econometric analysis toán kinh tế

B R I E F C O N T E N T SQ Chapter 1 Introduction 1Chapter 2 The Classical Multiple Linear Regression Model 7Chapter 3 Least Squares 19 Chapter 4 Finite-Sample Properties of the Least Sq

FIFTH EDITIONECONOMETRIC ANALYSIS

Q

William H Greene

New York University

Upper Saddle River, New Jersey 07458

iii

CIP data to come

Executive Editor: Rod Banister Editor-in-Chief: P J Boardman Managing Editor: Gladys Soto Assistant Editor: Marie McHale Editorial Assistant: Lisa Amato Senior Media Project Manager: Victoria Anderson Executive Marketing Manager: Kathleen McLellan Marketing Assistant: Christopher Bath

Managing Editor (Production): Cynthia Regan Production Editor: Michael Reynolds Production Assistant: Dianne Falcone Permissions Supervisor: Suzanne Grappi Associate Director, Manufacturing: Vinnie Scelta Cover Designer: Kiwi Design

Cover Photo: Anthony Bannister/Corbis Composition: Interactive Composition Corporation Printer/Binder: Courier/Westford

Cover Printer: Coral Graphics

Credits and acknowledgments borrowed from other sources and reproduced, with permission, in this textbook appear on appropriate page within text (or on page XX).

Copyright © 2003, 2000, 1997, 1993 by Pearson Education, Inc., Upper Saddle River, New Jersey, 07458 All rights reserved Printed in the United States of America This

publication is protected by Copyright and permission should be obtained from the publisher prior to any prohibited reproduction, storage in a retrieval system, or transmission in any form or by any means, electronic, mechanical, photocopying, recording, or likewise For information regarding permission(s), write to: Rights and Permissions Department.

Pearson Education LTD.

Pearson Education Australia PTY, Limited Pearson Education Singapore, Pte Ltd Pearson Education North Asia Ltd Pearson Education, Canada, Ltd Pearson Educación de Mexico, S.A de C.V.

Pearson Education–Japan Pearson Education Malaysia, Pte Ltd

10 9 8 7 6 5 4 3 2 1 ISBN 0-13-066189-9

iv

B R I E F C O N T E N T S

Q

Chapter 1 Introduction 1Chapter 2 The Classical Multiple Linear Regression Model 7Chapter 3 Least Squares 19

Chapter 4 Finite-Sample Properties of the Least Squares Estimator 41Chapter 5 Large-Sample Properties of the Least Squares and Instrumental

Variables Estimators 65Chapter 6 Inference and Prediction 93Chapter 7 Functional Form and Structural Change 116Chapter 8 Specification Analysis and Model Selection 148Chapter 9 Nonlinear Regression Models 162

Chapter 10 Nonspherical Disturbances—The Generalized

Regression Model 191Chapter 11 Heteroscedasticity 215Chapter 12 Serial Correlation 250Chapter 13 Models for Panel Data 283Chapter 14 Systems of Regression Equations 339Chapter 15 Simultaneous-Equations Models 378Chapter 16 Estimation Frameworks in Econometrics 425Chapter 17 Maximum Likelihood Estimation 468Chapter 18 The Generalized Method of Moments 525Chapter 19 Models with Lagged Variables 558Chapter 20 Time-Series Models 608

Chapter 21 Models for Discrete Choice 663Chapter 22 Limited Dependent Variable and Duration Models 756Appendix A Matrix Algebra 803

Appendix B Probability and Distribution Theory 845Appendix C Estimation and Inference 877

Appendix D Large Sample Distribution Theory 896

vii

Appendix E Computation and Optimization 919Appendix F Data Sets Used in Applications 946Appendix G Statistical Tables 953

References 959Author Index 000Subject Index 000

C O N T E N T S

Q

1.1 Econometrics 11.2 Econometric Modeling 11.3 Data and Methodology 41.4 Plan of the Book 5

2.1 Introduction 72.2 The Linear Regression Model 72.3 Assumptions of the Classical Linear Regression Model 10

2.3.1 Linearity of the Regression Model 11 2.3.2 Full Rank 13

2.3.3 Regression 14 2.3.4 Spherical Disturbances 15 2.3.5 Data Generating Process for the Regressors 16 2.3.6 Normality 17

2.4 Summary and Conclusions 18

3.1 Introduction 193.2 Least Squares Regression 19

3.2.1 The Least Squares Coefficient Vector 20 3.2.2 Application: An Investment Equation 21 3.2.3 Algebraic Aspects of The Least Squares Solution 24 3.2.4 Projection 24

3.3 Partitioned Regression and Partial Regression 263.4 Partial Regression and Partial Correlation Coefficients 283.5 Goodness of Fit and the Analysis of Variance 31

3.5.1 The Adjusted R-Squared and a Measure of Fit 34 3.5.2 R-Squared and the Constant Term in the Model 36 3.5.3 Comparing Models 37

3.6 Summary and Conclusions 38

ix

CHAPTER 4 Finite-Sample Properties of the Least Squares Estimator 41

4.1 Introduction 414.2 Motivating Least Squares 42

4.2.1 The Population Orthogonality Conditions 42 4.2.2 Minimum Mean Squared Error Predictor 43 4.2.3 Minimum Variance Linear Unbiased Estimation 44

4.3 Unbiased Estimation 444.4 The Variance of the Least Squares Estimator and the Gauss Markov

4.5 The Implications of Stochastic Regressors 474.6 Estimating the Variance of the Least Squares Estimator 484.7 The Normality Assumption and Basic Statistical Inference 50

4.7.1 Testing a Hypothesis About a Coefficient 50 4.7.2 Confidence Intervals for Parameters 52 4.7.3 Confidence Interval for a Linear Combination of Coefficients:

The Oaxaca Decomposition 53 4.7.4 Testing the Significance of the Regression 54 4.7.5 Marginal Distributions of the Test Statistics 55

4.8 Finite-Sample Properties of Least Squares 554.9 Data Problems 56

4.9.1 Multicollinearity 56 4.9.2 Missing Observations 59 4.9.3 Regression Diagnostics and Influential Data Points 60

4.10 Summary and Conclusions 61

5.1 Introduction 655.2 Asymptotic Properties of the Least Squares Estimator 65

5.2.1 Consistency of the Least Squares Estimator of β 66

5.2.2 Asymptotic Normality of the Least Squares Estimator 67 5.2.3 Consistency of s2and the Estimator of Asy Var[b] 69 5.2.4 Asymptotic Distribution of a Function of b: The Delta

Method 70 5.2.5 Asymptotic Efficiency 70

5.3 More General Cases 72

5.3.1 Heterogeneity in the Distributions of x i 72 5.3.2 Dependent Observations 73

5.4 Instrumental Variable and Two Stage Least Squares

Estimation 745.5 Hausman’s Specification Test and an Application to Instrumental Variable

Estimation 80

5.6 Measurement Error 83

5.6.1 Least Squares Attenuation 84 5.6.2 Instrumental Variables Estimation 86 5.6.3 Proxy Variables 87

5.6.4 Application: Income and Education and a Study of Twins 88

5.7 Summary and Conclusions 90

6.1 Introduction 936.2 Restrictions and Nested Models 936.3 Two Approaches to Testing Hypotheses 95

6.3.1 The F Statistic and the Least Squares Discrepancy 95 6.3.2 The Restricted Least Squares Estimator 99

6.3.3 The Loss of Fit from Restricted Least Squares 101

6.4 Nonnormal Disturbances and Large Sample Tests 1046.5 Testing Nonlinear Restrictions 108

6.6 Prediction 1116.7 Summary and Conclusions 114

7.1 Introduction 1167.2 Using Binary Variables 116

7.2.1 Binary Variables in Regression 116 7.2.2 Several Categories 117

7.2.3 Several Groupings 118 7.2.4 Threshold Effects and Categorical Variables 120 7.2.5 Spline Regression 121

7.3 Nonlinearity in the Variables 122

7.3.1 Functional Forms 122 7.3.2 Identifying Nonlinearity 124 7.3.3 Intrinsic Linearity and Identification 127

7.4 Modeling and Testing for a Structural Break 130

7.4.1 Different Parameter Vectors 130 7.4.2 Insufficient Observations 131 7.4.3 Change in a Subset of Coefficients 132 7.4.4 Tests of Structural Break with Unequal Variances 133

7.5 Tests of Model Stability 134

7.5.1 Hansen’s Test 134 7.5.2 Recursive Residuals and the CUSUMS Test 135 7.5.3 Predictive Test 137

7.5.4 Unknown Timing of the Structural Break 139

7.6 Summary and Conclusions 144

CHAPTER 8 Specification Analysis and Model Selection 148

8.1 Introduction 1488.2 Specification Analysis and Model Building 148

8.2.1 Bias Caused by Omission of Relevant Variables 148 8.2.2 Pretest Estimation 149

8.2.3 Inclusion of Irrelevant Variables 150 8.2.4 Model Building—A General to Simple Strategy 151

8.3 Choosing Between Nonnested Models 152

8.3.1 Testing Nonnested Hypotheses 153 8.3.2 An Encompassing Model 154 8.3.3 Comprehensive Approach—The J Test 154 8.3.4 The Cox Test 155

