Gpebook PDF

GAUSS Programming
for Econometricians and Financial Analysts
Kuan-Pin Lin
ETEXT
Los Angeles
www.etext.net
COMPUTATIONAL ECONOMETRICS
GAUSS Programming for Econometricians and Financial Analysts
Copyright  2001-2003 by K.–P. Lin
ISBN 0-9705314-3-5
Published by ETEXT Textbook Publisher, www.etext.net.
All rights reserved. No part of this book and the accompanying software may be reproduced,
stored in a retrieval system, translated or transcribed, in any form or by any means—
electronic, mechanical, photocopying, recording, or otherwise—without the prior written
permission of the copyright owner.
For permission requests or further information, see www.etext.net or email [email protected].
Printed in the United States of America.
Limit of Liability and Disclaimer of Warranty
Although every precaution has been taken in the preparation this book and the accompanying
software, the publisher and author make no representation or warranties with respect to the
accuracy or completeness of the contents, and specifically disclaim any implied warranties of
merchantability or fitness for any particular purpose, and shall in no event be liable for any
loss of profit or any damages arising out of the use of this book and the accompanying
software.
Trademarks
GAUSS is a trademark of Aptech Systems, Inc. GPE2 is a product name of Applied Data
Associates. All other brand names and product names used in this book are trademarks,
registered trademarks, or trade names of their respective holders.
ii
Preface
Computational Econometrics is an emerging field of applied economics which focuses on the
computational aspects of econometric methodology. To explore an effective and efficient
approach for econometric computation, GAUSS Programming for Econometricians and
Financial Analysts (GPE) was originally developed as the outcome of a faculty-student joint
project. The author developed the econometric program and used it in the classroom. The
students learned the subject materials and wrote about their experiences in using the program
and GAUSS.
We know that one of the obstacles in learning econometrics is the need to do computer
programming. Who really wants to learn a new programming language while at the same time
struggling with new econometric concepts? This is probably the reason that “easy-to-use”
packages such as RATS, SHAZAM, EVIEWS, and TSP are often used in teaching and
research. However, these canned packages are inflexible and do not allow the user sufficient
freedom in advanced modeling. GPE is an econometrics package running in the GAUSS
programming environment. You write simple codes in GAUSS to interact with GPE
econometric procedures. In the process of learning GPE and econometrics, you learn GAUSS
programming at your own pace and for your future development.
Still, it takes some time to become familiar with GPE, not to mention the GAUSS language.
The purpose of this GPE project is to provide hands-on lessons with illustrations on using the
package and GAUSS. GPE was first developed in 1991 and has since undergone several
updates and revisions. The first version of the project, code-named LSQ, started in the
summer of 1995 with limited functions of least squares estimation and prediction. This book
and CDROM represent a major revision of this work in progress, including linear and
nonlinear regression models, simultaneous linear equation systems, and time series analysis.
Here, in your hands, is the product of GPE. The best way to learn GPE is to read the book,
type in and run each lesson, and explore the sample programs and output. For your
convenience, all the lessons and data files are available on the distribution disk.
During several years of teaching econometrics using the GPE package, many students
contributed to the ideas and codes in GPE. Valuable feedback and suggestions were
incorporated into developing this book. In particular, the first LSQ version was a joint project
with Lani Pennington, who gave this project its shape. Special thanks are due to Geri
Manzano, Jennifer Showcross, Diane Malowney, Trish Atkinson, and Seth Blumsack for their
efforts in editing and proofreading many draft versions of the manuscript and program
lessons. As always, I am grateful to my family for their continuing support and understanding.
iii
Table of Contents
PREFACE................................................................................................................................iii
TABLE OF CONTENTS ............................................................................................................. v
I INTRODUCTION ................................................................................................................... 1
Why GAUSS?..................................................................................................................... 1
What is GPE? .................................................................................................................... 1
Using GPE......................................................................................................................... 2
II GAUSS BASICS .............................................................................................................. 5
Getting Started................................................................................................................... 5
An Introduction to GAUSS Language................................................................................ 7
Creating and Editing a GAUSS Program........................................................................ 17
Lesson 2.1 Let’s Begin ...............................................................................................................18
File I/O and Data Transformation................................................................................... 21
Lesson 2.2: File I/O ....................................................................................................................23
Lesson 2.3: Data Transformation................................................................................................25
GAUSS Built-In Functions............................................................................................... 26
Lesson 2.4: Data Analysis ..........................................................................................................32
Controlling Execution Flow ............................................................................................ 33
Writing Your Own Functions........................................................................................... 36
User Library .................................................................................................................... 40
GPE Package................................................................................................................... 41
III LINEAR REGRESSION MODELS ..................................................................................... 43
Least Squares Estimation ................................................................................................ 43
Lesson 3.1: Simple Regression...................................................................................................44
Lesson 3.2: Residual Analysis ....................................................................................................46
Lesson 3.3: Multiple Regression ................................................................................................48
Estimating Production Function...................................................................................... 50
Lesson 3.4: Cobb-Douglas Production Function ........................................................................51
Lesson 3.5: Testing for Structural Change..................................................................................55
Lesson 3.6: Residual Diagnostics ...............................................................................................58
IV DUMMY VARIABLES .................................................................................................... 63
Seasonality....................................................................................................................... 63
Lesson 4.1: Seasonal Dummy Variables.....................................................................................64
Lesson 4.2: Dummy Variable Trap.............................................................................................67
Structural Change............................................................................................................ 68
Lesson 4.3: Testing for Structural Change: Dummy Variable Approach ...................................68
V MULTICOLLINEARITY ..................................................................................................... 73
Detecting Multicollinearity.............................................................................................. 73
Lesson 5.1: Condition Number and Correlation Matrix..............................................................73
Lesson 5.2: Theil’s Measure of Multicollinearity.......................................................................75
Lesson 5.3: Variance Inflation Factors (VIF) .............................................................................77
Correction for Multicollinearity ...................................................................................... 78
Lesson 5.4: Ridge Regression and Principal Components..........................................................78
VI NONLINEAR OPTIMIZATION ......................................................................................... 81
Solving Mathematical Functions ..................................................................................... 81
Lesson 6.1: One-Variable Scalar-Valued Function.....................................................................82
Lesson 6.2: Two-Variable Scalar-Valued Function....................................................................85
Estimating Probability Distributions ............................................................................... 87
Lesson 6.3: Estimating Probability Distributions....................................................................... 88
Lesson 6.4: Mixture of Probability Distributions ....................................................................... 91
Statistical Regression Models .......................................................................................... 93
Lesson 6.5: Minimizing Sum-of-Squares Function.................................................................... 94
Lesson 6.6: Maximizing Log-Likelihood Function.................................................................... 96
VII NONLINEAR REGRESSION MODELS ............................................................................99
Nonlinear Least Squares.................................................................................................. 99
Lesson 7.1: CES Production Function...................................................................................... 100
Maximum Likelihood Estimation ................................................................................... 101
Lesson 7.2: Box-Cox Variable Transformation........................................................................ 104
Statistical Inference in Nonlinear Models...................................................................... 108
Lesson 7.3: Hypothesis Testing for Nonlinear Models ............................................................ 109
Lesson 7.4: Likelihood Ratio Tests of Money Demand Equation............................................ 112
VIII DISCRETE AND LIMITED DEPENDENT VARIABLES ..................................................113
Binary Choice Models.................................................................................................... 113
Lesson 8.1: Probit Model of Economic Education................................................................... 115
Lesson 8.2: Logit Model of Economic Education .................................................................... 119
Limited Dependent Variable Models ............................................................................. 121
Lesson 8.3: Tobit Analysis of Extramarital Affairs.................................................................. 122
IX HETEROSCEDASTICITY ..............................................................................................127
Heteroscedasticity-Consistent Covariance Matrix ........................................................ 127
Lesson 9.1: Heteroscedasticity-Consistent Covariance Matrix ................................................ 127
Weighted Least Squares ................................................................................................. 130
Lesson 9.2: Goldfeld-Quandt Test and Correction for Heteroscedasticity ............................... 130
Lesson 9.3: Breusch-Pagan and White Tests for Heteroscedasticity........................................ 132
Nonlinear Maximum Likelihood Estimation .................................................................. 134
Lesson 9.4: Multiplicative Heterscedasticity............................................................................ 135
X AUTOCORRELATION .....................................................................................................143
Autocorrelation-Consistent Covariance Matrix............................................................. 143
Lesson 10.1: Heteroscedasticity-Autocorrelation-Consistent Covariance Matrix .................... 144
Detection of Autocorrelation ......................................................................................... 146
Lesson 10.2: Tests for Autocorrelation .................................................................................... 147
Correction for Autocorrelation ...................................................................................... 149
Lesson 10.3: Cochrane-Orcutt Iterative Procedure .................................................................. 151
Lesson 10.4: Hildreth-Lu Grid Search Procedure .................................................................... 153
Lesson 10.5: Higher-Order Autocorrelation............................................................................. 154
Autoregressive and Moving Average Models: An Introduction ..................................... 157
Lesson 10.6: ARMA(1,1) Error Structure ................................................................................ 158
Nonlinear Maximum Likelihood Estimation .................................................................. 161
Lesson 10.7: Nonlinear ARMA Model Estimation .................................................................. 161
XI DISTRIBUTED LAG MODELS .......................................................................................167
Lagged Dependent Variable Models.............................................................................. 167
Lesson 11.1: Testing for Autocorrelation with Lagged Dependent Variable ........................... 167
Lesson 11.2: Instrumental Variable Estimation........................................................................ 170
Polynomial Lag Models ................................................................................................. 173
Lesson 11.3: Almon Lag Model Revisited............................................................................... 173
Autoregressive Distributed Lag Models......................................................................... 176
Lesson 11.4: Almon Lag Model Once More............................................................................ 176
XII GENERALIZED METHOD OF MOMENTS .....................................................................179
GMM Estimation of Probability Distributions............................................................... 179
Lesson 12.1 Gamma Probability Distribution .......................................................................... 181
vi
GMM Estimation of Econometric Models ......................................................................185
Lesson 12.2 A Nonlinear Rational Expectations Model ...........................................................188
Linear GMM ...................................................................................................................192
Lesson 12.3 GMM Estimation of U.S. Consumption Function ................................................192
XIII SYSTEM OF SIMULTANEOUS EQUATIONS ............................................................... 195
Linear Regression Equations System..............................................................................195
Lesson 13.1: Klein Model I ......................................................................................................197
Lesson 13.2: Klein Model I Reformulated................................................................................202
Seemingly Unrelated Regression Equations System (SUR) ............................................204
Lesson 13.3: Berndt-Wood Model............................................................................................204
Lesson 13.4: Berndt-Wood Model Extended............................................................................207
Nonlinear Maximum Likelihood Estimation...................................................................209
Lesson 13.5: Klein Model I Revisited ......................................................................................211
XIV UNIT ROOTS AND COINTEGRATION ....................................................................... 215
Testing for Unit Roots.....................................................................................................216
Lesson 14.1: Augmented
Dickey-Fuller Test for Unit Roots ...........................................................................................217
Testing for Cointegrating Regression.............................................................................223
Lesson 14.2: Cointegration Test: Engle-Granger Approach .....................................................225
Lesson 14.3: Cointegration Test: Johansen Approach ..............................................................230
XV TIME SERIES ANALYSIS ........................................................................................... 233
Autoregressive and Moving Average Models .................................................................234
Lesson 15.1: ARMA Analysis of Bond Yields.........................................................................235
Lesson 15.2: ARMA Analysis of U.S. Inflation .......................................................................239
Autoregressive Conditional Heteroscedasticity..............................................................240
Lesson 15.3 ARCH Model of U.S. Inflation.............................................................................243
Lesson 15.4 ARCH Model of Deutschemark-British Pound Exchange Rate ...........................245
XVI PANEL DATA ANALYSIS ........................................................................................ 251
Fixed Effects Model ........................................................................................................251
Lesson 16.1: One-Way Panel Data Analysis: Dummy Variable Approach ..............................253
Random Effects Model....................................................................................................256
Lesson 16.2: One-Way Panel Data Analysis: Deviation Approach ..........................................258
Lesson 16.3: Two-Way Panel Data Analysis............................................................................261
Seemingly Unrelated Regression System ........................................................................263
Lesson 16.4: Panel Data Analysis for Investment Demand: Deviation Approach....................264
Lesson 16.5: Panel Data Analysis for Investment Demand: SUR Method ...............................266
XVII LEAST SQUARES PREDICTION .............................................................................. 271
Predicting Economic Growth .........................................................................................271
Lesson 17.1: Ex-Post Forecasts and Forecast Error Statistics...................................................272
Lesson 17.2: Ex-Ante Forecasts ...............................................................................................277
EPILOGUE ................................................................................................................. 281
APPENDIX A GPE CONTROL VARIABLES ............................................................ 283
Input Control Variables..................................................................................................283
General Purpose Input Control Variables .................................................................................283
ESTIMATE Input Control Variables........................................................................................284
FORECAST Input Control Variables .......................................................................................292
Output Control Variables ...............................................................................................293
ESTIMATE Output Control Variables .....................................................................................293
FORECAST Output Control Variables ....................................................................................294
APPENDIX B GPE APPLICATION MODULES .......................................................... 297
Application Module B-1: GMM.GPE .............................................................................297
vii
Application Module B-2: JOHANSEN.GPE .................................................................. 299
Application Module B-3: PANEL1.GPE........................................................................ 300
Application Module B-4: PANEL2.GPE........................................................................ 303
APPENDIX C STATISTICAL TABLES ......................................................................313
Table C-1. Critical Values for the Dickey-Fuller Unit Root Test Based
on t-Statistic ................................................................................................................... 313
Table C-2. Critical Values for the Dickey-Fuller Unit Root Test Based
on F-Statistic.................................................................................................................. 315
Table C-3. Critical Values for the Dickey-Fuller Cointegration t-Statistic
τρ Applied on Regression Residuals.............................................................................. 316
Table C-4. Critical Values for Unit Root and Cointegration Tests Based
on Response Surface Estimates...................................................................................... 317
Table C-5: Critical Values for the Johansen’s Cointegration Likelihood
Ratio Test Statistics........................................................................................................ 319
REFERENCES ..........................................................................................................321
INDEX ...........................................................................................................................323
viii
I
Introduction
GAUSS Programming for Econometricians and Financial Analysts (GPE) is a
package of econometric procedures written in GAUSS, and this book is about
GAUSS programming for econometric analysis and applications using GPE. To
explore the computational aspects of applied econometrics, we adopt the
programming environment of GAUSS and GPE.
As you probably know, GAUSS is a programming language designed for matrix-

based operations and manipulations, suitable for high level statistical and
econometric computation. Many universities and research institutions have used
GAUSS in their econometrics curricula. Unfortunately, GAUSS is not an easy
language to learn and master, particularly for those without computer programming
experience. GPE is designed to provide access to the full power of GAUSS without
the intimidation of learning a new programming language. By using GPE, getting
acquainted with techniques for econometric analysis as well as the GAUSS
programming environment is fast and easy. This book was written so that you could
easily use GAUSS as a tool for econometric applications.
You cannot learn econometrics by just reading your textbook or by just writing
GAUSS code or programs. You must interact with the computer and textbook by
working through the examples. That is what this book is all about—learning by
doing.
Why GAUSS?
GAUSS is a programming language similar to C or Pascal. GAUSS code works on
matrices as the basis of a complete programming environment. It is flexible and
easily applies itself to any kind of matrix-based computation.
GAUSS comes with about 400 intrinsic commands ranging from file input/output
(I/O) and graphics to high-level matrix operations. There are many GAUSS libraries
and application packages, which take advantage of these built-in commands and
procedures for implementing accurate and efficient computations.
The use of libraries and packages hides complex programming details and simplifies
the interface with a set of extended procedures and control variables. For instance,
GAUSS supports publication quality graphics by use of a library which extends the
main system with a set of control variables manipulated on the defined graphic
procedures.
What is GPE?
GPE is a GAUSS package for linear and nonlinear regression useful for econometric
analysis and applications. GPE contains many econometric procedures controllable
by a few groups of global variables. It covers most basic econometric computations
GAUSS PROGRAMMING FOR ECONOMETRICIANS AND FINANCIAL ANALYSTS
including single linear equation estimation and prediction, systems of simultaneous

linear equations, nonlinear models, and time series analysis.
However, beyond econometric computation, GPE does not provide a user interface
for data input and output nor are there any procedures for data transformation. Both
of these operations and other related topics, which build the interaction between GPE
and the GAUSS programming environment, will be discussed in the next chapter on
GAUSS Basics. Using the GPE package in a GAUSS environment is first introduced
in Chapter III on linear least squares estimation and is the foundation of the rest of
the book.
Using GPE
This book and CDROM were developed based on the latest version of GAUSS for
Windows1. Before using the GPE package, it must be properly installed with your
GAUSS program. Install GPE according to the instructions given with the
distribution CD. Make sure that the version number of GPE matches with that of
your GAUSS program.2
Following the completion of GPE installation, the compiled GPE program named
GPE2.GCG should reside in the GAUSS directory. GPE2.GCG is a compiled
GAUSS program. It is an encoded binary file, which requires the correct version of
GAUSS. In addition, a GPE subdirectory is created and stores all the lesson
programs and data files. GPE is the working directory for all the empirical lessons.
By going through this book lesson by lesson, program files may be overwritten and
additional output files are generated. If you want a fresh start, just reinstall the GPE
package.
All the GPE lesson programs are written with direct reference to the GPE
subdirectory created during installation. Using the default GPE subdirectory is
convenient because all the lesson programs and data files are already there for you to
explore. Alternately, you may want to use a working diskette for the practice of
creating each lesson. If you don’t mind typing, using a working diskette is not only
portable but also a true hands-on experience. You need only to change the references
of the GPE subdirectory in each lesson program to the floppy drive your working
diskette resides on (a: is assumed). That is, in the beginning of each lesson program,
replace gpe\ with a:\. You may also need to copy the required data files to the
working diskette. A working diskette is recommended especially if you are using
GAUSS in a laboratory environment.
It is important to recognize that this book is not a GAUSS how-to manual or program
documentation, for which you are advised to consult GAUSS for Windows User
Guide and GAUSS Language References supplied from Aptech Systems. Also, this is
not a book on econometrics, although many fundamental formulas for econometric
computation are introduced in order to use the implemented algorithms and routines.
There are many textbooks on econometrics that describe the technical details. Rather,
this is a book on computational aspects of implementing econometric methods. We
provide step by step instruction using GPE and GAUSS, complete with explanations
and sample program codes. GAUSS program codes are given in small chunks in a
piece-meal construction. Each chunk, or lesson, offers hands-on practice for
1
This writing is based on GAUSS for Windows version 5.0.
2
GPE is also available for earlier versions of GAUSS.
2
INTRODUCTION
economic data analysis and econometric applications. Most examples can be used on
different computer platforms without modification.
Conventions Used in this Book

To distinguish our explanations from your typing, as seen on your video display, all
program code and output are in the monospace font Courier. For reference
purposes, each line of program code is numbered. Menu items in the Windows
interface, directory paths, file names, and key-stroke combinations are in bold. In
addition, the following icons are used to designate special information:
Extra notes and additional information are given here.
This warns of common mistakes causing programming errors.
Hints or remarks specific to GAUSS and GPE.3
A number of abbreviations for statistical and econometric terms are used in this text.
Although all are defined upon their first appearance, we provide a list of these
abbreviations below for reference purposes:
2SLS Two-stage least squares

3SLS Three-stage least squares
ACF Autocorrelation function
ADF Augmented Dickey-Fuller test
AIC Akaike Information Criterion
AOV Analysis of Variance
ARCH Autoregressive conditional heteroscedasticity
ARDL Autoregressive distributed lag
ARMA Autoregressive moving average
BFGS Broyden-Fletcher-Goldfarb-Shanno quasi-Newton optimization method
BHHH Berndt-Hall-Hall-Hausman maximum likelihood estimation method
BIC Schwartz Baysian Information Criterion
DF Dickey-Fuller test
DGP Data generating process
FIML Full information maximum likelihood
GARCH Generalized autoregressive conditional heteroscedasticity
GMM Generalized method of moments
IV Instrumental variable estimation
LIML Limited information maximum likelihood
LM Lagrangian multiplier
LR Likelihood Ratio
ML Maximum likelihood
OLS Ordinary least squares
PACF Partial autocorrelation function
3
We thank Aptech Systems for permission to use their GAUSS 3.2 “hammer on numbers”
icon.
3
QHC Quadratic hill-climbing optimization method

RSS Residual sum-of-squares
SUR Seemingly unrelated regressions
VAR Vector autoregression
VIF Variance Inflation Factors
4
II
GAUSS Basics
GAUSS is a high-level computer language suitable for mathematical and matrix-
oriented problem solving. It can be used to solve any kind of mathematical,
statistical, or econometric model. Since GAUSS is a computer language, it is
flexible. But it is also more difficult to learn than most canned (prewritten)
econometric programs such as EVIEWS, SHAZAM, and TSP.
In this chapter we begin with the basics of starting GAUSS for Windows. After
learning how to get in and out of GAUSS, we discuss much of the GAUSS language.
At the end of the chapter, we introduce the GPE (GAUSS Programming for
Econometricians and Financial Analysts) package and briefly describe its capacity
for econometric analysis and applications.
Getting Started
Start GAUSS for Windows in one of the following ways:
• Click the short-cut (an icon with GAUSS logo) on the desktop.
• From Start button at the lower left corner, from Programs menu, select and run
GAUSS.
• Use Windows Explorer or File Manager to locate the GAUSS directory4 and
execute the file GAUSS.EXE.
To quit and exit GAUSS for Windows, do either one of the following:
• Click and select File/Exit from the menu bar.

• Click on the “close” button (the box with the “X”) in the upper right-hand corner
of the GAUSS main window.
Windows Interface
If you are new to the GAUSS programming environment, you need to spend some
time to familiarize yourself with the GAUSS Windows interface. A good reference is
GAUSS for Windows User Guide. Or, from the menu bar go to Help/Contents to
learn about GAUSS and its Windows interface. Understanding the working function
of each button on the menu bar, toolbar (below the menu bar), and status bar (bottom
bar of the main window) is the crucial beginning of GAUSS programming.
Briefly, GAUSS for Windows runs in two modes: Command and Edit. Each mode
has its own window. The Command Input-Output window (or Command mode) is
4
GAUSS directory refers to the directory in which you have successfully installed the
GAUSS program in your computer. Assuming C: is your boot drive, by default installation,
the GAUSS directory may be C:\GAUSS, C:\GAUSS50 (for Version 5.0), or C:\GAUSSLT
(for Light Version 5.0). In the following, we refer to C:\GAUSS as the GAUSS directory.
for running single-line commands or program files. It is more commonly referred as

the interactive mode. The Edit window (or Edit mode) is for modifying or editing
program and data files. A file is created from the menu bar File/New. An existing
file can be open and edited from the menu bar File/Open. There is only one
Command Input-Output window, but you can open as many as Edit windows as
needed for the program, data, output, etc. The title of each Edit window consists of
the directory path and file name to indicate where the contents came from. From the
Run button on the menu bar, a program file is executed either from Run Main File
or Run Active File. Your program output can be displayed either in the Command
Input-Output window or in a separate Output window (if Window/Output Window
is selected from the menu bar). The ability to work simultaneously with multiple
program and data files in GAUSS allows straightforward monitoring of project
development. Screen displays of GAUSS Command and Edit Windows look like the
following (your screen may be slightly different because of differences in the
configuration setup of the Windows environment you use):
6
GAUSS BASICS
You may want to configure the programming environment to fit your taste as
desired. This is done from the menu bar button Configure in which you can change
the program setup and window properties. In the GAUSS programming
environment, you can also trace and debug a program file in the Debug window.
This is more suited for a programmer in developing a large program, which we will
not cover in this book.
An Introduction to GAUSS Language5

The rest of this chapter covers the basics of GAUSS language. It is written for
anyone who has no prior or only limited computer programming knowledge. Only
the basics of GAUSS programming are introduced, followed by discussions of more
advanced topics useful for econometric analysis. We aspire to promote a reasonable
proficiency in reading and understanding procedures that we will write in the
GAUSS language. If you are in a hurry to use the econometric package GPE for the
project at hand, you can skip the rest of this chapter and go directly to the next
chapter on linear regression models and least squares estimation. However, we
recommend that later, at your leisure, you come back for a thorough overview of the
GAUSS language.
We have seen that GAUSS commands are either written in the Command or Edit
mode. Command mode executes each line of code as it is written. Simple GAUSS
commands can be typed and executed (by pressing the carriage return or Enter key)
5
This session is written based on introductory materials for MathWorks’ MATLAB prepared
by William F. Sharpe for his finance course at Stanford (http://www.stanford.edu/~wfsharpe/
mia/mat/mia_mat3.htm). We thank Professor Sharpe for his helpful comments and
suggestions. Both GAUSS and MATLAB are matrix programming languages, and they are
syntactically similar.
7
line by line at the “>>” prompt in the Command window. In the beginning, to
introduce the basic commands and statements of GAUSS, we shall stay in the
Command Input-Output window and use the Command or interactive mode. If the
Output window is open, close it.
Matrices as Fundamental Objects

GAUSS is one of a few languages in which each variable is a matrix (broadly
constructed), and the language knows what the contents are and how big it is.
Moreover, the fundamental operators (e.g. addition, multiplication) are programmed
to deal with matrices when required. The GAUSS environment handles much of the
bothersome housekeeping that makes all this possible. Since so many of the
procedures required for economic and econometric computing involve matrices,
GAUSS proves to be an extremely efficient language for implementation and
computation.
First of all, each line of GAUSS code must end with a semi-colon (;).
Consider the following GAUSS expression:

