and reality. J. 435-460. Series B.D. B. J. [83] Shin. and S. 68. (1987): “Asymptotic properties of least squares estimators of cointegrating vectors.” The Annals of Statistics.H. 1055-1096. 48. 557-586. 190 . and D.A. [89] Stock. 89. and J. (2001): Empirical Likelihood. Stock (1997): “Instrumental variables regression with weak instruments. London: Chapman and Hall. Y. (1991): “Conﬁdence intervals for the largest autoregressive root in U. [77] Ramsey.B. J. (1987): Real and Complex Analysis.” Econometrica. macroeconomic time series. [86] Sims. and J.” Econometrica. 58.E.[71] Owen.” Econometrica. 91-115.” American Economic Review. D.B. [81] Sargan. Journal of the Royal Statistical Society. 599-608. J. 22. 28. [87] Staiger. [74] Politis. NY: Springer. S. (1972): “Money. [90] Stock. [84] Silverman.” Econometric Theory. 540-552. and J.” Econometric Theory.” Journal of Monetary Economics. J.M.W.S. Dickey (1984): “Testing for unit roots in autoregressive-moving average models of unknown order. 31. 181-240.H. [78] Rudin. Series B. D. 165-193. W. 2 6. [88] Stock.P. B. and M. 65.” Econometrica. income and causality. 7.J. 10. 350-371. [73] Phillips. [80] Shao. 393-415. 5. (1989): “Partially identiﬁed econometric models. New York: Chapman & Hall.A.H. 55. and D.C. 1035-1056. [72] Phillips.” Biometrika.A. [79] Said. [85] Sims.” Econometric Theory. 683-690.” Econometrica. (1969): “Tests for speciﬁcation errors in classical linear least-squares regression analysis. and J. Tu (1995): The Jackknife and Bootstrap.N. 3rd edition. J. 300-325. 1-48. Lawless (1994): “Empirical likelihood and general estimating equations. C. P.” Journal of the Royal Statistical Society. Jones (1991): “A reliable data-based bandwidth selection method for kernel density estimation. (1991): “Eﬀects of model selection on inference. [75] Potscher. 53. (1958): “The estimation of economic relationships using instrumental variables. [82] Sheather. Wright (2000): “GMM with weak identiﬁcation. Romano (1996): “The stationary bootstrap. Ouliaris (1990): “Asymptotic properties of residual based tests for cointegration. Art B.

[92] Tobin. [94] Wang. [97] Zellner. 1389-1404. Zivot (1998): “Inference on structural parameters in instrumental variables regression with weak instruments. (1984): Asymptotic Theory for Econometricians. H.” Econometrica.” Econometrica.” Econometrica. 57. (1958): “Estimation of relationships for limited dependent variables. [96] White. A. Academic Press. 66. J. 2 6. 817-838. (1962): “An eﬃcient method of estimating seemingly unrelated regressions. and tests for aggregation bias. (1980): “A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity. 48. [95] White. 24-36.” The Hague. 191 . and E. (1953): “Repeated least squares applied to complete equation systems.” Journal of the American Statistical Association. A. [93] Wald. 348-368. (1943): “Tests of statistical hypotheses concerning several parameters when the number of observations is large. Central Planning Bureau. H. mimeo. J. 54.” Transactions of the American Mathematical Society. 426-482. H.[91] Theil.

- Read and print without ads
- Download to keep your version
- Edit, email or read offline

Are you sure?

This action might not be possible to undo. Are you sure you want to continue?

CANCEL

OK

You've been reading!

NO, THANKS

OK

scribd

/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->