8.4 Model Selection Criteria 1598.5 Summary and Conclusions 160

9.1 Introduction 1629.2 Nonlinear Regression Models 162

9.2.1 Assumptions of the Nonlinear Regression Model 163 9.2.2 The Orthogonality Condition and the Sum of Squares 164 9.2.3 The Linearized Regression 165

9.2.4 Large Sample Properties of the Nonlinear Least Squares

Estimator 167 9.2.5 Computing the Nonlinear Least Squares Estimator 169

9.3 Applications 171

9.3.1 A Nonlinear Consumption Function 171 9.3.2 The Box–Cox Transformation 173

9.4 Hypothesis Testing and Parametric Restrictions 175

9.4.1 Significance Tests for Restrictions: F and Wald Statistics 175 9.4.2 Tests Based on the LM Statistic 177

9.4.3 A Specification Test for Nonlinear Regressions: The P E Test 178

9.5 Alternative Estimators for Nonlinear Regression Models 180

9.5.1 Nonlinear Instrumental Variables Estimation 181 9.5.2 Two-Step Nonlinear Least Squares Estimation 183 9.5.3 Two-Step Estimation of a Credit Scoring Model 186

9.6 Summary and Conclusions 189

10.1 Introduction 19110.2 Least Squares and Instrumental Variables Estimation 192

10.2.1 Finite-Sample Properties of Ordinary Least Squares 193 10.2.2 Asymptotic Properties of Least Squares 194

10.2.3 Asymptotic Properties of Nonlinear Least Squares 196

10.2.4 Asymptotic Properties of the Instrumental Variables

Estimator 196

10.3 Robust Estimation of Asymptotic Covariance Matrices 19810.4 Generalized Method of Moments Estimation 201

10.5 Efficient Estimation by Generalized Least Squares 207

10.5.1 Generalized Least Squares (GLS) 207 10.5.2 Feasible Generalized Least Squares 209

10.6 Maximum Likelihood Estimation 21110.7 Summary and Conclusions 212

11.1 Introduction 21511.2 Ordinary Least Squares Estimation 216

11.2.1 Inefficiency of Least Squares 217 11.2.2 The Estimated Covariance Matrix of b 217 11.2.3 Estimating the Appropriate Covariance Matrix for Ordinary

11.5 Weighted Least Squares When is Known 22511.6 Estimation When Contains Unknown Parameters 227

11.6.1 Two-Step Estimation 227 11.6.2 Maximum Likelihood Estimation 228 11.6.3 Model Based Tests for Heteroscedasticity 229

11.7 Applications 232

11.7.1 Multiplicative Heteroscedasticity 232 11.7.2 Groupwise Heteroscedasticity 235

11.8 Autoregressive Conditional Heteroscedasticity 238

11.8.1 The ARCH(1) Model 238 11.8.2 ARCH(q), ARCH-in-Mean and Generalized ARCH

Models 240 11.8.3 Maximum Likelihood Estimation of the GARCH Model 242 11.8.4 Testing for GARCH Effects 244

11.8.5 Pseudo-Maximum Likelihood Estimation 245

11.9 Summary and Conclusions 246

12.1 Introduction 25012.2 The Analysis of Time-Series Data 25312.3 Disturbance Processes 256

12.3.1 Characteristics of Disturbance Processes 256 12.3.2 AR(1) Disturbances 257

12.4 Some Asymptotic Results for Analyzing Time Series Data 259

12.4.1 Convergence of Moments—The Ergodic Theorem 260 12.4.2 Convergence to Normality—A Central Limit Theorem 262

12.5 Least Squares Estimation 265

12.5.1 Asymptotic Properties of Least Squares 265 12.5.2 Estimating the Variance of the Least Squares Estimator 266

12.6 GMM Estimation 26812.7 Testing for Autocorrelation 268

12.7.1 Lagrange Multiplier Test 269 12.7.2 Box and Pierce’s Test and Ljung’s Refinement 269 12.7.3 The Durbin–Watson Test 270

12.7.4 Testing in the Presence of a Lagged Dependent Variables 270 12.7.5 Summary of Testing Procedures 271

12.8 Efficient Estimation When Is Known 27112.9 Estimation When Is Unknown 273

12.9.1 AR(1) Disturbances 273 12.9.2 AR(2) Disturbances 274 12.9.3 Application: Estimation of a Model with Autocorrelation 274 12.9.4 Estimation with a Lagged Dependent Variable 277

12.10 Common Factors 27812.11 Forecasting in the Presence of Autocorrelation 27912.12 Summary and Conclusions 280

13.1 Introduction 28313.2 Panel Data Models 28313.3 Fixed Effects 287

13.3.1 Testing the Significance of the Group Effects 289 13.3.2 The Within- and Between-Groups Estimators 289 13.3.3 Fixed Time and Group Effects 291

13.3.4 Unbalanced Panels and Fixed Effects 293

13.7 Nonspherical Disturbances and Robust Covariance Estimation 314

13.7.1 Robust Estimation of the Fixed Effects Model 314

13.7.2 Heteroscedasticity in the Random Effects Model 316 13.7.3 Autocorrelation in Panel Data Models 317

13.8 Random Coefficients Models 31813.9 Covariance Structures for Pooled Time-Series Cross-Sectional

13.10 Summary and Conclusions 334

14.1 Introduction 33914.2 The Seemingly Unrelated Regressions Model 340

14.2.1 Generalized Least Squares 341 14.2.2 Seemingly Unrelated Regressions with Identical Regressors 343 14.2.3 Feasible Generalized Least Squares 344

14.2.4 Maximum Likelihood Estimation 347 14.2.5 An Application from Financial Econometrics:

The Capital Asset Pricing Model 351 14.2.6 Maximum Likelihood Estimation of the Seemingly Unrelated

Regressions Model with a Block of Zeros in the Coefficient Matrix 357

14.2.7 Autocorrelation and Heteroscedasticity 360

14.3 Systems of Demand Equations: Singular Systems 362

14.3.1 Cobb–Douglas Cost Function 363 14.3.2 Flexible Functional Forms: The Translog Cost Function 366

14.4 Nonlinear Systems and GMM Estimation 369

14.4.1 GLS Estimation 370 14.4.2 Maximum Likelihood Estimation 371 14.4.3 GMM Estimation 372

14.5 Summary and Conclusions 374

15.1 Introduction 37815.2 Fundamental Issues in Simultaneous-Equations Models 378

15.2.1 Illustrative Systems of Equations 378 15.2.2 Endogeneity and Causality 381 15.2.3 A General Notation for Linear Simultaneous Equations

Models 382

15.3 The Problem of Identification 385

15.3.1 The Rank and Order Conditions for Identification 389 15.3.2 Identification Through Other Nonsample Information 394 15.3.3 Identification Through Covariance Restrictions—The Fully

Recursive Model 394

15.4 Methods of Estimation 39615.5 Single Equation: Limited Information Estimation Methods 396

15.5.1 Ordinary Least Squares 396 15.5.2 Estimation by Instrumental Variables 397 15.5.3 Two-Stage Least Squares 398

15.5.4 GMM Estimation 400 15.5.5 Limited Information Maximum Likelihood and the k Class of

Estimators 401 15.5.6 Two-Stage Least Squares in Models That Are Nonlinear in

Variables 403

15.6 System Methods of Estimation 404

15.6.1 Three-Stage Least Squares 405 15.6.2 Full-Information Maximum Likelihood 407 15.6.3 GMM Estimation 409

15.6.4 Recursive Systems and Exactly Identified Equations 411

15.7 Comparison of Methods—Klein’s Model I 41115.8 Specification Tests 413

15.9 Properties of Dynamic Models 415

15.9.1 Dynamic Models and Their Multipliers 415 15.9.2 Stability 417

15.9.3 Adjustment to Equilibrium 418

15.10 Summary and Conclusions 421

16.1 Introduction 42516.2 Parametric Estimation and Inference 427

16.2.1 Classical Likelihood Based Estimation 428 16.2.2 Bayesian Estimation 429

16.2.2.a Bayesian Analysis of the Classical Regression Model 430 16.2.2.b Point Estimation 434

16.2.2.c Interval Estimation 435 16.2.2.d Estimation with an Informative Prior Density 435 16.2.2.e Hypothesis Testing 437

16.2.3 Using Bayes Theorem in a Classical Estimation Problem: The

Latent Class Model 439 16.2.4 Hierarchical Bayes Estimation of a Random Parameters Model

by Markov Chain Monte Carlo Simulation 444

16.3 Semiparametric Estimation 447

16.3.1 GMM Estimation in Econometrics 447 16.3.2 Least Absolute Deviations Estimation 448

16.3.3 Partially Linear Regression 450 16.3.4 Kernel Density Methods 452

16.6 Summary and Conclusions 466

17.1 Introduction 46817.2 The Likelihood Function and Identification of the Parameters 46817.3 Efficient Estimation: The Principle of Maximum Likelihood 47017.4 Properties of Maximum Likelihood Estimators 472

17.4.1 Regularity Conditions 473 17.4.2 Properties of Regular Densities 474 17.4.3 The Likelihood Equation 476 17.4.4 The Information Matrix Equality 476 17.4.5 Asymptotic Properties of the Maximum

Likelihood Estimator 476 17.4.5.a Consistency 477

17.4.5.b Asymptotic Normality 478 17.4.5.c Asymptotic Efficiency 479 17.4.5.d Invariance 480