C = A + B;
If both A and B are scalars (1 by 1 matrices), C will be a scalar equal to their sum. If
A and B are row vectors of identical length, C will be a row vector of the same
length. Each element of C will be equal to the sum of the corresponding elements of
A and B. Finally, if A and B are, say, 3×4 matrices, C will also be a 3×4 matrix, with
each element equal to the sum of the corresponding elements of A and B.
In short the symbol “+” means “perform a matrix addition.” But what if A and B are
of incompatible sizes? Not surprisingly, GAUSS will complain with a statement such
as:
(0) : error G0036 : matrices are not conformable
So the symbol “+” means “perform a matrix addition if you can and let me know if
you can’t.” Similar rules and interpretation apply to matrix operations such as “-”
(subtraction) and “*” (multiplication).
Assignment Statements
GAUSS uses a pattern common in many programming languages for assigning the
value of an expression to a variable. The variable name is placed on the left of an
equal sign and the expression on the right. The expression is evaluated and the result
assigned to the variable. In GAUSS, there is no need to declare a variable before
assigning a value to it. If a variable has previously been assigned a value, a number,
or a string, the new value overrides the predecessor. Thus if A and B are of size
20×30, the statement:
C = A + B;
creates a variable named C that is also 20×30 and fills it with the appropriate values
obtained by adding the corresponding elements in A and B. If C already existed and
was, say, 20×15 it would be replaced with the new 20×30 matrix. Therefore, matrix
8
GAUSS BASICS
variables in GAUSS are not fixed in size. In GAUSS, unlike some languages, there is
no need to pre-dimension or re-dimension variables. It all happens without any
explicit action on the part of the user.
Variable Names
The GAUSS environment is case insensitive. Typing variable names in uppercase,
lowercase, or a combination of both does not matter. That is, GAUSS does not
distinguish between uppercase and lowercase except inside double quotes. A variable
name can have up to 32 characters, including letters, numbers and underscores. The
first character must be alphabetic or an underscore. Therefore the variable
PersonalDisposableIncome is the same as personaldisposableincome.
While it is tempting to use long names for easy reading, small typing errors can mess
up your programs. If you do mistype a variable name, you may get lucky (e.g. the
system will complain that you have asked for the value of an undefined variable) or
you may not (e.g. you will assign the new value to a newly created variable instead
of the old one desired). In programming languages there are always tradeoffs. You
don’t have to declare variables in advance in GAUSS. This avoids a great deal of
effort, but it allows for the possibility that nasty and difficult-to-detect errors may
creep into your programs.
Showing Values
If at any time you wish to see the contents of a variable, just type its name. GAUSS
will do its best, although the result may extend beyond the Command or Output
window if the variable is a large matrix (remember that you can always resize the
window). If the variable, say x, is not defined or has not previously been given a
value, a message such as:
Undefined symbols:
x (0)
will appear.
GAUSS will not show you the result of an assignment statement unless you
specifically request for it. Thus if you type:
C = A + B;
No values will be shown although C is now assigned with values of the sum of A and
B. But, if you type:
C;
or, equivalently (though verbosely):

print C;
GAUSS will show you the value of C. It may be a bit daunting if C is, say, a 20 by 30
matrix. If the variable C is not of interest, and what you want to see is the result of A
plus B, simply type:
A + B;
9
That is, if an expression has no assignment operator (=), it will be assumed to be an

implicit print statement. Note that the value shown will be represented in
accordance with the format specified. If there is no explicit format used, by default
GAUSS will show the numeric value in 16 fields with 8 digits of precision.
Initializing Matrices
If a matrix is small enough, one can provide initial values by simply typing them in
the Command window. For example:
a = 3;
b = {1 2 3};
c = {4, 5, 6};
d = {1 2 3, 4 5 6};
Here, a is a scalar, b is a 1×3 row vector, c a 3×1 column vector, and d is a 2×3
matrix. Thus, typing
d;
produces:
1.0000000 2.0000000 3.0000000

4.0000000 5.0000000 6.0000000
The system for indicating matrix contents is very simple. Values separated by spaces
belong on the same row; those separated by commas are on separate rows. All values
are enclosed in brace brackets.
The alternative to creating a matrix using constants is to use the GAUSS built-in
command let. If dimensions are given, a matrix of that size is created. The
following statement creates a 2×3 matrix:
let d[2,3] = 1 2 3 4 5 6;
Note that dimensions of d are enclosed in square brackets, not curly brace
brackets. If dimensions are not given, a column vector is created:
let d = 1 2 3 4 5 6;
If curly braces are used, the let is optional. That is, the following two
expressions will create the same matrix d:
let d = {1 2 3, 4 5 6};
d = {1 2 3, 4 5 6};
Making Matrices from Matrices

The general scheme for initializing matrices can be extended to combine or
concatenate matrices. For example,
a = {1 2};
b = {3 4};
c = a~b;
print c;
gives a row vector:
10
GAUSS BASICS
1.0000000 2.0000000 3.0000000 4.0000000
While
a = {1 2 3};
b = {4 5 6};
d = a|b;
print d;
gives a 2×3 matrix:
1.0000000 2.0000000 3.0000000

4.0000000 5.0000000 6.0000000
Matrices can easily be pasted together in this manner, a process that is both simple
and easily understood by anyone reading a procedure. Of course, the sizes of the
matrices must be compatible. If they are not, GAUSS will tell you.
Note that by putting variables in brace brackets such as:

c = {a b};
or
d = {a,b};
will not work. It produces a syntax error message.
Using Portions of Matrices

Frequently one wishes to reference only a portion of a matrix. GAUSS provides
simple and powerful ways to do so. To reference a part of a matrix, give the matrix
name followed by square brackets with expressions indicating the portion desired.
The simplest case arises when only one element is wanted. For example, using
matrix d in the previous section,
d[1,2];
equals:
2.0000000
While
d[2,1];
equals:
4.0000000
In every case the first bracketed expression indicates the desired row (or rows), while
the second expression indicates the desired column (or columns). If a matrix is a
vector, a single expression may be given to indicate the desired element, but it is
often wise to give both row and column information explicitly.
11
The real power of GAUSS comes into play when more than a single element of a
matrix is wanted. To indicate “all the rows” use a dot for the first expression. To
indicate “all the columns,” use a dot for the second expression. Thus,
d[1,.];
equals:
1.0000000 2.0000000 3.0000000
That is, d[1,.] yields a matrix containing the entire first row of d. While,
d[.,2];
equals:
2.0000000
5.0000000
That is, d[.,2] yields a matrix containing the entire second column of d. In fact,
you may use any expression in this manner as long as it includes a valid row or
column numbers. For example,
d[2,2:3];
equals:
5.0000000 6.0000000
And
d[2,3:2];
equals:
6.0000000 5.0000000
Variables may also be used as subscripts. Thus,

z = {2,3};
d[2,z];
equals:
5.0000000 6.0000000
Another useful example is

d[1:2, 2:3];
which equals:
2.0000000 3.0000000
5.0000000 6.0000000
This is the same as

d[.,2:3];
Try the following:
12
GAUSS BASICS
d[.,1 3];
Recall that “.” is a wildcard symbol and may be used when indexing a matrix, rows,
or columns, to mean “any and all.”
Text Strings
GAUSS is wonderful with numbers. It deals with text too, but one can tell that its
heart isn’t in it.
A variable in GAUSS is one of two types: numeric or string. A string is like any
other variable, except the elements in it are interpreted as ASCII numbers. Thus the
number 32 represents a space, and the number 65 a capital A, etc. To create a string
variable, enclose a string of characters in double quotation marks. Thus:
stg = "This is a string";
The variable named stg is assigned a string of characters: “This is a string.” Since a
string variable is in fact a row vector of numbers, it is possible to create a list of
strings by creating a matrix in which each row or column is a separate string. As with
all standard matrices, each element of a string matrix can only have up to 8
characters long, which is exactly the 32-bit size number can hold. To print a string
matrix, the variable must be prefixed with a dollar sign ($). Thus the statement
x = {"ab", "cd"};
print $x;
produces:
ab
cd
While
x = {"ab" "cd"};
print $x;
produces:
ab cd
as always.
To see the importance of including the dollar sign in front of a variable, type:
print x;
and see what GAUSS gives you.
Matrix and Array Operations

The term “matrix operation” is used to refer to standard procedures such as matrix
multiplication, while the term “array operation” is reserved for element-by-element
computations.
13
Matrix Operations
Matrix transposition is as easy as adding a prime (apostrophe) to the name of the
matrix. Thus
x = {1 2 3};
print x';
produces:
1.0000000
2.0000000
3.0000000
To add two matrices of the same size, use the plus (+) sign. To subtract one matrix
from another of the same size, use a minus (-) sign. If a matrix needs to be “turned
around” to conform, use its transpose. Thus, if A is 3×4 and B is 4×3, the statement
C = A + B;
results in the message:
(0) : error G0036 : matrices are not conformable
While
C = A + B';
will get you a new 3×4 matrix C.
In GAUSS, there are some cases in which addition or subtraction works when the
matrices are of different sizes. If one is a scalar, it is added to or subtracted from all
the elements in the other. If one is a row vector and its size matches with the number
of columns in the other matrix, this row vector is swept down to add or subtract the
corresponding row elements of the matrix. Similarly, if one is a column vector and
its size matches with the number of rows in the other matrix, this column vector is
swept across to add or subtract the corresponding column elements of the matrix.
For instance,
x = {1 2 3};
y = {1 2 3, 4 5 6, 7 8 9};
x + y;
produces
2.0000000 4.0000000 6.0000000

5.0000000 7.0000000 9.0000000
8.0000000 10.000000 12.000000
While,
x' + y;
produces
2.0000000 3.0000000 4.0000000

6.0000000 7.0000000 8.0000000
10.000000 11.000000 12.000000
14
GAUSS BASICS
These situations are what we call “array operation” or element-by-element

compatibility to be discussed below. GAUSS does not make syntactical distinction
between matrix addition (subtraction) and array addition (subtraction).
Matrix multiplication is indicated by an asterisk (*), commonly regarded in

programming languages as a “times sign.” The usual rules of matrix multiplication
from linear algebra apply: the inner dimensions of the two matrices being multiplied
must be the same. If they are not, you will be told so. The one allowed exception is
the case in which one of the matrices is a scalar and one is not. In this instance, every
element of the non-scalar matrix is multiplied by the scalar, resulting in a new matrix
of the same size as the non-scalar matrix.
GAUSS provides two notations for matrix division which provide rapid solutions to
simultaneous equation or linear regression problems. They are better discussed in the
context of such problems later.
Array Operations
To indicate an array (element-by-element) multiplication, precede a standard
operator with a period (dot). Thus,
x = {1 2 3};
y = {4 5 6};
x.*y;
produces:
4.0000000 10.000000 18.000000
which is the “dot product” of two row vectors x and y.
You may divide all the elements in one matrix by the corresponding elements in
another, producing a matrix of the same size, as in:
C = A ./ B;
In each case, one of the operands may be a scalar or the matrices must be element-
by-element compatible. This proves handy when you wish to raise all the elements in
a matrix to a power. For example:
x = {1 2 3};
x.^2;
produces
1.0000000 4.0000000 9.0000000
GAUSS array operations include multiplication (.*), division (./) and

exponentiation (.^). Since the operation of exponentiation is obviously element-by-
element, the notation “.^” is the same as “^”. Array addition and subtraction are
discussed earlier using the same matrix operators “+” and “-”.
Logical and Relational Operations on Matrices

GAUSS offers six relational operators:
15
• LT or < Less than
• LE or <= Less than or equal to
• GT or > Greater than
• GE or >= Greater than or equal to
• EQ or == Equal
• NE or /= Not equal
Note carefully the difference between the double equality and the single equality.
Thus A==B should be read “A is equal to B,” while A=B should be read “A is
assigned the value of B.” The former is a logical relation, the latter an assignment
statement. For comparisons between character data and comparisons between strings,
these operators should be preceded by a dollar sign ($).
Whenever GAUSS encounters a relational operator, it produces a one (1) if the

expression is true and a zero (0) if the expression is false. Thus the statement,
x = 1 < 3;
print x;
produces:
1.0000000
While
x = 1 > 3;
print x;
produces:
0.0000000
Relational operators can be used on element-by-element compatible matrices. For

element-by-element comparisons of two matrices, the relational operator is preceded
by a dot (.). If the relational operator is not preceded by a dot (.), then the result is
always a scalar 1 (true) or 0 (false), based upon a comparison of all elements in the
two matrices. If the relational operator is preceded by a dot (.), then the operation is
performed element-by-element, resulting a matrix with ones in positions for which
the relation is true and zeros in positions for which the relation is false. Thus, for
example:
A = {1 2, 3 4};
A > 2;
produces:
0.0000000
This is because there is at least one element of A that is not greater than 2. While
16
GAUSS BASICS
A .> 2;
produces:
0.0000000 0.0000000
1.0000000 1.0000000
Similarly,
A = {1 2, 3 4};
B = {3 1, 2 2};
A > B;
produces:
0.0000000
While
A .> B;
produces:
0.0000000 1.0000000
1.0000000 1.0000000
You may also use logical operators of which we will only mention the frequently
used ones in passing:
• not
• and
• or
If the logical operator is preceded by a dot (.), the result will be a matrix of 1’s and
0’s based on an element-by-element logical comparison of two matrices. Each
operator works with matrices on an element-by-element basis and conforms to the
ordinary rules of logic, treating any non-zero element as true and a zero element as
false.
Relational and logical operators are used frequently with if statements (described
below) and scalar variables, as in more mundane programming languages. But the
ability to use them with matrices offers major advantages in statistical and
econometric applications.
Creating and Editing a GAUSS Program

So far, we have seen the working of GAUSS in the Command mode. That is, at the
“>>” prompt in the Command Input-Output window, you enter a statement and press
the carriage return (the Enter key) and the statement is immediately executed. Multi-
line commands can be entered by pressing CTRL Enter at the end of each line.
Then at the end of the final line in a multi-line command, press Enter. GAUSS
remembers all the variable names and their assigned values. Upon the execution of a
statement, the available result is displayed in the Command window (or in the
Output window if it is open). Given the power that can be packed into one GAUSS
17
statement, this is no small accomplishment. However, for many purposes it is

desirable to store a set of GAUSS statements for use when needed. The simplest
form of this approach is the creation and modification of a program file: a set of
commands in a file. You need to get to the Edit mode to create and edit the file. Once
such a file exists in the current directory, you can simply load and run the program
file. The statements stored in the file will then be executed, with the results
displayed.
GAUSS for Windows provides a consistent and convenient window interface for
program development. From the menu bar File/New (or by clicking on the blank
page icon from the toolbar), you can open a blank Edit window to create a file from
scratch. If the file exists, from the menu bar File/Open (or by clicking on the open
folder icon from the toolbar), then select the name of the file to load its contents into
the Edit window. You can also open a file in the Edit window by typing the file
name in the Command window, including the directory in which the file is stored.
This Edit window will then “pop up” and layer over the Command window. Note
that the title of the Edit window is the name of the file you open for editing. After
editing, selecting Run Active File from the Run menu button saves and runs the
program file, with outputs shown in the Command or Output window. If you are not
running the program file after editing, do not forget to save it.
A group of program and data files may be involved in a project. They can be created,
loaded, and edited each in their separate Edit windows. GAUSS keeps track of two
types of files: an active file and a main file. The active file is the file that is currently
displayed (in the front highlighted Edit windows). The main file is the file that is
executed to run the current job or project. An active program file can be executed,
and put in the main file list (that is, in the pull-down menu on the toolbar). The main
file list contains the program files you have been running (the results of which appear
in the Command window or in the Output window). Any files on the main file list
can be selected, edited, and executed repeatedly. The list of main files may be
retained or cleared anytime as you wish.
Many Edit/Run cycles are involved in the writing and testing of a GAUSS program.
The convention adopted in this book is that all example lessons (with only a few
exceptions such as the first one below) will be set up to have two files. The first
(program) file contains the GAUSS code, and the second (output) file will contain all
output from running the program in the first file. You will see not only the results in
the Command or Output window, but also the output is stored in a file you specified.
The benefit of using Edit mode is the ability to have a record of each line of code.
This is especially helpful when troubleshooting a long or complicated program.
Lesson 2.1 Let’s Begin

To get into the Edit mode, from the menu bar, select File/Open. Find and select the
file named lesson2.1 in the GPE subdirectory.
Alternatively, a file can be opened from the Command window by typing the file
name at the “>>” prompt:
edit gpe\lesson2.1;
Press Enter key to load gpe\lesson2.1 into the Edit window.
18
GAUSS BASICS
You are now ready for program editing. The full path of file name
c:\gauss\gpe\lesson2.1 (or something like that depending on your GAUSS
installation) shows up as the title of the Edit window. lesson2.1 is just the name of
the program file for the following exercise. GAUSS will create a file named
lesson2.1 in the c:\gauss\gpe directory if it does not already exist. If a file named
lesson2.1 does exist, GAUSS will simply bring the file to the Edit window. When
working on your own project, you should use the name of your file.
The purpose of this lesson is to demonstrate some basic matrix and array operations
in the GAUSS language we have learned so far and to familiarize you with the
Edit/Run dual mode operation of GAUSS. If you are typing the following lesson for
practice, do not type the line number in front of each line of code. The numbering
system is for reference and discussion only.
/*
** Lesson 2.1: Let’s Begin
*/
1 A = {1 2 3,
0 1 4,
0 0 1};
2 C = {2,7,1};
3 print"Matrix A" A;
4 print;
5 print "Matrix C" c;
6 print "A*C" a*c;
7 print "A.*C" a.*c;
8 print "A.*C'" a.*c';
9 end;
From the menu bar, click on the Run button and select Run Active File. This will
save and run the program. The name of the program file lesson2.1 appears in the
main file list located on the toolbar as a pull-down menu item. As of now, lesson2.1
is the active file. You can run, edit, compile, and debug the main file all by clicking
on the four buttons next to the main file list.
Each line of code must end with a semi-colon (;). In line 1, we have typed in the
numbers in matrix form to be easier to read. Spaces separate columns while commas
separate rows. Carriage return is not seen by GAUSS. That is,
A = {1 2 3, 0 1 4, 0 0 1};
is read by GAUSS in the same way as

A = {1 2 3,
0 1 4,
0 0 1};
The GAUSS command, print, is used to print output to the screen. You may have
wondered about the extra print statement in line 4. This creates an empty line
between matrix A and matrix C, making the output easier to read. The rest of
lesson2.1 demonstrates the difference between matrix multiplication (*) and
element-by-element array multiplication (.*) with matrices. In addition, the use of
matrix transpose notation (') is demonstrated.
After running lesson2.1, the following output should be displayed:
Matrix A
19
1.00000000 2.0000000 3.0000000

0.00000000 1.0000000 4.0000000
0.00000000 0.0000000 1.0000000
Matrix C
2.0000000
7.0000000
1.0000000
A*C
19.000000
11.000000
1.0000000
A.*C
2.0000000 4.0000000 6.0000000
0.00000000 7.0000000 28.000000
0.00000000 0.00000000 1.0000000
A.*C'
2.0000000 14.0000000 3.0000000
0.00000000 7.0000000 4.0000000
0.00000000 0.00000000 1.0000000
Notice that matrix multiplication requires that the number of columns in the first
matrix equals the number of rows in the second matrix. Element-by-element array
multiplication requires that both matrices have the same number of rows or columns.
It “sweeps across” each row, multiplying every element of matrix A by the
corresponding element in matrix C (line 7). Element-by-element array multiplication
is “swept down” each column if C is transposed first (C') into a horizontal row
vector as shown in line 8.
Programming Tips
Just a few comments on programming in general. Professional programmers judge
their work by two criteria: Does it do what it is supposed to? Is it efficient? We
would like to add a third criterion: Will you be able to understand what your program
is supposed to be doing six months from now? Adding a blank line between sections
in your program will not affect how it runs, but it will make reading your program
easier. Describing the function of each section within comment symbols will benefit
you not only in troubleshooting now, but also in understanding your program in the
future. To do so in GAUSS, put the comment statement between a pair of “at” (@)
signs or in between “/*” and “*/” symbols. Notice that the “*” is always adjacent
to the comment text. Everything between the sets of “@” signs or between “/*” and
“*/” symbols will be ignored by GAUSS. Comments can extend more than one line
as desired. The difference between these two kinds of comments is shown in the
following:
/* This kind of
/* comment */
can be nested */
@ This kind of comment cannot be nested @
Another important programming style observed throughout this book is that we will
keep each program small. Break down your problem into smaller tasks, and write
sub-programs for each task in separate blocks of a larger program or in separate
programs. Avoid long lines of coding. Write clear and readable code. Use indention
where applicable. Remember that programming is very fluid, and there are always
multiple routes to achieve any desired task.
20
GAUSS BASICS
File I/O and Data Transformation

File input and output operations (I/O) and data transformation in a GAUSS
programming environment are important prerequisites for econometric modeling and
statistical analysis. The file I/O chapter of GAUSS for Windows User Guide (see also
the on-line help from menu bar Help/Contents) describes various types of file
formats available in the GAUSS programming environment.
Most useful programs need to communicate and interact with peripheral devices such
as a file storage device, console display, printer, etc. A typical GAUSS program will
read input data from the keyboard or a file, perform the computation, show results on
the screen, and send outputs to a printer or store in a file.
GAUSS can handle at least three kinds of data formats: GAUSS data sets, GAUSS
matrix files, and text (or ASCII) files. The first two data file formats are unique and
efficient in GAUSS. For file transfer (import and export) between GAUSS and other
application software or across platforms, the text file format is preferred. Although
we do not limit the use of any particular file format, we focus here on text-formatted
file input and output. For the use of data set and matrix files, see GAUSS Language
References or on-line help from menu bar Help/References for more information.
Data Input
The most straightforward way to get information into GAUSS is to type it in the
Command window as we have been doing in the first part of this chapter. This
approach is useful for a small amount of data input. For example:
prices = {12.50 37.875 12.25};
assets = {"cash", "bonds", "stocks"};
holdings = {100 200,
300 400,
500 600};
For long series of data, it is recommended that your create a text file for the data
series using the GAUSS editor. That is, create the file and type the data in the Edit
window. Such a file should have numeric ASCII text characters, with each element
in a row separated from its neighbor with a space and each row on a separate line.
Now, we will introduce a text data file named longley.txt which comes with the GPE
package. If you installed GAUSS and GPE correctly, this data file should be located
in the GPE subdirectory of the GAUSS directory. The easiest way to bring it into the
Edit window is to click on the menu bar button File/Open and select the file name
longley.txt located in the GPE directory.
The alternative is typing the following in the Command window at the “>>”
prompt:
edit gpe\longley.txt;
and pressing Enter key.
The data matrix is arranged in seventeen rows and seven columns, and there are no
missing values. The first row contains only variable names, so it must not be
included in statistical operations. Each variable name is short, no longer than four
characters in this case. All values, except the first two columns (YEAR and PGNP),
21
are too large to handle easily. Scaling these variables may make interpreting the
resulting data easier. The bottom of the file contains the data source and description
for reference purposes. Of course, the descriptive information should not be included
for statistical analysis.
To load the data into GAUSS, the following statement:

load data[17,7] = gpe\longley.txt;
will create a matrix named data containing the data matrix.
Alternatively, it can be re-coded as the following two lines using GAUSS command
reshape to form the desired 17×7 matrix:
load data[]=gpe\longley.txt;
data=reshape(data,17,7);
Notice that the size of a matrix created with load must be equal to the size of the
file being loaded (not counting optional reference information at the bottom of the
file). If the matrix is larger than the actual file size, bogus data will be read. If it is
smaller, part of the data series will be discarded. In either case, computations will be
inaccurate.
Data Output
A simple way to output data is to display a matrix. This can be accomplished by
either giving its name in interactive mode or using the print function as we have
shown so far.
print data;
You can use the format statement to control the formats of matrices and numbers
printed out. For prettier output, the GAUSS function printfm can print a matrix
using different format for each column of the matrix.
If you want to save essentially everything that appears on your screen (i.e. the output
from your GAUSS program), issue the following command:
output file = [filename] [option];
where [filename] represents the name of a new file that will receive the
subsequent output. When using the command output file you must designate
one of three options in [option]: Reset, On, or Off. The option Reset clears
all the file contents so that each run of the program stores fresh output; On is
cumulative, each output is appended to the previous one; Off creates an output file,
but no data are directed to it. An output file is not created if none of these three
options is used. When you are through directing output, don’t forget to issue the
command:
output off;
You may want to examine the output files. To create a text file containing the data
from a matrix use output and print statements in combination. For example:
output file = gpe\output2.1 reset;
22
GAUSS BASICS
print data;
output off;
will save the matrix named data in the file named output2.1 in the directory
GPE.
Sending output to a printer is as easy as sending output to a file:
output file = lpt1 reset;
If you are in the Edit mode to write a program file, it is a good habit to end your
program with the statement:
end;
This will automatically perform output off and graciously close all the files still
open.
Lesson 2.2: File I/O

