Visit https://ebookfinal.

com to download the full version and

explore more ebook

A Guide to Modern Econometrics 5th Edition Marno

Verbeek

_____ Click the link below to download _____

https://ebookfinal.com/download/a-guide-to-modern-
econometrics-5th-edition-marno-verbeek/

Explore and download more ebook at ebookfinal.com

Here are some recommended products that might interest you.
You can download now and explore!

Introductory Econometrics A Modern Approach 5th Edition

Jeffrey M. Wooldridge

https://ebookfinal.com/download/introductory-econometrics-a-modern-
approach-5th-edition-jeffrey-m-wooldridge/

ebookfinal.com

A Guide to Econometrics 6th Edition Peter Kennedy

https://ebookfinal.com/download/a-guide-to-econometrics-6th-edition-
peter-kennedy/

ebookfinal.com

Introductory Econometrics A Modern Approach 2nd Edition

Jeffrey Wooldridge

https://ebookfinal.com/download/introductory-econometrics-a-modern-
approach-2nd-edition-jeffrey-wooldridge/

ebookfinal.com

Introductory Econometrics A Modern Approach 4th Edition

Jeffrey M. Wooldridge

https://ebookfinal.com/download/introductory-econometrics-a-modern-
approach-4th-edition-jeffrey-m-wooldridge/

ebookfinal.com
Applied Econometrics A Modern Approach Using Eviews and
Microfit Revised Edition Dimitrios Asteriou

https://ebookfinal.com/download/applied-econometrics-a-modern-
approach-using-eviews-and-microfit-revised-edition-dimitrios-asteriou/

ebookfinal.com

A Guide to Teaching Practice Revised 5th Edition Louis

Cohen

https://ebookfinal.com/download/a-guide-to-teaching-practice-
revised-5th-edition-louis-cohen/

ebookfinal.com

Trail guide to the body A hands on guide to locating

muscles bones and more 5th Ed 5th Edition Andrew Biel

https://ebookfinal.com/download/trail-guide-to-the-body-a-hands-on-
guide-to-locating-muscles-bones-and-more-5th-ed-5th-edition-andrew-
biel/
ebookfinal.com

The Will to Whatevs A Guide to Modern Life Eugene Mirman

https://ebookfinal.com/download/the-will-to-whatevs-a-guide-to-modern-
life-eugene-mirman/

ebookfinal.com

The Complete A Guide to PC Repair 5th Edition Cheryl A.

Schmidt

https://ebookfinal.com/download/the-complete-a-guide-to-pc-repair-5th-
edition-cheryl-a-schmidt/

ebookfinal.com
A Guide to Modern Econometrics 5th Edition Marno
Verbeek Digital Instant Download
Author(s): Marno Verbeek
ISBN(s): 9781119401100, 1119401100
Edition: 5
File Details: PDF, 2.88 MB
Year: 2017
Language: english
A Guide to
Modern
Econometrics
Fifth Edition

Marno Verbeek
Rotterdam School of Management, Erasmus University, Rotterdam
Copyright © 2017, 2012, 2008, 2004, 2000 John Wiley & Sons, Inc.

ISBN: 978-1-119-40115-5 (PBK)

ISBN: 978-1-119-40119-3 (EVALC)

Library of Congress Cataloging in Publication Data:

Names: Verbeek, Marno, author.

Title: A guide to modern econometrics / Marno Verbeek, Rotterdam School of
Management, Erasmus University, Rotterdam.
Description: 5th edition. | Hoboken, NJ : John Wiley & Sons, Inc., [2017] |
Includes bibliographical references and index. |
Identifiers: LCCN 2017015272 (print) | LCCN 2017019441 (ebook) | ISBN
9781119401100 (pdf) | ISBN 9781119401117 (epub) | ISBN 9781119401155 (pbk.)
Subjects: LCSH: Econometrics. | Regression analysis.
Classification: LCC HB139 (ebook) | LCC HB139 .V465 2017 (print) | DDC
330.01/5195—dc23
LC record available at https://lccn.loc.gov/2017015272
Contents

Preface xi

1 Introduction 1
1.1 About Econometrics 1
1.2 The Structure of This Book 3
1.3 Illustrations and Exercises 4

2 An Introduction to Linear Regression 6

2.1 Ordinary Least Squares as an Algebraic Tool 7
2.1.1 Ordinary Least Squares 7
2.1.2 Simple Linear Regression 9
2.1.3 Example: Individual Wages 11
2.1.4 Matrix Notation 11
2.2 The Linear Regression Model 12
2.3 Small Sample Properties of the OLS Estimator 15
2.3.1 The Gauss–Markov Assumptions 15
2.3.2 Properties of the OLS Estimator 16
2.3.3 Example: Individual Wages (Continued) 20
2.4 Goodness-of-Fit 20
2.5 Hypothesis Testing 23
2.5.1 A Simple t-Test 23
2.5.2 Example: Individual Wages (Continued) 25
2.5.3 Testing One Linear Restriction 25
2.5.4 A Joint Test of Significance of Regression
Coefficients 26
2.5.5 Example: Individual Wages (Continued) 28
2.5.6 The General Case 29
2.5.7 Size, Power and p-Values 30
2.5.8 Reporting Regression Results 32

k
 k
Trim Size: 7in x 10in Verbeek ftoc.tex V1 - 04/21/2017 3:53 P.M. Page iv

2.6 Asymptotic Properties of the OLS Estimator 33

2.6.1 Consistency 33
2.6.2 Asymptotic Normality 35
2.6.3 Small Samples and Asymptotic Theory 37
2.7 Illustration: The Capital Asset Pricing Model 39
2.7.1 The CAPM as a Regression Model 40
2.7.2 Estimating and Testing the CAPM 41
2.7.3 The World’s Largest Hedge Fund 43
2.8 Multicollinearity 44
2.8.1 Example: Individual Wages (Continued) 47
2.9 Missing Data, Outliers and Influential Observations 48
2.9.1 Outliers and Influential Observations 48
2.9.2 Robust Estimation Methods 50
2.9.3 Missing Observations 51
2.10 Prediction 53
Wrap-up 54
Exercises 55

3 Interpreting and Comparing Regression Models 60

3.1 Interpreting the Linear Model 60
3.2 Selecting the Set of Regressors 65
3.2.1 Misspecifying the Set of Regressors 65
3.2.2 Selecting Regressors 66
k k
3.2.3 Comparing Non-nested Models 71
3.3 Misspecifying the Functional Form 73
3.3.1 Nonlinear Models 73
3.3.2 Testing the Functional Form 74
3.3.3 Testing for a Structural Break 74
3.4 Illustration: Explaining House Prices 76
3.5 Illustration: Predicting Stock Index Returns 79
3.5.1 Model Selection 80
3.5.2 Forecast Evaluation 82
3.6 Illustration: Explaining Individual Wages 85
3.6.1 Linear Models 85
3.6.2 Loglinear Models 88
3.6.3 The Effects of Gender 91
3.6.4 Some Words of Warning 92
Wrap-up 93
Exercises 94

4 Heteroskedasticity and Autocorrelation 97

4.1 Consequences for the OLS Estimator 98
4.2 Deriving an Alternative Estimator 99
4.3 Heteroskedasticity 100
4.3.1 Introduction 100
4.3.2 Estimator Properties and Hypothesis Testing 103

k
 k
Trim Size: 7in x 10in Verbeek ftoc.tex V1 - 04/21/2017 3:53 P.M. Page v

4.3.3 When the Variances Are Unknown 104

4.3.4 Heteroskedasticity-consistent Standard Errors
for OLS 105
4.3.5 Multiplicative Heteroskedasticity 106
4.3.6 Weighted Least Squares with Arbitrary Weights 107
4.4 Testing for Heteroskedasticity 108
4.4.1 Testing for Multiplicative Heteroskedasticity 108
4.4.2 The Breusch–Pagan Test 109
4.4.3 The White Test 109
4.4.4 Which Test? 110
4.5 Illustration: Explaining Labour Demand 110
4.6 Autocorrelation 114
4.6.1 First-order Autocorrelation 116
4.6.2 Unknown 𝜌 118
4.7 Testing for First-order Autocorrelation 119
4.7.1 Asymptotic Tests 119
4.7.2 The Durbin–Watson Test 120
4.8 Illustration: The Demand for Ice Cream 121
4.9 Alternative Autocorrelation Patterns 124
4.9.1 Higher-order Autocorrelation 124
4.9.2 Moving Average Errors 125
4.10 What to Do When You Find Autocorrelation? 126
4.10.1 Misspecification 126
k k
4.10.2 Heteroskedasticity-and-autocorrelation-consistent
Standard Errors for OLS 128
4.11 Illustration: Risk Premia in Foreign Exchange Markets 129
4.11.1 Notation 129
4.11.2 Tests for Risk Premia in the 1-Month Market 131
4.11.3 Tests for Risk Premia Using Overlapping Samples 134
Wrap-up 136
Exercises 136

5 Endogenous Regressors, Instrumental Variables and GMM 139

5.1 A Review of the Properties of the OLS Estimator 140
5.2 Cases Where the OLS Estimator Cannot Be Saved 143
5.2.1 Autocorrelation with a Lagged Dependent Variable 143
5.2.2 Measurement Error in an Explanatory Variable 144
5.2.3 Endogeneity and Omitted Variable Bias 146
5.2.4 Simultaneity and Reverse Causality 148
5.3 The Instrumental Variables Estimator 150
5.3.1 Estimation with a Single Endogenous Regressor
and a Single Instrument 150
5.3.2 Back to the Keynesian Model 155
5.3.3 Back to the Measurement Error Problem 156
5.3.4 Multiple Endogenous Regressors 156
5.4 Illustration: Estimating the Returns to Schooling 157

k
 k
Trim Size: 7in x 10in Verbeek ftoc.tex V1 - 04/21/2017 3:53 P.M. Page vi

5.5 Alternative Approaches to Estimate Causal Effects 162

5.6 The Generalized Instrumental Variables Estimator 163
5.6.1 Multiple Endogenous Regressors with an Arbitrary
Number of Instruments 163
5.6.2 Two-stage Least Squares and the Keynesian Model
Again 167
5.6.3 Specification Tests 168
5.6.4 Weak Instruments 169
5.6.5 Implementing and Reporting Instrumental Variables
Estimators 170
5.7 Institutions and Economic Development 171
5.8 The Generalized Method of Moments 175
5.8.1 Example 175
5.8.2 The Generalized Method of Moments 177
5.8.3 Some Simple Examples 179
5.8.4 Weak Identification 180
5.9 Illustration: Estimating Intertemporal Asset Pricing Models 181
Wrap-up 184
Exercises 185

6 Maximum Likelihood Estimation and Specification Tests 187

6.1 An Introduction to Maximum Likelihood 188
k 6.1.1 Some Examples 188 k
6.1.2 General Properties 191
6.1.3 An Example (Continued) 194
6.1.4 The Normal Linear Regression Model 195
6.1.5 The Stochastic Frontier Model 197
6.2 Specification Tests 198
6.2.1 Three Test Principles 198
6.2.2 Lagrange Multiplier Tests 200
6.2.3 An Example (Continued) 203
6.3 Tests in the Normal Linear Regression Model 204
6.3.1 Testing for Omitted Variables 204
6.3.2 Testing for Heteroskedasticity 206
6.3.3 Testing for Autocorrelation 207
6.4 Quasi-maximum Likelihood and Moment Conditions Tests 208
6.4.1 Quasi-maximum Likelihood 208
6.4.2 Conditional Moment Tests 210
6.4.3 Testing for Normality 211
Wrap-up 212
Exercises 212

7 Models with Limited Dependent Variables 215

7.1 Binary Choice Models 216
7.1.1 Using Linear Regression? 216
7.1.2 Introducing Binary Choice Models 216
7.1.3 An Underlying Latent Model 219

k
 k
Trim Size: 7in x 10in Verbeek ftoc.tex V1 - 04/21/2017 3:53 P.M. Page vii

7.1.4 Estimation 219

7.1.5 Goodness-of-Fit 221
7.1.6 Illustration: The Impact of Unemployment Benefits
on Recipiency 223
7.1.7 Specification Tests in Binary Choice Models 226
7.1.8 Relaxing Some Assumptions in Binary Choice
Models 228
7.2 Multiresponse Models 229
7.2.1 Ordered Response Models 230
7.2.2 About Normalization 231
7.2.3 Illustration: Explaining Firms’ Credit Ratings 231
7.2.4 Illustration: Willingness to Pay for Natural Areas 234
7.2.5 Multinomial Models 237
7.3 Models for Count Data 240
7.3.1 The Poisson and Negative Binomial Models 240
7.3.2 Illustration: Patents and R&D Expenditures 244
7.4 Tobit Models 246
7.4.1 The Standard Tobit Model 247
7.4.2 Estimation 249
7.4.3 Illustration: Expenditures on Alcohol and Tobacco
(Part 1) 250
7.4.4 Specification Tests in the Tobit Model 253
7.5 Extensions of Tobit Models 256
k k
7.5.1 The Tobit II Model 256
7.5.2 Estimation 259
7.5.3 Further Extensions 261
7.5.4 Illustration: Expenditures on Alcohol and Tobacco
(Part 2) 262
7.6 Sample Selection Bias 265
7.6.1 The Nature of the Selection Problem 266
7.6.2 Semi-parametric Estimation of the Sample Selection
Model 268
7.7 Estimating Treatment Effects 269
7.7.1 Regression-based Estimators 271
7.7.2 Regression Discontinuity Design 274
7.7.3 Weighting and Matching 276
7.8 Duration Models 278
7.8.1 Hazard Rates and Survival Functions 278
7.8.2 Samples and Model Estimation 281
7.8.3 Illustration: Duration of Bank Relationships 283
Wrap-up 284
Exercises 285

8 Univariate Time Series Models 288

8.1 Introduction 289
8.1.1 Some Examples 289
8.1.2 Stationarity and the Autocorrelation Function 291

k
 k
Trim Size: 7in x 10in Verbeek ftoc.tex V1 - 04/21/2017 3:53 P.M. Page viii

8.2 General ARMA Processes 294

8.2.1 Formulating ARMA Processes 294
8.2.2 Invertibility of Lag Polynomials 297
8.2.3 Common Roots 298
8.3 Stationarity and Unit Roots 299
8.4 Testing for Unit Roots 301
8.4.1 Testing for Unit Roots in a First-order Autoregressive
Model 301
8.4.2 Testing for Unit Roots in Higher-Order Autoregressive
Models 304
8.4.3 Extensions 306
8.4.4 Illustration: Stock Prices and Earnings 307
8.5 Illustration: Long-run Purchasing Power Parity (Part 1) 309
8.6 Estimation of ARMA Models 313
8.6.1 Least Squares 314
8.6.2 Maximum Likelihood 315
8.7 Choosing a Model 316
8.7.1 The Autocorrelation Function 316
8.7.2 The Partial Autocorrelation Function 318
8.7.3 Diagnostic Checking 319
8.7.4 Criteria for Model Selection 319
8.8 Illustration: The Persistence of Inflation 320
k 8.9 Forecasting with ARMA Models 324 k
8.9.1 The Optimal Forecast 324
8.9.2 Forecast Accuracy 327
8.9.3 Evaluating Forecasts 329
8.10 Illustration: The Expectations Theory of the Term Structure 330
8.11 Autoregressive Conditional Heteroskedasticity 335
8.11.1 ARCH and GARCH Models 335
8.11.2 Estimation and Prediction 338
8.11.3 Illustration: Volatility in Daily Exchange Rates 340
8.12 What about Multivariate Models? 342
Wrap-up 343
Exercises 344

9 Multivariate Time Series Models 348

9.1 Dynamic Models with Stationary Variables 349
9.2 Models with Nonstationary Variables 352
9.2.1 Spurious Regressions 352
9.2.2 Cointegration 353
9.2.3 Cointegration and Error-correction Mechanisms 356
9.3 Illustration: Long-run Purchasing Power Parity (Part 2) 358
9.4 Vector Autoregressive Models 360
9.5 Cointegration: the Multivariate Case 364
9.5.1 Cointegration in a VAR 364
9.5.2 Example: Cointegration in a Bivariate VAR 366
9.5.3 Testing for Cointegration 367

k
 k
Trim Size: 7in x 10in Verbeek ftoc.tex V1 - 04/21/2017 3:53 P.M. Page ix

9.5.4 Illustration: Long-run Purchasing Power Parity

(Part 3) 370
9.6 Illustration: Money Demand and Inflation 372
Wrap-up 378
Exercises 379

10 Models Based on Panel Data 382

10.1 Introduction to Panel Data Modelling 383
10.1.1 Efficiency of Parameter Estimators 384
10.1.2 Identification of Parameters 385
10.2 The Static Linear Model 386
10.2.1 The Fixed Effects Model 386
10.2.2 The First-difference Estimator 388
10.2.3 The Random Effects Model 390
10.2.4 Fixed Effects or Random Effects? 394
10.2.5 Goodness-of-Fit 395
10.2.6 Alternative Instrumental Variables Estimators 396
10.2.7 Robust Inference 398
10.2.8 Testing for Heteroskedasticity and Autocorrelation 400
10.2.9 The Fama–MacBeth Approach 402
10.3 Illustration: Explaining Individual Wages 403
10.4 Dynamic Linear Models 405
k 10.4.1 An Autoregressive Panel Data Model 406 k
10.4.2 Dynamic Models with Exogenous Variables 411
10.4.3 Too Many Instruments 412
10.5 Illustration: Explaining Capital Structure 414
10.6 Panel Time Series 419
10.6.1 Heterogeneity 420
10.6.2 First Generation Panel Unit Root Tests 421
10.6.3 Second Generation Panel Unit Root Tests 424
10.6.4 Panel Cointegration Tests 425
10.7 Models with Limited Dependent Variables 426
10.7.1 Binary Choice Models 427
10.7.2 The Fixed Effects Logit Model 428
10.7.3 The Random Effects Probit Model 429
10.7.4 Tobit Models 431
10.7.5 Dynamics and the Problem of Initial Conditions 431
10.7.6 Semi-parametric Alternatives 433
10.8 Incomplete Panels and Selection Bias 433
10.8.1 Estimation with Randomly Missing Data 434
10.8.2 Selection Bias and Some Simple Tests 436
10.8.3 Estimation with Nonrandomly Missing Data 438
10.9 Pseudo Panels and Repeated Cross-sections 439
10.9.1 The Fixed Effects Model 440
10.9.2 An Instrumental Variables Interpretation 441
10.9.3 Dynamic Models 442
Wrap-up 444
Exercises 445

k
 k
Trim Size: 7in x 10in Verbeek ftoc.tex V1 - 04/21/2017 3:53 P.M. Page x

A Vectors and Matrices 450

A.1 Terminology 450
A.2 Matrix Manipulations 451
A.3 Properties of Matrices and Vectors 452
A.4 Inverse Matrices 453
A.5 Idempotent Matrices 454
A.6 Eigenvalues and Eigenvectors 454
A.7 Differentiation 455
A.8 Some Least Squares Manipulations 456