17.4.5.e Conclusion 480 17.4.6 Estimating the Asymptotic Variance of the Maximum

Likelihood Estimator 480 17.4.7 Conditional Likelihoods and Econometric Models 482

17.5 Three Asymptotically Equivalent Test Procedures 484

17.5.1 The Likelihood Ratio Test 484 17.5.2 The Wald Test 486

17.5.3 The Lagrange Multiplier Test 489 17.5.4 An Application of the Likelihood Based Test Procedures 490

17.6 Applications of Maximum Likelihood Estimation 492

17.6.1 The Normal Linear Regression Model 492 17.6.2 Maximum Likelihood Estimation of Nonlinear

Regression Models 496 17.6.3 Nonnormal Disturbances—The Stochastic Frontier Model 501 17.6.4 Conditional Moment Tests of Specification 505

17.7 Two-Step Maximum Likelihood Estimation 50817.8 Maximum Simulated Likelihood Estimation 51217.9 Pseudo-Maximum Likelihood Estimation and Robust Asymptotic

Covariance Matrices 51817.10 Summary and Conclusions 521

18.1 Introduction 52518.2 Consistent Estimation: The Method of Moments 526

18.2.1 Random Sampling and Estimating the Parameters of

Distributions 527 18.2.2 Asymptotic Properties of the Method of Moments

Estimator 531 18.2.3 Summary—The Method of Moments 533

18.3 The Generalized Method of Moments (GMM) Estimator 533

18.3.1 Estimation Based on Orthogonality Conditions 534 18.3.2 Generalizing the Method of Moments 536

18.3.3 Properties of the GMM Estimator 540 18.3.4 GMM Estimation of Some Specific Econometric Models 544

18.4 Testing Hypotheses in the GMM Framework 548

18.4.1 Testing the Validity of the Moment Restrictions 548 18.4.2 GMM Counterparts to the Wald, LM, and LR Tests 549

18.5 Application: GMM Estimation of a Dynamic Panel Data Model of

Local Government Expenditures 55118.6 Summary and Conclusions 555

19.1 Introduction 55819.2 Dynamic Regression Models 559

19.2.1 Lagged Effects in a Dynamic Model 560 19.2.2 The Lag and Difference Operators 562 19.2.3 Specification Search for the Lag Length 564

19.3 Simple Distributed Lag Models 565

19.3.1 Finite Distributed Lag Models 565 19.3.2 An Infinite Lag Model: The Geometric Lag Model 566

19.4 Autoregressive Distributed Lag Models 571

19.4.1 Estimation of the ARDL Model 572 19.4.2 Computation of the Lag Weights in the ARDL

Model 573 19.4.3 Stability of a Dynamic Equation 573 19.4.4 Forecasting 576

19.5 Methodological Issues in the Analysis of Dynamic Models 579

19.5.1 An Error Correction Model 579 19.5.2 Autocorrelation 581

19.5.3 Specification Analysis 582 19.5.4 Common Factor Restrictions 583

19.6 Vector Autoregressions 586

19.6.1 Model Forms 587 19.6.2 Estimation 588 19.6.3 Testing Procedures 589 19.6.4 Exogeneity 590 19.6.5 Testing for Granger Causality 592 19.6.6 Impulse Response Functions 593 19.6.7 Structural VARs 595

19.6.8 Application: Policy Analysis with a VAR 596 19.6.9 VARs in Microeconomics 602

19.7 Summary and Conclusions 605

20.1 Introduction 60820.2 Stationary Stochastic Processes 609

20.2.1 Autoregressive Moving-Average Processes 609 20.2.2 Stationarity and Invertibility 611

20.2.3 Autocorrelations of a Stationary Stochastic Process 614 20.2.4 Partial Autocorrelations of a Stationary Stochastic

Process 617 20.2.5 Modeling Univariate Time Series 619 20.2.6 Estimation of the Parameters of a Univariate Time

Series 621 20.2.7 The Frequency Domain 624

20.2.7.a Theoretical Results 625 20.2.7.b Empirical Counterparts 627

20.3 Nonstationary Processes and Unit Roots 631

20.3.1 Integrated Processes and Differencing 631 20.3.2 Random Walks, Trends, and Spurious Regressions 632 20.3.3 Tests for Unit Roots in Economic Data 636

20.3.4 The Dickey–Fuller Tests 637 20.3.5 Long Memory Models 647

20.4 Cointegration 649

20.4.1 Common Trends 653 20.4.2 Error Correction and VAR Representations 654 20.4.3 Testing for Cointegration 655

20.4.4 Estimating Cointegration Relationships 657 20.4.5 Application: German Money Demand 657

20.4.5.a Cointegration Analysis and a Long Run

Theoretical Model 659 20.4.5.b Testing for Model Instability 659

20.5 Summary and Conclusions 660

CHAPTER 21 Models for Discrete Choice 663

21.1 Introduction 66321.2 Discrete Choice Models 66321.3 Models for Binary Choice 665

21.3.1 The Regression Approach 665 21.3.2 Latent Regression—Index Function Models 668 21.3.3 Random Utility Models 670

21.4 Estimation and Inference in Binary Choice Models 670

21.4.1 Robust Covariance Matrix Estimation 673 21.4.2 Marginal Effects 674

21.4.3 Hypothesis Tests 676 21.4.4 Specification Tests for Binary Choice Models 679

21.4.4.a Omitted Variables 680 21.4.4.b Heteroscedasticity 680 21.4.4.c A Specification Test for Nonnested Models—Testing

for the Distribution 682 21.4.5 Measuring Goodness of Fit 683 21.4.6 Analysis of Proportions Data 686

21.5 Extensions of the Binary Choice Model 689

21.5.1 Random and Fixed Effects Models for Panel Data 689

21.5.1.a Random Effects Models 690 21.5.1.b Fixed Effects Models 695 21.5.2 Semiparametric Analysis 700 21.5.3 The Maximum Score Estimator (MSCORE) 702 21.5.4 Semiparametric Estimation 704

21.5.5 A Kernel Estimator for a Nonparametric Regression

Function 706 21.5.6 Dynamic Binary Choice Models 708

21.6 Bivariate and Multivariate Probit Models 710

21.6.1 Maximum Likelihood Estimation 710 21.6.2 Testing for Zero Correlation 712 21.6.3 Marginal Effects 712

21.6.4 Sample Selection 713 21.6.5 A Multivariate Probit Model 714 21.6.6 Application: Gender Economics Courses in Liberal

Arts Colleges 715

21.7 Logit Models for Multiple Choices 719

21.7.1 The Multinomial Logit Model 720 21.7.2 The Conditional Logit Model 723 21.7.3 The Independence from Irrelevant Alternatives 724 21.7.4 Nested Logit Models 725

21.7.5 A Heteroscedastic Logit Model 727 21.7.6 Multinomial Models Based on the Normal Distribution 727 21.7.7 A Random Parameters Model 728

21.7.8 Application: Conditional Logit Model for Travel

Mode Choice 729

21.8 Ordered Data 73621.9 Models for Count Data 740

21.9.1 Measuring Goodness of Fit 741 21.9.2 Testing for Overdispersion 743 21.9.3 Heterogeneity and the Negative Binomial

Regression Model 744 21.9.4 Application: The Poisson Regression Model 745 21.9.5 Poisson Models for Panel Data 747

21.9.6 Hurdle and Zero-Altered Poisson Models 749

21.10 Summary and Conclusions 752

22.1 Introduction 75622.2 Truncation 756

22.2.1 Truncated Distributions 757 22.2.2 Moments of Truncated Distributions 758 22.2.3 The Truncated Regression Model 760

22.3 Censored Data 761

22.3.1 The Censored Normal Distribution 762 22.3.2 The Censored Regression (Tobit) Model 764 22.3.3 Estimation 766

22.3.4 Some Issues in Specification 768

22.3.4.a Heteroscedasticity 768 22.3.4.b Misspecification of Prob[y* < 0] 770 22.3.4.c Nonnormality 771

22.3.4.d Conditional Moment Tests 772 22.3.5 Censoring and Truncation in Models for Counts 773 22.3.6 Application: Censoring in the Tobit and Poisson

Regression Models 774

22.4 The Sample Selection Model 780

22.4.1 Incidental Truncation in a Bivariate Distribution 781 22.4.2 Regression in a Model of Selection 782

22.4.3 Estimation 784 22.4.4 Treatment Effects 787 22.4.5 The Normality Assumption 789 22.4.6 Selection in Qualitative Response Models 790

22.5 Models for Duration Data 790

22.5.1 Duration Data 791 22.5.2 A Regression-Like Approach: Parametric Models

of Duration 792 22.5.2.a Theoretical Background 792 22.5.2.b Models of the Hazard Function 793 22.5.2.c Maximum Likelihood Estimation 794

22.5.2.d Exogenous Variables 796 22.5.2.e Heterogeneity 797 22.5.3 Other Approaches 798

22.6 Summary and Conclusions 801

A.1 Terminology 803A.2 Algebraic Manipulation of Matrices 803

A.2.1 Equality of Matrices 803 A.2.2 Transposition 804 A.2.3 Matrix Addition 804 A.2.4 Vector Multiplication 805 A.2.5 A Notation for Rows and Columns of a Matrix 805 A.2.6 Matrix Multiplication and Scalar Multiplication 805 A.2.7 Sums of Values 807