In Lesson 2.2, we will demonstrate how to direct program output to a file and how to
input data from a text file. In addition we will slice a matrix into column vectors,
which can be useful for working on individual variables. Columns (or rows) can be
joined as well. This is achieved through horizontal (or vertical) concatenation of
matrices or vectors.
Click on the menu bar button File/Open and select the file name lesson2.2 located in
the GPE directory.
Alternatively, at the “>>” prompt in the Command window, type:

edit gpe\lesson2.2;
and press Enter.
Make sure that the highlighted Edit window, with the title c:\gauss\gpe\lesson2.2, is
layered over the Command window and stays in the front. To run it, click on menu
bar button Run/Run Active File.
/*
** Lesson 2.2: File I/O
*/
1 output file = gpe\output2.2 reset;
2 load data[17,7] = gpe\longley.txt;
3 data = data[2:17,.];
4 PGNP = data[.,2];
5 GNP = data[.,3]/1000;
6 POPU = data[.,6]/1000;
7 EM = data[.,7]/1000;
8 X = PGNP~GNP~POPU~EM;
9 print X;
10 end;
For those of you who are using a working diskette (a:\ is assumed) and want to type
in the program, type these lines exactly as written. Misspellings, missing
semicolons, or improper spaces will all result in error messages. Be warned that each
23
type of bracket, { }, [ ], or( ), is interpreted differently. Errors commonly result

from using the wrong bracket.
The first line of the program code tells GAUSS to direct the output of this program
to a file named output2.2 located in the GPE subdirectory. If you want a printed
copy of your work, just change it to:
output file = lpt1 reset;
Let’s examine the code for data loading. In line 2, a matrix named data, containing
17 rows and 7 columns, is created using the GAUSS command load. A text file
located in the GPE subdirectory named longley.txt is then loaded into the variable
data.
Remember that the first row of data contains variable names. Chopping off the first
row, or indexing the matrix, is one way to remove these names from statistical
analysis. In line 3, the new data takes from the old data the second row through
the seventeenth row. After line 3, the matrix named data contains 16 rows and 7
columns. Now try to make some sense about what line 4 is doing. It assigns PGNP to
the second column of the modified data matrix. Notice that when a matrix is being
created, the brackets are to the left of the equal sign. When a matrix is indexed, the
brackets are to the right of the equal sign. In general, information is taken from the
right side of an equal sign and assigned to either a matrix or variable on the left side
of the equal sign.
The next few lines, 4 through 7, create new variables by picking the corresponding
columns of data. For easier handling of large numbers, quantity variables are
scaled down by 1000-fold: GNP is now in billions of 1954 dollars; POPU and EM are
in millions of persons. PGNP is kept as given. Note that only the variables needed for
study are named and identified.
We now have four variables (vectors) that have been scaled down to a workable size.
Statistical operations can be done on each variable separately, or they can be joined
together and then operated on with one command. Line 8 concatenates all of the four
variables horizontally with a “~” symbol, forming a new data matrix named X.
Line 9 prints the matrix X as follows:
83.000000 234.28900 107.60800 60.323000

88.500000 259.42600 108.63200 61.122000
88.200000 258.05400 109.77300 60.171000
89.500000 284.59900 110.92900 61.187000
96.200000 328.97500 112.07500 63.221000
98.100000 346.99900 113.27000 63.639000
99.000000 365.38500 115.09400 64.989000
100.00000 363.11200 116.21900 63.761000
101.20000 397.46900 117.38800 66.019000
104.60000 419.18000 118.73400 67.857000
108.40000 442.76900 120.44500 68.169000
110.80000 444.54600 121.95000 66.513000
112.60000 482.70400 123.36600 68.655000
114.20000 502.60100 125.36800 69.564000
115.70000 518.17300 127.85200 69.331000
116.90000 554.89400 130.08100 70.551000
24
GAUSS BASICS
If your output extends beyond your screen (in the Command or Output window),
you can resize the window for a better view. You can also try another font such as
New Courier, size 10, from the Configure button on the menu bar.
Lesson 2.3: Data Transformation

In Lesson 2.2 above, we have seen the utility of scaling data series to a more
manageable unit of measurement for analysis. For econometric applications, some
variables may be transformed for considerations of theoretical and empirical
interpretation. Exponential, logarithmic, and reciprocal transformations are
frequently used functional forms in econometrics. The data transformation chapter of
GAUSS for Windows User Guide emphasizes the use of GAUSS internal data sets.
Interested readers should refer to this chapter for more details.
Lesson 2.3 below demonstrates the use of logarithmic functional transformation as a

way to scale the size of each data series.
For those of you who are using a working diskette (a:\ is assumed), the first two
blocks of code in lesson2.2 can be used again. Duplicating and renaming lesson2.2
to lesson2.3 and then editing it will save typing and time. To do that, just start with
lesson2.2 in the Edit window and click on File/Save As. Since your working
diskette is in a:\, make sure that in the “Select File to Save…” dialog window the
“Save In:” line shows: “3 ½ Floppy (A):”. Type a:\lesson2.3 in the “File Name” line
and click on “Save.”
Here is the program lesson2.3:
/*
** Lesson 2.3: Data Transformation
*/
3 data = data[2:17,.];
4 PGNP = ln(data[.,2]);
5 GNP = ln(data[.,3]/1000);
6 POPU = ln(data[.,6]/1000);
7 EM = ln(data[.,7]/1000);
8 X = PGNP~GNP~POPU~EM;
9 print X;
10 end;
Running lesson2.3, the printout of matrix X looks like this:
4.4188406 5.4565554 4.6784950 4.0997135

4.4830026 5.5584715 4.6879660 4.1128719
4.4796070 5.5531689 4.6984146 4.0971905
4.4942386 5.6510812 4.7088904 4.1139347
4.5664294 5.7959818 4.7191683 4.1466365
4.5859874 5.8493219 4.7297743 4.1532265
4.5951199 5.9009516 4.7457492 4.1742180
4.6051702 5.8947113 4.7554763 4.1551417
4.6170988 5.9851169 4.7654847 4.1899426
4.6501436 6.0383004 4.7768857 4.2174025
4.6858281 6.0930482 4.7911932 4.2219899
4.7077268 6.0970535 4.8036111 4.1973974
4.7238417 6.1794036 4.8151555 4.2290940
4.7379513 6.2197966 4.8312534 4.2422472
25
4.7510006 6.2503092 4.8508733 4.2388921

4.7613189 6.3187771 4.8681573 4.2563359
Lines 4 through 7 introduce the logarithmic transformation of each variable in

addition to simple scaling. Line 8 concatenates all these variables into a matrix
named X. This is simply the logarithmic transformation of the data matrix presented
in the previous Lesson 2.2. In GAUSS, ln computes a natural log transformation of
a data matrix while log is a base 10 log transformation. We suggest the use of
natural log transformation for data scaling if needed.
GAUSS Built-In Functions

GAUSS has a large number of built-in functions or procedures—many of which are
very powerful. Without knowing it, you have been using some of them such as let,
print, load, and output. Most functions take some input arguments and return
some outputs. Before the outputs of a function are used, they must be retrieved. To
get all the outputs from a function, use a multiple assignment statement in which the
variables that are to receive the outputs are listed to the left of the equal sign,
separated by commas, and enclosed in brace brackets. The name of the function is on
the right of the equal sign, which takes input arguments separated by commas and
enclosed in round brackets. Typically, a function is called (initiated) in one of the
following two ways:
output1=functionName(input1);
{output1,output2,...}=functionName(input1,input2,...);
In case the function outputs are not of interest, the command call is used to call the
requested function or procedure without using any returned values. The syntax is,
call functionName(input1,input2,...);
Data Generating Functions

The following functions are particularly useful for creating a new matrix. Their
usage is explained by example:
• ones Creates a ones matrix, here a 2x4 matrix:

ones(2,4);
• zeros Creates a zeros matrix, here a 4x4 matrix:

zeros(4,4);
• eye Creates an identity matrix, here a 3x3 matrix:

eye(3);
• rndu Creates a matrix of uniform random numbers, here a 6x3

matrix:
rndu(6,3);
• rndn Creates a matrix of normal random numbers, here a 6x3

matrix:
rndn(6,3);
• seqa Creates a vector of additive sequence of numbers starting at a

given value and increasing with a given increment. For
instance,
26
GAUSS BASICS
seqa(0,0.1,10);
creates a 10x1 vector beginning at 0 and increasing with a 0.1
increment (i.e. 0, 0.1, … , 0.9).
• seqm Creates a vector of multiplicative sequence of numbers starting

at a given value and increasing by a given multiplier. For
instance,
seqm(2,2,10);
creates a 10x1 vector beginning at 2 and increasing with a
multiplier of 2 (i.e. 2, 4, … , 1032 or 210).
To convert or reshape an existing matrix to a new matrix of different size, use the
reshape function as in the following example: …
x=seqa(1,1,5);
print x;
y=reshape(x,5,5);
print y;
Creation of a sub-matrix based on some selection or deletion criteria is accomplished

by selif and delif functions, respectively. For example:
x=rndn(100,4);
y=selif(x, x[.,1] .> 0.5);
print y;
Equivalently,
y=delif(x, x[.,1] .<= 0.5);
print y;
There are other useful functions for vector or matrix conversion:
• vec Stacks columns of a matrix into a column vector.
• vech Stacks only the lower triangular portion of matrix into a

column vector.
• xpnd Expands a column vector into a symmetric matrix.
• submat Extracts a sub-matrix from a matrix.
• diag Retrieves the diagonal elements of a matrix.
• diagrv Replaces diagonal elements of a matrix.
Matrix Description Functions

To describe a matrix, such as a matrix x defined as
x=rndu(10,4);
the following functions can be used:
• rows Returns the number of rows of a matrix:

rows(x);
27
• cols Returns the number of columns of a matrix:

cols(x);
• maxc Returns the maximum elements of each column of a matrix:

maxc(x);
• minc Returns the minimum elements of each column of a matrix:

minc(x);
To find the maximum and minimum of a matrix, try these:

maxc(maxc(x));
minc(minc(x));
There are many other GAUSS functions like maxc and minc which work on the
columns of a matrix. For example:
x = {1, 2, 3};
y = sumc(x) + 10;
print y;
Since sumc computes the sum of each column of a matrix, this will produce:
16.000000
If a matrix is given as an argument to sumc function, the summation is applied

separately to each column, and a column vector of results is returned. Thus, typing
x = {1 2 3, 4 5 6};
sumc(x);
will result in:
5.0000000
7.0000000
9.0000000
To compute the cumulative sum of elements in each column of a matrix, use the
function cumsumc as follows:
cumsumc(x);
Similar to sumc and cumsumc, there are:
• prodc Computes the product of all elements in each column of a

matrix.
• cumprodc Computes the cumulative product of elements in each column

of a matrix.
We further list a few descriptive statistics functions which are applied to each
column of a matrix:
• meanc Computes the mean for each column of a matrix.
• median Computes the median for each column of a matrix.
28
GAUSS BASICS
• stdc Computes the standard error for each column of a matrix.
Matrix Sorting Functions

To sort a matrix in ascending order, use one of the following sort functions:
• sortc Sorts a matrix using quick sort algorithm.
• sorthc Sorts a matrix using heap sort algorithm.
These functions will sort the rows of a matrix with respect to a specified column.
That is, they will sort the elements of a column and will arrange all rows of the
matrix in the same order as the sorted column. The sort is in ascending order.
Another useful sort function, sortind, returns the sorted index of a column vector.
This can be used to sort several matrices in the same way that some other reference
matrix is sorted. For example,
x = {5, 2, 8};
idx = sortind(x);
y = x[idx];
print idx~y;
produces two columns containing the ordering index of the original x and the sorted
x:
2.0000000 2.0000000
1.0000000 5.0000000
3.0000000 8.0000000
Basic Matrix Computation

The following functions are related to several basic matrix computations:
• det Computes the determinant of a square matrix.
• inv Computes the inverse of a general square matrix.
• invpd Computes the inverse of a symmetric, positive definite square

matrix.
• corrx Computes a correlation matrix.
• vcx Computes a variance-covariance matrix.
• cond Computes the condition number of a matrix.
• rank Computes the rank of a matrix.
Solving Systems of Linear Equations

The problem is to find x from A*x = b, where A is a nxn matrix, b is a nx1 (or nxm
in general) matrix, and x has the same size as b. If A is a non-singular matrix, then x
= A-1b. Consider the following example:
29
a = { 6 8,
-2 4};
b = {2, 1};
x = inv(a)*b;
print x;
If the matrix A is symmetric positive definite, use the GAUSS function solpd to
solve for x. Note that the solpd function takes two arguments. The first is the
matrix on the right-hand side of the matrix equation, while the second is the matrix
being inverted. For example:
a = {40 40,
40 72}; @ a is symmetric positive definite @
b = {2, 1};
x = solpd(b,a);
print x;
Therefore, if the matrix A is nxk and n is equal or greater than k, solving x from A*x
= b is equivalent to solving x from (A'A)*x = (A'*b). In other words, x =
invpd(A'A)*(A'b). Using the solpd function:
a = { 6 8,
-2 4};
b = (2, 1};
x = solpd(a'b,a'a);
print x;
This is exactly the GAUSS division (/) operator for finding the least squares (LS)
solution of A*x = b:
x = b/a;
print x;
Another useful application of the solpd function is to find the inverse of a

symmetric positive definite matrix, which is equivalent to the invpd command:
x = solpd(eye(2),a'a); @ equivalent to invpd(a'a) @
print x;
Characteristic Roots and Vectors

Given a square matrix A, finding x from the characteristic equation (A-λI)x = 0 is a
two-step procedure: First, solve |A-λI| = 0 for λ (characteristic roots or eigenvalues).
Then, for each λ, solve (A-λI)x = 0 for x (characteristic vectors or eigenvectors).
Since we are interested in the case of real eigenvalues and eigenvectors, A is
assumed to be a real symmetric matrix only. Two GAUSS functions are used for the
computation of eigenvalues and eigenvectors of a real symmetric matrix:
• eigrs Computes the eigenvalues of a real symmetric matrix.
• eigrs2 Computes the eigenvalues and eigenvectors of a real

symmetric matrix.
Consider the following example:

a = {41 -23,
-23 13}; @ a is real and symmetric @
r = eigrs(a);
30
GAUSS BASICS
{r,v} = eigrs2(a);
print r~v;
We note that the function eigrs2 returns two values: the first is a vector of
eigenvalues, while the second is a matrix of the corresponding eigenvectors. The
returned results are listed to the left of the equal sign, enclosed in brace brackets.
Running the above block of codes, we have:
0.074175964 -0.48992502 -0.87176457

53.925824 -0.87176457 0.48992502
The first column of the matrix is the vector of two eigenvalues. The last two columns
of eigenvectors correspond to each of the two eigenvalues, respectively.
A useful application of eigenvalues is to compute the condition number of a data

matrix—an indicator of multicollinearity for a least squares regression model. Let X
be the data matrix of regressors. The condition number of X is computed as follows:
xx = x'x;
r = eigrs(xx);
cn = sqrt(maxc(r)./minc(r));
print cn;
The condition number, cn, is defined as the square root of the ratio of the largest
eigenvalue to the smallest. Compared with the GAUSS built-in function cond, the
identical result is:
print cond(x);
Not listed, but of great use, are the many functions that provide data plotting in two
or three dimensions, as well as a number of more specialized functions. To whet the
econometrician’s appetite, let’s name a few more in the following:
• pdfn Computes the standard normal probability density function

(pdf).
• cdfn Computes the complement of cumulative distribution function

(cdf) of standard normal distribution (i.e., the integral of
normal distribution in the lower tail).
• cdftc Computes the complement of cdf of t-distribution.
• cdffc Computes the complement of cdf of F-distribution.
• cdfchic Computes the complement of cdf of Chi-square distribution.
• gradp Computes the first derivative or gradient of a function.
• hessp Computes the second derivative or hessian of a function.
• intsimp Computes the integration of a function by Simpson’s method.
• dstat Computes descriptive statistics of a data matrix.
• ols Computes a typical least squares regression.
31
• eqsolve Solves a system of nonlinear equations.
• sqpsolve Solves the nonlinear programming problem using the

sequential quadratic programming method.
The full list of functions and information on each one can be obtained via GAUSS’s
on-line help system.
Lesson 2.4: Data Analysis
In Lesson 2.4 we write a GAUSS program to review what we have learned so far.
First we load the data matrix from the file longley.txt. Recall that the first row of this
data matrix consists of variable names, therefore it will not be used in statistical
calculations. We define y as the last (7th) column of the data matrix. In addition, we
select all values of the first 6 variables and add a column of ones (constant vector) to
form the matrix x.
First, we call the built-in function dstat to report the descriptive statistics of all
data series including y and x. Then the ordinary least squares (OLS) estimator of y
on x is computed. Finally, the data matrix x is checked for its condition number.
Here is the program:
/*
** Lesson 2.4: Data Analysis
*/
2 load x[17,7]=gpe\longley.txt;
3 y=x[2:17,7];
4 x=x[2:17,1:6]~ones(16,1);
5 call dstat(0,y~x);
6 b=y/x; @ b=invpd(x'x)*x'y=solvpd(x'x,x'y) @
7 print b;
8 xx=x'*x;
9 r=eigrs(xx);
10 cn=sqrt(maxc(r)./minc(r));
11 print cn cond(x);
12 end;
Note that in line 5, dstat is a GAUSS built-in procedure which when called prints
descriptive statistics of a data matrix into a table. The output is arranged row-wise
for each variable. In dstat, the first input argument 0 means that the data to be
used involve a matrix defined earlier in the program. In this case, it is the matrix y~x
defined in line 3 for y and line 4 for x.
Line 6 demonstrates a simple way to obtain the least squares estimator: b = y/x.
To compute the condition number of x, we first get the eigenvalues of x'x (line 9)
and then take the square root of the ratio of maximum and minimum eigenvalues
(line 10). The result of the formal calculation of the condition number should be the
same as that from calling the GAUSS built-in function cond. We leave the rest of
running the program and interpreting the results to you as an exercise.
32
GAUSS BASICS
Controlling Execution Flow

It is possible to do a great deal in GAUSS by simply executing statements involving
matrix expressions, one after the other. However, there are cases in which one
simply must substitute some non-sequential order. To facilitate this, GAUSS
provides several standard methods for controlling program flow. These are For
Loops, Do Loops, and If statements.
For Loops
The For Loop is easy to use. The most common use of a For Loop arises when a set
of statements is to be repeated a fixed number of times, as in:
for i (0, 9, 1);
.......
endfor;
where i is the counter integer followed by a pair of parentheses which enclose three
arguments. The first argument is the initial value of the counter, the second is its
final value, and the last is the increment value. The statements within the loop will be
executed 10 times starting from the counter i at value 0 through 9, each time with
increment of 1. Note that a For Loop ends with endfor statement.
There are fancier ways to use For Loops, but for our purposes, the standard one
suffices.
Do Loops
There are two kinds of Do Loops: do while and do until. The difference
between a do while loop and a do until loop is that the former will continue
the loop execution when the condition is true, while the latter will execute the loop
when the condition is false (or until the condition becomes true). A Do Loop always
ends with endo statement.
A do while loop contains statements to be executed as long as a stated condition

remains true, as in:
do while x <= 0.5;
.......
endo;
Similarly, a do until loop contains statements to be executed as long as a stated

condition remains false, as in:
do until x > 0.5;
.......
endo;
The statements break and continue are used within Do Loops to control
execution flow. When break is encountered, the execution will jump to the
33
statement following the endo. This terminates the loop. When continue is
encountered, the execution will jump to the top of the loop and reevaluate the do
while or do until expression. It reiterates the loop without executing any more
of the statements inside the loop. For the For Loops, both break and continue
statements work the same way as described for the Do Loops.
In contrast to the For Loop, there is no counter variable that is automatically

incremented in a Do Loop. If one is used, it must be set to its initial value before the
loop is entered and explicitly incremented or decremented inside the loop.
It is, of course, crucial that at some point a statement will be executed that will cause
the condition in the do while (or do until) statement to be false (true). If this
is not the case, you have created an infinite loop—one that will go merrily on until
you pull the plug.
For readability, it is sometimes useful to create variables for true and false, then use
them in a do while or do until loop. For example:
true = 1==1;
false = 1==0;
.....
done = false;
do while not done;
........
endo;
Of course, somewhere in the loop there should be a statement that will at some point
set done equal to true.
If Statements
An If statement provides a method for executing certain statements if a condition is
true and other statements (or none) if the condition is false. A complicated if
section can come with elseif and else statements, but it always ends with an
endif statement. For example:
if x > 0.5;
.......
elseif x > 0;
.......
else;
.......
endif;
In this case, if x is greater than 0.5 the first set of statements will be executed; if not,
x is checked again for a positive value. If x is greater than 0, the second set of
statements will be executed. Otherwise, the last set of statements will be executed.
A simpler version omits the elseif section, as in:
34
GAUSS BASICS
if x > 0.5;
.......
else;
.......
endif;
In this case, if x is greater than 0.5 the first set of statements will be executed; if not,
the second set will be executed. An even simpler version omits the else section, as
in:
If x > 0.5;
.......
endif;
Here, the statements will be executed if (and only if) x exceeds 0.5.
Nesting
All of these flow control structures allow nesting, in which one type of structure lies
within another. For example:
j = 1;
do until j > n;
for k (1,n,1);
if x[j,k] > 0.5;
x[j,k] = 1.5;
endif;
endfor;
j=j+1;
endo;
The indentation is for the reader’s benefit, but is highly recommended in this and
other situations for purposes of clarity. It is wise to pair up endo (endfor, endif)
statements with preceding do (for, if) statements in a last-come-first-served
manner. It is up to the programmer to ensure that this will give the desired results.
Indenting can help, but hardly guarantees success on every occasion.
While it is tempting for those with experience in traditional programming languages

to take the easy way out, using Do Loops for mathematical operations, this
temptation should be strenuously resisted in GAUSS. For example, instead of
value = 0;
j = 1;
do while j <= n;
value = value + price[j] * quantity[j];
j=j+1;
endo;
write:
value = price'*quantity;
The latter is more succinct, far clearer, and will run much, much faster. GAUSS
performs matrix operations at blinding speed, but is downright glacial at times when
35
loops are to be executed a great many times, since it must do a certain amount of
translation of each statement every time it is encountered.
A Practical Example
Do you know the accuracy of your computer’s numerical calculation? The following
example addresses this important problem. Suppose e is a known small positive
number, and the 5×4 matrix X is defined as follows:
1 1 1 1
e 0 0 0
0 e 0 0
0 0 e 0
0 0 0 e
Verify that the eigenvalues of X'X are 4+e^2, e^2, e^2, and e^2. How small a
value of e can your computer use and still successfully invert X'X? Try to make
some sense out of the following segment of code:
one=ones(1,4);
e=1.0;
do until e<1.0e-16;
x=one|(e.*eye(4));
print "e = " e;
print invpd(x'x);
e=e./10;
endo;
end;
Writing Your Own Functions