B Statistical and Distribution Theory 458

B.1 Discrete Random Variables 458
B.2 Continuous Random Variables 459
B.3 Expectations and Moments 460
B.4 Multivariate Distributions 461
B.5 Conditional Distributions 462
B.6 The Normal Distribution 463
B.7 Related Distributions 466

Bibliography 468

Index 488
k k

k
 Trim Size: 7in x 10in
k Verbeek fpref.tex V2 - 05/04/2017 4:58 P.M. Page xi

Preface

Emperor Joseph II: “Your work is ingenious. It’s quality work. And there are simply too
many notes, that’s all. Just cut a few and it will be perfect.”
Wolfgang Amadeus Mozart: “Which few did you have in mind, Majesty?”
from the movie Amadeus, 1984 (directed by Milos Forman)

The field of econometrics has developed rapidly in the last three decades, while the
k use of up-to-date econometric techniques has become more and more standard prac- k
tice in empirical work in many fields of economics. Typical topics include unit root
tests, cointegration, estimation by the generalized method of moments, heteroskedasticity
and autocorrelation consistent standard errors, modelling conditional heteroskedasticity,
causal inference and the estimation of treatment effects, models based on panel data,
models with limited dependent variables, endogenous regressors and sample selection.
At the same time econometrics software has become more and more user friendly and
up-to-date. As a consequence, users are able to implement fairly advanced techniques
even without a basic understanding of the underlying theory and without realizing poten-
tial drawbacks or dangers. In contrast, many introductory econometrics textbooks pay
a disproportionate amount of attention to the standard linear regression model under the
strongest set of assumptions. Needless to say that these assumptions are hardly satisfied in
practice (but not really needed either). On the other hand, the more advanced economet-
rics textbooks are often too technical or too detailed for the average economist to grasp the
essential ideas and to extract the information that is needed. This book tries to fill this gap.
The goal of this book is to familiarize the reader with a wide range of topics in modern
econometrics, focusing on what is important for doing and understanding empirical
work. This means that the text is a guide to (rather than an overview of) alternative
techniques. Consequently, it does not concentrate on the formulae behind each technique
(although the necessary ones are given) nor on formal proofs, but on the intuition behind
the approaches and their practical relevance. The book covers a wide range of topics
that is usually not found in textbooks at this level. In particular, attention is paid to
cointegration, the generalized method of moments, models with limited dependent
variables and panel data models. As a result, the book discusses developments in time
series analysis, cross-sectional methods as well as panel data modelling. More than
25 full-scale empirical illustrations are provided in separate sections and subsections,
taken from fields like labour economics, finance, international economics, consumer
behaviour, environmental economics and macro-economics. These illustrations carefully

k
 Trim Size: 7in x 10in
k Verbeek fpref.tex V2 - 05/04/2017 4:58 P.M. Page xii

discuss and interpret econometric analyses of relevant economic problems, and each of
them covers between two and nine pages of the text. As before, data sets are available
through the supporting website of this book. In addition, a number of exercises are of an
empirical nature and require the use of actual data.
This fifth edition builds upon the success of its predecessors. The text has been carefully
checked and updated, taking into account recent developments and insights. It includes
new material on causal inference, the use and limitations of p-values, instrumental vari-
ables estimation and its implementation, regression discontinuity design, standardized
coefficients, and the presentation of estimation results. Several empirical illustrations are
new or updated. For example, Section 5.7 is added containing a new illustration on the
causal effect of institutions on economic development, to illustrate the use of instrumental
variables. Overall, the presentation is meant to be concise and intuitive, providing refer-
ences to primary sources wherever possible. Where relevant, I pay particular attention to
implementation concerns, for example, relating to identification issues. A large number
of new references has been added in this edition to reflect the changes in the text. Increas-
ingly, the literature provides critical surveys and practical guides on how more advanced
econometric techniques, like robust standard errors, sample selection models or causal
inference methods, are used in specific areas, and I have tried to refer to them in the
text too.
This text originates from lecture notes used for courses in Applied Econometrics in the
M.Sc. programmes in Economics at K. U. Leuven and Tilburg University. It is written for
an intended audience of economists and economics students that would like to become
familiar with up-to-date econometric approaches and techniques, important for doing,
k understanding and evaluating empirical work. It is very well suited for courses in applied k
econometrics at the master’s or graduate level. At some schools this book will be suited
for one or more courses at the undergraduate level, provided students have a sufficient
background in statistics. Some of the later chapters can be used in more advanced courses
covering particular topics, for example, panel data, limited dependent variable models or
time series analysis. In addition, this book can serve as a guide for managers, research
economists and practitioners who want to update their insufficient or outdated knowledge
of econometrics. Throughout, the use of matrix algebra is limited.
I am very much indebted to Arie Kapteyn, Bertrand Melenberg, Theo Nijman and
Arthur van Soest, who all have contributed to my understanding of econometrics and
have shaped my way of thinking about many issues. The fact that some of their ideas
have materialized in this text is a tribute to their efforts. I also owe many thanks to
several generations of students who helped me to shape this text into its current form.
I am very grateful to a large number of people who read through parts of the manuscript
and provided me with comments and suggestions on the basis of the first three editions.
In particular, I wish to thank Niklas Ahlgren, Sascha Becker, Peter Boswijk, Bart
Capéau, Geert Dhaene, Tom Doan, Peter de Goeij, Joop Huij, Ben Jacobsen, Jan Kiviet,
Wim Koevoets, Erik Kole, Marco Lyrio, Konstantijn Maes, Wessel Marquering, Bertrand
Melenberg, Paulo Nunes, Anatoly Peresetsky, Francesco Ravazzolo, Regina Riphahn,
Max van de Sande Bakhuyzen, Erik Schokkaert, Peter Sephton, Arthur van Soest,
Ben Tims, Frederic Vermeulen, Patrick Verwijmeren, Guglielmo Weber, Olivier
Wolthoorn, Kuo-chun Yeh and a number of anonymous reviewers. Of course I retain
sole responsibility for any remaining errors. Special thanks go to Jean-Francois Flechet
for his help with many empirical illustrations and his constructive comments on many
early drafts. Finally, I want to thank my wife Marcella and our three children, Timo,
Thalia and Tamara, for their patience and understanding for all the times that my mind
was with this book when it should have been with them.

k
 Trim Size: 7in x 10in
k Verbeek c01.tex V2 - 04/22/2017 6:44 A.M. Page 1

1 Introduction

1.1 About Econometrics

Economists are frequently interested in relationships between different quantities, for
example between individual wages and the level of schooling. The most important job of
econometrics is to quantify these relationships on the basis of available data and using
statistical techniques, and to interpret, use or exploit the resulting outcomes appropriately.
k Consequently, econometrics is the interaction of economic theory, observed data and sta- k
tistical methods. It is the interaction of these three that makes econometrics interesting,
challenging and, perhaps, difficult. In the words of a seminar speaker, several years ago:
‘Econometrics is much easier without data’.
Traditionally econometrics has focused upon aggregate economic relationships.
Macro-economic models consisting of several up to many hundreds of equations
were specified, estimated and used for policy evaluation and forecasting. The recent
theoretical developments in this area, most importantly the concept of cointegration,
have generated increased attention to the modelling of macro-economic relationships
and their dynamics, although typically focusing on particular aspects of the economy.
Since the 1970s econometric methods have increasingly been employed in micro-
economic models describing individual, household or firm behaviour, stimulated by the
development of appropriate econometric models and estimators that take into account
problems like discrete dependent variables and sample selection, by the availability of
large survey data sets and by the increasing computational possibilities. More recently,
the empirical analysis of financial markets has required and stimulated many theoretical
developments in econometrics. Currently econometrics plays a major role in empirical
work in all fields of economics, almost without exception, and in most cases it is no
longer sufficient to be able to run a few regressions and interpret the results. As a result,
introductory econometrics textbooks usually provide insufficient coverage for applied
researchers. On the other hand, the more advanced econometrics textbooks are often too
technical or too detailed for the average economist to grasp the essential ideas and to
extract the information that is needed. Thus there is a need for an accessible textbook
that discusses the recent and relatively more advanced developments.

k
 Trim Size: 7in x 10in
k Verbeek c01.tex V2 - 04/22/2017 6:44 A.M. Page 2

2 INTRODUCTION

The relationships that economists are interested in are formally specified in mathemat-
ical terms, which lead to econometric or statistical models. In such models there is room
for deviations from the strict theoretical relationships owing to, for example, measure-
ment errors, unpredictable behaviour, optimization errors or unexpected events. Broadly,
econometric models can be classified into a number of categories.
A first class of models describes relationships between present and past. For example,
how does the short-term interest rate depend on its own history? This type of model, typ-
ically referred to as a time series model, usually lacks any economic theory and is mainly
built to get forecasts for future values and the corresponding uncertainty or volatility.
A second type of model considers relationships between economic quantities over a
certain time period. These relationships give us information on how (aggregate) economic
quantities fluctuate over time in relation to other quantities. For example, what happens
to the long-term interest rate if the monetary authority adjusts the short-term one? These
models often give insight into the economic processes that are operating.
Thirdly, there are models that describe relationships between different variables mea-
sured at a given point in time for different units (e.g. households or firms). Most of the
time, this type of relationship is meant to explain why these units are different or behave
differently. For example, one can analyse to what extent differences in household savings
can be attributed to differences in household income. Under particular conditions, these
cross-sectional relationships can be used to analyse ‘what if’ questions. For example, how
much more would a given household, or the average household, save if income were to
increase by 1%?
Finally, one can consider relationships between different variables measured for differ-
k ent units over a longer time span (at least two periods). These relationships simultane- k
ously describe differences between different individuals (why does person 1 save much
more than person 2?), and differences in behaviour of a given individual over time (why
does person 1 save more in 1992 than in 1990?). This type of model usually requires panel
data, repeated observations over the same units. They are ideally suited for analysing pol-
icy changes on an individual level, provided that it can be assumed that the structure of
the model is constant into the (near) future.
The job of econometrics is to specify and quantify these relationships. That is, econo-
metricians formulate a statistical model, usually based on economic theory, confront it
with the data and try to come up with a specification that meets the required goals. The
unknown elements in the specification, the parameters, are estimated from a sample of
available data. Another job of the econometrician is to judge whether the resulting model
is ‘appropriate’. That is, to check whether the assumptions made to motivate the estima-
tors (and their properties) are correct, and to check whether the model can be used for its
intended purpose. For example, can it be used for prediction or analysing policy changes?
Often, economic theory implies that certain restrictions apply to the model that is esti-
mated. For example, the efficient market hypothesis implies that stock market returns are
not predictable from their own past. An important goal of econometrics is to formulate
such hypotheses in terms of the parameters in the model and to test their validity.
The number of econometric techniques that can be used is numerous, and their valid-
ity often depends crucially upon the validity of the underlying assumptions. This book
attempts to guide the reader through this forest of estimation and testing procedures, not
by describing the beauty of all possible trees, but by walking through this forest in a
structured way, skipping unnecessary side-paths, stressing the similarity of the different
species that are encountered and pointing out dangerous pitfalls. The resulting walk is
hopefully enjoyable and prevents the reader from getting lost in the econometric forest.

k
 Trim Size: 7in x 10in
k Verbeek c01.tex V2 - 04/22/2017 6:44 A.M. Page 3

THE STRUCTURE OF THIS BOOK 3

1.2 The Structure of This Book

The first part of this book consists of Chapters 2, 3 and 4. Like most textbooks, it starts
with discussing the linear regression model and the OLS estimation method. Chapter 2
presents the basics of this important estimation method, with some emphasis on its valid-
ity under fairly weak conditions, while Chapter 3 focuses on the interpretation of the
models and the comparison of alternative specifications. Chapter 4 considers two partic-
ular deviations from the standard assumptions of the linear model: autocorrelation and
heteroskedasticity of the error terms. It is discussed how one can test for these phenom-
ena, how they affect the validity of the OLS estimator and how this can be corrected.
This includes a critical inspection of the model specification, the use of adjusted standard
errors for the OLS estimator and the use of alternative (GLS) estimators. These three
chapters are essential for the remaining part of this book and should be the starting point
in any course.
In Chapter 5 another deviation from the standard assumptions of the linear model is
discussed, which is, however, fatal for the OLS estimator. As soon as the error term in
the model is correlated with one or more of the explanatory variables, all good properties
of the OLS estimator disappear, and we necessarily have to use alternative approaches.
This raises the challenge of identifying causal effects with nonexperimental data. The
chapter discusses instrumental variable (IV) estimators and, more generally, the gen-
eralized method of moments (GMM). This chapter, at least its earlier sections, is also
recommended as an essential part of any econometrics course.
Chapter 6 is mainly theoretical and discusses maximum likelihood (ML) estimation.
k Because in empirical work maximum likelihood is often criticized for its dependence k
upon distributional assumptions, it is not discussed in the earlier chapters where alter-
natives are readily available that are either more robust than maximum likelihood or
(asymptotically) equivalent to it. Particular emphasis in Chapter 6 is on misspecification
tests based upon the Lagrange multiplier principle. While many empirical studies tend
to take the distributional assumptions for granted, their validity is crucial for consistency
of the estimators that are employed and should therefore be tested. Often these tests are
relatively easy to perform, although most software does not routinely provide them (yet).
Chapter 6 is crucial for understanding Chapter 7 on limited dependent variable models
and for a small number of sections in Chapters 8 to 10.
The last part of this book contains four chapters. Chapter 7 presents models that are
typically (though not exclusively) used in micro-economics, where the dependent vari-
able is discrete (e.g. zero or one), partly discrete (e.g. zero or positive) or a duration. This
chapter covers probit, logit and tobit models and their extensions, as well as models for
count data and duration models. It also includes a critical discussion of the sample selec-
tion problem. Particular attention is paid to alternative approaches to estimate the causal
impact of a treatment upon an outcome variable in case the treatment is not randomly
assigned (‘treatment effects’).
Chapters 8 and 9 discuss time series modelling including unit roots, cointegration and
error-correction models. These chapters can be read immediately after Chapter 4 or 5,
with the exception of a few parts that relate to maximum likelihood estimation. The
theoretical developments in this area over the last three decades have been substantial,
and many recent textbooks seem to focus upon it almost exclusively. Univariate time
series models are covered in Chapter 8. In this case, models are developed that explain an
economic variable from its own past. These include ARIMA models, as well as GARCH
models for the conditional variance of a series. Multivariate time series models that

k
 Trim Size: 7in x 10in
k Verbeek c01.tex V2 - 04/22/2017 6:44 A.M. Page 4

4 INTRODUCTION

consider several variables simultaneously are discussed in Chapter 9. These include

vector autoregressive models, cointegration and error-correction models.
Finally, Chapter 10 covers models based on panel data. Panel data are available if
we have repeated observations of the same units (e.g. households, firms or countries).
Over recent decades the use of panel data has become important in many areas of eco-
nomics. Micro-economic panels of households and firms are readily available and, given
the increase in computing resources, more manageable than in the past. In addition, it has
become increasingly common to pool time series of several countries. One of the reasons
for this may be that researchers believe that a cross-sectional comparison of countries
provides interesting information, in addition to a historical comparison of a country with
its own past. This chapter also discusses the recent developments on unit roots and coin-
tegration in a panel data setting. Furthermore, a separate section is devoted to repeated
cross-sections and pseudo panel data.
At the end of the book the reader will find two short appendices discussing mathemati-
cal and statistical results that are used in several places in the book. This includes a discus-
sion of some relevant matrix algebra and distribution theory. In particular, a discussion
of properties of the (bivariate) normal distribution, including conditional expectations,
variances and truncation, is provided.
In my experience the material in this book is too much to be covered in a single course.
Different courses can be scheduled on the basis of the chapters that follow. For example,
a typical graduate course in applied econometrics would cover Chapters 2, 3, 4 and parts
of Chapter 5, and then continue with selected parts of Chapters 8 and 9 if the focus is
k on time series analysis, or continue with Section 6.1 and Chapter 7 if the focus is on k
cross-sectional models. A more advanced undergraduate or graduate course may focus
attention on the time series chapters (Chapters 8 and 9), the micro-econometric chapters
(Chapters 6 and 7) or panel data (Chapter 10 with some selected parts from Chapters 6
and 7).
Given the focus and length of this book, I had to make many choices concerning which
material to present or not. As a general rule I did not want to bother the reader with
details that I considered not essential or not to have empirical relevance. The main goal
was to give a general and comprehensive overview of the different methodologies and
approaches, focusing on what is relevant for doing and understanding empirical work.
Some topics are only very briefly mentioned, and no attempt is made to discuss them at
any length. To compensate for this I have tried to give references in appropriate places to
other sources, including specialized textbooks, survey articles and chapters, and guides
with advice for practitioners.