A.2.8 A Useful Idempotent Matrix 808

A.3 Geometry of Matrices 809

A.3.1 Vector Spaces 809 A.3.2 Linear Combinations of Vectors and Basis Vectors 811 A.3.3 Linear Dependence 811

A.3.4 Subspaces 813 A.3.5 Rank of a Matrix 814 A.3.6 Determinant of a Matrix 816 A.3.7 A Least Squares Problem 817

A.4 Solution of a System of Linear Equations 819

A.4.1 Systems of Linear Equations 819 A.4.2 Inverse Matrices 820

A.4.3 Nonhomogeneous Systems of Equations 822 A.4.4 Solving the Least Squares Problem 822

A.5 Partitioned Matrices 822

A.5.1 Addition and Multiplication of Partitioned Matrices 823 A.5.2 Determinants of Partitioned Matrices 823

A.5.3 Inverses of Partitioned Matrices 823 A.5.4 Deviations from Means 824

A.5.5 Kronecker Products 824

A.6 Characteristic Roots and Vectors 825

A.6.1 The Characteristic Equation 825 A.6.2 Characteristic Vectors 826 A.6.3 General Results for Characteristic Roots and Vectors 826 A.6.4 Diagonalization and Spectral Decomposition of a Matrix 827 A.6.5 Rank of a Matrix 827

A.6.6 Condition Number of a Matrix 829 A.6.7 Trace of a Matrix 829

A.6.8 Determinant of a Matrix 830 A.6.9 Powers of a Matrix 830

A.6.10 Idempotent Matrices 832 A.6.11 Factoring a Matrix 832 A.6.12 The Generalized Inverse of a Matrix 833

A.7 Quadratic Forms and Definite Matrices 834

A.7.1 Nonnegative Definite Matrices 835 A.7.2 Idempotent Quadratic Forms 836 A.7.3 Comparing Matrices 836

A.8 Calculus and Matrix Algebra 837

A.8.1 Differentiation and the Taylor Series 837 A.8.2 Optimization 840

A.8.3 Constrained Optimization 842 A.8.4 Transformations 844

B.1 Introduction 845B.2 Random Variables 845

B.2.1 Probability Distributions 845 B.2.2 Cumulative Distribution Function 846

B.3 Expectations of a Random Variable 847B.4 Some Specific Probability Distributions 849

B.4.1 The Normal Distribution 849 B.4.2 The Chi-Squared, t, and F Distributions 851 B.4.3 Distributions With Large Degrees of Freedom 853 B.4.4 Size Distributions: The Lognormal Distribution 854 B.4.5 The Gamma and Exponential Distributions 855 B.4.6 The Beta Distribution 855

B.4.7 The Logistic Distribution 855 B.4.8 Discrete Random Variables 855

B.5 The Distribution of a Function of a Random Variable 856B.6 Representations of a Probability Distribution 858

B.7 Joint Distributions 860

B.7.1 Marginal Distributions 860 B.7.2 Expectations in a Joint Distribution 861 B.7.3 Covariance and Correlation 861 B.7.4 Distribution of a Function of Bivariate Random Variables 862

B.8 Conditioning in a Bivariate Distribution 864

B.8.1 Regression: The Conditional Mean 864 B.8.2 Conditional Variance 865

B.8.3 Relationships Among Marginal and Conditional

Moments 865 B.8.4 The Analysis of Variance 867

B.9 The Bivariate Normal Distribution 867B.10 Multivariate Distributions 868

B.10.1 Moments 868

B.10.2 Sets of Linear Functions 869 B.10.3 Nonlinear Functions 870

B.11 The Multivariate Normal Distribution 871

B.11.1 Marginal and Conditional Normal Distributions 871 B.11.2 The Classical Normal Linear Regression Model 872 B.11.3 Linear Functions of a Normal Vector 873

B.11.4 Quadratic Forms in a Standard Normal Vector 873 B.11.5 The F Distribution 875

B.11.6 A Full Rank Quadratic Form 875 B.11.7 Independence of a Linear and a Quadratic Form 876

C.1 Introduction 877C.2 Samples and Random Sampling 878C.3 Descriptive Statistics 878

C.4 Statistics as Estimators—Sampling Distributions 882C.5 Point Estimation of Parameters 885

C.5.1 Estimation in a Finite Sample 885 C.5.2 Efficient Unbiased Estimation 888

C.6 Interval Estimation 890C.7 Hypothesis Testing 892

C.7.1 Classical Testing Procedures 892 C.7.2 Tests Based on Confidence Intervals 895 C.7.3 Specification Tests 896

D.1 Introduction 896D.2 Large-Sample Distribution Theory 897

D.2.1 Convergence in Probability 897 D.2.2 Other Forms of Convergence and Laws of Large Numbers 900 D.2.3 Convergence of Functions 903

D.2.4 Convergence to a Random Variable 904 D.2.5 Convergence in Distribution: Limiting Distributions 906 D.2.6 Central Limit Theorems 908

D.2.7 The Delta Method 913

D.3 Asymptotic Distributions 914

D.3.1 Asymptotic Distribution of a Nonlinear Function 916 D.3.2 Asymptotic Expectations 917

D.4 Sequences and the Order of a Sequence 918

E.1 Introduction 919E.2 Data Input and Generation 920

E.2.1 Generating Pseudo-Random Numbers 920

E.2.2 Sampling from a Standard Uniform Population 921 E.2.3 Sampling from Continuous Distributions 921 E.2.4 Sampling from a Multivariate Normal Population 922 E.2.5 Sampling from a Discrete Population 922

E.2.6 The Gibbs Sampler 922

E.3 Monte Carlo Studies 923E.4 Bootstrapping and the Jackknife 924E.5 Computation in Econometrics 925

E.5.1 Computing Integrals 926 E.5.2 The Standard Normal Cumulative Distribution Function 926 E.5.3 The Gamma and Related Functions 927

E.5.4 Approximating Integrals by Quadrature 928 E.5.5 Monte Carlo Integration 929

E.5.6 Multivariate Normal Probabilities and Simulated

Moments 931 E.5.7 Computing Derivatives 933

E.6 Optimization 933

E.6.1 Algorithms 935 E.6.2 Gradient Methods 935 E.6.3 Aspects of Maximum Likelihood Estimation 939 E.6.4 Optimization with Constraints 941

E.6.5 Some Practical Considerations 942 E.6.6 Examples 943

P R E F A C E

Q

ANALYSIS

Econometric Analysis is intended for a one-year graduate course in econometrics for

social scientists The prerequisites for this course should include calculus, mathematical

statistics, and an introduction to econometrics at the level of, say, Gujarati’s Basic metrics (McGraw-Hill, 1995) or Wooldridge’s Introductory Econometrics: A Modern Approach [South-Western (2000)] Self-contained (for our purposes) summaries of the

Econo-matrix algebra, mathematical statistics, and statistical theory used later in the book aregiven in Appendices A through D Appendix E contains a description of numericalmethods that will be useful to practicing econometricians The formal presentation ofeconometrics begins with discussion of a fundamental pillar, the linear multiple regres-sion model, in Chapters 2 through 8 Chapters 9 through 15 present familiar extensions

of the single linear equation model, including nonlinear regression, panel data models,the generalized regression model, and systems of equations The linear model is usuallynot the sole technique used in most of the contemporary literature In view of this, the(expanding) second half of this book is devoted to topics that will extend the linearregression model in many directions Chapters 16 through 18 present the techniquesand underlying theory of estimation in econometrics, including GMM and maximumlikelihood estimation methods and simulation based techniques We end in the last fourchapters, 19 through 22, with discussions of current topics in applied econometrics, in-cluding time-series analysis and the analysis of discrete choice and limited dependentvariable models

This book has two objectives The first is to introduce students to applied metrics, including basic techniques in regression analysis and some of the rich variety

econo-of models that are used when the linear model proves inadequate or inappropriate

The second is to present students with sufficient theoretical background that they will

recognize new variants of the models learned about here as merely natural extensionsthat fit within a common body of principles Thus, I have spent what might seem to be

a large amount of effort explaining the mechanics of GMM estimation, nonlinear leastsquares, and maximum likelihood estimation and GARCH models To meet the secondobjective, this book also contains a fair amount of theoretical material, such as that onmaximum likelihood estimation and on asymptotic results for regression models Mod-ern software has made complicated modeling very easy to do, and an understanding ofthe underlying theory is important

I had several purposes in undertaking this revision As in the past, readers continue

to send me interesting ideas for my “next edition.” It is impossible to use them all, of

xxvii

course Because the five volumes of the Handbook of Econometrics and two of the Handbook of Applied Econometrics already run to over 4,000 pages, it is also unneces-

sary Nonetheless, this revision is appropriate for several reasons First, there are newand interesting developments in the field, particularly in the areas of microeconometrics(panel data, models for discrete choice) and, of course, in time series, which continuesits rapid development Second, I have taken the opportunity to continue fine-tuning thetext as the experience and shared wisdom of my readers accumulates in my files For thisrevision, that adjustment has entailed a substantial rearrangement of the material—themain purpose of that was to allow me to add the new material in a more compact andorderly way than I could have with the table of contents in the 4th edition The litera-ture in econometrics has continued to evolve, and my third objective is to grow with it.This purpose is inherently difficult to accomplish in a textbook Most of the literature iswritten by professionals for other professionals, and this textbook is written for studentswho are in the early stages of their training But I do hope to provide a bridge to thatliterature, both theoretical and applied