The power of GAUSS really comes into play when you add your own functions or
procedures to enhance the language. There are two kinds of user-defined functions in
GAUSS. Single-line functions that return one item can be defined with the fn
statement. A multi-line procedure is a group of GAUSS statements put together to
perform a given task. It is better to create a program file to hold the procedures for
future use. Procedures are declared with the proc statement. A single-line function
returns only one argument, while a multi-line procedure can return one or more
arguments. Once a function or procedure is written, debugged, and placed in the
library, it is for all practical purposes part of your version of GAUSS.
Single-Line Functions
A single-line function starts with a fn statement declaring the function, followed by
the name of the function with its arguments enclosed in parentheses. The “guts” of
the function are defined on the right-hand side of the equal sign, all in one line. It is
called the same way as GAUSS built-in functions. However, it returns only one
argument. For example:
fn value(p,q) = p'*q;
This value function takes two arguments, p and q, to produce the inner product of
them. The result may be a scalar, a vector, or a matrix. Of course, this will only work
if p and q vectors or matrices are compatible for matrix multiplication. A more
36
GAUSS BASICS
complex multi-line version (or procedure) could examine the sizes of these two
matrices p and q, then use transpositions, etc., as required.
It is important to note that the argument and output names used in a function are
strictly local variables that exist only within the function itself. Thus, in a program
one could write the following statement to use the value function defined above:
cost = value(price,quantity);
In this calling statement, the function value takes two arguments, price and
quantity, which become assigned to matrices p and q of the function,
respectively. This function returns an output, called cost, which is assigned to the
output argument (value). There is no need for the names to be the same in any
respect. Moreover, the function cannot change the original arguments in any way. It
can only return information via its output.
A powerful example of a single-line function for time series conversion is the

following:
fn qtoa1(x) = meanc(reshape(x,rows(x)/4,4)');
This function converts a quarterly time series into an annual series by taking the
average of every four data points. Of course, this function will work only if the input
data series starts from the first quarter, and it is designed to handle one series at a
time. That is, the input argument x is a column vector of quarterly series, and the
function returns a column vector of annual series. Note that if the last year does not
have a complete quarterly series of four data points for conversion, it is ignored.
As another example of defining and using a single-line function statement, consider

the following scalar-valued function of one variable:
f(x) = ln(x) – x2
The maximal of f(x) is found at x = ½ or 0.707. This can be checked by evaluating

its first and second derivatives, respectively:
f′(x) = 1/x –2x

f″(x) = -1/x2 –2
Let’s write these three single-line functions in GAUSS:

fn f(x) = ln(x) – x^2;
fn f1(x) = 1/x –2*x;
fn f2(x) = -1/(x^2) –2;
Now we check the maximum x = ½ for which f1(x) = 0 and f2(x) < 0:
xmax = sqrt(0.5);
f(xmax);
f1(xmax);
f2(xmax);
Remember the built-in procedures gradp and hessp serve the same purpose of
finding the first and second derivatives, the gradient vector, and hessian matrix of a
37
user defined function without writing their analytical forms f1 and f2 as above. Try
this:
gradp(&f,xmax);
hessp(&f,xmax);
The use of gradp and hessp procedures to numerically evaluate the first and
second derivatives of a function is particularly useful when the analytical forms of
derivatives are difficult to write. Consider the following function of two variables:
g(x) = (x12 + x2 – 11)2 + (x1 + x22 –7)2
With the 2×1 parameter vector x, the function is easily defined in GAUSS:
fn g(x) = (x[1]^2 + x[2] – 11)^2 + (x[1] + x[2]^2 –7)^2;
Writing out the analytical formulas of the first and second derivatives using single-
line functions may be difficult. For this function, there are four minima: (3, 2),
(3.5844, -1.8481), (-3.7793, -3.2832), and (-2.8051, 3.1313). Using gradp and
hessp to check them is easy.
At this point you may be tempted to try using a graph to find the minima of the
above function. GAUSS is not good at graphics. Nevertheless, there are functions
available to do publication quality graphics in GAUSS. See GAUSS for Windows
User Guide or the on-line help menu for more details.
Procedures
A procedure in GAUSS is basically a user-defined function which can be more than
one line and as complicated as necessary to perform a given task. Any GAUSS built-
in command or function may be used in a procedure, as well as any user-defined
function or other procedure. Procedures can refer to any global variable and declare
local variables within. The basic structure of a GAUSS procedure consists of the
following components:
1. proc statement Procedure declaration
2. local statement Local variable declaration
3. Body of procedure …
4. retp statement Return from procedure
5. endp statement End of procedure definition
There is always one proc statement and one endp statement in a procedure
definition. Anything that comes between these two statements is part of the
procedure. local and retp statements are optional, and may occur more than once
in a procedure. GAUSS does not allow nested procedure definitions. That is, a
procedure cannot be defined within another procedure.
Variables other than input and output arguments may be included in procedures as
needed. There are global and local variables. A global variable is already declared
38
GAUSS BASICS
and used outside the procedure. A local variable is only visible to the procedure and
has no existence outside the procedure. Indeed, a local variable in one procedure may
have the same name as a different local variable in another function or procedure; the
two will coexist with neither bothering the other.
A procedure can return multiple arguments of output through retp statements and
by specifying the number of returned items in the beginning of the proc statement.
As an example, the procedure version of the value function takes inputs of p and q
and returns the total (called s) and average (called m) values as follows:
proc (2) = value(p,q);
local s, m;
s = p'*q;
m = s./sumc(q);
retp(s,m);
endp;
In the proc statement, the syntax of an equal sign preceded with the number of
returned arguments enclosed in parentheses (that is, “(2)= ” in the above example)
is not needed for a procedure with single output argument (the default case).
To use the multiple output arguments of a procedure call, simply assign them names
in the declaration line, as in:
{sum,mean} = value(price,quantity);
Here, variables price and quantity are assigned to the input arguments p and q,
respectively. Similarly, sum and mean are assigned to the output arguments s and
m. All the input and output arguments are local variables.
Note that as with inputs, the correspondence between outputs in the calling statement
and the procedure itself is strictly by order. When the procedure has finished its
work, its output values are assigned to the variables in the calling statement.
If a procedure does not return any items or you want to discard the returned items,
just call it as we have demonstrated in the above example lessons. For example:
call value(price,quantity);
Now let’s extend the earlier single-line version of time series conversion function
qtoa1 to a multi-line procedure so that it can handle the conversion of a more
general data matrix. The working method of the following procedure qtoa is to take
a data matrix x which consists of quarterly data series in columns and convert it into
a matrix of the yearly averages. The procedure takes advantage of the previously
defined single-line function qtoa1 to compute the annual average series from each
column of the quarterly data matrix, all in a Do Loop. Here is the procedure:
proc qtoa(x);
local r,c,y,i;
r = rows(x);
c = cols(x);
y = qtoa1(x[.,1]);
i = 2;
do until i > c;
y = y~qtoa1(x[.,i]);
i = i+1;
endo;
39
retp(y);
endp;
Of course, the above time series conversion function and procedure are limited to a
quarterly data series. We can make them more flexible by specifying the number of
periods of each seasonal cycle as an input argument in addition to the data matrix.
The following function tss1 and procedure tss are essentially the same as qtoa1
and qtoa, respectively. The difference is that now the number of periods n for time
series conversion is specified as one of the input arguments. Depending on the
seasonal cycle of the data series, you can use the same procedure for either quarterly
or monthly conversion.
fn tss1(x,n) = meanc(reshape(x,rows(x)/n,n)');
proc tss(x,n);
local r,c,y,i;
r = rows(x);
c = cols(x);
y = tss1(x[.,1],n);
i = 2;
do until i > c;
y = y~tss1(x[.,i],n);
i = i+1;
endo;
retp(y);
endp;
As an exercise, write the procedures to compute the analytical first and second
derivatives of this scalar-valued function of two variables:
g(x) = (x12 + x2 – 11)2 + (x1 + x22 –7)2
User Library
The purpose of writing functions and procedures is to keep tasks organized and self-
contained. Each function or procedure will perform a few given tasks and nothing
else. To make programmers’ work easier, procedures allow programmers to build on
their previous work and on the work of others rather than starting over again and
again to perform related tasks. One way to organize the work is to collect a group of
functions and procedures into a program file and register the file and its contents
with the GAUSS library system. Note that you must have your own copy of GAUSS
installed on your own computer in order to access and modify the GAUSS library
facility. We will assume that your GAUSS comes with the User Library, to which
you can add your creative functions and procedures.
First, let’s put the function tss1 and procedure tss together in a file named
TSS.SRC (SRC is the default file extension name for GAUSS source codes,
although you can use any other name you want). Put the program file TSS.SRC in
the SRC subdirectory of GAUSS path. Next, we will add the following lines to the
library file USER.LCG located in the LIB directory of the GAUSS path:
TSS.SRC
tss1 : fn
tss : proc
40
GAUSS BASICS
Similar to the idea of using a dictionary, the function tss1 and procedure tss
defined in the program file TSS.SRC are now part of GAUSS library system, which
will be searched for name recognition every time GAUSS executes a program. You
can also add variable names as matrices or strings in the library. Refer to GAUSS
Language References or the on-line help system for more details on using and
maintaining the library system.
From now on, both tss1 and tss functions are an integral part of your version of
GAUSS. You have just extended the environment for GAUSS programming!
GPE Package
The other way of extending GAUSS is to use a package, which is a set of compiled
GAUSS libraries for special purposes. GAUSS Programming for Econometricians
and Financial Analysts (GPE) is a GAUSS package of econometric procedures. The
GAUSS command use is used to load a package at the beginning of your program.
For example,
use gpe2;
will load the GPE package (version 2) for econometric analysis and applications.
Note that use can only appear once and must occur at the top of a program.
GPE consists of three main procedures (estimate, forecast, and reset)

along with global control variables that modify the econometric routines. The
procedure estimate computes linear and nonlinear regressions, while forecast
performs least squares prediction. reset initializes global control variables to their
default values. Global control variables are of two types: input and output. Global
input variables control the execution behavior of the called procedure. For example,
they can modify estimate and forecast to use linear restrictions, weighted
variables, instrumental variables, lagged dependent and independent variables, etc.
Output global variables are the results of calling estimate and forecast. They
can be assigned to new variables for further analysis. Depending on the input global
variables used which control the econometric routines, not all the output global
variables will be available. The name of an input control variable starts with a single
underscore (for example, _b), while an output control variable starts with a double
underscore (for example, __b). Refer to Appendix A for a complete list of GPE
global control variables and their default or predefined values.
A template for a typical program using GPE is given below:
/*
** Comments on program title, purposes,
** and the usage of the program
*/
1 use gpe2; @ using GPE package (version 2) @

@ this must be the first executable statement @
. /*
. ** Writing output to file or sending it to printer:
. ** specify file name for output
*/
41
/*
** Loading data:
** read data series from data files.
*/
/*
** Generating or transforming data series:
** create and generate variables with
** data scaling or transformation
** (e.g. y and x are generated here and will be used below)
*/
call reset; @ initialize global variables @
/*
** Set input control variables for model estimation
** (e.g. _names for variable names, see Appendix A)
*/
call estimate(y,x); @ do model estimation @

@ variables y, x are generated earlier @
/*
** Retrieve output control variables for
** model evaluation and analysis
*/
/*
** Set more input control variables if needed,
** for model prediction
** (e.g. _b for estimated parameters)
*/
call forecast(y,x); @ do model prediction @
end; @ important: don’t forget this @
Using the GPE package in a GAUSS environment is the main focus of the rest of this
book, which begins in the next chapter on linear regression models. If you are
already familiar with linear least squares estimation, you can jump to the nonlinear
models discussion which begins in Chapter VI. The topic of simultaneous equation
systems is covered in Chapter XIII. In addition to many classical econometric
methods, modern approaches such as generalized method of moments (Chapter XII),
autoregressive conditional heteroscedasticity (Chapter XV), and panel data analysis
(Chapter XVI), are programmed and solved with GPE (version 2) for GAUSS.
Don’t forget that we are learning GAUSS as a tool to do econometrics. The package
GPE written in GAUSS acts as a bridge between the domain knowledge
(econometrics) and the programming environment (GAUSS). With this approach,
only a limited knowledge of computer programming is required in the beginning.
After gaining experience with GPE and GAUSS in general, you should be ready for
your own programming adventure in advanced econometrics, by either extending
GPE or writing new programs.
42
III
Linear Regression Models
GPE (GAUSS Programming for Econometricians and Financial Analysts) is a
GAUSS package for linear and nonlinear regressions useful for econometric analysis
and applications. The purpose of this chapter is to show you how to use GPE for
basic linear least squares estimation.
Least Squares Estimation

A linear regression model can be written either in a matrix form:
Y = Xβ + ε
or, in a vector form:
Y = X1β1 + X2β2 + … + XKβK + ε
where Y is the dependent variable, Xk is the k-th independent (explanatory) variable,

and βk is the corresponding parameter (coefficient), k = 1, 2, … , K. Given a sample
of N data observations, both Xk’s and Y are N-vectors of sample data. Denote X =
[X1 X2 … XK]. Typically, we assume the last column of data matrix X, that is XK, is
a constant one-vector. For parameter estimation of a linear regression model, the
random error term ε is assumed to be identically independently distributed (iid). For
statistical inference, an assumption of probability density will be necessary. For
example, ε is normally identically independently distributed (nid) with zero mean
and constant variance σ2.
The ordinary least squares regression amounts to the following estimation results:
b = (X'X)-1X'Y Estimator of β.
Var(b) = s2(X'X)-1 Estimated variance-covariance matrix of β.
e = Y - Xb Estimated errors ε or residuals.
s2 = e'e/N-K Estimated regression variance σ2. N is the number of
sample data; K the is the number of parameters.
Consider the simple case of regressing the dependent variable Y against one
independent variable X in addition to a constant:
Y = α + βX + ε
where ε is the difference between known Y and estimated Y, or the residual. The
parameter α is the intercept and β is the slope of the linear regression equation.
Continuing with the text data file longley.txt we used in Chapter II, Lesson 3.1
introduces the use of GPE to estimate a simple regression equation. Lesson 3.2
examines a set of regression statistics obtained from the simple regression. Lesson
3.3 is a multiple regression model.
Lesson 3.1: Simple Regression

This lesson estimates a simple regression relationship between employment EM in
millions of persons and RGNP or real GNP expressed in billions of 1954 dollars.
Note that the 1954 observation of PGNP is 100, which is the base year of the
deflator. All data series are read from the text data file longley.txt.
/*
** Lesson 3.1: Simple Regression
*/
4 data = data[2:17,.];
5 PGNP = data[.,2];
6 GNP = data[.,3]/1000;
7 EM = data[.,7]/1000;
8 RGNP = 100*GNP./PGNP;
9 call reset;
10 _names = {"EM","RGNP"};
11 call estimate(EM,RGNP);
12 end;
In order to use GPE package for econometric analysis and applications, the first
executable statement in your GAUSS program must be:
use gpe2;
This tells GAUSS where to look when GPE commands are used. However, remarks
enclosed in comment symbols are permitted before the first line of program code.
A linear regression model is estimated using the GPE econometric procedure,

estimate. Before calling estimate, it is recommended to initialize all the global
control variables first by calling the procedure reset. Between the two procedures,
reset and estimate, the required GPE input control variables for model
estimation are set.
In Lesson 3.1, line 9 initializes all the GPE global control variables by calling
reset procedure. Then, in line 10, the input control variable _names is defined to
be a list of character names for variables used in the regression (dependent variable
first, followed by independent variables in the order of appearance in the equation).
In this example, EM is the dependent variable, RGNP the independent variable, and
_names is a character vector of variable names as:
_names = {“EM”, “RGNP”};
Not starting GPE input control variables such as _names with an underline (_) is a
common mistake. GAUSS ignores them without warning or an error message. Your
program using GPE just will not work like it should. See Appendix A for more
information about the usage of _names and other input control variables.
44
LINEAR REGRESSION MODELS
If _names is not specified, then the default variable names are used for the
procedure estimate. That is, Y for the name of the dependent variable and X# for
the names of the independent variables (# indicates the number in sequence, i.e., 1,
2, … ).
The GPE econometric procedure estimate is called in line 11. It takes the
dependent variable as the first argument, and the list of independent variables as the
second. A constant vector for the estimated intercept term is automatically added to
the model estimation.
Now, let’s look at the output from running this program:

------------------------
Dependent Variable = EM
Estimation Range = 1 16
Number of Observations = 16
Mean of Dependent Variable = 65.317
Standard Error of Dependent Variable = 3.5120
R-Square = 0.97320 R-Square Adjusted = 0.97129

Standard Error of the Estimate = 0.59511
Log-Likelihood = -13.331
Log Ammemiya Prediction Criterion (APC) = -0.92023
Log Akaike Information Criterion (AIC) = -0.92154
Log Schwarz Bayesian Information Criterion (BIC) = -0.82497
Sum of Squares SS DF MSS F Prob>F

Explained 180.05 1 180.05 508.39 2.1048E-012
Residual 4.9582 14 0.35416
Total 185.01 15 12.334
Variable Estimated Standard t-Ratio Prob Partial

Name Coefficient Error 14 DF >|t| Regression
RGNP 0.058726 0.0026045 22.547 2.1048E-012 0.97320
CONSTANT 43.264 0.98931 43.732 2.2494E-016 0.99273
The basic output of estimate is presented in four blocks. The first block gives
general information about the regression. Goodness of fit of the estimated regression
and several model selection criteria are given in block two. Block three is the
standard Analysis of Variance (AOV) . The following discussion focuses on the last
block of output information. Values of the estimated coefficient, standard error, and
t-ratio for each variable are given row-wise. Reading the output for each variable
gives the estimated model as:
EM = 43.264 + 0.059 RGNP

s.e. (0.989) (0.0026)
t-ratio 43.7 22.5
Interpreting this output tells us that, on average for each one billion dollar increase of
RGNP (measured in 1954 value), there will be an increase of about 59 thousand in
people employed (EM).
Since the expected value of the residuals is zero, it is not in the estimated regression
equation. However, a list of error values for each observation is available for further
analysis to be discussed in later lessons.
45
Testing of the simple hypothesis that a given coefficient is equal to zero takes the
estimated coefficient’s t-ratio and compares it with the critical value from the
Student’s t distribution listed for the given degrees of freedom (DF). Prob > |t| is the
corresponding P-value, or the probability of a type II error (that is, the probability of
not rejecting the null (false) hypothesis that the corresponding coefficient equals
zero). We know that RGNP’s coefficient is statistically significant from its t-ratio of
22.5 and our chance of being wrong is 2×10-12, or very close to zero.
The partial regression coefficient measures the marginal contribution of the variable
when the effects of other variables have already been taken into account. For a linear
regression including only one independent variable, this is just the R-square (0.9732
in this case) of the regression.
The GPE estimate routine is the foundation of most models that you will use to
estimate a regression equation. Global control variables are then added in different
combinations to check, test, and hopefully correct the fitted regression. In the next
several lessons, further analysis of the regression is achieved through the use of input
control variables for the estimate procedure. Again, refer to Appendix A for a
more detailed description of GPE input control variables.
Lesson 3.2: Residual Analysis

How good are the results of the simple regression from Lesson 3.1? Of course, the
measurement of R-square is a popular yardstick to judge the goodness of fit. One
more tool for evaluating just how good an estimated regression fits a set of
observations is to analyze residuals, the difference between the actual and estimated
dependent variable. GPE provides several options in conjunction with the
estimate procedure to analyze and evaluate residuals. This lesson explores three
ways to perform residual analysis.
/*
** Lesson 3.2: Residual Analysis
*/
4 data = data[2:17,.];
5 PGNP = data[.,2];
6 GNP = data[.,3]/1000;
7 EM = data[.,7]/1000;
9 call reset; @ initialize control variables @

10 _rstat = 1; @ report residual statistics @
11 _rlist = 1; @ list residuals @
12 _rplot = 1; @ plot residuals @
13 _names = {"EM","RGNP"};
14 call estimate(EM,RGNP);
15 end;
Before using global control variables, as we have mentioned earlier, you need to call
the reset procedure once to initialize them. Calling reset returns all GPE global
control variables to their default setting.
46
The reset option used with an output file has nothing to do with a called
reset in GPE. The latter is a GPE procedure, while the former is an option
associated with the GAUSS command output. They are simply two different
concepts.
_rstat is the residual analysis tool most frequently used in conjunction with
GPE’s estimate. Setting the GPE input control variable to a non-zero value (the
convention is 1, meaning true or yes) provides a set of simple residual statistics.
These statistics are: squared correlation of the observed (actual) and predicted (fitted)
values of the dependent variable, sum-of-squared residuals, sum of absolute
residuals, sum of residuals, and the serial correlation coefficient, first-order Rho. The
well-known Durbin-Watson test statistic is useful for testing the presence of first-
order serial correlation. The output of residual statistics is:
Squared Correlation of Observed and Predicted = 0.97320
Sum of Squared Residuals = 4.9582
Sum of Absolute Residuals = 7.6446
Sum of Residuals = 5.32197E-012
First-Order Rho = 0.23785
Durbin-Watson Test Statistic = 1.4408
The option _rlist = 1 lists the observed (actual) and predicted (fitted) values of
the dependent variable. The residual is computed as the difference between the actual
and fitted values. Each observation of residual and its standard error is listed as well.
List of Observed, Predicted and Residuals
Obs Observed Predicted Residual Std Error
1 60.323 59.841 0.48195 0.52253
2 61.122 60.479 0.64314 0.53477
3 60.171 60.446 -0.27506 0.53419
4 61.187 61.938 -0.75124 0.55639
5 63.221 63.347 -0.12561 0.56955
6 63.639 64.037 -0.39762 0.57341
7 64.989 64.938 0.050585 0.57597
8 63.761 64.588 -0.82719 0.57531
9 66.019 66.329 -0.31005 0.57446
10 67.857 66.798 1.0588 0.57246
11 68.169 67.251 0.91781 0.56979
12 66.513 66.826 -0.31280 0.57232
13 68.655 68.439 0.21576 0.55933
14 69.564 69.110 0.45430 0.55112
15 69.331 69.565 -0.23401 0.54454
16 70.551 71.140 -0.58873 0.51511
Plotting is a quick way to evaluate the result of model estimation. Setting _rplot
= 1 will return a plot of estimated residuals, while setting _rplot = 2 produces
both the residual graph and fitted-vs.-actual dependent variable series. The graph is
shown in a separate window when running the program. By viewing the plot of
residuals, the correlation patterns in residuals may indicate a need to re-specify the
model.
47
Lesson 3.3: Multiple Regression

Evaluating output from Lesson 3.2 shows it was in general a good model, but could
adding more explanatory variables make it better? Let’s see what happens when an
additional variable, POPU (that is, population in millions of persons), is included in
the regression.
In the following, we add a few new twists to both the programs of Lesson 3.1 and
3.2. In addition to _rstat, _rplot, _rlist, this lesson introduces the use of
another input control variable, _vcov. By setting it to a non-zero value (i.e., 1), the
regression output will include a variance-covariance matrix as well as a correlation
matrix of the estimated coefficients. It is often useful to examine the relationship
among estimated coefficients in a multiple regression.
/*
** Lesson 3.3: Multiple Regression
*/
4 data = data[2:17,.];
5 PGNP = data[.,2];
6 GNP = data[.,3]/1000;
7 POPU = data[.,6]/1000;
8 EM = data[.,7]/1000;
10 call reset; @ initialize control variables @

11 _vcov = 1; @ print var-covar matrix @
12 _rstat = 1; @ report residual statistics @
13 _rlist = 1; @ list residuals @
14 _rplot = 2; @ plot data and residuals @
15 _names = {"EM","RGNP","POPU"};
16 call estimate(EM,RGNP~POPU);
17 end;
48
Remember how to use “~” to horizontally concatenate vectors or matrices?

RGNP~POPU is the data matrix of independent variables for this multiple regression.
Line 16 indicates that the dependent variable EM is the first argument of the
estimate procedure, followed by the data matrix of independent variables
RGNP~POPU as the second argument. The list of respective variable names used in
the regression is defined by _names, an input control variable, in line 15. Note that
a constant vector for the estimated intercept is automatically added to the model
estimation.
Including POPU has changed our estimated regression, but is it better? Analyzing the
following result will tell the story.
------------------------


Explained 180.26 2 90.131 246.81 4.5725E-011
Residual 4.7473 13 0.36518
Total 185.01 15 12.334

RGNP 0.068698 0.013386 5.1322 0.00019257 0.66954
POPU -0.086282 0.11353 -0.76001 0.46081 0.042542
CONSTANT 49.651 8.4631 5.8667 5.5306E-005 0.72585
Variance-Covariance Matrix of Coefficients

RGNP 0.00017918
POPU -0.0014897 0.012888
CONSTANT 0.10764 -0.95400 71.624
RGNP POPU CONSTANT
Correlation Matrix of Coefficients

RGNP 1.0000
POPU -0.98029 1.0000
CONSTANT 0.95017 -0.99293 1.0000
RGNP POPU CONSTANT

Sum of Residuals = -5.47779E-010
List of Observed, Predicted and Residuals

Obs Observed Predicted Residual Std Error
1 60.323 59.758 0.56493 0.51924
2 61.122 60.416 0.70616 0.53666
3 60.171 60.279 -0.10802 0.49591
49
4 61.187 61.925 -0.73787 0.56470

5 63.221 63.474 -0.25252 0.55371
6 63.639 64.178 -0.53860 0.55192
7 64.989 65.075 -0.086161 0.55650
8 63.761 64.568 -0.80739 0.58361
9 66.019 66.504 -0.48502 0.53598
10 67.857 66.937 0.92024 0.55198
11 68.169 67.319 0.85001 0.57167
12 66.513 66.692 -0.17851 0.55364
13 68.655 68.457 0.19823 0.56749
14 69.564 69.068 0.49565 0.55697
15 69.331 69.387 -0.055654 0.50068
16 70.551 71.036 -0.48547 0.50510
Our estimated multiple regression equation is
EM = 49.65 + 0.069 RGNP - 0.086 POPU

s.e. (8.46) (0.013) (0.114)
t-ratio 5.87 5.13 - 0.76
RGNP still has about the same influence on EM, as reported in the previous lesson.
Based on residual statistics, the model has a similar performance to the simple
regression without the POPU variable.
Pay special attention to the “Partial Regression Coefficient,” which gauges the
marginal contribution of each variable when the effects of other variables have
already been taken into account. In terms of model interpretation, the negative slope
coefficient of POPU is not what we would expect. POPU has an extremely low partial
regression coefficient and it is not statistically significant as seen by its near zero t-
ratio. Looking at the outputs of the variance-covariance matrix and the correlation
matrix of coefficients, the estimated coefficient of POPU has a relatively large
variance and it is strongly correlated with that of RGNP. Therefore, our regression is
better without POPU.
Estimating Production Function

So far this handbook has given cut and dried examples of what each GAUSS or GPE
command does. Now we will attempt to show the flexibility that GPE has to offer. In
the following sections, we will estimate the famous Cobb-Douglas production
function with time series of U.S. real output and inputs. To do so, a new data set
named cjx.txt is introduced; it also comes with the GPE package and is installed in
the GPE subdirectory. cjx.txt contains six annual data series from 1929 to 1967.
Let’s look at the contents of cjx.txt. The data matrix has 40 rows and 6 columns,
with no missing values. The first row contains variable names and should not be
included for analysis. The bottom of the file contains descriptive information and
also should not be used. Why is looking at each data file so important?
Troubleshooting problems caused by imperfect data files is very difficult. Error
messages caused by loading imperfect data are not always obvious and can show up
anywhere in your program.
Out of the cjx.txt data series we will use only the following selected variables:
GNP in constant dollars (X);

Number of persons, adjusted for hours of work and educational level (L1);
Capital stock, adjusted for rate utilization (K1).
50
To make our following presentation easier, L1 has been renamed L and K1 has been
renamed K. With the introduced notations, a simple two-input Cobb-Douglas
production function is written as
X = α Lβ1 Kβ2 exp(ε)
To transform the model into a more useful form, natural logarithms are taken on both
sides of the equation:
ln(X) = β0 + β1 ln(L) + β2 ln(K) + ε
where β0 = ln(α) is the intercept of the log model, and the slopes β1 and β2 are
interpreted as input elasticities. Econometric estimation of this Cobb-Douglas
production function is the focus of the following few lessons.
Lesson 3.4: Cobb-Douglas Production Function