1.3 Illustrations and Exercises

In most chapters a variety of empirical illustrations are provided in separate sections
or subsections. While it is possible to skip these illustrations essentially without losing
continuity, these sections do provide important aspects concerning the implementation of
the methodology discussed in the preceding text. In addition, I have attempted to provide
illustrations that are of economic interest in themselves, using data that are typical of
current empirical work and cover a wide range of different areas. This means that most
data sets are used in recently published empirical work and are fairly large, both in terms

k
 Trim Size: 7in x 10in
k Verbeek c01.tex V2 - 04/22/2017 6:44 A.M. Page 5

ILLUSTRATIONS AND EXERCISES 5

of number of observations and in terms of number of variables. Given the current state of
computing facilities, it is usually not a problem to handle such large data sets empirically.
Learning econometrics is not just a matter of studying a textbook. Hands-on experience
is crucial in the process of understanding the different methods and how and when to
implement them. Therefore, readers are strongly encouraged to get their hands dirty
and to estimate a number of models using appropriate or inappropriate methods, and
to perform a number of alternative specification tests. With modern software becoming
more and more user friendly, the actual computation of even the more complicated
estimators and test statistics is often surprisingly simple, sometimes dangerously simple.
That is, even with the wrong data, the wrong model and the wrong methodology,
programmes may come up with results that are seemingly all right. At least some
expertise is required to prevent the practitioner from such situations, and this book plays
an important role in this.
To stimulate the reader to use actual data and estimate some models, almost all data
sets used in this text are available through the website www.wileyeurope.com/college/
verbeek. Readers are encouraged to re-estimate the models reported in this text and check
whether their results are the same, as well as to experiment with alternative specifications
or methods. Some of the exercises make use of the same or additional data sets and pro-
vide a number of specific issues to consider. It should be stressed that, for estimation
methods that require numerical optimization, alternative programmes, algorithms or set-
tings may give slightly different outcomes. However, you should get results that are close
to the ones reported.
I do not advocate the use of any particular software package. For the linear regression
k k
model any package will do, while for the more advanced techniques each package has
its particular advantages and disadvantages. There is typically a trade-off between user-
friendliness and flexibility. Menu-driven packages often do not allow you to compute
anything other than what’s on the menu, but, if the menu is sufficiently rich, that may not
be a problem. Command-driven packages require somewhat more input from the user,
but are typically quite flexible. For the illustrations in the text, I made use of Eviews,
RATS and Stata. Several alternative econometrics programmes are available, including
MicroFit, PcGive, TSP and SHAZAM; for more advanced or tailored methods, econo-
metricians make use of GAUSS, Matlab, Ox, S-Plus and many other programmes, as
well as specialized software for specific methods or types of model. Journals like the
Journal of Applied Econometrics and the Journal of Economic Surveys regularly publish
software reviews.
The exercises included at the end of each chapter consist of a number of questions
that are primarily intended to check whether the reader has grasped the most important
concepts. Therefore, they typically do not go into technical details or ask for derivations
or proofs. In addition, several exercises are of an empirical nature and require the reader
to use actual data, made available through the book’s website.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 6

2 An Introduction to
Linear Regression

The linear regression model in combination with the method of ordinary least squares
(OLS) is one of the cornerstones of econometrics. In the first part of this book we
shall review the linear regression model with its assumptions, how it can be estimated,
k evaluated and interpreted and how it can be used for generating predictions and for k
testing economic hypotheses.
This chapter starts by introducing the ordinary least squares method as an algebraic tool,
rather than a statistical one. This is because OLS has the attractive property of providing
a best linear approximation, irrespective of the way in which the data are generated, or
any assumptions imposed. The linear regression model is then introduced in Section 2.2,
while Section 2.3 discusses the properties of the OLS estimator in this model under the
so-called Gauss–Markov assumptions. Section 2.4 discusses goodness-of-fit measures
for the linear model, and hypothesis testing is treated in Section 2.5. In Section 2.6,
we move to cases where the Gauss–Markov conditions are not necessarily satisfied
and the small sample properties of the OLS estimator are unknown. In such cases,
the limiting behaviour of the OLS estimator when – hypothetically – the sample size
becomes infinitely large is commonly used to approximate its small sample properties.
An empirical example concerning the capital asset pricing model (CAPM) is provided
in Section 2.7. Sections 2.8 and 2.9 discuss data problems related to multicollinearity,
outliers and missing observations, while Section 2.10 pays attention to prediction using
a linear regression model. Throughout, an empirical example concerning individual
wages is used to illustrate the main issues. Additional discussion on how to interpret the
coefficients in the linear model, how to test some of the model’s assumptions and how to
compare alternative models is provided in Chapter 3, which also contains three extensive
empirical illustrations.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 7

ORDINARY LEAST SQUARES AS AN ALGEBRAIC TOOL 7

2.1 Ordinary Least Squares as an Algebraic Tool

2.1.1 Ordinary Least Squares
Suppose we have a sample with N observations on individual wages and a number of
background characteristics, like gender, years of education and experience. Our main
interest lies in the question as to how in this sample wages are related to the other observ-
ables. Let us denote wages by y (the regressand) and the other K − 1 characteristics by
x2 , . . . , xK (the regressors). It will become clear below why this numbering of variables
is convenient. Now we may ask the question: which linear combination of x2 , . . . , xK and
a constant gives a good approximation of y? To answer this question, first consider an
arbitrary linear combination, including a constant, which can be written as

𝛽̃1 + 𝛽̃2 x2 + · · · + 𝛽̃K xK , (2.1)

where 𝛽̃1 , . . . , 𝛽̃K are constants to be chosen. Let us index the observations by i such
that i = 1, . . . , N. Now, the difference between an observed value yi and its linear
approximation is
yi − [𝛽̃1 + 𝛽̃2 xi2 + · · · + 𝛽̃K xiK ]. (2.2)

To simplify the derivations we shall introduce some shorthand notation. Appendix A

provides additional details for readers unfamiliar with the use of vector notation. The
special case of K = 2 is discussed in the next subsection. For general K we collect the
x-values for individual i in a vector xi , which includes the constant. That is,
k k
xi = (1 xi2 xi3 . . . xiK )

where  is used to denote a transpose. Collecting the 𝛽̃ coefficients in a K-dimensional

vector 𝛽̃ = (𝛽̃1 . . . 𝛽̃K ) , we can briefly write (2.2) as
̃
yi − xi 𝛽. (2.3)

Clearly, we would like to choose values for 𝛽̃1 , . . . , 𝛽̃K such that these differences
are small. Although different measures can be used to define what we mean by
‘small’, the most common approach is to choose 𝛽̃ such that the sum of squared
differences is as small as possible. In this case we determine 𝛽̃ to minimize the following
objective function:
∑N
̃
S(𝛽) ≡ ̃ 2.
(yi − xi 𝛽) (2.4)
i=1

That is, we minimize the sum of squared approximation errors. This approach is referred
to as the ordinary least squares or OLS approach. Taking squares makes sure that pos-
itive and negative deviations do not cancel out when taking the summation.
To solve the minimization problem, we consider the first-order conditions, obtained
̃ with respect to the vector 𝛽.
by differentiating S(𝛽) ̃ (Appendix A discusses some
rules on how to differentiate a scalar expression, like (2.4), with respect to a vector.)

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 8

8 AN INTRODUCTION TO LINEAR REGRESSION

This gives the following system of K conditions:

∑
N
−2 ̃ =0
xi (yi − xi 𝛽) (2.5)
i=1

or (N )
∑ ∑
N
xi xi 𝛽̃ = xi yi . (2.6)
i=1 i=1

These equations are sometimes referred to as normal equations. As this system has K
unknowns,
∑N one can obtain a unique solution for 𝛽̃ provided that the symmetric matrix

i=1 xi xi , which contains sums of squares and cross-products of the regressors xi , can
be inverted. For the moment, we shall assume that this is the case. The solution to the
minimization problem, which we shall denote by b, is then given by
(N )−1 N
∑  ∑
b= xi x i xi yi . (2.7)
i=1 i=1

By checking the second-order conditions, it is easily verified that b indeed corresponds

to a minimum of (2.4).
The resulting linear combination of xi is thus given by

ŷ i = xi b,
k k
which is the best linear approximation of y from x2 , . . . , xK and a constant. The phrase
‘best’ refers to the fact that the sum of squared differences between the observed values
yi and fitted values ŷ i is minimal for the least squares solution b.
In deriving the linear approximation, we have not used any economic or statistical
theory. It is simply an algebraic tool, and it holds irrespective of the way the data are
generated. That is, given a set of variables we can always determine the best linear
approximation of one variable using the other variables. The only assumption that
we
∑N had to make (which is directly checked from the data) is that the K × K matrix
i=1 xi xi is invertible. This says that none of the xik s is an exact linear combination of
the other ones and thus redundant. This is usually referred to as the no-multicollinearity
assumption. It should be stressed that the linear approximation is an in-sample
result (i.e. in principle it does not give information about observations (individuals)
that are not included in the sample) and, in general, there is no direct interpretation of
the coefficients.
Despite these limitations, the algebraic results on the least squares method are very use-
ful. Defining a residual ei as the difference between the observed and the approximated
value, ei = yi − ŷ i = yi − xi b, we can decompose the observed yi as

yi = ŷ i + ei = xi b + ei . (2.8)

This allows us to write the minimum value for the objective function as

∑
N
S(b) = e2i , (2.9)
i=1

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 9

ORDINARY LEAST SQUARES AS AN ALGEBRAIC TOOL 9

which is referred to as the residual sum of squares. It can be shown that the approximated
value xi b and the residual ei satisfy certain properties by construction. For example, if
we rewrite (2.5), substituting the OLS solution b, we obtain

∑
N
∑
N
xi (yi − xi b) = xi ei = 0. (2.10)
i=1 i=1

This means that the vector e = (e1 , . . . , eN ) is orthogonal1 to each vector of observa-
∑
tions on an x-variable. For example, if xi contains a constant, it implies that Ni=1 ei = 0.
That is, the average residual is zero. This is an intuitively appealing result. If the average
residual were nonzero, this would mean that we could improve upon the approximation
by adding or subtracting the same constant for each observation, that is, by changing b1 .
Consequently, for the average observation it follows that

ȳ = x̄  b, (2.11)
∑N ∑N
where ȳ = (1∕N) i=1 yi and x̄ = (1∕N) i=1 xi , a K-dimensional vector of sample
means. This shows that for the average observation there is no approximation error. Sim-
ilar interpretations hold for the other regressors: if the derivative
∑ of the sum of squared
approximation errors with respect to 𝛽̃k is positive, that is if Ni=1 xik ei > 0, it means that
we can improve the objective function in (2.4) by decreasing 𝛽̃k . Equation (2.8) thus
decomposes the observed value of yi into two orthogonal components: the fitted value
(related to xi ) and the residual.
k k
2.1.2 Simple Linear Regression
In the case where K = 2 we only have one regressor and a constant. In this case, the obser-
vations2 (yi , xi ) can be drawn in a two-dimensional graph with x-values on the horizontal
axis and y-values on the vertical one. This is done in Figure 2.1 for a hypothetical data
set. The best linear approximation of y from x and a constant is obtained by minimizing
the sum of squared residuals, which – in this two-dimensional case – equals the vertical
distances between an observation and the fitted value. All fitted values are on a straight
line, the regression line.
Because a 2 × 2 matrix can be inverted analytically, we can derive solutions for b1
and b2 in this special case from the general expression for b above. Equivalently, we
can minimize the residual sum of squares with respect to the unknowns directly. Thus
we have
∑
N
S(𝛽̃1 , 𝛽̃2 ) = (yi − 𝛽̃1 − 𝛽̃2 xi )2 . (2.12)
i=1

The basic elements in the derivation of the OLS solutions are the first-order conditions
𝜕S(𝛽̃1 , 𝛽̃2 ) ∑ N
= −2 (yi − 𝛽̃1 − 𝛽̃2 xi ) = 0, (2.13)
𝜕 𝛽̃1 i=1

∑
1
Two vectors x and y are said to be orthogonal if x y = 0, that is if i xi yi = 0 (see Appendix A).
2
In this subsection, xi will be used to denote the single regressor, so that it does not include the constant.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 10

10 AN INTRODUCTION TO LINEAR REGRESSION

.2

0
y

–.2

–.1 0 .1 .2 .3
x

Figure 2.1 Simple linear regression: fitted line and observation points.

𝜕S(𝛽̃1 , 𝛽̃2 ) ∑ N
= −2 xi (yi − 𝛽̃1 − 𝛽̃2 xi ) = 0. (2.14)
𝜕 𝛽̃
k 2 i=1 k
From (2.13) we can write

1∑ 1∑
N N
b1 = y − b2 x = ȳ − b2 x̄ , (2.15)
N i=1 i N i=1 i

where b2 is solved from combining (2.14) and (2.15). First, from (2.14) we write
(N )
∑
N
∑
N
∑ 2
xi yi − b1 xi − xi b2 = 0
i=1 i=1 i=1

and then substitute (2.15) to obtain

( )
∑
N
∑
N
xi yi − N x̄ ȳ − xi2 − N x̄ 2
b2 = 0
i=1 i=1

such that we can solve for the slope coefficient b2 as

∑N
(xi − x̄ )(yi − ȳ )
b2 = i=1∑N . (2.16)
i=1 (xi − x̄ )2
By dividing both numerator and denominator by N − 1 it appears that the OLS solution
b2 is the ratio of the sample covariance between x and y and the sample variance of x.
From (2.15), the intercept is determined so as to make the average approximation error
(residual) equal to zero.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 11

ORDINARY LEAST SQUARES AS AN ALGEBRAIC TOOL 11

2.1.3 Example: Individual Wages

An example that will appear at several places in this chapter is based on a sample of indi-
vidual wages with background characteristics, like gender, race and years of schooling.
We use a subsample of the US National Longitudinal Survey (NLS) that relates to 1987,
and we have a sample of 3294 young working individuals, of which 1569 are females.
The average hourly wage rate in this sample equals $6.31 for males and $5.15 for females.
Now suppose we try to approximate wages by a linear combination of a constant and a
0–1 variable denoting whether the individual is male. That is, xi = 1 if individual i is male
and zero otherwise. Such a variable that can only take on the values of zero and one is
called a dummy variable. Using the OLS approach the result is
ŷ i = 5.15 + 1.17xi .
This means that for females our best approximation is $5.15 and for males it is $5.15 +
$1.17 = $6.31. It is not a coincidence that these numbers are exactly equal to the sample
means in the two subsamples. It is easily verified from the results above that
b1 = ȳ f
b2 = ȳ m − ȳ f
∑ ∑
where ȳ m = i xi yi ∕ i xi is the sample average of the wage for males, and ȳ f =
∑ ∑
i (1 − xi )yi ∕ i (1 − xi ) is the average for females.

k 2.1.4 Matrix Notation k

Because econometricians make frequent use of matrix expressions as shorthand notation,
some familiarity with this matrix ‘language’ is a prerequisite to reading the economet-
rics literature. In this text, we shall regularly rephrase results using matrix notation,
and occasionally, when the alternative is extremely cumbersome, restrict attention
to matrix expressions only. Using matrices, deriving the least squares solution is
faster, but it requires some knowledge of matrix differential calculus. We introduce the
following notation:
⎛1 x12 . . . x1K ⎞ ⎛ x  ⎞ ⎛ y1 ⎞
⎜. . .. = .. , y = ⎜ .. ⎟ .
⎟ ⎜ 1
⎟
X = ⎜ .. .. . ⎟ ⎜.⎟ ⎜.⎟
⎜1 x . . . x ⎟ ⎜x  ⎟ ⎜y ⎟
⎝ N2 NK ⎠ ⎝ N⎠ ⎝ N⎠
So, in the N × K matrix X the ith row refers to observation i, and the kth column refers to
the kth explanatory variable (regressor). The criterion to be minimized, as given in (2.4),
can be rewritten in matrix notation using the fact that the inner product of a vector a with
itself (a a) is the sum of its squared elements (see Appendix A). That is,
̃ = (y − X 𝛽)
S(𝛽) ̃  (y − X 𝛽)
̃ = y y − 2y X 𝛽̃ + 𝛽̃ X  X 𝛽,
̃ (2.17)
from which the least squares solution follows from differentiating3 with respect to 𝛽̃ and
setting the result to zero:
̃
𝜕S(𝛽) ̃ = 0.
= −2(X  y − X  X 𝛽) (2.18)
𝜕𝛽̃
3
See Appendix A for some rules for differentiating matrix expressions with respect to vectors.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 12

12 AN INTRODUCTION TO LINEAR REGRESSION

Solving (2.18) gives the OLS solution

b = (X  X)−1 X  y, (2.19)

which is exactly the same as the one derived in ∑

(2.7) but now written in matrix notation.
Note that we again have to assume that X  X = Ni=1 xi xi is invertible, that is, there is no
exact (or perfect) multicollinearity.
As before, we can decompose y as

y = Xb + e, (2.20)

where e is an N-dimensional vector of residuals. The first-order conditions imply that

X  (y − Xb) = 0 or
X  e = 0, (2.21)

which means that each column of the matrix X is orthogonal to the vector of residuals.
With (2.19) we can also write (2.20) as

y = Xb + e = X(X  X)−1 X  y + e = ŷ + e (2.22)

so that the predicted value for y is given by

ŷ = Xb = X(X  X)−1 X  y = PX y. (2.23)

In linear algebra, the matrix PX ≡ X(X  X)−1 X  is known as a projection matrix (see
k Appendix A). It projects the vector y upon the columns of X (the column space of X). k
This is just the geometric translation of finding the best linear approximation of y
from the columns (regressors) in X. The matrix PX is also referred to as the ‘hat
matrix’ because it transforms y into ŷ (‘y hat’). The residual vector of the projection
e = y − Xb = (I − PX )y = MX y is the orthogonal complement. It is a projection of y
upon the space orthogonal to the one spanned by the columns of X. This interpretation
is sometimes useful. For example, projecting twice on the same space should leave the
result unaffected, so that it holds that PX PX = PX and MX MX = MX . More importantly,
it holds that MX PX = 0 as the column space of X and its orthogonal complement do
not have anything in common (except the null vector). This is an alternative way to
interpret the result that ŷ and e and also X and e are orthogonal. The interested reader is
referred to Davidson and MacKinnon (2004, Chapter 2) for an excellent discussion on
the geometry of least squares.

2.2 The Linear Regression Model

Usually, economists want more than just finding the best linear approximation of one vari-
able given a set of others. They want economic relationships that are more generally valid
than the sample they happen to have. They want to draw conclusions about what happens
if one of the variables actually changes. That is, they want to say something about values
that are not (yet) included in the sample. For example, we may want to predict the wage
of an individual on the basis of his or her background characteristics and determine how
it would be different if this person had more years of education. In this case, we want the
relationship that is found to be more than just a historical coincidence; it should reflect

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 13

THE LINEAR REGRESSION MODEL 13

a fundamental relationship. To do this it is assumed that there is a general relationship

that is valid for all possible observations from a well-defined population (e.g. all indi-
viduals with a paid job on a given date, or all firms in a certain industry). Restricting
attention to linear relationships, we specify a statistical model as

yi = 𝛽1 + 𝛽2 xi2 + · · · + 𝛽K xiK + 𝜀i (2.24)

or
yi = xi 𝛽 + 𝜀i , (2.25)

where yi and xi are observable variables and 𝜀i is unobserved and referred to as an error
term or disturbance term. In this context, yi is referred to as the dependent variable
and the variables in xi are called independent variables, explanatory variables, regressors
or – occasionally – covariates. The elements in 𝛽 are unknown population parameters.
The equality in (2.25) is supposed to hold for any possible observation, whereas we only
observe a sample of N observations. We consider this sample as one realization of all
potential samples of size N that could have been drawn from the same population. In this
way yi and 𝜀i (and often xi ) can be considered as random variables. Each observation
corresponds to a realization of these random variables. Again we can use matrix notation
and stack all observations to write

y = X𝛽 + 𝜀, (2.26)

where y and 𝜀 are N-dimensional vectors and X, as before, is of dimension N × K. Notice

k the difference between this equation and (2.20). k
In contrast to (2.8) and (2.20), (2.25) and (2.26) are population relationships, where 𝛽
is a vector of unknown parameters characterizing the population. The sampling process
describes how the sample is taken from the population and, as a result, determines the
randomness of the sample. In a first view, the xi variables are considered as fixed and
nonstochastic, which means that every new sample will have the same X matrix. In this
case one refers to xi as being deterministic. A new sample only implies new values for
𝜀i , or – equivalently – for yi . The only relevant case where the xi s are truly deterministic
is in a laboratory setting, where a researcher can set the conditions of a given experi-
ment (e.g. temperature, air pressure). In economics we will typically have to work with
nonexperimental data.4 Despite this, it is convenient and in particular cases appropriate
in an economic context to act as if the xi variables are deterministic. In this case, we
will have to make some assumptions about the sampling distribution of 𝜀i . A convenient
one corresponds to random sampling where each error 𝜀i is a random drawing from
the population distribution, independent of the other error terms. We shall return to this
issue below.
In a second view, a new sample implies new values for both xi and 𝜀i , so that each time
a new set of N observations for (yi , xi ) is drawn. In this case random sampling means
that each set (yi , xi ) is a random drawing from the population distribution. In this context,
it will turn out to be important to make assumptions about the joint distribution of xi and
𝜀i , in particular regarding the extent to which the distribution of 𝜀i is allowed to depend
upon X. The idea of a (random) sample is most easily understood in a cross-sectional
4
In recent years, the use of field experiments in economics has gained popularity, see, for example, Levitt and
List (2009).