This book is a broad survey of the field of econometrics This field grows tinually, and such an effort becomes increasingly difficult (A partial list of journals

con-devoted at least in part, if not completely, to econometrics now includes the Journal

of Applied Econometrics, Journal of Econometrics, Econometric Theory, Econometric Reviews, Journal of Business and Economic Statistics, Empirical Economics, and Econo- metrica.) Still, my view has always been that the serious student of the field must start somewhere, and one can successfully seek that objective in a single textbook This text

attempts to survey, at an entry level, enough of the fields in econometrics that a studentcan comfortably move from here to practice or more advanced study in one or morespecialized areas At the same time, I have tried to present the material in sufficientgenerality that the reader is also able to appreciate the important common foundation

of all these fields and to use the tools that they all employ

There are now quite a few recently published texts in econometrics Several havegathered in compact, elegant treatises, the increasingly advanced and advancing theo-retical background of econometrics Others, such as this book, focus more attention onapplications of econometrics One feature that distinguishes this work from its prede-cessors is its greater emphasis on nonlinear models [Davidson and MacKinnon (1993)

is a noteworthy, but more advanced, exception.] Computer software now in wide usehas made estimation of nonlinear models as routine as estimation of linear ones, and therecent literature reflects that progression My purpose is to provide a textbook treat-ment that is in line with current practice The book concludes with four lengthy chapters

on time-series analysis, discrete choice models and limited dependent variable models.These nonlinear models are now the staples of the applied econometrics literature Thisbook also contains a fair amount of material that will extend beyond many first courses

in econometrics, including, perhaps, the aforementioned chapters on limited dependentvariables, the section in Chapter 22 on duration models, and some of the discussions

of time series and panel data models Once again, I have included these in the hope ofproviding a bridge to the professional literature in these areas

I have had one overriding purpose that has motivated all five editions of this work.For the vast majority of readers of books such as this, whose ambition is to use, notdevelop econometrics, I believe that it is simply not sufficient to recite the theory ofestimation, hypothesis testing and econometric analysis Understanding the often subtle

background theory is extremely important But, at the end of the day, my purpose inwriting this work, and for my continuing efforts to update it in this now fifth edition,

is to show readers how to do econometric analysis I unabashedly accept the

unflatter-ing assessment of a correspondent who once likened this book to a “user’s guide toeconometrics.”

There are many computer programs that are widely used for the computations described

in this book All were written by econometricians or statisticians, and in general, allare regularly updated to incorporate new developments in applied econometrics Asampling of the most widely used packages and Internet home pages where you canfind information about them are:

E-Views www.eviews.com (QMS, Irvine, Calif.)

Gauss www.aptech.com (Aptech Systems, Kent, Wash.)

LIMDEP www.limdep.com (Econometric Software, Plainview, N.Y.)

Shazam shazam.econ.ubc.ca (Ken White, UBC, Vancouver, B.C.)

Stata www.stata.com (Stata, College Station, Tex.)

TSP www.tspintl.com (TSP International, Stanford, Calif.)Programs vary in size, complexity, cost, the amount of programming required of the user,

and so on Journals such as The American Statistician, The Journal of Applied metrics, and The Journal of Economic Surveys regularly publish reviews of individual

Econo-packages and comparative surveys of Econo-packages, usually with reference to particularfunctionality such as panel data analysis or forecasting

With only a few exceptions, the computations described in this book can be carriedout with any of these packages We hesitate to link this text to any of them in partic-

ular We have placed for general access a customized version of LIMDEP, which was

also written by the author, on the website for this text, http://www.stern.nyu.edu/

∼wgreene/Text/econometricanalysis.htm LIMDEP programs used for many of

the computations are posted on the sites as well

The data sets used in the examples are also on the website Throughout the text,these data sets are referred to “TableFn.m,” for example Table F4.1 The F refers toAppendix F at the back of the text, which contains descriptions of the data sets Theactual data are posted on the website with the other supplementary materials for thetext (The data sets are also replicated in the system format of most of the commonly

used econometrics computer programs, including in addition to LIMDEP, SAS, TSP, SPSS, E-Views, and Stata, so that you can easily import them into whatever program

you might be using.)

I should also note, there are now thousands of interesting websites containing ware, data sets, papers, and commentary on econometrics It would be hopeless toattempt any kind of a survey here But, I do note one which is particularly agree-ably structured and well targeted for readers of this book, the data archive for the

soft-Journal of Applied Econometrics This journal publishes many papers that are precisely

at the right level for readers of this text They have archived all the nonconfidentialdata sets used in their publications since 1994 This useful archive can be found athttp://qed.econ.queensu.ca/jae/

It is a pleasure to express my appreciation to those who have influenced this work I amgrateful to Arthur Goldberger and Arnold Zellner for their encouragement, guidance,and always interesting correspondence Dennis Aigner and Laurits Christensen werealso influential in shaping my views on econometrics Some collaborators to the earliereditions whose contributions remain in this one include Aline Quester, David Hensher,and Donald Waldman The number of students and colleagues whose suggestions havehelped to produce what you find here is far too large to allow me to thank them allindividually I would like to acknowledge the many reviewers of my work whose care-ful reading has vastly improved the book: Badi Baltagi, University of Houston: NealBeck, University of California at San Diego; Diane Belleville, Columbia University;Anil Bera, University of Illinois; John Burkett, University of Rhode Island; LeonardCarlson, Emory University; Frank Chaloupka, City University of New York; ChrisCornwell, University of Georgia; Mitali Das, Columbia University; Craig Depken II,University of Texas at Arlington; Edward Dwyer, Clemson University; Michael Ellis,Wesleyan University; Martin Evans, New York University; Ed Greenberg, WashingtonUniversity at St Louis; Miguel Herce, University of North Carolina; K Rao Kadiyala,Purdue University; Tong Li, Indiana University; Lubomir Litov, New York University;William Lott, University of Connecticut; Edward Mathis, Villanova University; MaryMcGarvey, University of Nebraska-Lincoln; Ed Melnick, New York University; ThadMirer, State University of New York at Albany; Paul Ruud, University of California atBerkeley; Sherrie Rhine, Chicago Federal Reserve Board; Terry G Seaks, University

of North Carolina at Greensboro; Donald Snyder, California State University at LosAngeles; Steven Stern, University of Virginia; Houston Stokes, University of Illinois

at Chicago; Dimitrios Thomakos, Florida International University; Paul Wachtel, NewYork University; Mark Watson, Harvard University; and Kenneth West, University

of Wisconsin My numerous discussions with B D McCullough have improved pendix E and at the same time increased my appreciation for numerical analysis I

Ap-am especially grateful to Jan Kiviet of the University of AmsterdAp-am, who subjected

my third edition to a microscopic examination and provided literally scores of tions, virtually all of which appear herein Chapters 19 and 20 have also benefited fromprevious reviews by Frank Diebold, B D McCullough, Mary McGarvey, and NageshRevankar I would also like to thank Rod Banister, Gladys Soto, Cindy Regan, MikeReynolds, Marie McHale, Lisa Amato, and Torie Anderson at Prentice Hall for theircontributions to the completion of this book As always, I owe the greatest debt to mywife, Lynne, and to my daughters, Lesley, Allison, Elizabeth, and Julianna

sugges-William H Greene

1 INTRODUCTION

Q

In the first issue of Econometrica, the Econometric Society stated that

its main object shall be to promote studies that aim at a unification of thetheoretical-quantitative and the empirical-quantitative approach to economicproblems and that are penetrated by constructive and rigorous thinking similar

to that which has come to dominate the natural sciences

But there are several aspects of the quantitative approach to economics, and

no single one of these aspects taken by itself, should be confounded with metrics Thus, econometrics is by no means the same as economic statistics Nor

econo-is it identical with what we call general economic theory, although a able portion of this theory has a definitely quantitative character Nor should

consider-econometrics be taken as synonomous [sic] with the application of mathematics

to economics Experience has shown that each of these three viewpoints, that

of statistics, economic theory, and mathematics, is a necessary, but not by itself

a sufficient, condition for a real understanding of the quantitative relations in

modern economic life It is the unification of all three that is powerful And it

is this unification that constitutes econometrics

Frisch (1933) and his society responded to an unprecedented accumulation of cal information They saw a need to establish a body of principles that could organizewhat would otherwise become a bewildering mass of data Neither the pillars nor theobjectives of econometrics have changed in the years since this editorial appeared.Econometrics is the field of economics that concerns itself with the application of math-ematical statistics and the tools of statistical inference to the empirical measurement ofrelationships postulated by economic theory

Econometric analysis will usually begin with a statement of a theoretical proposition.Consider, for example, a canonical application:

From Keynes’s (1936) General Theory of Employment, Interest and Money:

We shall therefore define what we shall call the propensity to consume as the

func-tional relationship f between X , a given level of income and C, the expenditure on consumption out of the level of income, so that C = f ( X).