We will first estimate the unrestricted influences of labor input L and capital input K
on real output X. Then, the economic theory of constant returns to scale (CRS) is
formulated and tested with the use of linear restrictions. Testing CRS is done by
restricting the summation of the two slope coefficients to 1 in the log specification of
the model. That is, β1 + β2 = 1. To test the hypothesis of CRS, we need to check how
residuals from both the first unrestricted and second restricted regressions compare.
/*
** Lesson 3.4: Cobb-Douglas Production Function
*/
1 use gpe2;
3 load data[40,6] = gpe\cjx.txt;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 call reset;
9 _names = {"X","L","K"};
10 call estimate(X,L~K);
11 _restr = {1 1 1};
13 end;
Optional residual analysis: _rstat=1, _rplot=2, _rlist=1, _vcov=1 may

be added to the program.
Before examining the output, let’s look at the programming style. This program is
efficient, that is, many actions are combined in few lines of code. Line 4 removes the
first row (variable names) of the data file and indexes the matrix data into a vector,
all in one step. Lines 5, 6, and 7 go one step further. In addition to indexing the
matrix, they take the natural logarithm of each variable.
Line 10 estimates the basic, unrestricted, least squares regression output:

51
------------------------
Dependent Variable = X

Log-Likelihood = 77.286

Explained 8.0310 2 4.0155 3332.2 1.3921E-041
Residual 0.043382 36 0.0012051
Total 8.0744 38 0.21248

L 1.4508 0.083228 17.431 3.9260E-019 0.89407
K 0.38381 0.048018 7.9930 1.7130E-009 0.63960
CONSTANT -3.9377 0.23700 -16.615 1.8332E-018 0.88464
The estimated model is:
ln(X) = - 3.94 + 1.45 ln(L) + 0.38 ln(K)

s.e. (0.24) (0.083) (0.048)
t-ratio - 16.62 17.43 7.99
Interpreting the estimation result of the log model takes into account that estimated
slope coefficients translate directly into elasticities. In other words, the influence of
labor L and capital K on output GNP (X) is expressed in terms of “percentage
change.” For every one percent increase in labor input L, GNP increases by 1.45
percent. When capital input K is increased by one percent, GNP increases 0.38
percent.
An adjusted R2 of .994 reveals a very good fit of the regression equation. Large t-
ratios and small P-values for all variables show that the chance of the elasticity
coefficients being zero is also small. Moreover, partial regression coefficients are
strong enough for both ln(L) and ln(K). The resulting model is the basis for many
lessons to come.
Now, let’s consider the theory of constant returns to scale often assumed in many
classical productivity studies. Restricted least squares is the technique used to
estimate models with linear restrictions. GPE’s estimate procedure can be
modified to perform restricted least squares with the use of the input control variable
_restr (see Appendix A for details).
The last part of the program (lines 11 and 12) demonstrates a simple example of
restricting β1 + β2 = 1, in order to test for CRS. To understand what line 11 is doing,
we need to describe a little matrix algebra. Linear restrictions on least squares
coefficients can be expressed by the equation
Rβ = q
52
where R is the restriction matrix specifying a set of linear relationships among

estimated coefficients. It tells which coefficients are to be restricted. β is the column
vector of estimated coefficients, and q is the column vector of values that the linear
combination of β’s are restricted to. In most cases, restrictions are imposed on slope
coefficients. Separating the intercept term from the slope coefficients, the matrix
representation of linear restrictions is rewritten as:
[ Rs R0 ]
 βs  = q
β0
where βs is the vector of slope coefficients and Rs is the restriction matrix

corresponding to βs. Similarly, β0 is the intercept coefficient and R0 corresponds to
β0. The input control variable _restr in GPE is implemented according to the
restrictions on slope coefficients defined below:
_restr = [ Rs q ]
Linear restrictions involving the intercept term will necessitate the explicit inclusion
of a constant column as part of the data matrix of independent variables and will
estimate the model without an intercept. This is done by setting the input control
variable _const = 0 before calling the estimate procedure.
Back to our CRS example, line 11:
_restr = {1 1 1};
Now let’s look at the matrix to the right of the equal sign, {1 1 1}. Each row of
_restr specifies a single restriction. Therefore, only one restriction is called out
(i.e. β1 + β2 = 1) in this example. The number of columns of _restr comes from
the number of slope coefficients βs plus the one column of the restricted value q. The
first two columns of 1’s in _restr(βs) select β1 and β2. When multiplied with
corresponding slope coefficients, it is the sum of β1 and β2. The last column of 1’s in
_restr(q) specifies that the resulting sum of β1 and β2 equals 1. In other words,
the _restr matrix calculates:
[1 1]
β1 = 1*β + 1*β = 1.
β2 1 2
Restricting β1 = β2 is the same as β1 - β2 = 0. Using the information from the

program lesson3.4, the GPE command for the restriction β1 - β2 = 0 is:
_restr = {1 -1 0};
That is, [ 1 -1 ]
β1 = 1*β -1*β = 0.
β2 1 2
More complicated cases may have several restrictions. To demonstrate such a

situation, assume a model with four slope coefficients β1, β2, β3, β4. Imposing the
restrictions β2 = 0 and β3 + β4 = 1 would use:
53
_restr ={0 1 0 0 0,
0 0 1 1 1};
β1
  0*β + 1*β + 0*β + 0*β 0
0  β 
That is, 
0 1 0 2
=
 1 2  = 
3 4
0 0 1 1 β
   0*β + 0*β + 1*β + 1*β  1
β 
3 1 2 3 4
Look at the output produced by restricted least squares estimation (line 12):
------------------------
WARNING: Linear Restrictions Imposed.

R-Square, AOV, SE, and t may not be reliable!
Wald F-Test for Linear Restrictions
F( 1, 36) Prob>F
427.66 1.4430E-021
Asymptotic Hypothesis Testing for Linear Restrictions
Chi-Sq DF Prob>Chi-Sq
Wald Chi-Square Test 427.66 1 5.2567e-095
Lagrange Multiplier Test 35.972 1 2.0018e-009
Likelihood Ratio Test 99.669 1 1.8008e-023


Explained 6.2806 1 6.2806 415.91 1.0185E-021
Residual 0.55874 37 0.015101
Total 8.0744 38 0.21248

L -0.15051 0.10802 -1.3933 0.17183 0.049854
K 1.1505 0.10802 10.651 8.0037E-013 0.75405
CONSTANT 0.95015 0.061657 15.410 1.1067E-017 0.86520
Before testing for CRS, let’s look at the format of the output. Notice the warning
near the top. Due to the imposed restrictions, standard statistics based on residuals
are reliable only when the restrictions are correct. The Wald test statistic of linear
restrictions is given along with its P-value, directly under the warning near the top of
the output. The Wald test statistic uses residual sum-of-squares (RSS) from both
unrestricted and restricted models to check if the stated restrictions yield a model that
is statistically different from the model without the restrictions.
The Wald test statistic is:
54
RSS* - RSS
J
RSS ∼ F(J, N-K)
N-K
where RSS* is the restricted residual sum-of-squares with β1 + β2 = 1; RSS is the

unrestricted residual sum-of-squares; J is the number of restrictions; K is the number
of variables, including the constant (do not confuse K with the variable name for
capital input in this program); and N is the number of total observations. That is,
0.55874 - 0.04338
1
0.04338 = 427.66 ~ F(1, 39-3)
39-3
At a 5% level of significance, the F critical value of 4.17 places our computed value,
427.66, in the right-tail rejection region of the F distribution. Together with a near
zero P-value for the Wald statistic, this result leads us to reject the linear restriction
β1 + β2 = 1.
In addition to Wald test for linear restrictions, large sample test statistics such as
Lagrange Multiplier test and Likelihood Ratio test are reported in the output. Refer
to econometrics textbooks about the derivation and application of these tests. All the
corresponding P-value of these statistics are very small pointing to the same
conclusion to reject the linear restriction of constant returns to scale. Based on this
simple two-input Cobb-Douglas specification of the production technology, the data
series from cjx.txt does not support the theory of constant returns to scale. As a
matter of fact, the U.S. production technology exhibited the pattern of increasing
returns to scale (i.e., β1 + β2 > 1) at least from 1929 to 1967.
Having thrown away the hypothesis of constant returns to scale, the next interesting
issue about the production function is presented. Is there any difference in factor
productivity between pre-war and post-war periods?
Lesson 3.5: Testing for Structural Change

The goal of this lesson is to determine if real output, measured by GNP from cjx.txt,
underwent structural change over the time period between 1929 to 1967, starting in
1948. That is, is there a difference between the estimated coefficients when one
regression spans the entire time period versus when two separate regressions are
estimated? One way to determine if there is a difference is to use the Chow test. The
Chow test compares the results of the regression of the entire time period (1929-
1967) against the regression data from 1929-1948 and then 1949-1967. If a
statistically significant difference is found, we can assume that there was a structural
change in productivity after 1948.
/*
** Lesson 3.5: Testing for Structural Change
*/
1 use gpe2;
4 year = data[2:40,1];
55
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 call reset;
9 _names = {"X", "L", "K"};
10 call estimate(X,L~K); @ whole sample @
11 _begin = 1; @ sub-sample: 1929-1948 @

12 _end = 20;
14 _begin = 21; @ sub-sample: 1949-1967 @

15 _end = 39;
17 end;
Run the above program to analyze the output. It calls estimate three times. The
first time it estimates the entire sample (the default case):
------------------------


Explained 8.0310 2 4.0155 3332.2 1.3921E-041
Residual 0.043382 36 0.0012051
Total 8.0744 38 0.21248

L 1.4508 0.083228 17.431 3.9260E-019 0.89407
K 0.38381 0.048018 7.9930 1.7130E-009 0.63960
CONSTANT -3.9377 0.23700 -16.615 1.8332E-018 0.88464
The first regression is named the restricted model because it restricts the entire time
period to having the same structure. The estimated restricted model is:
ln(X) = - 3.94 + 1.45 ln(L) + 0.38 ln(K)

s.e. (0.24) (0.083) (0.048)
t-ratio - 16.62 17.43 7.99
Sub-samples for the second and for the third regression estimations are controlled by
_begin and _end, set to the desired observation numbers, respectively. _begin
and _end allow regressions of varying sizes to be estimated from a single data
series. Since the default value of _begin is 1, it really is not necessary in line 11.
The next line _end = 20 tells GPE to use up to, and including, the 20th row of
data for estimate. Here is the output of the second regression equation:
------------------------
56


Explained 1.4399 2 0.71996 344.24 1.7649E-014
Residual 0.035555 17 0.0020915
Total 1.4755 19 0.077656

L 1.6167 0.20897 7.7367 5.7391E-007 0.77881
K 0.21967 0.22995 0.95530 0.35281 0.050947
CONSTANT -4.0576 0.35722 -11.359 2.3202E-009 0.88358
Notice that the estimation range is from 1 to 20, using 20 observations. Running the
regression using only the time period from 1929 to 1948 returns the following
model:
ln(X) = - 4.06 + 1.62 ln(L) + 0.22 ln(K)

s.e. (0.36) (0.21) (0.23)
t-ratio - 11.36 7.74 0.96
The third regression with _begin = 21 (line 14) and _end = 39 (line 15) tells
GPE to estimate the model from the 21st row of the data series up to, and including,
the last or the 39th row. Let’s look at the regression result:
------------------------


Explained 0.79237 2 0.39618 1888.5 1.0026E-019
Residual 0.0033566 16 0.00020979
Total 0.79572 18 0.044207

L 1.0090 0.14403 7.0054 2.9675E-006 0.75413
K 0.57909 0.055248 10.482 1.4222E-008 0.87288
CONSTANT -2.4981 0.53122 -4.7025 0.00023960 0.58021
57
Now, notice that the estimation range is from 21 to 39, using 19 observations.
Regressing only the time period from 1949 to 1967 returns the following model:
ln(X) = - 2.50 + 1.01 ln(L) + 0.58 ln(K)

s.e. (0.53) (0.14) (0.055)
t-ratio - 4.70 7.01 10.48
Back to the question at hand, was there a structural change in productivity between
the years of 1929 and 1967? We have processed our raw data using 1948 as the
break point, and only need to apply the formal Chow test on the results. Computing
the Chow test statistic as follows:
RSS* −(RSS1 + RSS2)

J ∼ F(K, N-2K)
RSS1 + RSS2
N − 2K
where RSS* is the restricted residual sum-of-squares for the whole sample (1929-
1967); RSS1 is the residual sum-of-squares for the first sub-sample (1929-1948);
RSS2 is the residual sum-of-squares for the second sub-sample (1949-1967); K is the
number of variables, including constant, in each regression (again, do not confuse K
with the variable name for capital input in this program); and N is the number of
observations for the whole sample.
Plugging in the numbers yields:
0.04338 − (0.03555 + 0.00336 )

3 = 1.27 ~ F(3, 39-6)
0.03555 + 0.00336
39 − 6
At a 5% level of significance, comparing the Chow test statistic (1.27) against the F
critical value of 2.92 leads us to conclude that, based on the Cobb-Douglas
specification, there was no structural change in productivity between 1929 and 1967.
Lesson 3.6: Residual Diagnostics

From the previous two lessons on the study of the Cobb-Douglas production function
using U.S. input and output data series from cjx.txt, we concluded that production
technology did not exhibit constant returns to scale. As a matter of fact, from 1929 to
1967, there was a pattern of increasing returns to scale (see Lesson 3.4). By dividing
the sample into pre-war (1929-1948) and post-war (1949-1967) periods, we did not
find any structural differences in these two periods (see Lesson 3.5). It is better to
estimate the production function using the whole sample.
This returns us to our original least squares estimation of the model:
ln(X) = -3.94 + 1.45 ln(L) + 0.38 ln(K)
or in exponential form:
1.45 0.38
X = 0.02 L K
58
For the purpose of statistical inference, the log-model is assumed to be normally

distributed. In other words, X is log-normally distributed. Do the estimated errors, or
residuals, in fact follow a normal distribution? Are there any dominant observations
of residuals that distort the distribution of the remaining residuals? The former is a
question about the underlying normality assumption of the model, while the latter
relates to issues of influential observations and outliers.
Besides examining standard residual statistics and plotting residual series, GPE
offers a set of diagnostic information to check the characteristics of residuals in
depth.
The first half of the program below is the same as that of lesson3.4. After removing
_rplot and setting _rstat to typical values, we add the following two lines:
_bjtest = 1;
_rlist = 2;
Setting _bjtest = 1 (meaning yes or true) will carry out the Bera-Jarque
normality test on the residuals.
We have seen the use of _rlist = 1 which lists each observation of the residuals
and their standard errors, in addition to observed (actual) and predicted (fitted) data
series. With _rlist = 2, in addition to residuals and their standard errors, useful
information on influential observations and outliers is available.
/*
** Lesson 3.6: Residual Diagnostics
*/
1 use gpe2;
2 output file = gpe\output3.6;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 names = {"X", "L", "K"};
9 call reset;
10 _rstat = 1;
11 _rlist = 2; @ check influential obs. @
12 _bjtest = 1; @ normality test @
14 end;
Running the above program, the output file output3.6 is generated. For model
evaluation, we now refer to output3.6. After reporting basic residual statistics,
the Bera-Jarque Wald test for normality computes the statistic based on the
measurements of skewness and kurtosis for the residuals as follows:
Bera-Jarque Wald Test for Normality
Asymptotic Standard Error of Residuals = 0.033352
Skewness of Residuals = 0.84226
Kurtosis of Residuals = 4.7072
Chi-Sq( 2) Prob>Chi-Sq
0.0093379
59
The resulting test statistic follows the Chi-squared probability distribution with 2
degrees of freedom. The computed value of 9.35 for the Bera-Jarque test statistic is
far greater than the critical value for either a 5 or 10 percent level of significance (or,
less than 1% in its P-value). The null hypothesis of residual normality is rejected!
For a perfect normal distribution, residual skewness should be 0 and residual kurtosis
should be 3. The rejection of normality is not a surprise. However, non-normal
residuals render potential problems for statistical inference.
The last part of the output reports the regression diagnostics for influential
observations and outliers:
Residual Diagnostics for Influential Observations and Outliers

(Valid for Checking Ordinary Least Squares Residuals Only)
Standardized Studentized
Obs Leverage Residual Residual DFFITS
1 0.048570 -0.37556 -0.37103 -0.083832
2 0.069668 -1.2819 -1.2938 -0.35406
3 0.092468 -1.2069 -1.2149 -0.38779
4 0.20093 -0.41784 -0.41300 -0.20710
5 0.18527 -0.97098 -0.97019 -0.46265
6 0.12035 -0.61196 -0.60656 -0.22435
7 0.088917 -0.23985 -0.23669 -0.073941
8 0.063757 0.30955 0.30563 0.079757
9 0.064645 -0.076541 -0.075477 -0.019842
10 0.067310 1.3562 1.3727 0.36876
11 0.057612 1.4644 1.4889 0.36815
12 0.048564 1.2914 1.3039 0.29458
13 0.045681 0.54239 0.53701 0.11749
14 0.049876 -0.51380 -0.50848 -0.11650
15 0.044225 0.036933 0.036417 0.0078335
16 0.048269 1.9447 2.0269 0.45646
17 0.049727 3.2828 3.8671 0.88462
18 0.14309 0.41797 0.41313 0.16882
19 0.13632 -2.0428 -2.1422 -0.85110
20 0.11090 -1.8036 -1.8646 -0.65853
21 0.064225 -0.61213 -0.60673 -0.15895
22 0.055357 0.035430 0.034935 0.0084570
23 0.056146 -0.79457 -0.79042 -0.19278
24 0.050215 -0.95665 -0.95549 -0.21970
25 0.043698 -0.69982 -0.69478 -0.14852
26 0.032275 0.32336 0.31930 0.058311
27 0.037287 0.43459 0.42964 0.084554
28 0.042054 -0.44591 -0.44089 -0.092378
29 0.046087 -0.29348 -0.28972 -0.063681
30 0.056159 0.42446 0.41957 0.10235
31 0.058795 0.010363 0.010218 0.0025538
32 0.064299 -0.41384 -0.40902 -0.10722
33 0.078586 0.11346 0.11190 0.032679
34 0.084366 0.36577 0.36132 0.10968
35 0.083668 0.75412 0.74951 0.22648
36 0.094336 0.69948 0.69443 0.22412
37 0.097121 0.56208 0.55667 0.18257
38 0.10356 0.017775 0.017526 0.0059569
39 0.11562 -0.91018 -0.90796 -0.32830
Mean 0.076923 -0.0072089 0.0067883 -0.024796
To check for influential observations and outliers, we first take a look at the column
“Leverage,” which measures the influence of each observation on the regressors. We
check for leverage which is greater than 2×(K/N) where K is the number of
estimated coefficients and N is the number of observations. In this case, 2×(K/N) =
2×(3/39) = 0.154. Observations 4 and 5 (leverage 0.201 and 0.185, respectively) are
quite influential.
60
Standardized (or normalized) residuals should follow a standardized normal

distribution, provided that the Bera-Jarque test statistic confirms the distribution
assumption. Unfortunately, this is not the case as shown by the test results above.
Observations 17 and 19 (3.283 and –2.043, respectively) are greater than 2 standard
errors from the mean.
A more robust measure of outliers uses the studentized residuals (or standardized
predicted residuals) which follows the Student’s t distribution with N-K-1 degrees of
freedom. Given the critical value of 1.69 at a 5% level of significance, observations
16, 17, 19, and 20 are candidates for outliers.
The last column, “DFFITS,” measures the contribution of each observation to the
0.5
prediction of the model. The cutoff value 2×(K/N) is suggested (that is, 0.555 in
the case of this Cobb-Douglas production model). The contribution of observations
17, 19, and 20 are rather large.
Materials of this lesson on influential observations and outliers can be found in Judge
et al. (1988) and Maddala (1988). In summary, for our study of the Cobb-Douglas
production function, the model is sensitive to the use of data near the end of World
War II (i.e., observations 17, 19, and 20). The model may be better explained
without them.
61
IV
Dummy Variables
Dummy variables are widely used in econometrics to isolate sub-group effects in a
given sample. These sub-groups may be geographical regions, yearly quarters,
gender, or periods in time. How dummy variables are used in regression estimation
determines in which way the sub-groups differ. The so-called dummy variables
themselves remain vectors of ones and zeros. A one indicates the presence of a given
characteristic, while a zero indicates its absence. In most cases, one less dummy
variable is used than there are sub-groups. Estimated regressions from these sub-
groups may have an additive difference, a multiplicative difference, or a combined
additive and multiplicative difference. An additive difference refers to a parallel shift
in the level of an estimated regression. This shift is reflected in a change of the
intercept term, while the other coefficients remain unchanged. The slope coefficients
will vary with their associated multiplicative dummy variables. The estimated
changes in slope coefficients among sub-groups are measured by the coefficients of
multiplicative dummy variables. A combined additive and multiplicative difference
in sub-groups is achieved by a change in all coefficients, both intercept and slope
terms.
Since entire chapters on dummy variables are written in excellent academic

references detailing the interpretation of the results from using dummy variables, we
will only give brief interpretations of our results. Instead, we explore in detail the
generation of dummy variables using GAUSS.
Seasonality
Determining seasonal patterns in time series data is one application of dummy
variables. A new text data file named almon.txt will be used to study quarterly
seasonality. It has three variables. The first column is the date, in years and quarters
(YEARQT). The second column is capital expenditures in millions of dollars
(CEXP). The last column holds capital appropriations in millions of dollars (CAPP).
The basic Almon model describes the simple relationship between capital
expenditures and appropriations as follows:
CEXP = β0 + β1 CAPP + ε
There are 60 observations in total, although Almon’s original study used the first 36
observations from 1953 to 1961. Lesson 4.1 is devoted to the study of seasonal
differences with Almon’s quarterly time series on capital appropriations and
expenditures.
Does the use of dummy variables matter? Lesson 4.1 continues the hypothesis testing
procedure for significant differences in quarterly seasonality. It is achieved by
comparing regression results from restricted (without seasonal dummy variables) and
unrestricted (with seasonal dummy variables) least squares.
In Lesson 4.2, the notorious problem of the “dummy variable trap” is discussed with
an alternative use and interpretation of dummy variables in conjunction with the
regression without intercept.
Lesson 4.1: Seasonal Dummy Variables

In this lesson, we introduce the use of additive dummy variables to remove
seasonality in Almon’s quarterly time series data.
First of all, seasonality implies that the best-fitting regression for each season
(quarter) may be different. In other words, the intercept and slope terms that provide
the best fit for one quarter may not provide the best fit for different quarters. Before
generating the seasonal dummy variable matrix, you need to have some idea of what
it should look like. It has a repeating set pattern of four columns, one for each
quarter. Consider all 60 observations of time series data in almon.txt; a pattern of
0’s and 1’s is created to represent one cycle of seasonality (that is, one year). The
pattern is reshaped into a 4-column matrix with the desired 60 rows:
pattern = {1 0 0 0,
0 1 0 0,
0 0 1 0,
0 0 0 1};
D = reshape(pattern,60,4);
q1 = D[.,1];
q2 = D[.,2];
q3 = D[.,3];
To avoid perfect collinearity with the constant column associated with the intercept,
only three columns of the dummy variable matrix D will be used. That is, four
quarters are indicated with only three dummies: q1, q2, and q3. Lesson 4.2 on the
dummy variable trap explains why we must do this.
Now the model with three quarterly dummy variables is:
CEXP = β0 + β1CAPP + δ1 Q1 + δ2 Q2 + δ3 Q3 + ε
Or, in four equations (one for each quarter) as follows:
Quarter 1: CEXP = (β0 + δ1) + β1CAPP + ε

Quarter 4: CEXP = β0 + β1CAPP + ε
We also will address the significance of seasonal differences in the model by testing
when the three coefficients δ1, δ2, and δ3 are jointly equal to zero. This is a test
procedure for the presence of seasonality in the model we will examine later.
/*
** Lesson 4.1: Seasonal Dummy Variables
*/
1 use gpe2;
3 load almon[61,3] = gpe\almon.txt;
4 cexp = almon[2:61,2];
5 capp = almon[2:61,3];
64
DUMMY VARIABLES
6 qt = almon[2:61,1];
7 pattern = {1 0 0 0,
0 1 0 0,
0 0 1 0,
0 0 0 1};
8 D = reshape(pattern,60,4);
9 q1 = D[.,1]; @ quarterly seasonal dummies @
10 q2 = D[.,2];
11 q3 = D[.,3];
12 call reset;
13 _names = {"cexp", "capp", "q1", "q2", "q3"};
14 call estimate(cexp,capp~q1~q2~q3);
15 _restr = {0 1 0 0 0,
0 0 1 0 0,
0 0 0 1 0};
16 call estimate(cexp,capp~q1~q2~q3);
17 end;
The estimation is carried out with three quarter dummy variables named q1, q2, and
q3. The fourth quarter is the base case, and the coefficients of three dummy
variables identify the additive differences from that of the fourth quarter, or the
intercept term.
There are many ways to generate the dummy variables other than the suggested use
of the reshape command. A simple alternative is to rely on the quarter indicator
qt, appearing in the first column of the data file almon.txt. Lines 7 through 11 of
the above program can be replaced by the following three lines:
q1 = (qt%10) .== 1;
q2 = (qt%10) .== 2;
q3 = (qt%10) .== 3;
The modulo division “%” returns the remainder of the integer division, and the
notation “.==” in GAUSS performs element-by-element equality comparison. In
other words, each line compares the last digit of qt to a given quarter, placing a one
in the dummy variable if the comparison turns out to be true.
GAUSS has its own commands for creating dummy variables: dummy, dummybr,
dummydn. The command dummy creates a matrix of dummy variables by breaking
a vector of data into multiple groups. To make sense of this example, lines 7 and 8
of the above program may be replaced by the following:
seasons = {1,2,3,4};
D = dummy(qt%10, seasons);
where the column vector seasons containing four quarter indicators is used to
compare with the last digit of the variable qt. The GAUSS command dummy
creates a matrix of four columns of dummy variables, D. It compares each data
observation of qt%10 to the breakpoints designated in the vector seasons. If the
data are in the range designated, a one is placed in the corresponding element of
matrix D, if not, a zero is placed.
Running the program lesson4.1 will produce two sets of regression results in the
output file output4.1. The first estimated model looks like this:
CEXP = 670.93 + 0.737 CAPP - 13.69 Q1 - 50.60 Q2 - 31.53 Q3
65
s.e. (230.9) (0.053) (204.5) (204.3) (204.2)

t-ratio 2.91 13.95 - 0.067 - 0.248 - 0.154
We can also write the estimated model as four separate equations, one for each
quarter:
Quarter 1: CEXP = (670.93 - 13.69) + 0.737 CAPP