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 14

14 AN INTRODUCTION TO LINEAR REGRESSION

context, where interest lies in a large and fixed population, for example all UK house-
holds in January 2015, or all stocks listed at the New York Stock Exchange on a given
date. In a time series context, different observations refer to different time periods, and it
does not make sense to assume that we have a random sample of time periods. Instead,
we shall take the view that the sample we have is just one realization of what could
have happened in a given time span and the randomness refers to alternative states of the
world. In such a case we will need to make some assumptions about the way the data are
generated (rather than the way the data are sampled).
It is important to realize that without additional restrictions the statistical model in
(2.25) is a tautology: for any value of 𝛽 one can always define a set of 𝜀i s such that
(2.25) holds exactly for each observation. We thus need to impose some assumptions
to give the model a meaning. A common assumption is that the expected value of 𝜀i
given all the explanatory variables in xi is zero, that is, E{𝜀i |xi } = 0. Usually, people
refer to this assumption by saying that the explanatory variables are exogenous. Under
this assumption it holds that
E{yi |xi } = xi 𝛽, (2.27)

so that the (population) regression line xi 𝛽 describes the conditional expectation of yi
given the values for xi . The coefficients 𝛽k measure how the expected value of yi is
affected if the value of xik is changed, keeping the other elements in xi constant (the
ceteris paribus condition). Economic theory, however, often suggests that the model in
(2.25) describes a causal relationship, in which the 𝛽 coefficients measure the changes
in yi caused by a ceteris paribus change in xik . In such cases, 𝜀i has an economic inter-
k k
pretation (not just a statistical one) and imposing that it is uncorrelated with xi , as we do
by imposing E{𝜀i |xi } = 0, may not be justified. Because in many applications it can be
argued that unobservables in the error term are related to observables in xi , we should
be cautious interpreting our regression coefficients as measuring causal effects. We shall
come back to these issues in Section 3.1 and, in more detail, in Chapter 5 (‘endogenous
regressors’).
Now that our 𝛽 coefficients have a meaning, we can try to use the sample (yi , xi ), i =
1, . . . , N, to say something about them. The rule that says how a given sample is translated
into an approximate value for 𝛽 is referred to as an estimator. The result for a given
sample is called an estimate. The estimator is a vector of random variables, because the
sample may change. The estimate is a vector of numbers. The most widely used estimator
in econometrics is the ordinary least squares (OLS) estimator. This is just the ordinary
least squares rule described in Section 2.1 applied to the available sample. The OLS
estimator for 𝛽 is thus given by
(N )−1
∑ ∑
N
b= xi xi xi yi . (2.28)
i=1 i=1

Because we have assumed an underlying ‘true’ model (2.25), combined with a sampling
scheme, b is now a vector of random variables. Our interest lies in the true unknown
parameter vector 𝛽, and b is considered an approximation to it. Whereas a given sample
only produces a single estimate, we evaluate the quality of it through the properties of
the underlying estimator. The estimator b has a sampling distribution because its value
depends upon the sample that is taken (randomly) from the population.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 15

SMALL SAMPLE PROPERTIES OF THE OLS ESTIMATOR 15

It is extremely important to understand the difference between the estimator b and the
true population coefficients 𝛽. The first is a vector of random variables, the outcome of
which depends upon the sample that is employed (and, in the more general case, upon the
estimation method that is used). The second is a set of fixed unknown numbers, character-
izing the population model (2.25). Likewise, the distinction between the error terms 𝜀i and
the residuals ei is important. Error terms are unobserved, and distributional assumptions
about them are necessary to derive the sampling properties of estimators for 𝛽. We will
see this in the next section. The residuals are obtained after estimation, and their values
depend upon the estimated value for 𝛽 and therefore depend upon the sample and the
estimation method. The properties of the error terms 𝜀i and the residuals ei are not the
same and occasionally very different. For example, (2.10) is typically not satisfied when
the residuals are replaced by the error terms. Empirical papers are often rather sloppy in
their terminology, referring to the error terms as being ‘residuals’ or using the two terms
interchangeably. In this text, we will be more precise and use ‘error term’ or occasionally
‘disturbance term’ for 𝜀i and ‘residuals’ for ei .

2.3 Small Sample Properties of the OLS Estimator

2.3.1 The Gauss–Markov Assumptions
Whether or not the OLS estimator b provides a good approximation to the unknown
parameter vector 𝛽 depends crucially upon the assumptions that are made about the dis-
k tribution of 𝜀i and its relation to xi . A standard case in which the OLS estimator has k
good properties is characterised by the Gauss–Markov conditions. Later, in Section 2.6,
Chapter 4 and Section 5.1, we shall consider weaker conditions under which OLS still
has some attractive properties. For now, it is important to realize that the Gauss–Markov
conditions are not all strictly needed to justify the use of the ordinary least squares esti-
mator. They just constitute a simple case in which the small sample properties of b are
easily derived.
For the linear regression model in (2.25), given by

yi = xi 𝛽 + 𝜀i ,

the Gauss–Markov conditions are

E{𝜀i } = 0, i = 1, . . . , N (A1)

{𝜀1 , . . . , 𝜀N } and {x1 , . . . , xN } are independent (A2)

V{𝜀i } = 𝜎 2 , i = 1, . . . , N (A3)
cov{𝜀i , 𝜀j } = 0, i, j = 1, . . . , N, i ≠ j. (A4)

Assumption (A1) says that the expected value of the error term is zero, which means
that, on average, the regression line should be correct. Assumption (A3) states that all
error terms have the same variance, which is referred to as homoskedasticity, while
assumption (A4) imposes zero correlation between different error terms. This excludes
any form of autocorrelation. Taken together, (A1), (A3) and (A4) imply that the error
terms are uncorrelated drawings from a distribution with expectation zero and constant

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 16

16 AN INTRODUCTION TO LINEAR REGRESSION

variance 𝜎 2 . Using the matrix notation introduced earlier, it is possible to rewrite these
three conditions as
E{𝜀} = 0 and V{𝜀} = 𝜎 2 IN , (2.29)

where IN is the N × N identity matrix. This says that the covariance matrix of the vector of
error terms 𝜀 is a diagonal matrix with 𝜎 2 on the diagonal. Assumption (A2) implies that
X and 𝜀 are independent. Loosely speaking, this means that knowing X does not tell us
anything about the distribution of the error terms in 𝜀. This is a fairly strong assumption.
It implies that
E{𝜀|X} = E{𝜀} = 0 (2.30)

and
V{𝜀|X} = V{𝜀} = 𝜎 2 IN . (2.31)

That is, the matrix of regressor values X does not provide any information about the
expected values of the error terms or their (co)variances. The two conditions (2.30) and
(2.31) combine the necessary elements from the Gauss–Markov assumptions needed for
the results below to hold. By conditioning on X, we may act as if X were nonstochastic.
The reason for this is that the outcomes in the matrix X can be taken as given without
affecting the properties of 𝜀, that is, one can derive all properties conditional upon X.
For simplicity, we shall take this approach in this section and Section 2.5. Under the
Gauss–Markov assumptions (A1) and (A2), the linear model can be interpreted as the
conditional expectation of yi given xi , that is, E{yi |xi } = xi 𝛽. This is a direct implication
k of (2.30). k

2.3.2 Properties of the OLS Estimator

Under assumptions (A1)–(A4), the OLS estimator b for 𝛽 has several desirable properties.
First of all, it is unbiased. This means that, in repeated sampling, we can expect that the
OLS estimator is on average equal to the true value 𝛽. We formulate this as E{b} = 𝛽.
It is instructive to see the proof:

E{b} = E{(X  X)−1 X  y} = E{𝛽 + (X  X)−1 X  𝜀}

= 𝛽 + E{(X  X)−1 X  𝜀} = 𝛽.

In the second step we have substituted (2.26). The final step is the essential one and
follows from
E{(X  X)−1 X  𝜀} = E{(X  X)−1 X  }E{𝜀} = 0,

because, from assumption (A2), X and 𝜀 are independent and, from (A1), E{𝜀} = 0.
Note that we did not use assumptions (A3) and (A4) in the proof. This shows that the
OLS estimator is unbiased as long as the error terms are mean zero and independent of
all explanatory variables, even if heteroskedasticity or autocorrelation are present. We
shall come back to this issue in Chapter 4. If an estimator is unbiased, this means that its
probability distribution has an expected value that is equal to the true unknown parameter
it is estimating.
In addition to knowing that we are, on average, correct, we would also like to make
statements about how (un)likely it is to be far off in a given sample. This means we

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 17

SMALL SAMPLE PROPERTIES OF THE OLS ESTIMATOR 17

would like to know the distribution of b (around its mean 𝛽). First of all, the variance of
b (conditional upon X) is given by
(N )−1
2  −1
∑ 
V{b|X} = 𝜎 (X X) = 𝜎 2
xi x i , (2.32)
i=1

which, for simplicity, we shall denote by V{b}. The K × K matrix V{b} is a variance–
covariance matrix, containing the variances of b1 , b2 , . . . , bK on the diagonal, and their
covariances as off-diagonal elements. The proof is fairly easy and goes as follows:

V{b} = E{(b − 𝛽)(b − 𝛽) } = E{(X  X)−1 X  𝜀𝜀 X(X  X)−1 }

= (X  X)−1 X  (𝜎 2 IN )X(X  X)−1 = 𝜎 2 (X  X)−1 .

Without using matrix notation the proof goes as follows:

{( )−1 } ( )−1 { }( )−1
∑  ∑ ∑  ∑ ∑ 
V{b} = V xi x i xi 𝜀i = xi x i V xi 𝜀i xi x i
i i i i i
( )−1 ( )( )−1 ( )−1
∑ ∑ ∑ ∑
= xi xi 𝜎2 xi xi xi xi = 𝜎2 xi xi . (2.33)
i i i i

This requires assumptions (A1)–(A4).

k The last result is collected in the Gauss–Markov theorem, which says that under k
assumptions (A1)–(A4) the OLS estimator b is the best linear unbiased estimator for 𝛽.
In short we say that b is BLUE for 𝛽. To appreciate this result, consider the class of linear
unbiased estimators. A linear estimator is a linear function of the elements in y and can
be written as b̃ = Ay, where A is a K × N matrix. The estimator is unbiased if E{Ay} = 𝛽.
(Note that the OLS estimator is obtained for A = (X  X)−1 X  .) Then the theorem states
that the difference between the covariance matrices of b̃ = Ay and the OLS estimator b is
always positive semi-definite. What does this mean? Suppose we are interested in some
linear combination of 𝛽 coefficients, given by d 𝛽, where d is a K-dimensional vector.
Then the Gauss–Markov result implies that the variance of the OLS estimator d b for d 𝛽
̃ that is,
is not larger than the variance of any other linear unbiased estimator d b,
̃ ≥ V{d b} for any vector d.
V{d b}

As a special case this holds for the kth element and we have

V{b̃ k } ≥ V{bk }.

Thus, under the Gauss–Markov assumptions, the OLS estimator is the most accurate
(linear) unbiased estimator for 𝛽. More details on the Gauss–Markov result can be found
in Greene (2012, Section 4.3).
To estimate the variance of b we need to replace the unknown error variance 𝜎 2 with
an estimate. An obvious candidate is the sample variance of the residuals ei = yi − xi b,
that is,
1 ∑ 2
N
s̃2 = e (2.34)
N − 1 i=1 i

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 18

18 AN INTRODUCTION TO LINEAR REGRESSION

(recalling that the average residual is zero). However, because ei is different from 𝜀i ,
it can be shown that this estimator is biased for 𝜎 2 . An unbiased estimator is given by

1 ∑ 2
N
s2 = e. (2.35)
N − K i=1 i

This estimator has a degrees of freedom correction as it divides by the number of obser-
vations minus the number of regressors (including the intercept). An intuitive argument
for this is that K parameters were chosen so as to minimize the residual sum of squares
and thus to minimize the sample variance of the residuals. Consequently, s̃ 2 is expected to
underestimate the variance of the error term 𝜎 2 . The estimator s2 , with a degrees of free-
dom correction, is unbiased under assumptions (A1)–(A4); see Greene (2012, Section
4.3) for a proof. The variance of b can thus be estimated by
(N )−1
∑
V{b} = s2 (X  X)−1 = s2 xi xi . (2.36)
i=1

The estimated variance of an element bk is given by s2 ckk , where ckk is the (k, k) element in
(Σi xi xi )−1 . The square root of this estimated variance is usually referred to as the standard
error of bk . We shall denote it as se(bk ). It is the estimated standard deviation of bk and is
a measure for the accuracy of the estimator. Under assumptions (A1)–(A4), it holds that
√
se(bk ) = s ckk . When the error terms are not homoskedastic or exhibit autocorrelation,
k the standard error of the OLS estimator bk will have to be computed in a different way k
(see Chapter 4).
In general the expression for the estimated covariance matrix in (2.36) does not allow
derivation of analytical expressions for the standard error of a single element bk . As an
illustration, however, let us consider the regression model with two explanatory variables
and a constant:
yi = 𝛽1 + 𝛽2 xi2 + 𝛽3 xi3 + 𝜀i .

In this case it is possible to derive that the variance of the OLS estimator b2 for 𝛽2 is
given by
[N ]−1
𝜎2 ∑
V{b2 } = 2
(xi2 − x̄ 2 ) 2
,
1 − r23 i=1

where r23 is the sample correlation coefficient between xi2 and xi3 , and x̄ 2 is the sample
average of xi2 . We can rewrite this as
[ ]−1
𝜎2 1 1 ∑
N
V{b2 } = 2 N N
(xi2 − x̄ 2 ) 2
. (2.37)
1 − r23 i=1

This shows that the variance of b2 is driven by four elements. First, the term in square
brackets denotes the sample variance of x2 : more variation in the regressor values leads
to a more accurate estimator. Second, the term N1 is inversely related to the sample size:
having more observations increases precision. Third, the larger the error variance 𝜎 2 ,
the larger the variance of the estimator. A low value for 𝜎 2 implies that observations

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 19

SMALL SAMPLE PROPERTIES OF THE OLS ESTIMATOR 19

are typically close to the regression line, which obviously makes it easier to estimate it.
Finally, the variance is driven by the correlation between the regressors. The variance of
b2 is inflated if the correlation between xi2 and xi3 is high (either positive or negative). In
the extreme case where r23 = 1 or −1, xi2 and xi3 are perfectly correlated and the above
variance becomes infinitely large. This is the case of perfect collinearity, and the OLS
estimator in (2.7) cannot be computed (see Section 2.8).
Assumptions (A1)–(A4) state that the error terms 𝜀i are mutually uncorrelated, are
independent of X, have zero mean and have a constant variance, but do not specify
the shape of the distribution. For exact statistical inference from a given sample of N
observations, explicit distributional assumptions have to be made.5 The most common
assumption is that the errors are jointly normally distributed.6 In this case the uncorrelat-
edness of (A4) is equivalent to independence of all error terms. The precise assumption is
as follows:
𝜀 ∼ N (0, 𝜎 2 IN ), (A5)

saying that the vector of error terms 𝜀 has an N-variate normal distribution with mean
vector 0 and covariance matrix 𝜎 2 IN . Assumption (A5) thus replaces (A1), (A3) and (A4).
An alternative way of formulating (A5) is

𝜀i ∼ NID(0, 𝜎 2 ), (A5 )

which is a shorthand way of saying that the error terms 𝜀i are independent drawings from
a normal distribution (‘normally and independently distributed’, or n.i.d.) with mean zero
k and variance 𝜎 2 . Even though error terms are unobserved, this does not mean that we are k
free to make any assumption we like. For example, if error terms are assumed to follow
a normal distribution, this means that yi (for given values of xi ) also follows a normal
distribution. Clearly, we can think of many variables whose distribution (conditional upon
a given set of xi variables) is not normal, in which case the assumption of normal error
terms is inappropriate. Fortunately, not all assumptions are equally crucial for the validity
of the results that follow and, moreover, the majority of the assumptions can be tested
empirically; see Chapters 3, 4 and 6.
To make things simpler, let us consider the X matrix as fixed and deterministic or,
alternatively, let us work conditionally upon the outcomes X. Then the following result
holds. Under assumptions (A2) and (A5) the OLS estimator b is normally distributed
with mean vector 𝛽 and covariance matrix 𝜎 2 (X  X)−1 , that is,

b ∼ N (𝛽, 𝜎 2 (X  X)−1 ). (2.38)

The proof of this follows directly from the result that b is a linear combination of all
𝜀i and is omitted here. The result in (2.38) implies that each element in b is normally
distributed, for example
bk ∼ N (𝛽k , 𝜎 2 ckk ), (2.39)

where, as before, ckk is the (k, k) element in (X  X)−1 . These results provide the basis for
statistical tests based upon the OLS estimator b.