The amount that the community spends on consumption depends (i) partly onthe amount of its income, (ii) partly on other objective attendant circumstances, and

1

(iii) partly on the subjective needs and the psychological propensities and habits ofthe individuals composing it The fundamental psychological law upon which we areentitled to depend with great confidence, both a priori from our knowledge of humannature and from the detailed facts of experience, is that men are disposed, as a ruleand on the average, to increase their consumption as their income increases, but not

by as much as the increase in their income.1That is, dC/dX is positive and less

than unity

But, apart from short period changes in the level of income, it is also obvious that

a higher absolute level of income will tend as a rule to widen the gap between incomeand consumption. These reasons will lead, as a rule, to a greater proportion of

income being saved as real income increases

The theory asserts a relationship between consumption and income, C = f ( X), and claims

in the third paragraph that the marginal propensity to consume (MPC), dC /dX, is between

0 and 1 The final paragraph asserts that the average propensity to consume (APC), C / X, falls as income rises, or d( C / X)/dX = (MPC − APC)/ X < 0 It follows that MPC < APC The most common formulation of the consumption function is a linear relationship, C =

α + β X, that satisfies Keynes’s “laws” if β lies between zero and one and if α is greater

than zero

These theoretical propositions provide the basis for an econometric study Given an propriate data set, we could investigate whether the theory appears to be consistent withthe observed “facts.” For example, we could see whether the linear specification appears to

ap-be a satisfactory description of the relationship ap-between consumption and income, and, if

so, whetherα is positive and β is between zero and one Some issues that might be

stud-ied are (1) whether this relationship is stable through time or whether the parameters of therelationship change from one generation to the next (a change in the average propensity tosave, 1—APC, might represent a fundamental change in the behavior of consumers in theeconomy); (2) whether there are systematic differences in the relationship across differentcountries, and, if so, what explains these differences; and (3) whether there are other factorsthat would improve the ability of the model to explain the relationship between consumptionand income For example, Figure 1.1 presents aggregate consumption and personal income

in constant dollars for the U.S for the 10 years of 1970–1979 (See Appendix Table F1.1.)Apparently, at least superficially, the data (the facts) are consistent with the theory The rela-tionship appears to be linear, albeit only approximately, the intercept of a line that lies close

to most of the points is positive and the slope is less than one, although not by much.Economic theories such as Keynes’s are typically crisp and unambiguous Models

of demand, production, and aggregate consumption all specify precise, deterministic

relationships Dependent and independent variables are identified, a functional form isspecified, and in most cases, at least a qualitative statement is made about the directions

of effects that occur when independent variables in the model change Of course, themodel is only a simplification of reality It will include the salient features of the rela-tionship of interest, but will leave unaccounted for influences that might well be presentbut are regarded as unimportant No model could hope to encompass the myriad essen-tially random aspects of economic life It is thus also necessary to incorporate stochasticelements As a consequence, observations on a dependent variable will display varia-tion attributable not only to differences in variables that are explicitly accounted for,but also to the randomness of human behavior and the interaction of countless minorinfluences that are not It is understood that the introduction of a random “disturbance”into a deterministic model is not intended merely to paper over its inadequacies It is

1 Modern economists are rarely this confident about their theories More contemporary applications generally begin from first principles and behavioral axioms, rather than simple observation.

800 C 850 900 950

FIGURE 1.1 Consumption Data, 1970–1979.

essential to examine the results of the study, in a sort of postmortem, to ensure that theallegedly random, unexplained factor is truly unexplainable If it is not, the model is, infact, inadequate The stochastic element endows the model with its statistical proper-ties Observations on the variable(s) under study are thus taken to be the outcomes of

a random process With a sufficiently detailed stochastic structure and adequate data,the analysis will become a matter of deducing the properties of a probability distri-bution The tools and methods of mathematical statistics will provide the operatingprinciples

A model (or theory) can never truly be confirmed unless it is made so broad as toinclude every possibility But it may be subjected to ever more rigorous scrutiny and,

in the face of contradictory evidence, refuted A deterministic theory will be dated by a single contradictory observation The introduction of stochastic elementsinto the model changes it from an exact statement to a probabilistic description aboutexpected outcomes and carries with it an important implication Only a preponder-ance of contradictory evidence can convincingly invalidate the probabilistic model, andwhat constitutes a “preponderance of evidence” is a matter of interpretation Thus, theprobabilistic model is less precise but at the same time, more robust.2

invali-The process of econometric analysis departs from the specification of a cal relationship We initially proceed on the optimistic assumption that we can obtainprecise measurements on all the variables in a correctly specified model If the idealconditions are met at every step, the subsequent analysis will probably be routine.Unfortunately, they rarely are Some of the difficulties one can expect to encounter arethe following:

theoreti-2 See Keuzenkamp and Magnus (1995) for a lengthy symposium on testing in econometrics.

• The data may be badly measured or may correspond only vaguely to the variables

in the model “The interest rate” is one example

• Some of the variables may be inherently unmeasurable “Expectations” are a case

in point

• The theory may make only a rough guess as to the correct functional form, if itmakes any at all, and we may be forced to choose from an embarrassingly longmenu of possibilities

• The assumed stochastic properties of the random terms in the model may bedemonstrably violated, which may call into question the methods of estimationand inference procedures we have used

• Some relevant variables may be missing from the model

The ensuing steps of the analysis consist of coping with these problems and attempting

to cull whatever information is likely to be present in such obviously imperfect data.The methodology is that of mathematical statistics and economic theory The product

is an econometric model

The connection between underlying behavioral models and the modern practice ofeconometrics is increasingly strong Practitioners rely heavily on the theoretical tools

of microeconomics including utility maximization, profit maximization, and marketequilibrium Macroeconomic model builders rely on the interactions between economicagents and policy makers The analyses are directed at subtle, difficult questions thatoften require intricate, complicated formulations A few applications:

• What are the likely effects on labor supply behavior of proposed negative incometaxes? [Ashenfelter and Heckman (1974).]

• Does a monetary policy regime that is strongly oriented toward controllinginflation impose a real cost in terms of lost output on the U.S economy?

[Cecchetti and Rich (2001).]

• Did 2001’s largest federal tax cut in U.S history contribute to or dampen theconcurrent recession? Or was it irrelevant? (Still to be analyzed.)

• Does attending an elite college bring an expected payoff in lifetime expectedincome sufficient to justify the higher tuition? [Krueger and Dale (2001) andKrueger (2002).]

• Does a voluntary training program produce tangible benefits? Can these benefits

be accurately measured? [Angrist (2001).]

Each of these analyses would depart from a formal model of the process underlying theobserved data

The field of econometrics is large and rapidly growing In one dimension, wecan distinguish between theoretical and applied econometrics Theorists develop newtechniques and analyze the consequences of applying particular methods when the as-sumptions that justify them are not met Applied econometricians are the users of thesetechniques and the analysts of data (real world and simulated) Of course, the distinction

is far from clean; practitioners routinely develop new analytical tools for the purposes of

the study that they are involved in This book contains a heavy dose of theory, but it is rected toward applied econometrics I have attempted to survey techniques, admittedlysome quite elaborate and intricate, that have seen wide use “in the field.”

di-Another loose distinction can be made between microeconometrics and econometrics The former is characterized largely by its analysis of cross section andpanel data and by its focus on individual consumers, firms, and micro-level decision mak-ers Macroeconometrics is generally involved in the analysis of time series data, usually

macro-of broad aggregates such as price levels, the money supply, exchange rates, output, and

so on Once again, the boundaries are not sharp The very large field of financial metrics is concerned with long-time series data and occasionally vast panel data sets,but with a very focused orientation toward models of individual behavior The analysis

econo-of market returns and exchange rate behavior is neither macro- nor microeconometric

in nature, or perhaps it is some of both Another application that we will examine inthis text concerns spending patterns of municipalities, which, again, rests somewherebetween the two fields

Applied econometric methods will be used for estimation of important quantities,analysis of economic outcomes, markets or individual behavior, testing theories, and forforecasting The last of these is an art and science in itself, and (fortunately) the subject

of a vast library of sources Though we will briefly discuss some aspects of forecasting,our interest in this text will be on estimation and analysis of models The presentation,where there is a distinction to be made, will contain a blend of microeconometric andmacroeconometric techniques and applications The first 18 chapters of the book arelargely devoted to results that form the platform of both areas Chapters 19 and 20 focus

on time series modeling while Chapters 21 and 22 are devoted to methods more suited

to cross sections and panels, and used more frequently in microeconometrics Save forsome brief applications, we will not be spending much time on financial econometrics.For those with an interest in this field, I would recommend the celebrated work by

Campbell, Lo, and Mackinlay (1997) It is also necessary to distinguish between time series analysis (which is not our focus) and methods that primarily use time series data.

The former is, like forecasting, a growth industry served by its own literature in manyfields While we will employ some of the techniques of time series analysis, we will spendrelatively little time developing first principles

The techniques used in econometrics have been employed in a widening variety

of fields, including political methodology, sociology [see, e.g., Long (1997)], health nomics, medical research (how do we handle attrition from medical treatment studies?)environmental economics, transportation engineering, and numerous others Practi-tioners in these fields and many more are all heavy users of the techniques described inthis text

The remainder of this book is organized into five parts:

1. Chapters 2 through 9 present the classical linear and nonlinear regression models

We will discuss specification, estimation, and statistical inference

2. Chapters 10 through 15 describe the generalized regression model, panel data

applications, and systems of equations.