Quarter 4: CEXP = 670.93 + 0.737 CAPP
We have estimated the linear relationship between capital expenditures (CEXP) and
appropriations (CAPP) with varying intercept terms to represent the seasonal
differences in the model. Is there a real or significant difference among the four
estimated regression equations? Analyzing both the t-ratios and the P-values reveals
that the coefficients of dummy variables are not statistically significantly different
from zero. Furthermore, the partial regression values are very small for the dummy
variables. A more formal procedure is to test the hypothesis that all of the
coefficients of dummy variables are jointly equal to zero. The hypothesis is that δ1 =
0, δ2 = 0, and δ3 = 0 hold simultaneously. The GPE input control variable _restr
(line 15) defines Almon’s equation with the three quarterly dummy variables jointly
equaling zero.
_restr = {0 1 0 0 0,
0 0 1 0 0,
0 0 0 1 0};
Then, restricted least squares estimation is carried out in line 16. Here is the second
set of estimation results in which the coefficients of three quarterly dummy variables
are restricted to zero:

------------------------
Dependent Variable = CEXP

F( 3, 55) Prob>F
0.023130 0.99518
Wald Chi-Square Test 0.069391 3 0.99524
Lagrange Multiplier Test 0.075604 3 0.99459
Likelihood Ratio Test 0.075652 3 0.99459

Log Ammemiya Prediction Criterion (APC) = 12.681
Log Akaike Information Criterion (AIC) = 12.734
Log Schwarz Bayesian Information Criterion (BIC) = 12.908

Explained 6.1072E+007 1 6.1072E+007 205.70 9.9398E-021
Residual 1.7220E+007 58 2.9689E+005
66
DUMMY VARIABLES
Total 7.8291E+007 59 1.3270E+006

CAPP 0.73684 0.051375 14.342 9.9398E-021 0.78006
Q1 -6.1284E-014 0.00000 0.00000 0.00000 0.00000
Q2 2.6290E-013 0.00000 0.00000 0.00000 0.00000
Q3 -1.2079E-013 0.00000 0.00000 0.00000 0.00000
CONSTANT 646.48 184.48 3.5044 0.00088917 0.17474
By comparing regression results from restricted (without seasonal dummy variables)

and unrestricted (with seasonal dummy variables) least squares, the computed Wald
test statistic for the above three linear restrictions is a negligible 0.02, implying
insignificant seasonal variation in the model. It is further confirmed with an
extremely large P-value 0.99 for all the test statistics of linear restrictions. We can
conclude safely that there is no difference in the estimated Almon equations for the
different quarters.
Lesson 4.2: Dummy Variable Trap

Here comes a technical question. When we estimated Almon’s model in Lesson 4.1,
we only explicitly included three quarterly dummy variables in the regression
equation. Why would we drop the fourth dummy variable? If you keep the constant
term and use a dummy variable for each group, your program will generate an error
message similar to the following:
C:\GAUSS\SRC\GPE2.SRC(1383) : error G0121 : Matrix not positive definite

Currently active call: _lsqest [1383]
This condition is called the “dummy variable trap.” The dummy variable trap gets
just about everyone at some time. Understanding how the dummy variable trap
happens will make avoiding it easier. Remember that a typical regression equation
contains a constant vector of ones associated with the intercept coefficient. Now, if
there is a dummy variable for each group, summing all the dummy variables together
equals one. The problem of perfect collinearity exists! Dropping one dummy
variable is not the only solution to stay out of the “trap.” The alternative is to include
all dummy variables but to estimate the regression without the intercept term. In
GPE, regression estimation without intercept is carried out by setting the input
control variable:
_const = 0;
The following program is a slightly modified version of the previous program. It

includes all four quarterly dummy variables, but the model is estimated without the
intercept term.
/*
** Lesson 4.2: Dummy Variable Trap
*/
1 use gpe2;
4 cexp = almon[2:61,2];
5 capp = almon[2:61,3];
6 qt = almon[2:61,1];
7 pattern = {1 0 0 0,
67
0 1 0 0,
0 0 1 0,
0 0 0 1};
8 D = reshape(pattern,60,4);
9 call reset;
10 _const = 0; @ regression without intercept @
11 _names = {"cexp","capp","q1","q2","q3","q4"};
12 call estimate(cexp,capp~D);
13 end;
Run this program, and refer to the output file output4.2 for details. The important
catch is the statement of line 10:
_const = 0;
Without it, you will fall into the “dummy variable trap”! The estimated model can be
summarized as follows:
CEXP = 0.737 CAPP + 657.23 Q1 + 620.32 Q2 + 639.40 Q3 + 670.93 Q4

s.e. (0.053) (222.4) (222.7) (222.4) (230.9)
t-ratio 13.95 2.95 2.74 2.80 2.91
The interpretation of the coefficients associated with four quarter dummy variables is
directly reflected as the intercept values of each equation:

A careful eye will see that these results are the same as those of the first regression
equation in Lesson 4.1 using three dummies and a constant term.
Structural Change
In the next lesson, we will use a dummy variable approach to estimate and test for
structural change previously studied in the production function of Lesson 3.5. Recall
that a simple Cobb-Douglas production function was estimated using time series of
U.S. real output (X), labor (L) and capital (K) inputs obtained from the data file
cjx.txt. The question was, is there a change in both intercept and slope terms during
post-war expansion after 1948? In Lesson 3.5, a Chow test was formulated and
performed with two separate samples: 1929-1948 and 1949-1967. The alternative
approach is to use a dummy variable for sample separation, and check for the
difference in intercept and slope terms of the regression for each sub-sample. To
check for the intercept difference, the use of an additive dummy variable would
suffice. To check for the slope difference, a multiplicative dummy variable
associated with each explanatory variable should be used.
Lesson 4.3: Testing for Structural Change: Dummy Variable Approach

When breaking the time series data in cjx.txt into two sub-groups, only one dummy
variable named D is created (one less than the number of sub-groups). D is a vector
whose entries are equal to zero for all observations in the first time period (1929-
1948), and one for all observations in the second period (1949-1967). One way to
68
DUMMY VARIABLES
create D is to compare a given vector available in the original time series to a set
value. In this lesson we create a dummy variable D by comparing each observation
in the vector YEAR to the value 1948. For every observation greater than 1948, D is
set to one, otherwise D is set to zero. Notice that the dot (.) before the “>” means
element-by-element greater-than comparison:
D = year.>1948;
If the number of continuing observations designed for the base and alternative
situations are known, concatenating a vector of zeros vertically to a vector of ones is
a simple method of creating the dummy variable D. In this case,
D = zeros(20,1)|ones(19,1);
/*
** Lesson 4.3: Testing for Structural Change
** Dummy Variable Approach
*/
1 use gpe2;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 D = year.>1948;
9 DL = D.*L;
10 DK = D.*K;
11 call reset;
12 _names = {"X","L","K","DL","DK","D"};
13 call estimate(X,L~K~DL~DK~D);
14 _restr = {0 0 1 0 0 0,
0 0 0 1 0 0,
0 0 0 0 1 0};
15 call estimate(X,L~K~DL~DK~D);
16 end;
Line 8 creates the additive dummy variable named D. Lines 9 and 10 use D to set up
multiplicative dummy variables in association with the other two explanatory
variables L and K, respectively. Thus, for entries of D equal to one, the corresponding
entry in DL equals L and the corresponding entry in DK equals K. Otherwise, the
entries of DL and DK are zeros. The three dummy variables, one additive and two
multiplicative, are added to estimate in line 13. In this example, our model can be
written in two ways. It may be written with two separate regressions, one for years
before 1948, and one for the years after. This example demonstrates how to construct
both situations into one combined regression as follows:
X = β0 + β1 L + β2 K + δ0 D + δ1 DL + δ2 DK + ε
When D equals zero (that is, for the period 1929-1948), we have what is called the
base case. When D equals one (that is, 1949-1967), the estimated coefficients of the
dummy variables are added to the estimated coefficients of the independent variables
including the constant vector. In other words,
For 1929-1948, X = β0 + β1 L + β2 K + ε;
69
For 1949-1967, X = (β0 + δ0) + (β1 + δ1) L + (β2 + δ2)K + ε
Run the program so that we can check out the first estimated regression from the
output:

------------------------


Explained 8.0355 5 1.6071 1362.9 3.2814E-037
Residual 0.038911 33 0.0011791
Total 8.0744 38 0.21248

L 1.6167 0.15690 10.304 7.6101E-012 0.76288
K 0.21967 0.17266 1.2723 0.21217 0.046758
DL -0.60772 0.37578 -1.6172 0.11535 0.073433
DK 0.35942 0.21672 1.6584 0.10670 0.076935
D 1.5595 1.2876 1.2111 0.23444 0.042559
CONSTANT -4.0576 0.26822 -15.128 2.1113E-016 0.87397
Interpreting the output gives us the estimated model;
X = - 4.06 + 1.62 L + 0.22 K - 0.61 DL + 0.36 DK + 1.56 D

s.e. (0.27) (0.16) (0.17) (0.38) (0.22) (1.29)
t-ratio - 15.1 10.3 1.27 - 1.62 1.66 1.21
In terms of two separate samples:
For 1929 - 1948, X = -4.06 + 1.62 L + 0.22 K (base case)

For 1949 - 1967, X = (-4.06 + 1.56) + (1.62 -0.61) L + (0.22 + 0.36) K
or, X = -2.50 + 1.01 L + 0.58 K
One look at t-ratios and P-values tells us that the dummy variables are not
statistically significant. To test for the structural change, we need to verify that the
coefficients of both additive and multiplicative dummy variables are all zero. In
other words, we must show that δ0 = 0, δ1 = 0, and δ2 = 0 jointly. The GPE input
control variable _restr, defines these three dummy variables in a 3 by 6 matrix as
shown in line 14:
_restr = {0 0 1 0 0 0,
0 0 0 1 0 0,
0 0 0 0 1 0};
Line 15 estimates the restricted model (restricting all coefficients associated with
dummy variables to zeros) in which no structural change is assumed. Here is the
result of the restricted least squares estimation:
70
DUMMY VARIABLES

------------------------

F( 3, 33) Prob>F
1.2639 0.30275


Explained 8.0310 2 4.0155 3332.2 1.3921E-041
Residual 0.043382 36 0.0012051
Total 8.0744 38 0.21248

L 1.4508 0.083228 17.431 3.9260E-019 0.89407
K 0.38381 0.048018 7.9930 1.7130E-009 0.63960
DL -5.1890E-012 0.00000 0.00000 0.00000 0.00000
DK 2.3426E-012 0.00000 0.00000 0.00000 0.00000
D 1.6888E-011 1.7378E-007 9.7182E-005 0.99992 2.6234E-010
CONSTANT -3.9377 0.23700 -16.615 1.8332E-018 0.88464
Comparing the result of Chow test presented in Lesson 3.5 to the above output shows
an identical computed Wald F-test statistic of 1.27 for three linear restrictions on
dummy variables. In other words, based on the Cobb-Douglas log specification of
the production function, there is no reason to believe that there was a structural
change in output productivity between the years of 1929 and 1967. Both Lesson 3.5
(sample separation approach) and Lesson 4.3 (dummy variable approach) reach the
same conclusion. However, to a careful eye, there are subtle differences in the
estimated standard errors and t-ratios for the regression coefficients obtained from
these two approaches. Why?
71
V
Multicollinearity
Multicollinearity is a data problem due to a group of highly correlated explanatory
variables used in the regression equation. The consequence of multicollinearity is
large standard errors of the coefficient estimates. The size of these errors suggest that
there are too many explanatory variables and some of them may not be needed. Then
the question is how to identify and treat the irrelevant explanatory variables in the
regression.
The famous Longley data are known for the problem of multicollinearity. Instead of
constructing a meaningful model, we will demonstrate a hypothetical relationship
with the dependent variable (EM), regressed against a set of four other variables
(YEAR, PGNP, GNP, and AF).
Detecting Multicollinearity
Given the regression equation:
EM = β0 + β1 YEAR + β2 PGNP + β3 GNP + β4 AF + ε
the focus of this chapter is to examine how closely the four explanatory variables
(YEAR, PGNP, GNP, and AF) are related. Lessons 5.1, 5.2, and 5.3 address the
techniques of detecting multicollinearity. These include: condition number and
correlation matrix (Lesson 5.1), Theil’s measure of multicollinearity (Lesson 5.2),
and Variance Inflation Factors (Lesson 5.3).
Lesson 5.1: Condition Number and Correlation Matrix

We have seen the regression outputs from previous lessons, which include a column
of partial correlation coefficients. Technically, it is computed using the formula:
t2
2
t +DF
where t is the vector of t-ratios of coefficient estimates, and DF is the degrees of

freedom of the estimated regression. The partial correlation coefficient, as its name
suggests, measures the marginal or partial contribution of the designated variable
when the influence of other variables is already considered. The smaller the partial
correlation coefficient, the less important the variable would be in the regression
equation. It is a useful approach to identify the irrelevant variable as a candidate for
deletion.
Another useful tool to check for the problem of multicollinearity is the data
correlation matrix, which describes the simple pair-wise correlation among all the
variables used in the regression. The built-in GAUSS command corrx can do
exactly that, but the GPE package offers the convenience of a data correlation matrix
by setting the following input control variable:
_corr = 1;
The condition number of a normalized data matrix of explanatory variables is

computed as well. The normalization is necessary so that the condition number is
independent of the unit of measurement for each variable. A large condition number
indicates multicollinearity. Large values of pair-wise correlation coefficients may
hint at the source of the problem.
Using the Longley data, the following program estimates the model with the
dependent variable (EM) regressed against a set of four other variables (YEAR,
PGNP, GNP, and AF). The problem of multicollinearity is detected by examining the
partial regression coefficients, as well as the condition number and correlation matrix
of the explanatory variables.
/*
** Lesson 5.1: Condition Number and Correlation Matrix
*/
1 use gpe2;
4 data = data[2:17,.];
5 year = data[.,1];
6 pgnp = data[.,2];
7 gnp = data[.,3];
8 af = data[.,5];
9 em = data[.,7];
10 call reset;
11 _corr = 1; @ cond# and correlation matrix @
12 _names = {"em","year","pgnp","gnp","af"};
13 call estimate(em,year~pgnp~gnp~af);
14 end;
Running the above program returns the following output:

------------------------
Mean of Dependent Variable = 65317.


Explained 1.8011E+008 4 4.5028E+007 101.11 1.3458E-008
Residual 4.8987E+006 11 4.4534E+005
Total 1.8501E+008 15 1.2334E+007

YEAR -576.46 433.49 -1.3298 0.21049 0.13850
74
MULTICOLLINEARITY
PGNP -19.768 138.89 -0.14233 0.88940 0.0018381

GNP 0.064394 0.019952 3.2275 0.0080515 0.48638
AF -0.010145 0.30857 -0.032878 0.97436 9.8262E-005
CONSTANT 1.1691E+006 8.3590E+005 1.3986 0.18949 0.15098
Condition Number of Explanatory Variables = 15824.

Correlation Matrix of Dependent and Explanatory Variables
EM 1.0000
YEAR 0.97133 1.0000
PGNP 0.97090 0.99115 1.0000
GNP 0.98355 0.99527 0.99159 1.0000
AF 0.45731 0.41725 0.46474 0.44644 1.0000
EM YEAR PGNP GNP AF
With the exception of the variable GNP, small partial regression coefficients are
strong indications of irrelevant explanatory variables. The added information from
the use of the input control variable _corr = 1 (line 11) includes the condition
number and correlation matrix of the explanatory variables. The correlation
coefficients between the dependent variable and each independent variable are given
in the first column of the correlation matrix. These measure the individual effect of
each independent variable on the dependent variable. With the exception of the
variable AF, the explanatory variables have a rather high correlation with the
dependent variable. However, these variables are also highly correlated among
themselves, as seen from the rest of the correlation matrix. In addition, the condition
number of explanatory variables is extremely large, suggesting severe
multicollinearity for this set of variables.
Lesson 5.2: Theil’s Measure of Multicollinearity

Similar to the concept of partial regression coefficients, Theil’s measure of
multicollinearity uses the R-square from multiple partial regressions to determine if
multicollinearity is present in a regression equation.
Theil’s measure of multicollinearity is a formula derived from
R2 – ∑j=2,…,K (R2 – R-j2)
where R2 is the R-square (that is, coefficient of determination) of the full model,
including all explanatory variables. R-j2 is the R-square of the same regression model
excluding the j-th explanatory variable. Therefore, the difference R2 - R-j2 measures
the net contribution of the j-th explanatory variable in terms of R-square. K is the
number of explanatory variables of the full regression, in which the first one is the
constant term. Notice that the index j for summation does not count the constant
term. In the ideal case of no multicollinearity, Theil’s measure equals or is close to
zero.
The first regression in the following program (lines 10-13) estimates the full model
with dependent variable (EM) on a set of four independent variables (YEAR, PGNP,
GNP, and AF). The rest of the program (lines 14-33) estimates four regression
equations; each corresponds to the partial model with one of the independent
variables removed. The R-squares from the full model and from the four partial
models are then used to compute the Theil’s measure of multicollinearity.
Instead of showing the lengthy results of each regression estimation, we explain the
use of output control variables in GPE for keeping track of the information from
75
each of the regression runs. The use of an output control variable is first introduced
in line 13. In GPE, output control variables take on new values each time
estimate or forecast is called. An output control variable is identified with a
name beginning with a double underscore ( __ ). For example, __r2 is the value of
R-square computed in the previous estimation. Therefore, in line 13, assigning __r2
to a variable named r2 allows us to use that value later in the program. See
Appendix A for a complete list of output control variables available in GPE.
/*
** Lesson 5.2: Theil’s Measure of Multicollinearity
*/
1 use gpe2;
4 data = data[2:17,.];
5 year = data[.,1];
6 pgnp = data[.,2];
7 gnp = data[.,3];
8 af = data[.,5];
9 em = data[.,7];
10 call reset;
13 r2 =__r2;
14 call reset;
15 print"Partial Regression 1: EM = PGNP GNP AF";
16 _names = {"em","pgnp","gnp","af"};
17 call estimate(em,pgnp~gnp~af);
18 r2x1 = __r2;
19 print"Partial Regression 2: EM = YEAR GNP AF";
20 _names = {"em","year","gnp","af"};
21 call estimate(em,year~gnp~af);
22 r2x2 = __r2;
23 print"Partial Regression 3: EM = YEAR PGNP AF";
24 _names = {"em","year","pgnp","af"};
25 call estimate(em,year~pgnp~af);
26 r2x3 = __r2;
27 print"Partial Regression 4: EM = YEAR GNP PGNP";
28 _names = {"em","year","gnp","pgnp"};
29 call estimate(em,year~gnp~pgnp);
30 r2x4 = __r2;
31 print "Theil’s Measure of Multicollinearity =";;
32 print r2-sumc(r2-(r2x1|r2x2|r2x3|r2x4));
33 end;
From four partial regressions, we repeat the use of output variable __r2, to keep
track of the R-square of each regression. By renaming each __r2 and subtracting it
from the R-square of the full model, these net differences are concatenated and then
summed using a GAUSS command sumc (see line 32). Running the program, the
output displays the results of all the regressions before the line:
Theil’s Measure of Multicollinearity = 0.94414
In summary, the near unity of the Theil’s measure confirms the problem of
multicollinearity.
76
MULTICOLLINEARITY
Lesson 5.3: Variance Inflation Factors (VIF)

Relating to the correlation matrix of explanatory variables, Variance Inflation
Factors (VIF) indicate the ratio of a variable’s actual variance to the perfect variance
of zero collinearity. VIF is defined as:
1
1- Rj2
It can be used to detect multicollinearity, where Rj2 is the R-square from regressing
the j-th explanatory variable on all the other explanatory variables. A near unity Rj2
and hence a high value of VIF indicates a potential problem of multicollinearity with
the j-th variable.
The following program computes VIF for each explanatory variable through a set of
four auxiliary regressions similar to the procedure used in computing Theil’s
measure of multicollinearity.
/*
** Lesson 5.3: Variance Inflation Factors (VIF)
*/
1 use gpe2;
4 data = data[2:17,.];
5 year = data[.,1];
6 pgnp = data[.,2];
7 gnp = data[.,3];
8 af = data[.,5];
9 em = data[.,7];
10 call reset;
11 print "Aux Regression 1: YEAR = PGNP GNP AF";
12 y = year;
13 x = pgnp~gnp~af;
14 _names = {"year","pgnp","gnp","af"};
15 call estimate(y,x);
16 r2x1 = __r2;
17 print "Aux Regression 2: PGNP = YEAR GNP AF";
18 y = pgnp;
19 x = year~gnp~af;
20 _names = {"pgnp","year","gnp","af"};
22 r2x2 = __r2;
23 print "Aux Regression 3: GNP = YEAR PGNP AF";
24 y = gnp;
25 x = year~pgnp~af;
26 _names = {"gnp","year","pgnp","af"};
28 r2x3 = __r2;
29 print "Aux Regression 4: AF = YEAR GNP PGNP";
30 y = af;
31 x = year~gnp~pgnp;
32 _names = {"af","year","gnp","pgnp"};
34 r2x4 = __r2;
35 r2=r2x1|r2x2|r2x3|r2x4;
36 print "Variance Inflation Factors:";
37 print " Model R-Square VIF";;
38 print seqa(1,1,4)~r2~(1/(1-r2));
39 end;
77
The first part of the program performs four auxiliary regression estimations. Each
corresponds to the regression of one selected explanatory variable against the rest of
the others. Only the R-squares from the four estimated regressions are of interest in
computing the VIF. First, these values are retained using the output variable __r2,
then they are concatenated into a vector (line 35) for calculating the VIF of each
variable (line 38). Based on the R-square measure of each auxiliary regression, VIF
for each explanatory variable is reported as follows:
Variance Inflation Factors:

Model R-Square VIF
1.0000 0.99303 143.46
2.0000 0.98678 75.671
3.0000 0.99245 132.46
4.0000 0.35616 1.5532
Again, all explanatory variables except variable AF (Model 4) have higher than
normal values of VIF, indicating a severe problem of multicollinearity.
Correction for Multicollinearity

What to do with the problem of multicollinearity? Besides tweaking the appropriate
explanatory variables and data transformation, techniques such as ridge regression
and principal components are suggested in the literature. The ridge regression
approach seeks to find a set of “stable” coefficient estimates with a “shrinkage
parameter,” while the principal components approach is to extract a smaller number
of independent variables (principal components) that explain most of the variation of
regressors. The resulting coefficient estimates from these methods are biased and
difficult to interpret, even though they may be more precise (efficient) than their
ordinary least squares counterparts. Since multicollinearity is a data problem, data
adjustment and variable transformation should be considered in favor of mechanical
correction methods for estimation. Nevertheless, the following lesson illustrates the
correction mechanics of ridge regression and principal components.
Lesson 5.4: Ridge Regression and Principal Components

GPE’s estimate procedure does not offer either ridge regression or principal
components. Given the estimates obtained from ordinary least squares, however, it is
rather straightforward to implement the computation of ridge regression and
principal components. Many standard econometrics textbooks outline the formula.
For examples, see Judge, et al. (1988, Chap. 21) and Greene (1999, Chap. 6).
Given a shrinkage parameter of r>0, the coefficient estimates br of ridge regression

are related to the ordinary least squares estimate b of β in the regression equation Y
= Xβ + ε as follows:
br = (I + r(X'X)-1)-1b
Therefore, the corresponding estimated variance-covariance matrix Var(br) is:
Var(br) = (I + r(X'X)-1)-1Var(b) (I + r(X'X)-1)-1