5
Later we shall see that for approximate inference in large samples this is not necessary.
6
The distributions used in this text are explained in Appendix B.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 20

20 AN INTRODUCTION TO LINEAR REGRESSION

2.3.3 Example: Individual Wages (Continued)

Let us now turn back to our wage example. We can formulate a (fairly trivial) econometric
model as
wagei = 𝛽1 + 𝛽2 malei + 𝜀i ,

where wagei denotes the hourly wage rate of individual i and malei = 1 if i is male and
0 otherwise. Imposing that E{𝜀i } = 0 and E{𝜀i |malei } = 0 gives 𝛽1 the interpretation of
the expected wage rate for females, while E{wagei |malei = 1} = 𝛽1 + 𝛽2 is the expected
wage rate for males. Thus, 𝛽2 is the expected wage differential between an arbitrary male
and female. These parameters are unknown population quantities, and we may wish to
estimate them. Assume that we have a random sample, implying that different observa-
tions are independent. Also assume that 𝜀i is independent of the regressors, in particular,
that the variance of 𝜀i does not depend upon gender (malei ). Then the OLS estimator for 𝛽
is unbiased and its covariance matrix is given by (2.32). The estimation results are given
in Table 2.1. In addition to the OLS estimates, identical to those presented before, we now
also know something about the accuracy of the estimates, as reflected in the reported stan-
dard errors. We can now say that our estimate of the expected hourly wage differential 𝛽2
between males and females is $1.17 with a standard error of $0.11. Combined with the
normal distribution, this allows us to make statements about 𝛽2 . For example, we can test
the hypothesis that 𝛽2 = 0. If this hypothesis is true, the wage differential between males
and females in our sample is nonzero only by chance. Section 2.5 discusses how to test
hypotheses regarding 𝛽.
k k

2.4 Goodness-of-Fit
Having estimated a particular linear model, a natural question that comes up is: how
well does the estimated regression line fit the observations? A popular measure for the
goodness-of-fit of a regression model is the proportion of the (sample) variance of y that
is explained by the model. This variable is called the R2 (R squared) and is defined as
̂ y } 1∕(N − 1) ∑N (̂y − ȳ )2
V{̂ i i=1 i
2
,
̂V{y } 1∕(N − 1) ∑N (y − ȳ )2
R = = (2.40)
i i=1 i
∑
where ŷ i = xi b and ȳ = (1∕N) i yi denotes the sample mean of yi . Note that ȳ also
corresponds to the sample mean of ŷ i , because of (2.11).

Table 2.1 OLS results wage equation

Dependent variable: wage

Variable Estimate Standard error
constant 5.1469 0.0812
male 1.1661 0.1122
s = 3.2174 R2 = 0.0317 F = 107.93

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 21

GOODNESS-OF-FIT 21

From the first-order conditions (compare (2.10)) it follows directly that

∑
N
ei xik = 0, k = 1, . . . , K.
i=1
∑
Consequently, we can write yi = ŷ i + ei , where i ei ŷ i = 0. In the most relevant case
where the model contains an intercept term, it holds that

V{y ̂ y } + V{e
̂ } = V{̂ ̂ }, (2.41)
i i i

̂ } = s̃2 . Using this, the R2 can be rewritten as

where V{e i
∑
̂ }
V{e 1∕(N − 1) Ni=1 e2i
i
2
R =1− =1− ∑ . (2.42)
̂ }
V{y 1∕(N − 1) Ni=1 (yi − ȳ )2
i

Equation (2.41) shows how the sample variance of yi can be decomposed into the sum of
the sample variances of two orthogonal components: the predictor ŷ i and the residual ei .
The R2 thus indicates which proportion of the sample variation in yi is explained by
the model.
If the model of interest contains an intercept term, the two expressions for R2 in (2.40)
and (2.42) are equivalent. Moreover, in this case it can be shown that 0 ≤ R2 ≤ 1. Only
if all ei = 0 does it hold that R2 = 1, whereas the R2 is zero if the model does not explain
anything in addition to the sample mean of yi . That is, the R2 of a model with just an
k intercept term is zero by construction. In this sense, the R2 indicates how much better the k
model fits the data than a trivial model with only a constant term.
From the results in Table 2.1, we see that the R2 of the very simple wage equation is
only 0.0317. This means that only approximately 3.2% of the variation in individ-
ual wages can be attributed to gender differences. Apparently, many other observable
and unobservable factors affect a person’s wage besides gender. This does not auto-
matically imply that the model that was estimated in Table 2.1 is incorrect or
useless: it just indicates the relative (un)importance of gender in explaining individual
wage variation.
In the exceptional cases where the model does not contain an intercept term, the
2
two
∑N expressions for R are not equivalent. The reason is that (2.41) is violated because
2
e
i=1 i is no longer equal to zero. In this situation it is possible that the R computed
from (2.42) becomes negative. An alternative measure, which is routinely computed by
some software packages if there is no intercept, is the uncentred R2 , which is defined as
∑N ∑N
̂ 2i
i=1 y
2
i=1 ei
uncentred R2 = ∑N = 1 − ∑N . (2.43)
2 2
i=1 yi i=1 yi

Generally, the uncentred R2 is higher than the standard R2 .

Because the R2 measures the explained variation in yi , it is also sensitive to the defini-
tion of this variable. For example, explaining wages is different to explaining log wages,
and the R2 s will be different. Similarly, models explaining consumption, changes in

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 22

22 AN INTRODUCTION TO LINEAR REGRESSION

consumption or consumption growth will not be directly comparable in terms of their

R2 s. It is clear that some sources of variation are much harder to explain than others.
For example, variation in aggregate consumption for a given country is usually easier
to explain than the cross-sectional variation in consumption over individual households.
Consequently, there is no absolute benchmark to say that an R2 is ‘high’ or ‘low’. A value
of 0.2 may be high in certain applications but low in others, and even a value of 0.95 may
be low in certain contexts.
Sometimes the R2 is suggested to measure the quality of the econometric model,
whereas it measures nothing more than the quality of the linear approximation. As the
OLS approach is developed to give the best linear approximation, irrespective of the
‘true’ model and the validity of its assumptions, estimating a linear model by OLS
will always give the best R2 possible. Any other estimation method, and we will see
several below, will lead to lower R2 values even though the corresponding estimator
may have much better statistical properties under the assumptions of the model. Even
worse, when the model is not estimated by OLS the two definitions (2.40) and (2.42)
are not equivalent and it is not obvious how an R2 should be defined. For later use, we
shall present an alternative definition of the R2 , which for OLS is equivalent to (2.40)
and (2.42), and for any other estimator is guaranteed to be between zero and one. It is
given by
(∑ )2
N
(y
i=1 i − ̄
y )(̂
y i − ̄
y )
R2 = corr2 {yi , ŷ i } = (∑ ) (∑ ), (2.44)
N N
k (y
i=1 i − ̄
y ) 2 (̂
i=1 iy − ̄
y ) 2
k

which denotes the squared (sample) correlation coefficient between the actual and fitted
values. Using (2.41) it is easily verified that, for the OLS estimator, (2.44) is equivalent
to (2.40). Written in this way, the R2 can be interpreted to measure how well the variation
in ŷ i relates to variation in yi . Despite this alternative definition, the R2 reflects the quality
of the linear approximation and not necessarily that of the statistical model in which
we are interested. Accordingly, the R2 is typically not the most important aspect of our
estimation results.
Another drawback of the R2 is that it will never decrease if the number of regressors
is increased, even if the additional variables have no real explanatory power. A common
way to solve this is to correct the variance estimates in (2.42) for the degrees of freedom.
This gives the so-called adjusted R2 , or R̄ 2 , defined as

∑
1∕(N − K) Ni=1 e2i
R̄ 2 = 1 − ∑ . (2.45)
1∕(N − 1) Ni=1 (yi − ȳ )2

This goodness-of-fit measure has some punishment for the inclusion of additional
explanatory variables in the model and therefore does not automatically increase when
regressors are added to the model (see Chapter 3). In fact, it may decline when a variable
is added to the set of regressors. Note that, in extreme cases, the R̄ 2 may become
negative. Also note that the adjusted R2 is strictly smaller than R2 unless K = 1 and the
model only includes an intercept.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 23

HYPOTHESIS TESTING 23

2.5 Hypothesis Testing

Under the Gauss–Markov assumptions (A1)–(A4) and normality of the error terms (A5),
we saw that the OLS estimator b has a normal distribution with mean 𝛽 and covariance
matrix 𝜎 2 (X  X)−1 . We can use this result to develop tests for hypotheses regarding the
unknown population parameters 𝛽. Starting from (2.39), it follows that the variable
bk − 𝛽k
z= √ (2.46)
𝜎 ckk
has a standard normal distribution (i.e. a normal distribution with mean 0 and variance
1). If we replace the unknown 𝜎 by its estimate s, this is no longer exactly true. It can
be shown7 that the unbiased estimator s2 defined in (2.35) is independent of b and has a
Chi-squared distribution with N − K degrees of freedom. In particular,8

(N − K)s2 ∕𝜎 2 ∼ 𝜒 2N−K . (2.47)

Consequently, the random variable

bk − 𝛽k
tk = √ (2.48)
s ckk
is the ratio of a standard normal variable and the square root of an independent Chi-
squared variable and therefore follows Student’s t distribution with N − K degrees of
k freedom. The t distribution is close to the standard normal distribution except that it has k
fatter tails, particularly when the number of degrees of freedom N − K is small. The
larger the N − K, the more closely the t distribution resembles the standard normal, and
for sufficiently large N − K the two distributions are identical.

2.5.1 A Simple t-Test

The result above can be used to construct test statistics and confidence intervals. The
general idea of hypothesis testing is as follows. Starting from a given hypothesis, the
null hypothesis, a test statistic is computed that has a known distribution under the
assumption that the null hypothesis is valid. Next, it is decided whether the computed
value of the test statistic is unlikely to come from this distribution, which indicates that
the null hypothesis is unlikely to hold. Let us illustrate this with an example. Suppose we
have a null hypothesis that specifies the value of 𝛽k , say H0 : 𝛽k = 𝛽k0 , where 𝛽k0 is a specific
value chosen by the researcher. If this hypothesis is true, we know that the statistic
bk − 𝛽k0
tk = (2.49)
se(bk )
has a t distribution with N − K degrees of freedom. If the null hypothesis is not true, the
alternative hypothesis H1 : 𝛽k ≠ 𝛽k0 holds. The quantity in (2.49) is a test statistic and is
computed from the estimate bk , its standard error se(bk ), and the hypothesized value 𝛽k0

7
The proof of this is beyond the scope of this text. The basic idea is that a sum of squared normals is Chi-squared
distributed (see Appendix B).
8
See Appendix B for details about the distributions in this section.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 24

24 AN INTRODUCTION TO LINEAR REGRESSION

under the null hypothesis. If the test statistic realizes a value that is very unlikely under
the null distribution, we reject the null hypothesis. This corresponds to having very large
values for tk , either positive or negative. To be precise, one rejects the null hypothesis if
the probability of observing a value of |tk | or larger is smaller than a given significance
level 𝛼, often 5%. From this, one can define the critical values tN−K;𝛼∕2 using

P{|tk | > tN−K;𝛼∕2 } = 𝛼.

For N − K not too small, these critical values are only slightly larger than those of the
standard normal distribution, for which the two-tailed critical value for 𝛼 = 0.05 is 1.96.
Consequently, at the 5% level the null hypothesis will be rejected if

|tk | > 1.96.

The above test is referred to as a two-sided test because the alternative hypothesis
allows for values of 𝛽k on both sides of 𝛽k0 . Occasionally, the alternative hypothesis is
one-sided, for example: the expected wage for a man is larger than that for a woman.
Formally, we define the null hypothesis as H0 : 𝛽k ≤ 𝛽k0 with alternative H1 : 𝛽k > 𝛽k0 . Next
we consider the distribution of the test statistic tk at the boundary of the null hypothesis
(i.e. under 𝛽k = 𝛽k0 , as before) and we reject the null hypothesis if tk is too large (note that
large values for bk lead to large values for tk ). Large negative values for tk are compatible
with the null hypothesis and do not lead to its rejection. Thus for this one-sided test the
critical value is determined from
k P{tk > tN−K;𝛼 } = 𝛼. k

Using the standard normal approximation again, we reject the null hypothesis at the 5%
level if
tk > 1.64.

Regression packages typically report the following t-value:

bk
tk = ,
se(bk )
sometimes referred to as the t-ratio, which is the point estimate divided by its standard
error. The t-ratio is the t-statistic one would compute to test the null hypothesis that
𝛽k = 0, which may be a hypothesis that is of economic interest as well. If it is rejected,
it is said that ‘bk differs significantly from zero’, or that the corresponding variable ‘xik
has a statistically significant impact on yi ’. Often we simply say that (the effect of) ‘xik is
statistically significant’. If an explanatory variable is statistically significant, this does
not necessarily imply that its impact is economically meaningful. Sometimes, particu-
larly with large data sets, a coefficient can be estimated very accurately, and we reject
the hypothesis that it is zero, although the economic magnitude of its effect is very small.
Conversely, if a variable is insignificant this does not necessarily mean that it has no
impact. Insignificance can result from absence of the effect, or from imprecision, par-
ticularly if the sample is small or exhibits little variation. It is good practice to pay
attention to the magnitude of the estimated coefficients as well as to their statistical sig-
nificance. Confidence intervals are also very useful, as they combine information about
the economic magnitude of an effect as well as its precision.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 25

HYPOTHESIS TESTING 25

A confidence interval can be defined as the interval of all values for 𝛽k0 for which the
null hypothesis that 𝛽k = 𝛽k0 is not rejected by the t-tests. Loosely speaking, given the
estimate bk and its associated standard error, a confidence interval gives a range of values
that are likely to contain the true value 𝛽k . It is derived from the fact that the following
inequalities hold with probability 1 − 𝛼:
bk − 𝛽k
−tN−K;𝛼∕2 < < tN−K;𝛼∕2 , (2.50)
se(bk )
or
bk − tN−K;𝛼∕2 se(bk ) < 𝛽k < bk + tN−K;𝛼∕2 se(bk ). (2.51)
Consequently, using the standard normal approximation, a 95% confidence interval
(setting 𝛼 = 0.05) for 𝛽k is given by the interval
[bk − 1.96se(bk ), bk + 1.96se(bk )]. (2.52)
In repeated sampling, 95% of these intervals will contain the true value 𝛽k which is a
fixed but unknown number (and thus not stochastic). Shorter intervals (corresponding to
lower standard errors) are obviously more informative, as they narrow down the range of
plausible values for the true parameter 𝛽k .

2.5.2 Example: Individual Wages (Continued)

From the results in Table 2.1 we can compute t-ratios and perform simple tests.
k For example, if we want to test whether 𝛽2 = 0, we construct the t-statistic as the k
estimate divided by its standard error to get t = 10.38. Given the large number of
observations, the appropriate t distribution is virtually identical to the standard normal
one, so the 5% two-tailed critical value is 1.96. This means that we clearly reject the
null hypothesis that 𝛽2 = 0. That is, we reject that in the population the expected wage
differential between males and females is zero. We can also compute a confidence
interval, which has bounds 1.17 ± 1.96 × 0.11. This means that with 95% confidence we
can say that over the entire population the expected wage differential between males and
females is between $0.95 and $1.39 per hour. Our sample thus provides a reasonably
accurate estimate of the wage differential, suggesting that an economically meaningful
difference exists between (average) wages for males and females.

2.5.3 Testing One Linear Restriction

The test discussed above involves a restriction on a single coefficient. Often, a hypothesis
of economic interest implies a linear restriction on more than one coefficient, such as9
𝛽2 + 𝛽3 + · · · + 𝛽K = 1. In general, we can formulate such a linear hypothesis as
H0 ∶ r1 𝛽1 + · · · + rK 𝛽K = r 𝛽 = q, (2.53)
for some scalar value q and a K-dimensional vector r. We can test the hypothesis in (2.53)
using the result that r b is the BLUE for r 𝛽 with variance V{r b} = r V{b}r. Replacing
9
For example, in a Cobb–Douglas production function, written as a linear regression model in logs, constant
returns to scale corresponds to the sum of all slope parameters (the coefficients for all log inputs) being equal
to one.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 26

26 AN INTRODUCTION TO LINEAR REGRESSION

𝜎 2 in the covariance matrix V{b} by its estimate s2 produces the estimated covariance
matrix, denoted ̂ 
√ as V{b}. Consequently, the standard error of the linear combination r b
  ̂ 
is se(r b) = r V{b}r. As b is K-variate normal, r b is normal as well (see Appendix B),
so we have
r b − r 𝛽
∼ tN−K , (2.54)
se(r b)
which is a straightforward generalization of (2.48).10 The test statistic for H0 follows as
r b − q
t= , (2.55)
se(r b)
which has a tN−K distribution under the null hypothesis. At the 5% level, absolute values
of t in excess of 1.96 (the normal approximation) lead to rejection of the null. This repre-
sents the most general version of the t-test. Any modern software package will provide
easy ways to calculate (2.55), with (2.49) as a special case.