3. Chapters 16 through 18 present general results on different methods of estimationincluding maximum likelihood, GMM, and simulation methods Various

estimation frameworks, including non- and semiparametric and Bayesianestimation are presented in Chapters 16 and 18

4. Chapters 19 through 22 present topics in applied econometrics Chapters 19 and 20are devoted to topics in time series modeling while Chapters 21 and 22 are aboutmicroeconometrics, discrete choice modeling, and limited dependent variables

5. Appendices A through D present background material on tools used ineconometrics including matrix algebra, probability and distribution theory,estimation, and asymptotic distribution theory Appendix E presents results oncomputation Appendices A through D are chapter-length surveys of the toolsused in econometrics Since it is assumed that the reader has some previoustraining in each of these topics, these summaries are included primarily for thosewho desire a refresher or a convenient reference We do not anticipate that theseappendices can substitute for a course in any of these subjects The intent of thesechapters is to provide a reasonably concise summary of the results, nearly all ofwhich are explicitly used elsewhere in the book

The data sets used in the numerical examples are described in Appendix F The actualdata sets and other supplementary materials can be downloaded from the website forthe text,

www.prenhall.com/greene

2 THE CLASSICAL MULTIPLE LINEAR REGRESSION

MODEL Q

An econometric study begins with a set of propositions about some aspect of theeconomy The theory specifies a set of precise, deterministic relationships among vari-ables Familiar examples are demand equations, production functions, and macroeco-nomic models The empirical investigation provides estimates of unknown parameters

in the model, such as elasticities or the effects of monetary policy, and usually attempts tomeasure the validity of the theory against the behavior of observable data Once suitablyconstructed, the model might then be used for prediction or analysis of behavior Thisbook will develop a large number of models and techniques used in this framework

The linear regression model is the single most useful tool in the econometrician’s

kit Though to an increasing degree in the contemporary literature, it is often onlythe departure point for the full analysis, it remains the device used to begin almost allempirical research This chapter will develop the model The next several chapters willdiscuss more elaborate specifications and complications that arise in the application oftechniques that are based on the simple models presented here

The multiple linear regression model is used to study the relationship between a

depen-dent variable and one or more independepen-dent variables The generic form of the linear

regression model is

y = f (x1 , x2, , x K ) + ε

where y is the dependent or explained variable and x1 , , x K are the independent

or explanatory variables One’s theory will specify f (x1, x2, , x K ) This function is

commonly called the population regression equation of y on x1 , , x K In this

set-ting, y is the regressand and x k , k= 1, , K, are the regressors or covariates The

underlying theory will specify the dependent and independent variables in the model

It is not always obvious which is appropriately defined as each of these—for

exam-ple, a demand equation, quantity = β1 + price × β2 + income × β3 + ε, and an inverse demand equation, price = γ1 + quantity × γ2 + income × γ3 + u are equally valid rep-

resentations of a market For modeling purposes, it will often prove useful to think interms of “autonomous variation.” One can conceive of movement of the independent

7

variables outside the relationships defined by the model while movement of the dent variable is considered in response to some independent or exogenous stimulus.1The termε is a random disturbance, so named because it “disturbs” an otherwise

depen-stable relationship The disturbance arises for several reasons, primarily because wecannot hope to capture every influence on an economic variable in a model, no matterhow elaborate The net effect, which can be positive or negative, of these omitted factors

is captured in the disturbance There are many other contributors to the disturbance

in an empirical model Probably the most significant is errors of measurement It iseasy to theorize about the relationships among precisely defined variables; it is quiteanother to obtain accurate measures of these variables For example, the difficulty ofobtaining reasonable measures of profits, interest rates, capital stocks, or, worse yet,flows of services from capital stocks is a recurrent theme in the empirical literature

At the extreme, there may be no observable counterpart to the theoretical variable.The literature on the permanent income model of consumption [e.g., Friedman (1957)]provides an interesting example

We assume that each observation in a sample(y i , x i 1 , x i 2 , , x i K ), i = 1, , n, is

generated by an underlying process described by

y i = x i 1 β1+ x i 2 β2+ · · · + x i K β K + ε i The observed value of y i is the sum of two parts, a deterministic part and the randompart, ε i Our objective is to estimate the unknown parameters of the model, use thedata to study the validity of the theoretical propositions, and perhaps use the model to

predict the variable y How we proceed from here depends crucially on what we assume

about the stochastic process that has led to our observations of the data in hand

Example 1.1 discussed a model of consumption proposed by Keynes and his General Theory (1936) The theory that consumption, C, and income, X , are related certainly seems consistent

with the observed “facts” in Figures 1.1 and 2.1 (These data are in Data Table F2.1.) Ofcourse, the linear function is only approximate Even ignoring the anomalous wartime years,

consumption and income cannot be connected by any simple deterministic relationship.

The linear model, C = α + β X, is intended only to represent the salient features of this part

of the economy It is hopeless to attempt to capture every influence in the relationship Thenext step is to incorporate the inherent randomness in its real world counterpart Thus, we

write C = f ( X, ε), where ε is a stochastic element It is important not to view ε as a catchall

for the inadequacies of the model The model including ε appears adequate for the data

not including the war years, but for 1942–1945, something systematic clearly seems to bemissing Consumption in these years could not rise to rates historically consistent with theselevels of income because of wartime rationing A model meant to describe consumption inthis period would have to accommodate this influence

It remains to establish how the stochastic element will be incorporated in the equation

The most frequent approach is to assume that it is additive Thus, we recast the equation

in stochastic terms: C = α + β X + ε This equation is an empirical counterpart to Keynes’s

theoretical model But, what of those anomalous years of rationing? If we were to ignoreour intuition and attempt to “fit” a line to all these data—the next chapter will discuss

at length how we should do that—we might arrive at the dotted line in the figure as our bestguess This line, however, is obviously being distorted by the rationing A more appropriate

1 By this definition, it would seem that in our demand relationship, only income would be an independent variable while both price and quantity would be dependent That makes sense—in a market, price and quantity

are determined at the same time, and do change only when something outside the market changes We will

return to this specific case in Chapter 15.

1949

1950

FIGURE 2.1 Consumption Data, 1940–1950.

specification for these data that accommodates both the stochastic nature of the data andthe special circumstances of the years 1942–1945 might be one that shifts straight down

in the war years, C = α + β X + dwaryears δ w + ε, where the new variable, dwaryearsequals one in1942–1945 and zero in other years and w < ∅.

One of the most useful aspects of the multiple regression model is its ability to identifythe independent effects of a set of variables on a dependent variable Example 2.2describes a common application

A number of recent studies have analyzed the relationship between earnings and tion We would expect, on average, higher levels of education to be associated with higherincomes The simple regression model

however, neglects the fact that most people have higher incomes when they are older thanwhen they are young, regardless of their education Thus, β2 will overstate the marginalimpact of education If age and education are positively correlated, then the regression modelwill associate all the observed increases in income with increases in education A betterspecification would account for the effect of age, as in

It is often observed that income tends to rise less rapidly in the later earning years than inthe early ones To accommodate this possibility, we might extend the model to

We would expectβ3to be positive andβ4to be negative

The crucial feature of this model is that it allows us to carry out a conceptual experimentthat might not be observed in the actual data In the example, we might like to (and could)compare the earnings of two individuals of the same age with different amounts of “education”even if the data set does not actually contain two such individuals How education should be

measured in this setting is a difficult problem The study of the earnings of twins by Ashenfelterand Krueger (1994), which uses precisely this specification of the earnings equation, presents

an interesting approach We will examine this study in some detail in Section 5.6.4

A large literature has been devoted to an intriguing question on this subject Education

is not truly “independent” in this setting Highly motivated individuals will choose to pursuemore education (for example, by going to college or graduate school) than others By thesame token, highly motivated individuals may do things that, on average, lead them to havehigher incomes If so, does a positiveβ2that suggests an association between income andeducation really measure the effect of education on income, or does it reflect the effect ofsome underlying effect on both variables that we have not included in our regression model?

We will revisit the issue in Section 22.4

Let the column vector xk be the n observations on variable x k , k = 1, , K, and semble these data in an n × K data matrix X In most contexts, the first column of X is

as-assumed to be a column of 1s so thatβ1is the constant term in the model Let y be the

n observations, y1, , y n, and letε be the column vector containing the n disturbances.

TABLE 2.1 Assumptions of the Classical Linear Regression Model

A1 Linearity: y i = xi 1 β1+ xi 2 β2+ · · · + xi K β K + εi The model specifies a linear relationship

between y and x1, , x K

A2 Full rank: There is no exact linear relationship among any of the independent variables

in the model This assumption will be necessary for estimation of the parameters of themodel

A3 Exogeneity of the independent variables: E [ ε i | xj 1 , x j 2 , , x j K]= 0 This states that

the expected value of the disturbance at observation i in the sample is not a function of the

independent variables observed at any observation, including this one This means that theindependent variables will not carry useful information for prediction ofε i

A4 Homoscedasticity and nonautocorrelation: Each disturbance,ε ihas the same finitevariance,σ2and is uncorrelated with every other disturbance,ε j This assumption limits thegenerality of the model, and we will want to examine how to relax it in the chapters tofollow

A5 Exogenously generated data: The data in(x j 1 , x j 2 , , x j K ) may be any mixture of

constants and random variables The process generating the data operates outside theassumptions of the model—that is, independently of the process that generatesε i Note that

this extends A3 Analysis is done conditionally on the observed X.

A6 Normal distribution: The disturbances are normally distributed Once again, this is a

convenience that we will dispense with after some analysis of its implications

The model in (2-1) as it applies to all n observations can now be written

y = x1β1+ · · · + xK β K + ε, (2-2)

or in the form of Assumption 1,

ASSUMPTION: y= Xβ + ε. (2-3)

A NOTATIONAL CONVENTION.