= s2(I + r(X'X)-1)-1X'X (I + r(X'X)-1)-1
78
MULTICOLLINEARITY
Where Var(b)is the estimated variance-covariance matrix and s2 is the regression

variance of the ordinary least squares estimates. By varying the shrinkage parameter
r, we can find the most “stable” coefficient estimates.
The method of principal components is to extract sufficient variation of independent

variables to explain the dependent variable of a regression equation. Let X be the
data matrix of the explanatory variables, including the constant term. Principal
components of X are derived from linear combinations of characteristic vectors of
X'X. We will use only the principal components which satisfy a minimum size
requirement of the characteristic roots of X'X. Let V be the matrix of such
characteristic vectors. Then, the coefficient estimates of principal components bpc are
related to the least squares estimates b as follows
bpc = VV'b
and the corresponding estimated variance-covariance matrix Var(bpc) is:
Var(bpc) = (VV')Var(b)(VV')
lesson5.4 is a GAUSS program which implements the ridge regression and principal
components based on the hypothetical regression equation with the Longley data as
described in the previous lessons 5.1 to 5.3. After obtaining the ordinary least
squares result, we introduce several GAUSS commands to perform ridge regression
and principal components estimation. For detailed explanations of the GAUSS
commands used therein, refer to the GAUSS Command References or consult the
on-line help menu.
/*
** Lesson 5.4: Ridge Regression and Principal Components
*/
1 use gpe2;
4 data = data[2:17,.];
5 year = data[.,1];
6 pgnp = data[.,2];
7 gnp = data[.,3];
8 af = data[.,5];
9 em = data[.,7];
10 call reset;
/* explanatory variables including constant */

13 x = year~pgnp~gnp~af~ones(rows(year),1);
/* ridge regression */
14 r=0.3;
15 a = invpd(eye(cols(x))+r*invpd(x'x));
16 br = a*__b;
17 vbr = a*__vb*a';
18 print "Ridge Regression Model:";
19 print " Coefficient Std Error";;
20 print br~sqrt(diag(vbr));
/* Principal Components */
@ compute char. roots and vectors of X'X @
21 {r,v}=eigrs2(x'x);
22 v = selif(v',r.>0.1)';
23 bpc = v*v'__b;
24 vbpc = v*v'__vb*v*v';
79
25 print;
26 print "Principal Components Model:";
27 print " Coefficient Std Error";;
28 print bpc~sqrt(diag(vbpc));
29 end;
First we estimate the equation using ordinary least squares (lines 10-12), then from
line 13 on we focus on the data matrix of explanatory variables including the
constant term, to perform ridge regression and principal components. Ridge
regression is obtained for a shrinkage parameter of 0.3 (lines 14-20). We could try
several small positive values for a shrinkage parameter to find the most “stable”
coefficient estimates. The following output is the result of ridge regression:
Ridge Regression Model:

Coefficient Std Error
29.750 3.8485
-99.844 124.85
0.043571 0.013233
0.16452 0.28191
2.5061 1.7759
We continue on to perform the principal components method (lines 21-28). To

compute the principal components of the data matrix X of regressors, we pick only
the characteristic vectors which correspond to the characteristic roots of X'X with
values greater than 0.1 (line 22). We notice that the computation of characteristic
roots and vectors and therefore the result of principal components are sensitive to the
scale of measurement used in the data matrix.
The resulting estimates of principal components are:
Principal Components Model:

Coefficient Std Error
29.784 3.8893
-100.92 126.19
0.043684 0.013372
0.16499 0.28202
0.022451 0.010768
As we can see from the above example, the computation of ridge regression and
principal components is easy, but the interpretation of the resulting coefficient
estimates will be difficult.
80
VI
Nonlinear Optimization
To find an optimal (maximal or minimal) solution of a scalar-valued function is at
the core of econometric methodology. The technique of least squares estimation is an
example of solving the nonlinear “sum-of-squares” objective function. For a linear
regression model, the exact solution is derived using the analytical formula of matrix
algebra. However, the problem may be more complicated if the regression equation
is nonlinear in the parameters. In this case, approximation or iterative methods of
nonlinear optimization will be necessary. We will consider only the case of
unconstrained optimization. In most cases, simple equality constraints can be
substituted into the objective function so that the problem is essentially the
unconstrained one. Nonlinear optimization with inequality constraints is difficult,
though not impossible.
From elementary differential calculus, an optimal solution can be found by setting

the first derivatives (gradient) of the objective function with respect to the function’s
variables equal to zero and solving for the variables. Then to verify a minimum
(maximum), the second derivative’s matrix (hessian) must be positive (negative)
definite. When the functional form is simple, the analytical approach of working out
the derivatives is useful. The techniques of numerical approximation and iteration
are more practical for solving large and complex optimization problems. GPE offers
practical techniques for finding numeric solutions of a general nonlinear objective
function.
Although the nonlinear optimization of GPE is designed with statistical or

econometric problems in mind, it can be used for solving mathematical functions as
well. The first step is to define the objective function as mathematical or statistical.
The next step is to solve the function through numerical optimization by calling the
procedure estimate.
Solving Mathematical Functions

Without going into the details of optimization theory, we first describe how to define
and write an objective function suitable for solving its optimum (or optima). Recall
that a simple function can be defined with a single-line fn statement in GAUSS (see
Chapter II). Since GPE is designed for econometric or statistical problem-solving, an
objective function is defined with a set of sample data and a vector of unknown
parameters. Typically a one-line fn or multi-line proc statement is declared as:
fn FunctionName(Data,Parameters) = …;
or
proc FunctionName(Data,Parameters);
…
endp;
where FunctionName is the name of the function, Data are the sample
observations of data series, and Parameters are the parameters or coefficients of
the function. For a statistical model, both Data and Parameters are used to
define the function FunctionName. For a mathematical function, only the
Parameters matter, therefore Data can be set to 0 (or a dummy value) in this
case.
The actual computation of solving the function FunctionName is carried out by

calling the procedure estimate, as follows:
call estimate(&FunctionName,Data);
Here, &FunctionName denotes the code address (holding place) of the function
FunctionName we declared earlier, which itself is defined with Data (a set of
sample data) and Parameters (a vector of initial values of parameters).
Using GPE for nonlinear functional optimization (or estimation), the following input
control variables are required:
• _nlopt
• _b
• _iter
The GPE input control variable _nlopt defines the type of optimization problem
involved. _nlopt=0 indicates a minimization problem, while _nlopt=1 indicates
a maximization problem. Since numerical iteration is used for solving a nonlinear
model, the solution found can be at best a local one. The input variable _b provides
the initial guess of parameters as the starting point of iterations. Different starting
values of _b may lead to different (local) solutions. In an effort to find a global
solution for the function, several different values of _b should be tried. The variable
_iter sets the maximal number of iterations allowed for a particular problem.
Usually we keep _iter low for testing the function. When the function is debugged
and ready for solving, _iter should be set large enough to ensure the convergence
of an iterative solution.
Calling the procedure estimate for nonlinear model estimation (or optimization)
is similar to the case of the linear regression model. The differences are that under
nonlinear estimation or optimization, the first argument of estimate is now an
address for the objective function and the second argument (for the data matrix) is
more forgiving in its structure. Remember that the objective function must be
defined with both data and parameters before calling the estimate procedure.
Lesson 6.1: One-Variable Scalar-Valued Function

Consider a scalar-valued function of one variable,
f(x) = ln(x) – x2
The single maximum of f(x) is found at x = ½ . First we translate the mathematical

function f(x) into a GAUSS fn statement as in line 3 of the following program:
82
NONLINEAR OPTIMIZATION
/*
** Lesson 6.1: One-Variable Scalar-Valued Function
** f(x) = ln(x) – x^2
*/
1 use gpe2;
2 output file=output6.1 reset;
3 fn f(data,x)=ln(x)-x^2;
4 call reset;
5 _nlopt=1;
6 _iter=100;
7 _b=0.5;
8 call estimate(&f,0);
9 end;
Line 5 indicates the maximization problem involved, and line 6 sets the iteration
limit for finding the solution. The estimation (maximization, in particular) of
function f starts with the initial value of x at 0.5 as shown in line 7. The GPE input
variable _b controls the starting value of iteration. Notice that here we do not use
sample data or parameter names in defining the function and its maximization.
Running the above lesson program, we obtain the following result:
Non-Linear Optimization: Maximization Problem

---------------------------------------------
Number of Parameters = 1
Maximum Number of Iterations = 100

Step Size Search Method = 0
Convergence Criterion = 0
Tolerance = 0.001
Initial Result:
Function Value = -0.94315
Parameters = 0.50000
Using Steepest-Ascent Algorithm

Iteration = 1 Step Size = 0.2500 Value = -0.85018
Final Result:
Iterations = 6 Evaluations = 38
Gradient Vector = -4.2549e-006
Hessian Matrix = -4.0000
Starting at x = 0.5 with function value – 0.94315, it takes six iterations to reach the
convergence of a solution. The solution 0.70711 is indeed a maximum with function
value – 0.84657, where the gradient is almost zero at – 4.2549e-06 and the hessian is
negative at – 4.0.
83
You may want to define the function’s analytical derivatives and use them for
solving the function. For this example, they are:
fn f1(data,x) = 1/x – 2*x;
fn f2(data,x) = -1/(x^2) – 2;
The functions f1 and f2 are the first and second derivatives of f, respectively. It
may be necessary to write a multi-line procedure for the derivatives of a more
complicated function. To solve the function with analytical derivatives, just
concatenate the first and second derivatives together and set it to the input control
variable _derviv before calling the procedure estimate as below:
_deriv = &f1|&f2;
call estimate(&f,0);
There is no need to use both first and second derivatives. Using only the first
derivative will work. That is,
_deriv = &f1;
The use of analytical derivatives will speed up the computation and increase the
numerical precision of the solution. However, for a complicated function, it is often
a difficult task to write and code the analytical formulas of derivatives.
The bare-bones program of Lesson 6.1 does not take advantage of the many options
available in GPE to fine tune the optimization process. For a simple problem, as the
one shown above, the default settings of the optimization method (i.e., steepest-
ascent method) and convergence criteria (i.e., convergence in function value and
solution relative to the tolerance level of 0.001) may be acceptable.
We now explain some of the GPE input control variables, which provide the option
to select one of many optimization methods and control its behavior in order to find
the optimal solution for a more complicated and difficult function. These control
variables are:
• _method
• _step
• _conv
• _tol
• _restart
By default, _method is set to 0 if we do not specify any method of optimization in

the program. For a mathematical problem, the default method is the steepest descent
or ascent method. For a statistical model, it is the Gauss-Newton method for
nonlinear least squares, or the steepest descent method for maximum likelihood
(ML) estimation. The following lists the more sophisticated optimization methods
available :
_method=1 Broyden-Fletcher-Goldfarb-Shanno (BFGS) quasi-Newton method.

_method=2 Davidon-Fletcher-Powell (DFP) quasi-Newton method.
_method=3 Greenstadt method.
_method=4 Newton-Raphson method.
_method=5 Quadratic hill-climbing (QHC) method.
_method=6 Modified quadratic hill-climbing method.
84
All the optimization or estimation methods should be combined with a line search to
determine the step size of optimization for each iteration. The default line search
method is a simple cutback method (_step=0). Setting _step=1 causes the
quadratic step size to be used in the search. Readers interested in a more detailed
discussion and comparison of different optimization methods should check the
references (e.g., Quandt, 1983; Judge, et al., 1985, Appendix B; Greene, 1999,
Chapter 5) for details.
The other optional input variables control the accuracy and convergence of the
solution. The variable _tol sets the tolerance level of convergence. Typically _tol
is a small number (default value 0.001). The variable _conv checks for two
consecutive iterations to reach convergence, relative to the tolerance level. When
_conv=0 (default), only the function values and solutions are checked for
convergence with _tol; when _conv=1, the convergence of function values,
solutions, and zero gradients are checked with _tol. Finally, the variable
_restart sets the number of times to restart the computation when the function
value fails to improve. A maximum of 10 restarts is allowed, with no restart as the
default (_restart=0).
As will be demonstrated in many example lessons below, we use all sorts of different
optimization methods or algorithms for different types of problems. It is not unusual
that a different (local) solution may be found due to the particular algorithm in use.
Although there is no clear indication which method should be used for what type of
problem, we recommend a mixed bag of optimization methods in conjunction with a
variety of options controlling the numerical optimization. It is a matter of
experimentation to find the best suite of solution tools for a particular problem.
Lesson 6.2: Two-Variable Scalar-Valued Function

This example demonstrates the use of GPE input control variables. We now consider
a two-variable scalar-valued function:
g(x) = g(x1,x2) = (x12 + x2 – 11)2 + (x1 + x22 –7)2.
There are four minima, (3,2), (3.5844, -1.8481), (-3.7793, -3.2832), and
(-2.8051, 3.1313) with the same function value 0, although we can only find one
minimum at a time. With various initial starting values of the variables, we are able
to find all of the four solutions. Also, the maximal function value 181.62 is found at
the solution (-0.27084, -0.92304). Be warned that sometimes the solutions are
difficult to find because there are several saddle points, (0.08668, 2.88430),
(3.38520, 0.07358), and (-3.07300, -0.08135), in the way. Here is the program:
/*
** Lesson 6.2: Two-Variable Scalar-Valued Function
** g(x) = (x[1]^2+x[2]-11)^2 + (x[1]+x[2]^2-7)^2
*/
1 use gpe2;
3 fn g(data,x)=(x[1]^2+x[2]-11)^2+(x[1]+x[2]^2-7)^2;
4 call reset;
5 _nlopt=0;
6 _method=1;
85
7 _iter=100;
8 _step=1;
9 _conv=1;
10 _b={3,-2};
11 call estimate(&g,0);
12 end;
Line 3 defines the one-line objective function g. Again, data is not used for
defining such a function, in which only the vector of parameters x matters. In this
example, a version of the quasi-Newton method (BFGS, i.e., _method=1) is used
(line 6) for optimization. It takes seven iterations to find one of the four minima
(3.58, -1.85) from the initial starting point _b=(3, -2) given in line 10. Run this
program, and refer to the output file output6.2 for more details.
For pedagogical purposes, we write out the procedures for analytical first and second
derivatives g1 and g2, although we do not use them in the above lesson. We note
that g1 is a row-vector gradient and g2 is a hessian matrix:
proc g1(data,x); @ 1st derivative of g(x) @
local f1,f2;
f1=4*x[1]*(x[1]^2+x[2]-11)+2*(x[1]+x[2]^2-7);
f2=2*(x[1]^2+x[2]-11)+4*x[2]*(x[1]+x[2]^2-7);
retp(f1~f2);
endp;
proc g2(data,x); @ 2nd derivative of g(x) @

local f11,f22,f12;
f11=12*x[1]^2+4*x[2]-42;
f22=4*x[1]+12*x[2]^2-26;
f12=4*(x[1]+x[2]);
retp((f11~f12)|(f12~f22));
endp;
If both derivatives g1 and g2 were used in the optimization, we need only to set
_deriv = &g1|&g2;
before calling the procedure estimate(&g,0).
By changing the initial values of the parameters in line 10 of lesson6.2, all solutions
may be found. We suggest the following values and the corresponding minima to
which they converge. Try them out:
Initial Value Function

(Line 10) Minimum Value
(3, -2) (3.58, -1.85) 0
(-3, 2) (-2.81, 3.13) 0
(-3, -2) (-3.78, -3.28) 0
(2, 2) (3, 2) 0
Unfortunately, without knowing the solution ahead of time, the search is rather a
blind process. The general rule of thumb is to try as many different initial values as
possible. As an exercise, modify the program of Lesson 6.2 to find the maximum
(-0.27, -0.92) with function value 181.62. Hint: Try _nlopt=1 (line 5) and
_b={0,0} (line 10).
86
Estimating Probability Distributions

The main use of nonlinear optimization in GPE is statistical model estimation, in
which the underlying probability distribution of the random variables is estimated.
The characteristics of a random variable (e.g., mean and variance, etc.) may be
evaluated through the joint probability density of a finite sample. This joint density
function, or the likelihood function, is defined as the product of N independent
density functions f(Xi,θ) of sample observations Xi (i=1,2,…,N) and an unknown
parameter vector θ. That is, ∏i=1,2,…,N f(Xi,θ), or equivalently in log form:
ll(θ) = ∑i=1,2,…,N ln f(Xi,θ)
The problem is to maximize the log-likelihood function ll(θ) so that the solution θ
characterizes the probability distribution of the random variable X under
consideration. To find the θ that maximizes ll(θ) is the essence of maximum
likelihood estimation. The corresponding variance-covariance matrix of θ is derived
from the information matrix (negatives of the expected values of the second
derivatives) of the log-likelihood function as follows:
∂2ll   -1
Var(θ) = -E
 ∂θ∂θ' 
The familiar example is the likelihood function derived from a normal probability
distribution:
1  (X − µ )2 
f ( X, θ ) = exp 2 
2πσ2  2σ 
where θ = (µ,σ2) represents the distribution parameters. It is straightforward to show

1
that the maximum likelihood solution is µ = E(X) = N ∑i=1,…,N Xi (the sample mean),
1
and σ2 = Var(X) = N ∑i=1,…,N (Xi- µ)2 (the sample variance).
Another example is based the log-normal distribution of X (or equivalently, normal

distribution of ln(X)) defined as:
1 1  (ln(X ) − µ )2 
f ( X, θ ) = exp  
2πσ2 X  2σ 2 
1 1
with the solution µ = N ∑i=1,…,N ln(Xi) and σ2 = N ∑i=1,…,N (ln(Xi)- µ)2, the
corresponding mean and variance of X are E(X) = exp(µ+σ2/2) and Var(X) =
exp(2µ+σ2) [exp(σ2)-1], respectively. Many economic variables are described with a
log-normal instead of a normal probability distribution. If µ is re-parameterized in
terms of a set of non-random variables Z and additional parameters β, µ = Zβ for
example, we get the statistical regression model, to be discussed in the next section.
87
Of course, maximum likelihood estimation is not limited to models with normal or

log-normal probability distribution. In many situations, the probability distribution of
a random variable may be non-normal. For example, to estimate the gamma
distribution of a nonnegative random variable X ≥ 0, the distribution function is
λρ
f (X, θ) = exp −λX X ρ−1
Γ(ρ)
where θ = (λ, ρ) is the parameter vector with λ > 0 and ρ > 0. The mean of X is ρ/λ,
and the variance is ρ/λ2. Many familiar distributions, such as the exponential and
Chi-square distributions, are special cases of the gamma distribution.
As with the normal distribution, the technique of maximum likelihood can be used to
estimate the parameters of the gamma distribution. Sampling from N independent
observations from the gamma distribution, the log-likelihood function is:
ll(θ) = N [ρ ln(λ) – lnΓ(ρ)] – λ ∑i=1,2,…,N Xi + (ρ-1) ∑i=1,2,…,N ln(Xi)
With the normal, log-normal, and gamma probability distributions, the

characteristics of the random variable X may be described in terms of the estimated
mean and variance for each probability distribution as follows:
Normal Log-Normal Gamma

Distribution Distribution Distribution
Mean µ exp(µ+σ2/2) ρ/λ
E(X)
Variance σ2 exp(2µ+σ2)[exp(σ2)-1] ρ/λ2
Var(X)
1 1
Where: µ =N ∑i=1,…,N Xi µ =N ∑i=1,…,N ln(Xi)
1 1
σ2=N ∑i=1,…,N (Xi- µ)2 σ2=N ∑i=1,…,N (ln(Xi)- µ)2
Lesson 6.3: Estimating Probability Distributions

In the following we use the hypothetical income data series of Greene (1999,
Chapter 4, Table 4.1), and estimate its mean and variance under the assumption of
three probability distributions. The data are replicated in the text file yed20.txt. In
lesson6.3 below, these 20 observations of two variables INCOME and
EDUCATION are loaded first. Only the variable INCOME scaled by a factor 10
(here it is called x) will be analyzed. Data scaling is useful for nonlinear model
estimation. We estimate the parameters of three probability distributions (normal,
log-normal, and gamma) by maximizing the corresponding log-likelihood function.
/*
** Lesson 6.3: Estimating Probability Distributions
** See Greene (1999), Chapter 4
*/
1 use gpe2;
3 load data[21,2]=gpe\yed20.txt;
4 x=data[2:21,1]/10; @ income data: scaling may be helpful @
88
/* normal probability distribution: b[1]=mu, b[2]=sigma */

5 fn llfn(x,b)=sumc(ln(pdfn((x-b[1])/b[2])./b[2]));
6 fn llfln(x,b)=sumc(ln(pdfn((ln(x)-b[1])/b[2])./(b[2].*x)));
/* gamma probability distribution: b[1]=rho, b[2]=lambda */

7 fn pdfg(x,b)=((b[2]^b[1])./gamma(b[1])).*exp(-b[2]*x).*x^(b[1]-1);
8 fn llfg(x,b)=sumc(ln(pdfg(x,b)));
9 call reset;
10 _nlopt=1;
11 _method=4;
12 _iter=100;
13 _b={3.0,2.0};
14 call estimate(&llfn,x);
15 _b={1.0,0.5};
16 call estimate(&llfln,x);
17 _b={2.0,0.5};
18 call estimate(&llfg,x);
19 end;
By definition, the log-likelihood function is just the summation of the logarithmic

probability distribution function over the sample observations. Based on a normal
probability distribution, line 5 defines the corresponding log-likelihood function in
which the unknown parameters are µ and σ. Similarly, line 6 is the log-likelihood
function of the underlying log-normal distribution. We note that the GAUSS built-in
normal probability density function, pdfn, is used to compute the log-likelihood
functions of normal and log-normal distributions. For the case of the gamma
distribution, the definition of probability distribution function is given in line 7,
which uses the built-in gamma function of GAUSS.
For all cases of maximum likelihood estimation of probability distributions, we use

the Newton-Raphson optimization method (_method=4), for up to 100 iterations
(_iter=100) as shown in lines 10 through 12. To estimate the parameters µ and σ
of the underlying normal distribution, we maximize the corresponding log-likelihood
function as carried out in line 14, with initial values 3.0 and 2.0 for µ and σ (line 13).
The estimation result below shows that final solution of (µ, σ) is obtained at (3.1278,
2.1809) with the log-likelihood function value – 43.974.

---------------------------------------------
Assuming Maximum Likelihood Function

Tolerance = 0.001
Initial Result:
Parameters = 3.0000 2.0000
Using Newton-Raphson Algorithm

Parameters = 3.1323 2.1747
Parameters = 3.1278 2.1809
Parameters = 3.1278 2.1809
89
Final Result:
Parameters = 3.1278 2.1809
Gradient Vector = -0.00010836 0.00040791
Asymptotic Asymptotic
Parameter Std. Error t-Ratio
X1 3.1278 0.48766 6.4139
X2 2.1809 0.34483 6.3247
For the case of log-normal distribution, starting from the initial values of (µ, σ) at
(1.0, 0.5) in line 15, the maximum likelihood solution is found at (0.9188, 0.6735) as
shown in line 16. The maximal value of log-likelihood function is –38.849. Here is
the output:

---------------------------------------------

Tolerance = 0.001
Initial Result:
Parameters = 1.0000 0.50000

Parameters = 0.89823 0.67385
Parameters = 0.91884 0.67317
Parameters = 0.91880 0.67349
Final Result:
Parameters = 0.91880 0.67349
Gradient Vector = -0.0018227 0.027819
X1 0.91880 0.15053 6.1039
X2 0.67349 0.10637 6.3317
For the gamma distribution, the estimation of parameters λ and ρ is implemented in

line 18 starting with the initial values 0.5 and 2.0 for λ and ρ, respectively (line 17).
The maximum likelihood estimator of (λ, ρ) is obtained at (0.7707, 2.4106) where
the log-likelihood function value is –39.324. The output looks like this:

---------------------------------------------

90
Tolerance = 0.001
Initial Result:
Parameters = 2.0000 0.50000

Parameters = 2.2115 0.71252
Parameters = 2.4112 0.77079
Parameters = 2.4106 0.77070
Final Result:
Parameters = 2.4106 0.77070
Gradient Vector = -0.0036550 0.0078771
X1 2.4106 0.71610 3.3663
X2 0.77070 0.25442 3.0293
To summarize the statistical characteristics of the random variable INCOME

(divided by 10) under consideration, we compute the estimated mean and variance of
the maximum likelihood estimates of parameters from the normal, log-normal, and
gamma probability distributions. We also compare the maximal values of the log-
likelihood functions associated with these probability distributions. It is interesting to
note that the variable INCOME is more likely drawn from log-normal or gamma
distributions, as their log-likelihood function values are greater than that of the
normal distribution.
Normal Log-Normal Gamma

Distribution Distribution Distribution
Mean 3.1278 3.1443 3.1278
Variance 4.7563 5.6745 4.0584
Log-likelihood -43.974 -38.849 -39.324
Lesson 6.4: Mixture of Probability Distributions

It is possible that a random variable is drawn from a mixture of probability
distributions (two or more, same or different types of distributions). For simple
exploration, consider X distributed with a mixture of two normal distributions:
1  (X − µ 1 ) 2 
f 1 ( X, µ 1 , σ 1 ) = exp 2 ,
2πσ 1
2
 2σ 1 
1  (X − µ 2 )2 
f 2 ( X, µ 2 , σ 2 ) = exp 2 .
2πσ 2
2
 2σ 2 
Then the likelihood function is
f(X,θ) = λ f1(X,µ1,σ1) + (1-λ) f2(X,µ2,σ2)
91
where λ is the probability that an observation is drawn from the first distribution
f1(X,µ1,σ1), and 1-λ is the probability of that drawn from the second distribution. θ =
(µ1,µ2,σ1,σ2,λ) is the unknown parameter vector that must be estimated.
Continuing from the previous example, suppose each observation of the variable
INCOME is drawn from one of two different normal distributions. There are five
parameters, the first two are the mean and standard error of the first normal
distribution, while the second pair of parameters corresponds to the second
distribution. The last parameter is the probability that the data are drawn from the
first distribution. Lines 12 to 17 of lesson6.4 below define the log-likelihood
function for the mixture of two normal distributions.
/*
** Lesson 6.4: Mixture of Two Normal Distributions
** See Greene (1999), Chapter 4
*/
1 use gpe2;
4 x=data[2:21,1]/10; @ income data: scaling may help @
5 call reset;
6 _nlopt=1;
7 _method=5;
8 _iter=100;
9 _b={3,3,2,2,0.5};
10 call estimate(&llf,x);
11 end;
/*
mixture of two normal distributions
mu1=b[1], mu2=b[2]
se1=b[3], se2=b[4]
prob.(drawn from the 1st distribution)=b[5]
*/
12 proc llf(x,b);
13 local pdf1,pdf2;
14 pdf1=pdfn((x-b[1])/b[3])/b[3];
15 pdf2=pdfn((x-b[2])/b[4])/b[4];
16 retp(sumc(ln(b[5]*pdf1+(1-b[5])*pdf2)));
17 endp;
The problem is to maximize the log-likelihood function (_nlopt=1 in line 6) using

the QHC algorithm (_method=5 in line 7) with the initial estimates of the
parameters given in line 9. After 11 iterations, we obtain the following result (to save
space, we print only the first and last iterations):