2.5.4 A Joint Test of Significance of Regression Coefficients

A standard test that is typically automatically supplied by a regression package is a test
for the joint hypothesis that all coefficients, except the intercept 𝛽1 , are equal to zero.
We shall discuss this procedure slightly more generally by testing the null that J of the K
coefficients are equal to zero (J < K). Without loss of generality, assume that these are
k the last J coefficients in the model k
H0 ∶ 𝛽K−J+1 = · · · = 𝛽K = 0. (2.56)

The alternative hypothesis in this case is that H0 is not true, that is, at least one of these
J coefficients is not equal to zero.
The easiest test procedure in this case is to compare the sum of squared residuals of the
full model with the sum of squared residuals of the restricted model (which is the model
with the last J regressors omitted). Denote the residual sum of squares of the full model
by S1 and that of the restricted model by S0 . If the null hypothesis is correct, one would
expect that the sum of squares with the restriction imposed is only slightly larger than
that in the unrestricted case. A test statistic can be obtained by using the following result,
which we present without proof. Under the null hypothesis and assumptions (A1)–(A5)
it holds that
S0 − S1
∼ 𝜒 2J . (2.57)
𝜎2
From earlier results we know that (N − K)s2 ∕𝜎 2 = S1 ∕𝜎 2 ∼ 𝜒 2N−K . Moreover, under the
null hypothesis it can be shown that S0 − S1 and s2 are independent. Consequently, we
can define the following test statistic:
(S0 − S1 )∕J
F= . (2.58)
S1 ∕(N − K)

10
The statistic is the same if r is a K-dimensional vector of zeros with a 1 on the kth position.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 27

HYPOTHESIS TESTING 27

Under the null hypothesis, F has an F distribution with J and N − K degrees of freedom,
J
denoted as FN−K . If we use the definition of the R2 from (2.42), we can also write this
F-statistic as
(R21 − R20 )∕J
F= , (2.59)
(1 − R21 )∕(N − K)

where R21 and R20 are the usual goodness-of-fit measures for the unrestricted and the
restricted models, respectively. This shows that the test can be interpreted as testing
whether the increase in R2 moving from the restricted model to the more general model
is significant.
It is clear that in this case only very large values for the test statistic imply rejection
of the null hypothesis. Despite the two-sided alternative hypothesis, the critical values
J
FN−K;𝛼 for this test are one-sided and defined by the following equality:

P{F > FN−K;𝛼

J
} = 𝛼,
where 𝛼 is the significance level of the test. For example, if N − K = 60 and J = 3 the
critical value at the 5% level is 2.76. The resulting test is referred to as the F-test.
In most applications the estimators for different elements in the parameter vector will
be correlated, which means that the explanatory powers of the explanatory variables over-
lap. Consequently, the marginal contribution of each explanatory variable, when added
last, may be quite small. Hence, it is perfectly possible for the t-tests on each variable’s
coefficient to be insignificant, while the combined F-test for a number of these coeffi-
k cients is highly significant. That is, it is possible that the null hypothesis 𝛽1 = 0 is as such k
not unlikely, that the null 𝛽2 = 0 is not unlikely, but that the joint null 𝛽1 = 𝛽2 = 0 is quite
unlikely to be true. As a consequence, in general, t-tests on each restriction separately may
not reject, while a joint F-test does. The converse is also true: it is possible that individ-
ual t-tests do reject the null, while the joint test does not. The section on multicollinearity
below illustrates this point. Section 3.6 provides an empirical illustration.
A special case of this F-test is sometimes misleadingly referred to as the model test,11
where one tests the significance of all regressors, that is, one tests H0 ∶ 𝛽2 = 𝛽3 = · · · =
𝛽K = 0, meaning that all partial slope coefficients are equal to zero. The appropriate test
statistic in this case is
(S − S1 )∕(K − 1)
F= 0 , (2.60)
S1 ∕(N − K)
∑
where S1 is the residual sum of squares of the model, that is S1 = i e2i , and S0 is the
residual
∑ sum of squares of the restricted model containing only an intercept term, that is,
S0 = i (yi − ȳ )2 .12 Because the restricted model has an R2 of zero by construction, the
test statistic can also be written as
R2 ∕(K − 1)
F= . (2.61)
(1 − R2 )∕(N − K)
11
This terminology is misleading as it does not in any way test whether the restrictions imposed by the model
are correct. The only thing tested is whether all coefficients, excluding the intercept, are equal to zero, in
which case one would have a trivial model with an R2 of zero. As shown in (2.61), the test statistic associated
with the model test is simply a function of the R2 .
12
Using the definition of the OLS estimator, it is easily verified that the intercept term in a model without
regressors is estimated as the sample average ȳ . Any other choice would result in a larger S value.

k
 Trim Size: 7in x 10in
k Verbeek c02.tex V3 - 04/21/2017 3:58 P.M. Page 28

28 AN INTRODUCTION TO LINEAR REGRESSION

This F-statistic is routinely provided by the majority of all regression packages. Note that
it is a simple function of the R2 of the model. If the test based on F does not reject the null
hypothesis, one can conclude that the model performs rather poorly: a ‘model’ with just
an intercept term would not do significantly worse. However, the converse is certainly not
true: if the test does reject the null, one cannot conclude that the model is good, perfect,
valid or correct. An alternative model may perform much better. Chapter 3 pays more
attention to this issue.

2.5.5 Example: Individual Wages (Continued)

The fact that we concluded previously that there was a significant difference between
expected wage rates for males and females does not necessarily point to discrimination.
It is possible that working males and females differ in terms of their characteristics, for
example their years of schooling. To analyse this, we can extend the regression model
with additional explanatory variables, for example schooli , which denotes the years of
schooling, and experi , which denotes experience in years. The model is now interpreted
to describe the conditional expected wage of an individual given his or her gender, years
of schooling and experience, and can be written as

wagei = 𝛽1 + 𝛽2 malei + 𝛽3 schooli + 𝛽4 experi + 𝜀i .

The coefficient 𝛽2 for malei now measures the difference in expected wage between a
male and a female with the same schooling and experience. Similarly, the coefficient 𝛽3
k for schooli gives the expected wage difference between two individuals with the same k
experience and gender where one has an additional year of schooling. In general, the
coefficients in a multiple regression model can only be interpreted under a ceteris paribus
condition, which says that the other variables that are included in the model are constant.
Estimation by OLS produces the results given in Table 2.2. The coefficient for malei
now suggests that, if we compare an arbitrary male and female with the same years of
schooling and experience, the expected wage differential is $1.34 compared with $1.17
before. With a standard error of $0.11, this difference is still statistically highly signifi-
cant. The null hypothesis that schooling has no effect on a person’s wage, given gender
and experience, can be tested using the t-test described previously, with a test statistic
of 19.48. Clearly the null hypothesis is rejected at any reasonable level of significance.
The estimated wage increase from one additional year of schooling, keeping years of
experience fixed, is $0.64. It should not be surprising, given these results, that the joint
hypothesis that all three partial slope coefficients are zero, that is, wages are not affected

Table 2.2 OLS results wage equation

Dependent variable: wage

Variable Estimate Standard error t-ratio
constant −3.3800 0.4650 −17.2692
male 1.3444 0.1077 12.4853
school 0.6388 0.0328 19.4780
exper 0.1248 0.0238 5.2530
s = 3.0462 R2 = 0.1326 R̄ 2 = 0.1318 F = 167.63

k
Exploring the Variety of Random
Documents with Different Content
 CHAPTER XXVII.
The Malabar eagle, and sago manufactory.

My wife the next morning began the conversation. She told me the
children had been good and diligent, that all of them had ascended
Cape Disappointment together, that they had gathered wood and
made some torches for the night; and, what I could scarcely think
possible, they had ventured to fell an immense palm-tree, to the top
of which Ernest had again climbed for amusement; and afterwards
they undertook the laborious and perilous work of bringing it down,
which they also happily executed. This monarch of the forests lay
prostrate on the ground, and covered a space of at least seventy
feet in length. To effect their purpose, Ernest had got up the tree a
second time, with a long rope which he fastened tight to the top of
it. As soon as he had come down again, he and Fritz worked with
the axe and saw, to cut it through. When it was nearly divided, they
cautiously managed its fall with the rope, and in this manner they
succeeded. But during their excursion and labours, a somewhat
unfortunate visit took place. A numerous group of monkeys had
found their way to the hut: every drop of the palm-wine we had
preserved in our gourd-shells, these intrusive gentry contrived to
swallow; they had upset and thrown about all the potatoes, stolen
the cocoa-nuts, and disturbed, nay even almost wholly demolished,
the order of the branches and other contrivances we had recourse to
for our hut; so that my children on their return were employed a full
hour in repairing the damage effected by those mischievous
creatures. Fritz was in high spirits with the luck he had met with and
secured; he brought me on his wrist a young bird of prey, of the
most beauteous plumage; he had taken it from a nest in one of the
rocks near the hill at Cape Disappointment. Very young as the bird
was, it had already all its feathers, though they had not yet received
their full colouring; but it was evidently not an European eagle, and I
supposed it to be the beautiful eagle of Malabar16. I viewed it with
the admiration it was entitled to:—meeting with one of these birds is
thought a lucky omen; and as this species of eagle is neither large
nor expensive in its food, I was desirous to train it like a falcon to
pursue smaller birds. Fritz had already covered its eyes and fastened
the foot with packthread; I advised him to hold it often, and for a
length of time, on his hand, and to tame it with hunger as falconers
do.
When we had all finished our narratives, my wife began her usual
lamentations upon the subject of so many living and devouring
animals being brought to her, and which she said must in no long
time, from the food they required, become burthensome to us. I
consoled her with some difficulty, by observing that the buffalo
would be a good substitute for the ass; and I established as an
invariable law, that he who wished to have a useful animal in his
service, should also have the care of keeping it. It is a cruelty, said I,
to deprive a feeling creature of its liberty, for a transient pleasure or
amusement to ourselves, and to make it suffer; we ought, on the
contrary, to recompense it for the privation it would have to endure,
by kind treatment and feeding it well. I declare then, that he who
neglects the animals committed to his charge shall no longer be
thought worthy of having so important a care intrusted to him. This
warning intimidated my hearers; for man is a creature so formed for
society, and so dependent upon it for his comforts and content, that,
when by accidents deprived of intercourse with his own species, he
attaches himself to the brute creation with even warm affection. This
satisfied the kind temper of my wife; and as she was herself very
partial to animals, she promised her sons to take the charge of one
and all of the animals upon herself whenever their absence from
home should render this necessary. We communicated to each other
our hopes that such attentions would form them to more active,
regular, and constant habits of affection and mildness towards us. I
have said, I frequently noticed that treating animals with kindness
tends to produce the benign effects I have been speaking of.
I next ordered a fire to be lighted and a quantity of green wood to
be put on it, for the purpose of raising a thick smoke, over which I
meant to hang the buffalo meat I had salted, to dry and preserve it
for our future use; and this was done accordingly. We fixed pieces of
it on long forked poles. I cut away what appeared not quite fresh,
and saw our little eagle feast upon it voraciously. The young buffalo
was beginning to browse, but the cow’s milk was still given it; and
Jack succeeded in making his little jackal drink some occasionally,
likewise. We added to the buffalo’s meal, whose appetite we found
to be enormous, a heap of sliced potatoes, the whole of which he
greedily devoured; and this led us to conclude that the pain from the
wound in his nose was subsided, and that he would soon become
tame.
Supper-time was now arrived, and we did not fail to acquit
ourselves at it, as well, to say the least, as our cattle, seasoning our
repast with lively anecdotes and affectionate conversation. The
combat with the buffalos excited much raillery; but we saw that it
was no less intended for compliment also, nor was our design to
make boots and shoes of their hoofs passed over in silence; the
candle-moulds too came in for a share in these their playful attacks
upon what they called the wonderful feats we had performed. Jack
persisted in defending himself, and this he did so cleverly as to point
the laughter at the adversary. The arrangements for this night were
much the same as for the preceding. We left our meat suspended
over the smoke of our fires during our sleep; we tied the young
buffalo by the side of the cow, and were pleased to see them agree
and bid fair to live in peace together. The dogs were set upon the
watch. Fritz resolved to go to bed with his eaglet fastened on his
wrist, and its eyes still bound: it remained in this state throughout
the night without disturbing its master. The time of repose elapsed
so calmly that none of us awoke to keep in the torchlights, which
now for the first time the industry of the boys had supplied us with,
and we did not get up till after sun-rise. Directly after a moderate
breakfast I chaunted the accustomed summons for our setting out;
but my young ones had some projects in their heads, and neither
they nor their mother were just then in the humour to obey me.
Let us reflect a little first, said my wife: as we had so much
difficulty in felling the palm-tree, would it not be a pity to lose our
labour by leaving it in this place? It is the one from which Ernest cut
out the famous cabbage: as it was thus deprived of future growth
and bearing fruit, we wished to reap benefit from it in another way;
Ernest assures me it is a sago-tree; if so, the pith would be an
excellent ingredient for our soups. I request you will examine it, and
let us see if in any way we can turn it to account.
I found she was in the right: but in that case it was necessary to
employ a day in the business, since to open from one end to the
other the trunk of a tree sixty feet long was not a trivial task. I
assented however with some readiness, as, independent of the use
of the farinaceous pith, I could, by emptying it, obtain two
handsome and large troughs for the conveyance of water from
Jackal’s river to my wife’s kitchen garden at Tent House, and thence
to my new plantations of trees.
Fritz.—One of the halves, father, will answer that purpose, and the
other will serve as a conduit for our little stream from Falcon’s Nest
into my pretty bason lined with tortoise-shell; we then shall be
constantly regaled with the agreeable view of a fountain close to our
dwelling:—I fancy it now before my eyes and that I see its course.
And I, for my part, said Ernest, long for a sight of the sago formed
into small grains as I have seen it in Europe. Can you, father, make it
up into that sort of composition?
With your help I think I could. Come, let us set up a famous
manufactory of vermicelli and macaroni! and what other delicacy
shall I say?
Francis.—Oh! yes, papa, I intreat you, macaroni! it is so nice, I will
assist in making that.
 Little glutton! answered I, you would, I imagine, rather assist in
eating up all we can make. I do not however promise that our
manufactory will produce such good and well made articles as those
of Genoa and Naples; but let us first make the sago-paste, and
afterwards work it up for such purposes of magnificence as in our
profound wisdom we may afterwards be led to desire. Have we not
one of our manioc graters here at hand?
Yes certainly, replied Ernest. We were even thinking of making
some of them here for our amusement, but we found we were not
likely to want employment: and he accordingly scampered away to
fetch it me, while the rest crowded round me. Patience, children,
patience, exclaimed I; we are not yet in readiness to use it, many
other matters are previously requisite: in the first place you must
assist me to raise this palm-tree from the ground, and it must be
done by fixing at each end two small cross pieces or props to
support it; to split it open as it lies, would be a work of too much
labour: this done, I shall want several wooden wedges to keep the
cleft open while I am sawing it, and afterwards a sufficient quantity
of water. There is the difficulty, said my wife; our Falcon’s Stream is
too far off, and we have not yet discovered any spring in the
neighbourhood of this place.
Ernest.—That is of no consequence, mother; I have seen
hereabouts so great an abundance of the plants which contain
water, that we need not be at a loss; for they will fully supply us, if I
could only contrive to get vessels enough to hold it.
We now produced the enormous reeds we had brought home,
which being hollow would answer the purpose of vessels: and as
some time was required to draw off the water from such small
tubes, he and Francis at once set to work; they cut a number of the
plants, which they placed slantingly over the brim of a vessel, and
whilst that was filling they were preparing another. The rest of us
got round the tree, and with our united strength we soon succeeded
in raising the heavy trunk, and the top of it was then sawed off. We
next began to split it through the whole length, and this the softness
of the wood enabled us to effect with little trouble. We soon reached
the pith or marrow that fills up the middle of the trunk the whole of
its length. When divided, we laid one half on the ground, and we
pressed the pith together with our hands so as to make temporary
room for the pith of the other half of the trunk, which rested still on
the props. We wished to empty it entirely, that we might employ it
as a kneading-trough, leaving merely enough of the pith at both
ends to prevent a running out, and then we proceeded to form our
paste. We had fastened the grater at one end, for the purpose of
squeezing the paste through the small holes as soon as it was made.
My young manufacturers with stripped arms joyfully fell to work,
and really surpassed my expectation; they brought water in
succession and poured it gradually into the trough, whilst we mixed
it with the flour. In a short time the paste appeared sufficiently
fermented; I then made an aperture at the bottom of the grater on
its outside, and pressed the paste strongly with my hand; the
farinaceous parts passed with ease through the small holes of the
grater, and the ligneous parts blended with particles of wood, which
did not pass, were thrown aside in a heap, in the hope that
mushrooms, &c. might spring from them. My boys were in readiness
to receive in the reed vessels, what fell from the grater, and
conveyed it directly to their mother, whose business was to spread
out the small grains in the sun upon sail cloth for the purpose of
drying them. The subsequent process was the making of vermicelli,
by working up the paste into a thicker consistence and pressing it
more forcibly against the perforations of the grater; they passed
through in slender rolls of different lengths, which were quickly dried
by means of a gentle fire. To remunerate our toil my wife promised
to dress us an excellent dish of this new manufacture, with some
Dutch cheese, similar to Italian macaroni. Thus we procured a good
supply of a wholesome and pleasant food; and should have had a
larger stock of it, had we not been restricted as to time; but the
privilege of renewing the process at pleasure, by felling a sago-tree,
added to some impatience to take home our two pretty conduits and
employ them as proposed, prompted us to expedite the business.
The paste which remained was thrown upon the mushroom-bed,
and watered well to promote a fermentation.
We employed ourselves the remainder of the evening in loading
the cart with our tools and the two halves of the tree. Night coming
on, we retired to our hut, where we enjoyed our usual repose, and
early next morning were ready to return to Falcon’s Stream. Our
buffalo now commenced his service, yoked with the cow; he
supplied the want of the ass, and was very tractable: it is true I led
him by the cord which passed through his nose, and thus I
restrained him within the bounds of his duty whenever he was
disposed to deviate.
We returned the same way as we came, in order to load the cart
with a provision of berries, wax, and elastic gum. I gave up looking
for eggs this day, as I was very desirous to get back as soon as
possible to Falcon’s Stream, to look after the beasts we had left
there. I sent forward Fritz and Jack as a van-guard, with one of the
dogs; they were to cut a commodious and secure road through the
bushes for our cart. The two water conductors, which were very
long, produced numerous difficulties and somewhat impeded our
progress. My sons well performed their task, and we reached with
tolerable speed and without any accident the wax and gum trees,
where we halted to place our sacks filled with berries in the cart.
The elastic gum had not yielded as much as I expected, from the too
rapid thickening caused by an ardent sun, and an incrustation
formed over the incision: we obtained however about a quart, which
sufficed for the experiment of the impenetrable boots I had so long
desired.
We set out again, still preceded by our pioneers, who cleared the
way for us through the little wood of guavas. Suddenly we heard a
dreadful noise which came from our van-guard, and beheld Fritz and
Jack hastening towards us. I began now to fear a tiger or panther
was near at hand, or had perhaps attacked them. Turk began to
bark so frightfully, and Ponto running up to him joined in so hideous
a yell, that I prepared myself, not without terror, for a bloody
conflict. I advanced at the head of my troop, who expressed their
determination to follow me to the assistance of those I thought in
danger, and my high-mettled dogs ran furiously up to a thicket,
where they stopped, and with their noses to the ground and almost
breathless strove to enter it: I had no doubt some terrible animal
was lurking there; and Fritz, who had seen it through the leaves,
confirmed my suspicion; he said it was about the size of the young
buffalo, and that its hair was black and shaggy. I was going to fire at
it promiscuously in the thicket, when Jack, who had thrown himself
on his face on the ground to have a better view of the animal, got
up in a fit of laughter—It is only, exclaimed he, dame pig, that has
played us another trick—our old sow, who is never tired of playing
off her tricks upon us. He had hardly spoken, when the grunting of
the concealed monster justified the assertion made by Jack. Half
vexed, half laughing, we broke into the midst of the thicket, where
in reality we found our sow stretched supinely on the earth, but by
no means in a state of dreary solitude; the good matron had round
her seven little creatures, which had been littered a few days, and
were sprawling about contending with each other for the best place
near their mother for a hearty meal. This discovery gave us
considerable satisfaction, and we all greeted the corpulent creature,
who seemed to recollect and welcome us with a sociable kind of
grunting, while she licked her young without any ceremony or show
of fear. We rewarded her docility with potatoes, sweet acorns, and
manioc bread; for the boys one and all readily consented to go
without themselves for her accommodation: they felt obliged to her
for the new family she had given them, and anticipated ideally the
pleasure of beholding a nice crisp suckling pig before the fire. But
their mother censured their greedy impatience and cruelty, in
already thinking about roasting the poor animals on their very
entrance into life. A general consultation took place—should this
swinish family be left where we found it, or conveyed to Falcon’s
Stream? Fritz voted for their being all left to run at large like the wild
boars in Europe, that he might have the sport of hunting them. My
wife, like a good housewife, proposed that two of them at least
should be domesticated for breeding; and as to the old sow, as she
was always running away, it would be better to kill her when done
suckling, and she would afford a large provision of salt meat;—and
her opinion was adopted. For the moment they were suffered to
keep quiet possession of their retreat; we resolved to rear three of
them, and allow the other four to take to the woods, where they
might be hunted if they injured our plantations.
We then, so many adventures ended, pursued our road to Falcon’s
Stream, and arrived there in safety and content;—so true it is, that
home is always dear and sacred to the heart, and anticipated with
delight. All was in due order, and our domestic animals welcomed
our return in their own jargon and manner, but which did not fail to
be expressive of their satisfaction in seeing us again. We threw them
some of the food we knew them to be the most partial to, which
they greedily accepted, and then voluntarily went back to their usual
stand. It was necessary to practise a measure dictated by prudence,
which was to tie up again the buffalo and jackal, to inure them by
degrees to confinement; and the handsome Malabar eagle shared
the same fate: Fritz deemed himself remarkably clever in placing it
near the parrot on the root of a tree; he fastened it with a piece of
packthread of sufficient length to allow it free motion, and
uncovered its eyes: till then the bird had been tolerably quiet; but
the instant it was restored to light, it fell into a species of rage that
surprised us; it proudly raised its head; its feathers became ruffled,
and its eye-balls seemed to whirl in their orbits and dart out vivid
lightnings. All the poultry was terrified and fled; but the poor
luckless parrot was too near the sanguinary creature to escape.
Before we were aware of the danger, it was seized and mangled by
the formidable hooked beak of the eagle in an almost inconceivably
short space of time. Fritz vented his anger in loud, severe, and
passionate reproaches: he would have killed the murderer on the
spot, had not Ernest run up and entreated him to spare its life:
Parrots, said he, we shall find in plenty, but never perhaps so
beauteous, so magnificent a bird as this eagle, which, as father
observes, we may train for hawking. You may too blame only
yourself, continued Ernest, for the parrot’s death;—why did you
uncover the eagle’s eyes? If you had consulted me, I could have told
you that falconers keep them covered six weeks, till they are
completely tamed. But now, brother, let me for a certain time have
the care of him, let me manage the unruly fellow; he shall soon, in
consequence of the methods I shall use, be as tractable and
submissive as a new-born puppy; say then that you consent, I will
pledge myself you shall not be disappointed.
Yes, forsooth, said Fritz, but I shall not give you my eagle neither;
it is mine, and I will not part with it to anyone; I can bring it up
myself, only tell me how; it will be very unkind not to teach me the
way—Father, is not Ernest very ungenerous to keep his management
of eagles a secret, and wanting to bring it up himself?—Ernest, by
the by, had signified his refusal by a shake of the head.
Gently, gently, children, said I: Fritz, listen to a little moral fable.
“A dog placed himself on a truss of straw, which he considered as his
property. An ass and an ox extremely hungry entreated permission
to eat some of it, as it could not serve him for food; but the surly
selfish cur would not even suffer them to approach it. Envious
animal, said the ox, eat thy straw, or permit it to be eaten by
another; but the dog was deaf to their entreaties, and dismissed
them with snarling selfishness.” Now tell me, boy, is not your
conduct on this occasion similar to the selfish dog’s? You know not
how to tame your eagle, you a few minutes ago would even have
inflicted death upon it, and now you refuse to give it to Ernest, who
promises to bring it up, because you are envious of his knowing
more than you do: as he has reflected more than you on what he
has read, it is but just he should wish to derive some advantage
from his reading and knowledge; you cannot do less than make him
a civil offer of something for his secret, if you persist in not resigning
your eagle. Should Ernest after this, have the generosity to impart it
without the reward, the greater will be his merit, and I shall be
pleased with you both.
Right, father, replied Fritz briskly; well, I will give him my monkey
if he chooses to have it.—An eagle is a more noble and heroical
animal; as I found it, I mean to keep it; but you will teach me to
tame it? What say you, Ernest? Do you accept my offer?
 Ernest.—With all my heart; but I value the heroism you talk of,
very little: I for my part prefer being learned; you shall be knight of
the eagle if you think proper, and I your historiographer and poet. I
will write a grand epic poem to record the splendid deeds of the
knight and his eagle.
Fritz.—Poor jeering this, master Ernest! but, no matter; take the
monkey, and teach me to tame my eagle: How shall I render it more
docile and quiet?
Ernest.—I have read somewhere that the Caribs puff tobacco-
smoke into the nostrils of the birds of prey and of the parrots they
catch, until they are giddy and almost senseless;—this stupefaction
over, they are no longer wild and untractable.
Fritz.—And this is the great witchcraft! Tobacco-smoke! it was
worth puffing off to be sure! Your secret is not equivalent to the
property of my monkey, is it, Father?
Why not? If the bargain is fair, as I think it is, it should by all
means hold good; if not, Ernest will not require any thing for useless
counsel; but I have reason to think well of the method, and
therefore approve of it. Bees may be stupefied in the same manner,
so as to take them without resistance, or how could the honey-comb
be obtained? The notion I assure you, boys, is far from bad.
Francis.—Oh! then there is a way to hinder the naughty bees from
stinging poor little boys who want just to taste their honey? Good
papa, pray go and smoke awhile opposite that hole in the tree, and
make these stinging gentry sleep long enough for us to take away at
least half of their honey without being devoured by them.
Fritz having consented to the experiment, took some tobacco and
a pipe, of which we had found plenty in the sailors’ chests, and
began to smoke, at the same time gradually approaching the unruly
bird. As soon as it was somewhat composed, he replaced the fillet
over the eyes, and smoked close to its beak and nostrils so
effectually, that it became motionless on the spot, and had the exact
air of a stuffed bird. Fritz thought it dead, and was inclined to be
angry with his brother; but I told him it would not hold on the perch
if it were lifeless, and that its head alone was affected,—and so it
proved. The favourite came to itself by degrees, and made no noise
when its eyes were unbound; it looked at us with an air of surprise,
but void of fury, and grew tamer and calmer every day. The monkey
was adjudged unanimously to Ernest, who took possession of it the
same evening, and made it lie down near him. We all passed an
excellent night in our green castle, and in our good beds, which we
returned to with abundant satisfaction.
 CHAPTER XXVIII.
Origin of some European fruit-trees.—Bees.