Henceforth, to avoid a possibly confusing and cumbersome notation, we will use a

boldface x to denote a column or a row of X Which applies will be clear from the

context In (2-2), xk is the kth column of X Subscripts j and k will be used to denote

columns (variables) It will often be convenient to refer to a single observation in (2-3),which we would write

y i = x

Subscripts i and t will generally be used to denote rows (observations) of X In (2-4), xi

is a column vector that is the transpose of the ith 1 × K row of X.

Our primary interest is in estimation and inference about the parameter vectorβ.

Note that the simple regression model in Example 2.1 is a special case in which X has

only two columns, the first of which is a column of 1s The assumption of linearity of theregression model includes the additive disturbance For the regression to be linear inthe sense described here, it must be of the form in (2-1) either in the original variables

or after some suitable transformation For example, the model

deter-of the two parts is directly observed becauseα and β are unknown.

The linearity assumption is not so narrow as it might first appear In the regression

context, linearity refers to the manner in which the parameters and the disturbance enter

the equation, not necessarily to the relationship among the variables For example, the

equations y = α + βx + ε, y = α + β cos(x) + ε, y = α + β/x + ε, and y = α + β ln x + ε are all linear in some function of x by the definition we have used here In the examples, only x has been transformed, but y could have been as well, as in y = Ax β e ε, which is alinear relationship in the logs of x and y; ln y = α + β ln x + ε The variety of functions

is unlimited This aspect of the model is used in a number of commonly used functional

forms For example, the loglinear model is

ln y = β1+ β2ln X2+ β3ln X3+ · · · + β K ln X K + ε.

This equation is also known as the constant elasticity form as in this equation, the

elasticity of y with respect to changes in x is ∂ ln y/∂ ln x = β , which does not vary

with x k The log linear form is often used in models of demand and production Differentvalues ofβ produce widely varying functions.

Data on the U.S gasoline market for the years 1960—1995 are given in Table F2.2 inAppendix F We will use these data to obtain, among other things, estimates of the income,own price, and cross-price elasticities of demand in this market These data also present aninteresting question on the issue of holding “all other things constant,” that was suggested

in Example 2.2 In particular, consider a somewhat abbreviated model of per capita gasolineconsumption:

ln( G /pop) = β1+ β2ln income + β3ln price G + β4ln P newcars + β5ln P usedcars + ε.

This model will provide estimates of the income and price elasticities of demand for gasolineand an estimate of the elasticity of demand with respect to the prices of new and used cars.What should we expect for the sign ofβ4? Cars and gasoline are complementary goods, so ifthe prices of new cars rise, ceteris paribus, gasoline consumption should fall Or should it? Ifthe prices of new cars rise, then consumers will buy fewer of them; they will keep their usedcars longer and buy fewer new cars If older cars use more gasoline than newer ones, thenthe rise in the prices of new cars would lead to higher gasoline consumption than otherwise,not lower We can use the multiple regression model and the gasoline data to attempt toanswer the question

A semilog model is often used to model growth rates:

ln y t = x

t β + δt + ε t

In this model, the autonomous (at least not explained by the model itself) proportional,

per period growth rate is d ln y /dt = δ Other variations of the general form

an approximation is likely to be useful only over a small range of variation of theindependent variables The translog model discussed in Example 2.4, in contrast, hasproved far more effective as an approximating function

Modern studies of demand and production are usually done in the context of a flexible tional form Flexible functional forms are used in econometrics because they allow analysts

func-to model second-order effects such as elasticities of substitution, which are functions of the

second derivatives of production, cost, or utility functions The linear model restricts these toequal zero, whereas the log linear model (e.g., the Cobb–Douglas model) restricts the inter-esting elasticities to the uninteresting values of –1 or+1 The most popular flexible functional

form is the translog model, which is often interpreted as a second-order approximation to

an unknown functional form [See Berndt and Christensen (1973).] One way to derive it is

as follows We first write y = g( x1, , x K ) Then, ln y = ln g( .) = f ( .) Since by a trivial transformation xk = exp(ln xk) , we interpret the function as a function of the logarithms of the x’s Thus, ln y = f (ln x , , ln x )

Now, expand this function in a second-order Taylor series around the point x= [1, 1, , 1]

so that at the expansion point, the log of each variable is a convenient zero Then

ln y = f (0) +

K



k=1[∂ f (·)/∂ ln x k]| lnx=0ln x k

+12

We will see in Chapter 14 how this feature is studied in practice

Despite its great flexibility, the linear model does not include all the situations weencounter in practice For a simple example, there is no transformation that will reduce

y = α + 1/(β1+ β2x ) + ε to linearity The methods we consider in this chapter are not

appropriate for estimating the parameters of such a model Relatively straightforwardtechniques have been developed for nonlinear models such as this, however We shalltreat them in detail in Chapter 9

Assumption 2 is that there are no exact linear relationships among the variables

ASSUMPTION: X is an n × K matrix with rank K. (2-5)

Hence, X has full column rank; the columns of X are linearly independent and there

are at least K observations [See (A-42) and the surrounding text.] This assumption is

known as an identification condition To see the need for this assumption, consider an

example

Suppose that a cross-section model specifies

where total income is exactly equal to salary plus nonlabor income Clearly, there is an exact

linear dependency in the model Now let

where a is any number Then the exact same value appears on the right-hand side of C if

If there are fewer than K observations, then X cannot have full rank Hence, we make

the (redundant) assumption that n is at least as large as K.

In a two-variable linear model with a constant term, the full rank assumption means

that there must be variation in the regressor x If there is no variation in x, then all our

observations will lie on a vertical line This situation does not invalidate the otherassumptions of the model; presumably, it is a flaw in the data set The possibility that

this suggests is that we could have drawn a sample in which there was variation in x,

but in this instance, we did not Thus, the model still applies, but we cannot learn about

it from the data set in hand

There is a subtle point in this discussion that the observant reader might have noted

In (2-7), the left-hand side states, in principle, that the mean of eachε i conditioned on

all observations x i is zero This conditional mean assumption states, in words, that no

observations on x convey information about the expected value of the disturbance.

It is conceivable—for example, in a time-series setting—that although xi might

pro-vide no information about E [ ε i |·], x j at some other observation, such as in the next

time period, might Our assumption at this point is that there is no information about

E [ε i| ·] contained in any observation xj Later, when we extend the model, we willstudy the implications of dropping this assumption [See Woolridge (1995).] We willalso assume that the disturbances convey no information about each other That is,

E [ε i | ε1, , εi –1 , ε i+1, , ε n] = 0 In sum, at this point, we have assumed that thedisturbances are purely random draws from some population

The zero conditional mean implies that the unconditional mean is also zero, since

E [ε i]= Ex[E [ ε i | X]] = Ex[0]= 0.

Since, for eachε i , Cov[E [ ε i | X], X] = Cov[ε i , X], Assumption 3 implies that Cov[ε i , X]=

0 for all i (Exercise: Is the converse true?)

In most cases, the zero mean assumption is not restrictive Consider a two-variablemodel and suppose that the mean of

(α + µ) + βx + (ε – µ) Letting α= α + µ and ε = ε–µ produces the original model.

For an application, see the discussion of frontier production functions in Section 17.6.3

But, if the original model does not contain a constant term, then assuming E [ ε i] = 0

could be substantive If E [ ε i] can be expressed as a linear function of xi, then, as before, atransformation of the model will produce disturbances with zero means But, if not, thenthe nonzero mean of the disturbances will be a substantive part of the model structure.This does suggest that there is a potential problem in models without constant terms As

a general rule, regression models should not be specified without constant terms unlessthis is specifically dictated by the underlying theory.2 Arguably, if we have reason tospecify that the mean of the disturbance is something other than zero, we should build itinto the systematic part of the regression, leaving in the disturbance only the unknownpart ofε Assumption 3 also implies that

The fourth assumption concerns the variances and covariances of the disturbances:

Var[ε i | X] = σ2, for all i = 1, , n,

and

Cov[ε i , ε j | X] = 0, for all i

Constant variance is labeled homoscedasticity Consider a model that describes the

prof-its of firms in an industry as a function of, say, size Even accounting for size, measured indollar terms, the profits of large firms will exhibit greater variation than those of smallerfirms The homoscedasticity assumption would be inappropriate here Also, survey data

on household expenditure patterns often display marked heteroscedasticity, even after

accounting for income and household size

Uncorrelatedness across observations is labeled generically nonautocorrelation In

Figure 2.1, there is some suggestion that the disturbances might not be truly independentacross observations Although the number of observations is limited, it does appearthat, on average, each disturbance tends to be followed by one with the same sign This

“inertia” is precisely what is meant by autocorrelation, and it is assumed away at this

point Methods of handling autocorrelation in economic data occupy a large proportion

of the literature and will be treated at length in Chapter 12 Note that nonautocorrelation

does not imply that observations y i and y j are uncorrelated The assumption is that

deviations of observations from their expected values are uncorrelated.

2Models that describe first differences of variables might well be specified without constants Consider y t – y t–1.

If there is a constant termα on the right-hand side of the equation, then y tis a function ofαt, which is an

explosive regressor Models with linear time trends merit special treatment in the time-series literature We will return to this issue in Chapter 19.