We commenced early the next day a business which we had long

determined to engage in: it was to plant bamboos close to all the
young trees, to support them effectually in their growth. We quitted
our tree with great alertness, having our cart loaded with canes and
a large pointed iron to dig holes in the ground. We left my wife this
time with only her dear little Francis, requesting them to prepare us
a plentiful dinner, and to include the palm-tree cabbage and the
sago-macaroni mixed with some Dutch cheese; in addition to these
performances, they volunteered the melting some of the wax berries
for our store of candles.
We did not take the buffalo with us, as I wished to give it a day’s
rest for its nose to heal up; the cow was sufficient for drawing the
load of light bamboo canes. Before setting out, we gave the buffalo
a few handfuls of salt, to ingratiate ourselves with our horned
companion; and this treat pleased him so highly, that he showed by
many signs his inclination of accompanying us; and to prevent this,
we were compelled to fasten him securely till we were out of sight.
We began our work at the entrance of the avenue which we had
formed, and nearest to Falcon’s Stream. The walnut, chesnut, and
cherry-trees we had planted in a regular line and at equal distances,
we found disposed to bend considerably to one side, seemingly as
they had been directed by the wind. Being the strongest, I took the
task of making holes with the implement, upon myself, which, as the
soil was light, I easily performed, taking care to go deep enough to
fix the stake firmly. In the mean time the boys selected the
bamboos, cut them of equal lengths, and pointed the ends to go into
the ground. When they were well fixed, we threw up the earth
compactly about them, and fastened the saplings by the branches to
them with some long, straight tendrils of a plant which we found
near the spot. In the midst of our exertions we entered with much
detail into a conversation respecting the culture of trees. Till then my
boys had only thought of eating fruits, without giving themselves
much trouble about their production; but now their curiosity was
excited, and they questioned me so closely on the subject, that I
was somewhat at a loss in answering them. I communicated
however, with pleasure, all within my knowledge: I perceived that
the occasion was particularly favourable for rendering my lessons
instructive and truly profitable, as the occupation we were engaged
in gave me an opportunity to illustrate the lesson by means of the
objects before our eyes. I shall briefly relate the summary of our
discourse, which may be of use to young persons designed for an
agricultural life.
Fritz.—Are the young trees that we have recently planted, and just
propped up, wild or cultivated?
Jack.—A pretty question indeed! Do you imagine that trees are
tamed like buffalos and eagles? You would perhaps teach them to
stoop politely when we come to gather their fruit, to avoid giving us
the trouble of reaching to their branches.
Fritz.—You suppose, friend Jack, that your speech is very witty,
while in reality it is only nonsense. Do you think then that all created
beings are brought up alike? If so, papa should punish the first
occurrence of ill behaviour in you, by passing a cord through your
nose, like the buffalo, to render you more considerate and
compliant.
Ernest.—We should soon, if this plan were adopted, see brother
Jack with a bridle round his neck.
Father.—All of you, children, would be exposed to such discipline
as this, if there were no other means: but the observation of Fritz is
perfectly just; men are not brought up like brutes, nor are brutes
trained like plants, though the education of all ever tends to the
same end—that of subjecting the will to the yoke of necessity and
duty, and to make them walk upright: for did not these trees bend to
every wind till we raised and supported them. Every creature is
capable of improvement, if not susceptible of perfection, that is to
say, with care and cultivation, every being may become better, and
acquire virtues and qualities which, left to himself and nature, he
would be destitute of; thus I render our buffalo, and you your jackal,
tame and manageable, by making them feel the power of man over
brutes; thus I strive, my dear children, to lead you on towards
perfection, by cultivating your understanding, giving you, as far as I
am able, good lessons and good examples; so these trees, which at
first were mere wild stocks, produced from kernels or seeds, and
bearing only small imperfect fruits, have been made to produce
those of an excellent quality by grafting on them a superior species.
Come here, examine this branch; it is easy to see it has been
inserted into this other branch; all the rest were lopped away, and
this alone preserved; the whole of the sap or nutritious juice flowed
then to this point, and thence diffused itself; and thus the wild
sapling is become a fine fruitful tree, producing as good fruit as that
from which the graft was taken.—Such is the process of ingrafting.
Fritz.—There are many things in it I do not yet understand; I have
often heard of the inoculation of trees—is that the same as
ingrafting them?
Father.—Nearly so; to inoculate is, I believe, the appropriate term
for the operation, and graft that for the branch, or part inserted.
According to the different species of trees, different methods and
seasons are chosen; different terms are also adopted; and when the
first insertion fails, a second, and even a third attempt is made.
Jack.—Can good fruits be grafted on every kind of tree, such as
firs, or oaks?
Father.—No, my son; trees of an homogeneous kind must be
selected.
 Jack.—Ah! now that’s a fruit I am a stranger to; homogeneous—is
it very good? are they to be found here? I long to taste them.
Father.—It is not a fruit, my dear child: it is, I confess, rather too
learned a word, which I was wrong to use without explanation. It is
derived or comes from the Greek, and signifies of the same nature,
or bearing great similitude; thus an apple, a pear, a quince-tree, may
be ingrafted on each other, because their wood and their seeds are
homogeneous or resemble each other. It is the same with a variety
of fruits having stones and kernels; the cherry, plum, peach, and
almond-tree: to attempt to graft any of these on a pine or oak,
would be useless; so with the best education every man is not
constituted to be learned, to form an artist or a general.
Fritz.—You said, father, that the wild stocks or trees, produced
only from seed, bear but indifferent fruits; how is it, then, with all
those in our island—our good cocoas and guavas were not grafted,
as no gardeners or other persons were here to perform the
operation?
Father.—Your remark is just; but I spoke only of our fruit-trees in
Europe, where, almost without exception, the fruits require to be
meliorated by a better soil, ingrafting, and culture. No doubt kind
Providence has meant to indemnify these burning climes for many
inconveniences, in bestowing on them palm-trees and other
agreeable fruits that grow spontaneously and without trouble or
labour on the soil.
Ernest.—I comprehend all this: yet one thing puzzles me; how
were the scions and grafts of the best kinds first procured in Europe?
Jack.—What a question! From those who had them, certainly.
Ernest.—And what a reply! I answer. And where did those from
whom they were obtained get them? I wish to know where the first
branches of the best species were had, before any persons had
undertaken the husbandry of trees, or thought of ingrafting them. All
trees, I presume, were wild originally .....
Jack.—Indeed! ...... What think you then of the terrestrial
paradise? Do you not believe that excellent fruits of all kinds were
there? And might not the scions and grafts you are so puzzled
about, have been taken there to any number?
Father.—My dear young pupil, if you had read the Bible with
attention, you would have seen that our father Adam was driven out
of the terrestrial paradise for having eaten of one of those goodly
fruits, contrary to the positive prohibition of God; and as he and his
wife Eve were then alone on the earth, none could go and take
grafts in that beautiful garden, which moreover was not in Europe;
therefore the inquiry of Ernest is just and sensible. Good fruit-trees
are doubtless natives of some part of the earth, where they bear
spontaneously, in their natural climate, as good fruits as those we
raise in ours with care and art. Such trees were torn from their
native soil when young, and transplanted into Europe, where, by the
assiduous attentions of the gardener, they prospered, and furnished
grafts for their multiplication; for the European climate is so little
fitted to the natural production of good fruits, that the best tree,
propagated from mere seed, soon resumes the wild state, and
requires to be grafted. Gardeners usually collect a number of tender
shoots or saplings in an inclosure, which they first raise by seed and
afterwards ingraft; these inclosures are called nurseries, where such
plants are purchased, and where all these shipped for our use were
procured.
Fritz.—Do you know, father, the native country of all these trees?
Father.—Of most of them, I think. The vine I have ventured to
plant near our tree at Falcon’s Stream, grows only in the temperate
zone; it neither thrives in very cold countries nor under the torrid
zone, though it generally prefers the south to the north. The vine is
of antient date; for we find in the Bible that Noah was acquainted
with the use of it. It seems then that the vine is a native of Asia
Minor and Armenia; and it appears to have been brought at a very
remote period into Egypt, Greece, and most parts of Europe. The
fabulous accounts in mythology of Bacchus, relate, no doubt, to the
propagation of the vine. Italy probably received it from the emigrant
Greeks and from the Romans, who became masters of the known
world; thence it was carried into Gaul, Spain, Germany, and those
parts of Switzerland in which it promised to thrive: perhaps the
Phoenicians too had previously transported it to some of the above
mentioned countries.
The boys speaking together.—And the apple, pear, chesnut,
walnut, almond, peach, and mulberry-trees?
Father.—Patience, patience, prattlers! Can I tell you every thing in
a breath? And pray speak one after the other, gentlemen.—Fruits
with shells or pods, such as the nut, almond, and chesnut, generally
called glands or kernels by the Romans, come from the East; but
that is too general a term,—for the East being the quarter in which
the sun rises, includes too many regions. Chesnuts were called by
the antients glandes Sardes, from Syria, a province of Asia Minor,
and they received their present name from a Grecian town, near
which they were cultivated in abundance. As to walnuts, they were
formerly named glandes Jovis, (Jupiter’s acorns or kernels) by the
Romans; they are originally from Persia, and were spread through
Europe by the Roman conquests. The great filberd-tree, bearing the
Portugal nut or filberd, is originally from Pontus, a country of Asia
Minor, and has been transplanted in the same way as the preceding.
The origin of the almond-tree is uncertain; it is found wild in Asia
and Africa: its fruit bore the name of thalos in Greece, because it
had been first transplanted from an island of that name in the
Archipelago.
Jack.—And cherries, papa,—cherries I like so much? Such
quantities of them are in Switzerland, on the high roads and every
where, that I think them natives of my country.
Father.—Not so, my friend; but of all exotic fruits it is certainly the
one that thrives the most with us. They derived their name from
Cerasus, another town of the Pontus already mentioned to you, and
if I recollect right, the place is called Chirisarda. The celebrated
Roman general Lucullus, after his victory over Mithridates king of
Pontus, was the first who transported them into Europe, seventy
years before the birth of Christ.
 Ernest.—I have not read any thing about cherries in Eutropius,
where I should have been glad to find some account of them.
Father.—I am of your opinion, Ernest, that historians would have
done as well to give the names of those who procured an agreeable
fruit, as in recording the destroyers of mankind.
In this kind of instructive conversation we had got to the end of
our alley of trees, which looked all the better for the uprightness we
had restored them to. This accomplished, we crossed Family Bridge
on our way to the southern plantation of trees, in order to raise and
prop them also. We were delighted with the view of beautiful
orange, citron, and pomegranate trees, that had all taken root and
were thriving to our satisfaction, as well as the pistachio and
mulberry trees. Some of these were in blossom, and inspired us with
the most pleasing hope. We quickly set to work, and my sons with
increased curiosity renewed their inquiries concerning the origin of
these delicious fruits.
Ah! how charming the country must be where such fruits as these
grow spontaneously! exclaimed Fritz.
Considered as confined to this question only, said I, this country,
no doubt, may be termed propitious; but in some other respects it
has likewise its unfavourable side. All the fruits abounding with acid
and refreshing juices, are natives of the torrid or burning zone, or of
parts of the temperate zone most adjacent to it; they seem to have
been designed to cool the mass of blood, and keep off the
inflammatory diseases so frequent in these climates. Orange and
lemon trees are certainly, as I think, from Medea and Assyria; the
Romans called them mala Medica, or Medean apples; they were
brought by the Persians to Albina, and thence into Sicily and the
island of Malta, where the best species of orange trees is still found;
they were afterwards transplanted to Italy, and in succession to
many other parts of Europe. Pomegranates were named mala
Punicea, (Punic apples,) and doubtless were brought by the
Phoenicians and Carthaginians into the southern provinces of
Europe. I have no positive knowledge of the original growth of
pistachios. You are now, children, tolerably well informed on the
subject of fruit trees.
All.—No no, dear father, since you have gone so far, and are so
kind as not to be tired of instructing us, pray tell us the primitive
country of all the trees we have to straighten and prop; the
description amuses us exceedingly.
I am glad of it, and all I know is entirely at your service; but
hearing in this sudden and rapid succession, the species, qualities,
and countries of such a variety of fruits, you will, I fear, overload
your memory and forget the whole.
Fritz.—Pardon me, father, every one will recollect perfectly what
relates to his favourite fruit, and we shall often talk on the subject in
our walks under the shade of the very trees so principally concerned.
Father.—Well, be it so! Thus it is in fact with all the sciences; we
do not easily forget what we wish to know, what is connected with
our desires and hopes;—propose your questions then, and I will
answer you as fully as I am able.
Fritz.—Well then, father, from what part do olives come originally?
Father.—From Armenia and Palestine: Authors of antiquity say that
Hercules brought the first into Europe, and planted them on Mount
Olympus; they were gradually cultivated throughout Greece, and
especially in the territory of Athens, whence they came into Italy,
and from thence were spread over the south of France, and in Spain,
where they are cultivated with the greatest care, on account of the
excellent oil which is extracted from them. Figs belong to the same
native soil; they were brought from Lydia and the isle of Chios into
the Archipelago, in the time of Cato the Elder: they were
transplanted into Gaul a long while after by the emperor Julian, who
was previously prefect or governor of that Roman province. Peaches
are from Persia, and were first named mala Persica (Persian apples).
In Pliny’s time, who lived under the emperor Vespasian, they were
still a novelty in Italy: the family of apricots from Armenia came
amongst the Romans about the same time.
 But now let us go on to plums, your favourite fruit:—Whence do
they originate? Some of the inferior kinds are probably European;
but those of the finest qualities are from other parts: they have
reached us from Damascus, a town of Syria, from which their name
is borrowed. In course of time the crusaders brought several species
of them into Europe, and prunes most likely were among them.
Your favourite apples and pears now claim, I think, a few
observations. We find them first noticed in Greek authors, under the
denomination of Peloponnesian fruits; they were obtained from that
country by the Romans; they found also several kinds of them in
Syria and Alexandria. These two fruits, of all others offer the
greatest varieties, and no doubt several of them are the result of
culture, or of the influence of soil and climate. It is in general a law
of nature, that care and attentive management produce in the same
species a more considerable number of shades and varieties, than is
met with in the wild uncultivated state, in which much sameness
prevails throughout. Man in the rude state of nature is nearly alike
every where, and is destitute of those diversities of character that
naturally unfold themselves in his civilized condition. Every class of
animals and plants, which man renders subservient to his use by
cultivation and due care, exhibits beyond comparison greater
differences in the same species, more varieties, more families, than
those which remain in their native state, remote from and unaided
by his fostering care:—it would appear as if Providence had thus
meant to encourage and reward activity and labour.
Jack.—But, father, you left off with the apples too soon; do pray
resume their history: I wish you may tell us they are of Swiss or
German origin; they are so useful as a fruit, keep so well through
the winter, and may be eaten raw as well as dressed.
Father.—This refreshing fruit, my apple-eating boy, is not a native
of Switzerland or Germany, as you desire it should be, but comes to
us from more favoured climates; at least this is the case with the
best sorts of them. We have a number of wild pear and apple trees,
the fruit of which is crabbed, harsh, and scarcely eatable; whether
they were so originally, or have degenerated, remains to be
determined. As I have said, none of these valuable fruits are
indigenous or native in the colder parts of Europe: yet this
ungrateful and rough climate it is that operates on the European so
as to distinguish him from the inhabitants of the other parts of the
world, by his intelligence, his fitness for toil, and his skill in
agriculture. There exist abundant means and facilities for rendering
man effeminate and indolent, but necessity and want stimulate him
to industry and useful inventions; and by these blessings the
inconveniences of climate are amply compensated.
Jack.—I dare say you are quite right, father;—but tell me where
then do apples come from?
Father.—From the eastern countries, my son; and it is to the
victories of the Romans we are indebted for some of the best kinds,
which have been diversified by experiments, ingrafting, or in other
words the influence of soil and labour and intelligence.
Fritz.—Quince and mulberry-trees are the last we have to inquire
about; and then, father, we will cease our importunities for the
present.
Father.—It is almost time, I must confess. Mulberry-trees are in
general from Asia; they have, I presume, been cultivated more for
the sake of their leaves, on which silkworms feed, than for their
fruit: however, it cannot be denied that the juicy berry of the dark-
coloured mulberry-tree merits to be held in some estimation, and the
white-coloured, whose fruit is small and indifferent, contributes to
the production of the finest silk. The quince-tribe must have taken
its name from the town of Cydonia in the isle of Crete; the Romans
called them pyrus Cydonæ. On the quince-tree may be most
successfully grafted pear-trees designed to be afterwards planted as
espaliers.
Fritz.—But why is it thought right to stunt the growth of a fine
tree, and force it to remain diminutive?
Father.—This, in several respects, is useful; wall-trees, being
sheltered on one side, bear earlier and more choice fruits; it is easier
to defend them from insects; their fruits are more conveniently
gathered. The tree, giving less shade, is not so injurious to the
culinary plants that are near it.—Are not these substantial reasons?
Jack.—Then I must ask, why are not all trees set in this way?
That would not be a judicious plan by any means; an espalier
takes up too much ground; besides, trees with high stems produce
more fruit, they form orchards; a crop of hay too may be raised
under them, whereas espaliers serve in general as fences or
boundaries in gardens.
This is a compendium of our morning’s conversation, in the course
of which we finished our work in the completest manner. Towards
noon, a keen appetite hastened our return to Falcon’s Stream, where
we found an excellent and plentiful dinner prepared by our good and
patient steward, of which the palm-tree cabbage was the chief dish.
We all agreed that to eat of a better or more delicate food was
impossible; and Ernest, who had procured it, received the thanks of
all the board.
When the sharpness of hunger was appeased, a new subject was
introduced which I and my wife had been seriously revolving for
some time; she found it difficult and even dangerous to ascend and
descend our tree with a rope ladder: we never went there but on
going to-bed, and each time felt an apprehension that one of the
children, who scrambled up like cats, might make a false step and
perhaps be lamed for ever. Bad weather might come on and compel
us for a long time together to seek an asylum in our aërial
apartment, and consequently to ascend and descend oftener.
My wife addressed me constantly on the subject, incessantly
asking whether my inventive genius could not suggest some easier
and less perilous mode of getting to our dwelling. I smiled at her
implicit confidence that I could accomplish wonders: I assured her
that if I were an enchanter or magician no desire of hers should
remain ungratified, and that with a single touch of my wand I would
instantly produce for her a commodious firm stair-case of perfect
workmanship; but that not being the case, I acknowledged myself at
a loss for the means to effect such an accommodation for her: still
her reiterated appeals and my own anxiety had often made me
reflect if the thing were really possible? A stair-case on the outside
was not to be thought of, the considerable height of the tree
rendered that impracticable, as I had nothing to rest it on, and
should be at a loss to find beams to sustain it; but I had for some
time formed the idea of constructing winding stairs within the
immense trunk of the tree, if it should happen to be hollow, or I
could contrive to make it so: Francis had excited this idea in
speaking of the bees.
Did you not tell me, dear wife, said I, that there is a hole in the
trunk of this enormous tree of ours, in which a swarm of bees is
lodged?
Without doubt, answered she; it was there little Francis was so
severely stung in attempting to thrust in a stick; look at it yourself,
you will see the bees go in and come out in throngs.
Then, replied I, we have only to examine how far this excavation
goes, whether it extends to the roots, and what the circumference of
it is; this done, we shall have gained the first difficult step in favour
of our stair-case.
All my children seized the idea with ardour; they sprang up, and
prepared themselves to climb the tops of the roots like squirrels, to
succeed in striking at the trunk with axes, and to judge from the
sound how far it was hollow; but they soon paid dearly for their
attempt: the whole swarm of bees, alarmed at the noise made
against their dwelling, issued forth, buzzing with fury, attacked the
little disturbers, began to sting them, stuck to their hair and clothes,
and soon put them to flight, bearing along with them their enemies,
and uttering lamentable cries. My wife and I had some trouble to
stop the course of this uproar, and cover their little wounds with
fresh earth to allay the smart. Jack, whose temper was on all
occasions rash, had struck exactly upon the bees’ nest, and was
more severely attacked by them than the rest; it was necessary, so
serious was the injury, to cover the whole of his face with linen. The
less active Ernest got up the last, and was the first to run off when
he saw the consequences, and thus avoided any further injury than
a sting or two; but some hours elapsed before the other boys could
open their eyes or be in the least relieved from the acute pain that
had been inflicted. When they grew a little better, the desire of being
avenged of the insects that had so roughly used them had the
ascendant in their minds: they teased me to hasten the measures
for getting every thing in readiness for obtaining possession of their
honey. The bees in the mean time were still buzzing furiously round
the tree. I prepared tobacco, a pipe, some clay, chisels, hammers,
&c. I took the large gourd long intended for a hive, and I fitted a
place for it by nailing a piece of board on a branch of the tree; I
made a straw roof for the top to screen it from the sun and rain; and
as all this took up more time than I was aware of, we deferred the
attack of the fortress to the following day, and got ready for a sound
sleep, which completed the cure of my little wounded patients.
 CHAPTER XXIX.
Victory over the bees;—winding stair-case; training
of various animals; divers manufactures; fountain,
&c.

Next morning almost before dawn all were up and in motion; the
bees had returned to their cells, and I stopped the passages with
clay, leaving only a sufficient aperture for the tube of my pipe. I then
smoked as much as was requisite to stupefy without killing the little
warlike creatures. Not having a cap with a mask, such as bee-
catchers usually wear, nor even gloves, this precaution was
necessary. At first a humming was heard in the hollow of the tree,
and a noise like a gathering tempest, which died away by degrees.
All was become calm, and I withdrew my tube without the
appearance of a single bee. Fritz had got up by me: we then began
with a chisel and a small axe to cut out of the tree, under the bees’
hole of entrance, a piece three feet square. Before it was entirely
separated, I repeated the fumigation, lest the stupefaction produced
by the first smoking should have ceased, or the noise we had been
just making revived the bees. As soon as I supposed them quite
lulled again, I separated from the trunk the piece I had cut out,
producing as it were the aspect of a window, through which the
inside of the tree was laid entirely open to view; and we were filled
at once with joy and astonishment on beholding the immense and
wonderful work of this colony of insects. There was such a stock of
wax and honey, that we feared our vessels would be insufficient to
contain it. The whole interior of the tree was lined with fine honey-
combs: I cut them off with care, and put them in the gourds the
boys constantly supplied me with. When I had somewhat cleared the
cavity, I put the upper combs, in which the bees had assembled in
clusters and swarms, into the gourd which was to serve as a hive,
and placed it on the plank I had purposely raised. I came down,
bringing with me the rest of the honey-combs, with which I filled a
small cask, previously well washed in the stream. Some I kept out
for a treat at dinner; and had the barrel carefully covered with cloths
and planks, that the bees, when attracted by the smell, might be
unable to get at it. We then sat round the table, and regaled
ourselves plentifully with the delicious and odoriferous treat of the
honey. Having finished our meal, my wife put by the remainder; and
I proposed to my sons to go back to the tree, in order to prevent the
bees from swarming again there on being roused from their stupor,
as they would not have failed to do, but for the precaution I took of
passing a board at the aperture, and burning a few handfuls of
tobacco on it, the smell and smoke of which drove them back from
their old abode, whenever they attempted to return to it. At length
they desisted from approaching it, and became gradually reconciled
to their new residence, where their queen no doubt had settled
herself. I took this opportunity to relate to my children all I had read
in the interesting work by Mr. Huber of Geneva17 of the queen-bee,
this beloved and respected mother of her subjects, who are all her
children, and who take care of and guard her, work for her, nourish
the rising swarms, make the cells in which they are to lodge, prepare
others of a different structure, as well as nutriment for the young
queens destined to lead forth the fresh colonies: and I entered into
all those details which celebrated observers, and particularly the one
we have just mentioned, have described so interestingly. These
accounts highly entertained my youthful auditory, who almost
regretted having molested by their depredation the repose of a fine
peaceable kingdom that had flourished so long without interruption
in the huge trunk. As to me, it so well suited my intended stair-case,
that I readily adopted the prevailing moral amongst conquerors, who
dispense with scruples when the seizing a country is convenient to
their policy, and I resolved to take full possession next day. In the
mean time I advised all to watch during the night, over the whole
provision of honey obtained while the bees were torpid, who when
recovered would not fail to be troublesome, and come in legions to
get back to their property. That we might not be ourselves injured
by so much fatigue, we went and threw ourselves on our beds, and
in our clothes, to take a short doze before the hour of retreat; we
were lulled to sleep with their buzzing, which had quite ceased when
we awoke at the coming on of night; they had remained quiet in the
gourd or suspended in clusters from some branches: without
concerning ourselves about them, we went promptly to business;
the cask of honey was emptied into a kettle, except a few prime
combs which we kept for daily consumption; the remainder mixed
with a little water was set over a gentle fire and reduced to a liquid
consistence, strained and squeezed through a bag, and afterwards
poured back into the cask, which was left upright and uncovered all
night to cool. In the morning the wax was entirely separated, and
had risen to the surface in a compact and solid cake that was easily
removed; beneath was the purest, most beautiful and delicate honey
that could be seen: the cask was then carefully headed again, and
put into cool ground near our wine vessels; and now we promised
ourselves an abundant supply of an agreeable article for desserts.
This task accomplished, I mounted to revisit the hive, and found
every thing in order; the bees going forth in swarms and returning
loaded with wax, from which I judged they were forming fresh
edifices in their new dwelling place. I was surprised to see the
numbers that had occupied the trunk of the tree find room in the
gourd; but on looking round me, I perceived a part of them collected
in a cluster upon a branch, and I thence concluded a young queen
was amongst them.
On perceiving this, I procured another gourd, into which I shook
them and placed it by the former: thus I had the satisfaction of
obtaining at an easy rate two fine hives of bees in activity.
 We soon after these operations proceeded to examine the inside
of the tree. I sounded it with a pole from the opening I had made
towards the top; and a stone fastened to a string served us to sound
the bottom, and thus to ascertain the height and depth of the cavity.
To my great surprise the pole penetrated without any resistance to
the branches on which our dwelling rested, and the stone descended
to the roots. The trunk, it appeared, had wholly lost its pith, and
most of its wood internally; nothing therefore was more practicable
than to fix winding stairs in this capacious hollow, that should reach
from top to bottom. It seems that this species of tree, like the willow
in our climates, receives nourishment through the bark; for it did not
look decayed, and its far-extended branches were luxuriant and
beautiful in the extreme. I determined to begin our construction that
very day. The undertaking appeared at first beyond our powers; but
intelligence, patience, time, and a firm resolution vanquished all
obstacles. We were not disposed to relax in any of these requisites;
and I was pleased to find opportunities to keep my sons in continual
action, and their minds and bodies were all the better for exertion.
They grew tall, strong, and were too much engaged to regret, in
ignoble leisure, any of their past enjoyments in Europe.
We began to cut into the side of the tree, towards the sea, a door-
way equal in dimensions to the door of the captain’s cabin, which we
had removed with all its frame-work and windows; by means of
which we should at once be guarded against every attack on that
side. We next cleared away from the cavity all the rotten wood, and
rendered the interior even and smooth, leaving sufficient thickness
for cutting out resting-places for the winding stairs, without injuring
the bark. I then fixed in the centre, the trunk of a tree ten or twelve
feet high and a foot thick, completely stripped of its branches, in
order to carry my winding staircase round it: on the outside of this
trunk, and the inside of the cavity of our own tree, we formed
grooves, so calculated as to correspond with the distances at which
the boards were to be placed to form the stairs. These were
continued till I had got to the height of the trunk round which they
turned. The window I had opened at the top to take out the honey
gave light enough. I made a second aperture below, and a third
above it, and thus completely lighted the whole ascent. I also
effected an opening near our room, that I might more conveniently
finish the upper part of the stair-case. A second trunk was fixed
upon the first, and firmly sustained with screws and transverse
beams. It was surrounded like the other with stairs cut slopingly;
and thus we happily effected the stupendous undertaking of
conducting it to the level of our bed-chamber. Here I made another
door directly into it; and I then found I could add nothing further to
my design. If my staircase was not in strict conformity to the rules of
architecture, it at least answered the purpose it was built for, that of
conducting us with safety and shelter to our nocturnal residence. To
render it more solid and agreeable, I closed the spaces between the
stairs with plank. I then fastened two strong ropes, the one
descending the length of the little tree, the other along the side of
the large one, to assist in case of slipping. I fixed the sash-windows
taken from the captain’s cabin in the apertures we had made to give
light to the stairs; and when the whole was complete, it was so
pretty, solid, and convenient, that we were never tired of going up
and coming down it; and I fear I must add, for the sake of truth,
with no small admiration of our united talents. I must, however,
candidly own, that we succeeded in this arduous attempt by mere
dint of efforts, patience, industry, and time; for it occupied us for
several weeks together with no intermission. It more than once
reminded me of the wise system of education of the philosopher of
Geneva, J. J. Rousseau; and particularly where he recommends that
boys of all classes in society should learn a trade, and especially that
of a carpenter. How happy should I have been in our circumstances
to have known this trade myself, and to have taught it to my eldest
son! I cannot too earnestly exhort all fathers to put their sons in
early possession of a resource which, though it may not become of
the first necessity, has, at all events, the advantage of making a
young man stronger, and more dexterous; of filling up many of the
dangerous idle hours of ardent youth; and of being able, in maturer
age, if it be unnecessary to work ourselves, at least to overlook the
workmen we employ. I am not an enthusiast for the system of
Welcome to our website – the ideal destination for book lovers and
knowledge seekers. With a mission to inspire endlessly, we offer a
vast collection of books, ranging from classic literary works to
specialized publications, self-development books, and children's
literature. Each book is a new journey of discovery, expanding
knowledge and enriching the soul of the reade

Our website is not just a platform for buying books, but a bridge
connecting readers to the timeless values of culture and wisdom. With
an elegant, user-friendly interface and an intelligent search system,
we are committed to providing a quick and convenient shopping
experience. Additionally, our special promotions and home delivery
services ensure that you save time and fully enjoy the joy of reading.

Let us accompany you on the journey of exploring knowledge and

personal growth!

ebookfinal.com