---------------------------------------------

Tolerance = 0.001
Initial Result:
92
Parameters = 3.0000 3.0000 2.0000 2.0000 0.50000
Using Quadratic Hill-Climbing Algorithm

Parameters = 3.0542 3.0542 2.1325 2.1325 0.50000
…
Parameters = 2.0495 5.7942 0.81222 2.2139 0.71204
Final Result:
Parameters = 2.0495 5.7942 0.81222 2.2139 0.71204
Gradient Vector = 1.1441e-005 0.00000 2.1870e-005 5.1351e-006 -
4.7899e-005
X1 2.0495 0.24421 8.3922
X2 5.7942 1.4988 3.8659
X3 0.81222 0.19424 4.1816
X4 2.2139 0.90617 2.4431
X5 0.71204 0.14456 4.9254
With the maximum log-likelihood function value of –38.309, the variable INCOME
when drawn from the mixtures of two different normal probability distributions is as
convincing as when that variable is drawn from a single non-normal distribution
(log-normal or gamma) as demonstrated in Lesson 6.3.
Statistical Regression Models

Estimating probability distributions of a random variable is interesting, but
econometric modeling focuses on statistical causal relationships within a group of
variables. The GPE package is designed specifically for statistical model estimation.
If the sample data, in addition to the parameters, are used to define the scalar-valued
objective function, estimate assumes nonlinear least squares for a minimization
problem and maximum likelihood for a maximization problem. In addition to
reporting the optimization process and outcome, the estimated results are interpreted
according to these two types of statistical regression models. If your problem is
neither a least squares nor a maximum likelihood, your own scrutiny into the
classical statistical interpretation of the model is necessary.
For a statistical model, the estimated variance-covariance matrix of the parameters

may be requested by setting a positive value to the input control variable _vcov.
Typically, the estimated variance-covariance matrix is derived from the
approximated hessian (_vcov=1). A recalculated exact hessian matrix is used
instead if we set _vcov=2. Nevertheless, the estimated variance-covariance matrix
may only be meaningful in the contexts of nonlinear least squares and maximum
likelihood problems.
As with the linear regression model, values of several output control variables are
available after nonlinear least squares or maximum likelihood estimation:
• __b Estimated parameters or solution
• __vb Estimated variance of the parameters
93
• __e Estimated regression residuals
• __rss Sum-of-squares function value at minimum
• __ll Log-likelihood function value at maximum
• __g Gradient vector of objective function at optimum
• __h Hessian matrix of objective function at optimum
Appendix A, GPE Control Variables, lists and explains the usage of these input and
output control variables.
Lesson 6.5: Minimizing Sum-of-Squares Function

The following example is taken from Judge, et al., (1988, Chapter 12, p.512).
Consider a CES production function of a single output (Q) using two factors, labor
(L) and capital (K), as follows:
ln(Q) = β1 + β4 ln (β2Lβ3 + (1-β2)K β3) + ε
where ε is the error term and β’s are the unknown parameters. The data matrix X =
(L, K, Q) is available in the text file judge.txt. The method of least squares
estimation is to find the vector β = (β1,β2,β3,β4) so that the sum-of-squared errors
S(β) = ε'ε is minimized.
/*
** Lesson 6.5: Minimizing Sum-of-Squares Function
** Estimating a CES Production Function
** See Judge, et al. (1988), Chapter 12
*/
1 use gpe2;
3 load x[30,3]=gpe\judge.txt;
4 call reset;
5 _nlopt=0;
6 _method=5;
7 _iter=100;
8 _tol=1.0e-5;
9 _vcov=1;
10 _b={1.0,0.5,-1.0,-1.0};
11 call estimate(&cessse,x);
12 end;
/* Objective Function */
13 proc cessse(data,b); @ sum-of-squares function @
14 local l,k,q,e;
15 l=data[.,1];
16 k=data[.,2];
17 q=data[.,3];
18 e=ln(q)-b[1]-b[4]*ln(b[2]*l^b[3]+(1-b[2])*k^b[3]);
19 retp(sumc(e^2));
20 endp;
94
The objective function cessse, sum-of-squared errors, is defined as a procedure in

lines 13 through 20. It is evident that both data matrix data and parameter vector b
are used to specify the functional form. The address of function &cessse is used in
calling estimate in line 11, where x is the data matrix (see also line 3). Line 5
indicates that the problem is to minimize the objective function cessse, and line 6
requests the quadratic hill-climbing (QHC, i.e., _method=5) optimization method.
In line 8, the convergence of function value and solutions are checked relative to a
smaller tolerance level of 0.00001. The vector of initial parameter values _b is given
in line 10, which is rather a random guess. Finally, at the end of optimization, the
variance-covariance matrix of estimated parameters will be computed because
_vcov=1, as shown in line 9.
We keep the definition of objective function cessse outside (beyond the end
statement) of the main program. There is no strict rule dictating where to place the
functions you define. Putting the function or procedure outside of the main program
makes the function accessible to other procedures you write for other purposes.
The final solution is found after 36 iterations. To save space, we report only the final
result of the iterations. The output file output6.5 contains the details of all the
iterations for reference.
In your program, setting _print=0 will suppress the printing of iteration outputs
to the file and the screen.
Non-Linear Optimization: Minimization Problem

---------------------------------------------
Assuming Nonlinear Least Squares Function

Tolerance = 1e-005
Initial Result:
Function Value = 37.097
Parameters = 1.0000 0.50000 -1.0000 -1.0000
Final Result:
Parameters = 0.12449 0.33668 -3.0109 -0.33631
Gradient Vector = 2.6755e-006 4.6166e-007 2.5664e-006 1.7166e-006
Hessian Matrix =
60.000 -5.7563 35.531 295.65
-5.7563 19.377 -3.4569 -23.595
35.531 -3.4569 35.461 298.10
295.65 -23.595 298.10 2509.4
Asymptotic
X1 0.12449 0.074644 1.6678
X2 0.33668 0.10809 3.1147
X3 -3.0109 2.2904 -1.3145
X4 -0.33631 0.26823 -1.2538
95
Asymptotic Variance-Covariance Matrix

X1 0.0055717
X2 -0.0013729 0.011684
X3 -0.065566 0.16893 5.2462
X4 0.0071194 -0.019797 -0.61389 0.071948
X1 X2 X3 X4
Both the gradient and hessian of the solution confirm that the solution indeed
minimizes the sum-of-squares objective function at the value 1.761. The estimated
model is presented as follows:
ln(Q) = 0.125 - 0.336 ln (0.337 L –3.01 + 0.663 K –3.01)

s.e. (0.075) (0.268) (0.108) (2.29)
Lesson 6.6: Maximizing Log-Likelihood Function

The same problem can be estimated by maximizing the likelihood objective function.
Assuming the model error ε follows a normal probability distribution with zero mean
and constant variance σ2, the log output ln(Q) is normally distributed with the
following distribution function:
 1 
N
 ε (X, β )′ ε (X, β )
  exp 
 2   2σ 2 
 2πσ 
where N is the sample size, and ε(X, β) = ln(Q) - β1 - β4 ln (β2L β3 + (1-β2)K β3). The
corresponding log-likelihood function of the unknown parameters θ = (β,σ) is
written as
 ε (X, β )   ε (X, β ) 
'
( )
ll (θ ) = − N 2 ln(2π ) − N 2 ln σ 2 − 1 2 
 σ   σ 
The program below follows the same basic structure as in the previous lesson. The
relevant modifications of lesson6.6 include changing the objective function in the
call estimate to cesll (line 11) and setting the variable _nlopt=1 (line 5).
The objective log-likelihood function cesll is defined from lines 13 to 21. In
addition to β, the standard error of the model σ must be estimated simultaneously.
Line 10 sets out the initial values of θ = (β,σ).
/*
** Lesson 6.6: Maximizing Log-Likelihood Function
** Estimating a CES Production Function
** See Judge, et al. (1988), Chapter 12
*/
1 use gpe2;
4 call reset;
5 _nlopt=1;
6 _method=5;
7 _iter=100;
8 _tol=1.0e-5;
9 _vcov=1;
10 _b={1.0,0.5,-1.0,-1.0,1.0};
11 call estimate(&cesll,x);
96
12 end;
/* Objective Function */
13 proc cesll(data,b); @ log-likelihood function @
14 local l,k,q,e,n;
15 l=data[.,1];
16 k=data[.,2];
17 q=data[.,3];
18 e=ln(q)-b[1]-b[4]*ln(b[2]*l^b[3]+(1-b[2])*k^b[3]);
19 n=rows(e);
20 retp(-0.5*n*(ln(2*pi)+ln(b[5]^2))-0.5*sumc((e./b[5])^2));
21 endp;
Solving the maximization problem of a classical log-likelihood function, the

standard error parameter σ is shown to be dependent on β. That is, if β is known, σ2
is solved as:
σ2(β) = ε(X,β)'ε(X,β)/Ν
Therefore, the same maximum log-likelihood may be expressed with the

concentrated log-likelihood function instead:
ll*(β) = -N/2 (1+ ln(2π) – ln(N)) –N/2 ln(ε(X,β)'ε(X,β))
The advantage of using the concentrated log-likelihood function is that there is one
less parameter (that is, σ) to estimate directly.
Running the program lesson6.6, we obtain the following result (again, the details of
interim iterations can be found in the output file output6.6):

---------------------------------------------

Tolerance = 1e-005
Initial Result:
Parameters = 1.0000 0.50000 -1.0000 -1.0000 1.0000

Final Result:
Parameters = 0.12449 0.33667 -3.0109 -0.33630 0.24229
Gradient Vector = 8.5614e-006 1.0552e-006 -8.2596e-007 7.3948e-006 -
3.6658e-006
Hessian Matrix =
-511.05 49.030 -302.63 -2518.3 0.063432
49.030 -165.04 29.444 200.98 0.020503
-302.63 29.444 -302.02 -2539.1 -0.037482
-2518.3 200.98 -2539.1 -21375. -2.6556
0.063432 0.020503 -0.037482 -2.6556 -1021.5
Asymptotic
97
X1 0.12449 0.074646 1.6677

X2 0.33667 0.10811 3.1142
X3 -3.0109 2.2911 -1.3142
X4 -0.33630 0.26830 -1.2535
X5 0.24229 0.031295 7.7421

X1 0.0055721
X2 -0.0013740 0.011687
X3 -0.065602 0.16904 5.2493
X4 0.0071233 -0.019807 -0.61423 0.071984
X5 -1.5793E-005 4.5442E-005 0.0014036 -0.00016456 0.00097936
X1 X2 X3 X4 X5
It is no surprise that the solution is identical to the one obtained from minimizing the
sum-of-squares function in Lesson 6.5. In addition, the estimated standard error of
the normal distribution is found to be 0.2423, or σ2 = 0.0587. This also confirms the
minimal sum-of-squares S(β) = Nσ2 = 1.761.
Minimizing sum-of-squares and maximizing log-likelihood are popular techniques

for econometric model estimation. In the next chapter on nonlinear regression
models, for the convenience of classical econometric analysis, it is only the
functional form of model error ε(X,β) that we will need to specify. The objective of
either minimizing a sum-of-squares or maximizing a log-likelihood is readily
available once the specific error structure is given. Nevertheless, as we have learned
from this chapter, being able to work on the scalar-valued objective function directly
is useful when dealing with difficult optimization problems.
98
VII
Nonlinear Regression Models
Many economic and econometric problems can be formulated as optimization
(minimization or maximization) problems. In econometrics, sum-of-squares
minimization and log-likelihood maximization are standard in empirical model
estimation. In the previous chapter, we defined a scalar-valued objective function to
minimize (maximize) and interpreted the parameter estimates in accordance with the
classical least squares (maximum likelihood) model. This approach is flexible
enough to encompass many different econometric models. In many situations,
however, it becomes troublesome to write out the objective function in detail. It is
more desirable to present only the functional form which defines the model directly,
such as
F(Z,β) = ε
where Z is the data matrix, β is the parameter vector, and ε is the error term. Both Z
and β are used to define the functional form of the model (that is, the error structure).
The data matrix Z can be further decomposed as Z = [Y, X] where Y consists of
endogenous (dependent) variables and X is a list of predetermined (independent)
variables. For a classical single regression equation, Y = f(X,β) + ε or ε = Y - f(X,β).
The special case of linear model is simply ε = Y - Xβ.
Nonlinear Least Squares

The functional form F(Z,β) = ε is of interest in econometric modeling. Consider the
sum-of-squares objective function:
S(β) = ε'ε
A nonlinear least squares estimator b of β is computed from the first-order condition

for minimization (or the zero gradient condition) as follows:
∂S(b ) ∂β = 2ε ′(∂ε ∂β ) = 0
In addition, the hessian matrix of second derivatives evaluated at b,
∂ 2 S(b )  ′  ∂ 2 ε i 
= 2(∂ε ∂β ) (∂ε ∂β ) + ∑i =1, 2,..., N ε i  

∂β∂β ′   ∂β∂β ′ 
must be positive definite to guarantee the minimum solution. The estimated

variance-covariance matrix of the parameters is derived from the expected values of
the hessian:
−1
  ∂ 2S(b )  −1
Var(b ) = s 2 ½E  = s 2 (∂ε ∂β )′ (∂ε ∂β )
 
  ∂β∂β′ 
where s2 = e'e/N is the estimated regression variance σ2, e = F(Z,b) is the estimated
error, and N is the sample size used for estimation. It becomes clear that only the
information concerning the functional form F(Z,β) and its first and second
derivatives is needed to carry out the nonlinear least squares estimation of b, e, s2,
and Var(b).
It is straightforward to generalize the technique of nonlinear least squares to consider

weighted model errors. Denote the weighting scheme w = w(Z,β). The variable w is
a scalar or a vector, which in turn may depend on part or all of the data and
parameters. Define the weighted error terms as ε* = w ε. Then the model is estimated
by minimizing the sum-of-squared weighted errors: S*(β) = ε*’ε*.
Lesson 7.1: CES Production Function

Let’s return to the example of Lesson 6.5 estimating a nonlinear CES production
function. Instead of defining and minimizing the sum-of-squares objective function,
this example demonstrates the use of the error or residual function for model
estimation. A residual function is usually more intuitive and easier to write than a
sum-of-squares function. Since the computation of least squares relies on the
residuals and their derivatives, the advantage of increased numerical precision is
another reason of working with the residuals directly.
The setup of input control variables is the same as in Lesson 6.5. The difference is
the use of the residual function (instead of the sum-of-squares objective function) in
calling the estimate procedure (line 12). The residual function ces is defined in
the block from line 14 to line 20.
/*
** Lesson 7.1: CES Production Function Revisited
** Judge, et al. (1988), Chapter 12
*/
1 use gpe2;
2 output file=gpe\output7.1 reset;
4 call reset;
5 _nlopt=0; @ NLSQ: SSE minimization @

6 _method=5; @ optimization method @
7 _iter=100;
8 _tol=1.0e-5;
9 _conv=1;
10 _vcov=1;
11 _b={1.0,0.5,-1.0,-1.0};
12 call estimate(&ces,x);
13 end;
14 proc ces(data,b); @ residual function @

15 local l,k,q;
16 l=data[.,1];
17 k=data[.,2];
18 q=data[.,3];
19 retp(ln(q)-b[1]-b[4]*ln(b[2]*l^b[3]+(1-b[2])*k^b[3]));
20 endp;
100
NONLINEAR REGRESSION MODELS
The regression output duplicates that of Lesson 6.5, and it is available in the output
file output7.1.
Maximum Likelihood Estimation

To explore the idea of maximum likelihood estimation in the econometric context,
we assume a classical normal probability distribution for the independent model
error ε = F(Z,β). That is, ε is normally independently distributed with zero mean and
constant variance σ2. Then the probability density function of Y (recall that Z = [Y,
X]) is written as:
1  F(Z, β)2 
exp   | J (Z, β) |
 2σ 
2
2πσ 2
where J(Z,β) = ∂ε/∂Y is the Jacobian of the transformation from ε to Y. Sampling

over N independent observations, the log-likelihood function of the unknown
parameter vector θ = (β,σ) is:
′
 F(Z, β )   F(Z, β ) 
( )
ll (θ ) = − N 2 ln(2π ) − N 2 ln σ 2 − 1 2  + ∑i =1, 2,..., N ln | J (Zi , β ) |
 σ   σ 
The technique of maximum likelihood estimation is to find the θ that maximizes the
log-likelihood function ll(θ). Usually the computation is performed by substituting
out the variance estimate σ2 = ε'ε/N = F(Z,β)'F(Z,β)/N. Then the following
concentrated log-likelihood function is maximized with respect to the parameter
vector β:
′
ll * (β ) = − N 2 [1 + ln(2π) − ln(N )] − N 2 ln F(Z, β ) F(Z, β ) + ∑i =1, 2,..., N ln | J (Z i , β ) |
 
Define ε* = ε/[(∏i=1,2,…,N|Ji|)1/N] or equivalently F*(Z,β) = F(Z,β)/[(∏i=1,2,…,N|Ji|)1/N],

where Ji = J(Zi,β). Then the last two terms of the above concentrated log-likelihood
function can be combined and the function is rewritten as:
′
ll * (β ) = − N 2 [1 + ln(2π) − ln(N )] − N 2 ln F * (Z, β ) F * (Z, β )
 
where F*(Z,β) = ε* is the weighted error, with the weight being the inverse of the
geometric mean of Jacobians (that is, 1/[(∏i=1,2,…,N|Ji|)1/N]). Therefore, maximizing
the concentrated log-likelihood function ll*(β) is equivalent to minimizing the
corresponding sum-of-squared weighted errors S*(β) = ε*'ε*.
The maximum likelihood estimator b of β is obtained from solving from the first-
order condition (recall that S* = ε*'ε* and ε* = F*(Z,β)):
∂ll * (b ) ∂β = − 1 2 (N S *)(∂S * ∂β ) = −(N S *)[ε *′ (∂ε * ∂β )] = 0
We must also check that the hessian matrix is negative definite (the second-order
condition for maximization) at b:
101
∂ 2 ll * (b )  ′ ∂ 2S * 
= − 1 2 (N S *)(1 S *)(∂S * ∂β ) (∂S * ∂β ) − 
∂β∂β ′  ∂β∂β ′ 
Since ∂S*/∂β = 0, at the maximum the corresponding negative definite hessian

matrix is simply:
∂ 2 ll * (b )   ∂ 2 S * 
= −(N S *)1 2  
∂β∂β ′   ∂β∂β ′ 
 ′  ∂ 2 ε i * 
= −(N S *)(∂ε * ∂β ) (∂ε * ∂β ) + ∑i =1, 2,..., N ε i *′  

  ∂β∂β ′ 
Thus the estimated variance-covariance matrix of maximum likelihood estimator b,

defined as the inverse of negative expected hessian, is:
−1
  ∂ 2 ll * (b )  −1
Var(b ) = − E 
 = s 2 * (∂ε * ∂β )′ (∂ε * ∂β )
 
  ∂β∂β ′ 
where s2* = e*'e*/N and e* = F*(Z,b) = F(Z,b)/[(∏i=1,2,…,N|Ji|)1/N]. Therefore, if the

objective is to maximize the log-likelihood function under the assumption of
normally independently distributed model error, then we need only to be concerned
with the residual function ε = F(Z,β) and its associated Jacobian transformations
J(Z,β).
We now introduce the GPE input variable _jacob, we have not yet introduced,
which controls the use of Jacobians in deriving the objective function (log-likelihood
or sum-of-squares) from the residuals. Notice that a Jacobian transformation is
nothing but a function of data and parameters. If you define a Jacobian function of
your own, then _jacob should be set to the location (address) of the function. An
example of a Jacobian function is given later in Lesson 7.2 on Box-Cox variable
transformation. If you do not wish to write out the Jacobian analytically, you may set
_jacob = 1;
Then the numerical Jacobian is computed for each sample observation, which is
usually a time consuming process. In case of requesting numerical Jacobians, the
first column of the data matrix used to define the residuals must be the dependent
variable Y (recall that J(Z,β) = ∂ε/∂Y and Z = [Y,X]).
Here, based on Lesson 7.1 above, we insert the following statement before calling
the estimate procedure in line 12:
_jacob = 0;
Although it is not necessary (_jacob=0 by default), _jacob=0 is used here to

make sure that you understand the implication of Jacobian terms in the log-
likelihood function. As a matter of fact, for the classical model of the CES
production function, there is no need to consider the vanishing Jacobian terms in the
log-likelihood function.
102
Lesson7.1 may be modified to carry out maximum likelihood estimation instead.

Change the type of problem from minimization to maximization in line 5:
_nlopt = 1;
It is no surprise that the empirical results are identical for both techniques of
nonlinear least squares and maximum likelihood.
If you ask for the numerical Jacobians to be computed in this example:

_jacob = 1;
You must be sure that the first column of the data matrix data used to define the
residual function ces(data,b) corresponds to the dependent variable of the
model, ln(Q) in this case. As it was presented in Lesson 7.1, this rule is not
followed. You may want to correct the data matrix and rewrite the procedure
ces(data,b) so that you can use the numerical Jacobians which are all ones. The
estimation result should not be affected.
Box-Cox Variable Transformation

A nonlinear regression equation may involve nonlinearity in both parameters and
variables. The Box-Cox variable transformation is a classic example of a nonlinear
model in econometrics. The so-called Box-Cox transformation of a data variable X is
defined by
X(λ) = (Xλ-1)/λ
Although the range of λ can cover the whole set of real numbers, -2 ≤ λ ≤ 2 is the
area of interest in many econometric applications. λ = 2 corresponds to a quadratic
transformation, while λ = ½ is a square-root transformation. A linear model
corresponds to λ =1, and the logarithmic transformation is the limiting case where λ
approaches 0 (by L’Hôspital’s rule, limλ−>0 (Xλ-1)/λ = ln(X)).
The value of the power transformation parameter λ may not have to be the same for
each variable in the model. In particular, the dependent variable and independent
variables as a group may need different Box-Cox transformations. Let β = (α,λ,θ) be
the vector of unknown parameters for a regression model:
ε = F(Z,β) = F(Y,X,β) = Y(θ) – X(λ)α
or, equivalently,
Y(θ) = X(λ)α + ε
To estimate the parameters, we assume the error ε is normally independently

distributed with zero mean and constant variance σ2. The log-likelihood function of
the random variable Y is
[ ( )] ′
ll (β ) = − N 2 ln(2π ) + ln σ 2 − 1 2 F(Z, β ) F(Z, β ) / σ 2 + (θ − 1)∑i =1, 2,..., N ln(| Yi |)
103
Z = [Y,X], β = (α,λ,θ), and for each data observation i, the Jacobian term of the
function is derived as J(Yi,θ) = ∂εi/∂Yi = Yi(θ-1). By substituting out the variance σ2 =
ε'ε/N, the concentrated log-likelihood function is
ll * (β ) = − N 2 [1 + ln(2π) − ln(N )] − N 2 ln F(Z, β )′ F(Z, β ) + (θ − 1)∑ ln(| Yi |)

  i =1, 2 ,..., N
′
= − N 2 [1 + ln(2π ) − ln(N )] − N 2 ln F * (Z, β ) F * (Z, β )
 
where F*(Z,β) = ε* = ε/[(∏i=1,2,…,N |Yi|)(θ−1)/N]. Given the values of Box-Cox

transformation parameters θ and λ, a wide range of model specifications are
possible. Of course, θ and λ must be estimated jointly with α. An efficient estimator
of the parameter vector β = (α,λ,θ) is obtained by maximizing the above
concentrated log-likelihood function. It is equivalent to minimizing the sum-of-
squared weighted errors S(β) = ε*'ε*, where ε* = w ε and w = w(Y,θ) = 1/[(∏i=1,2,…,N
|Yi|)(θ−1)/N].
Lesson 7.2: Box-Cox Variable Transformation

The following example of U.S. money demand equation is taken from Greene (1999,
Chapter 10):
M(θ) = α0 + α1 R(λ) + α2 Y(λ) + ε
As described in Greene’s Example 10.9, M is the real money stock M2, R is the
discount interest rate, and Y is real GNP. money.txt is the data text file consisting of
these variables. Several variations of the Box-Cox transformation may be estimated
and tested for selecting the most appropriate functional form of the money demand
equation:
1. θ −> 0 ln(M) = α0 + α1 R(λ) + α2 Y(λ) + ε

2. θ=1 M = α0 + α1 R(λ) + α2 Y(λ) + ε
3. λ −> 0 M(θ) = α0 + α1 ln(R) + α2 ln(Y) + ε
4. λ=1 M(θ) = α0 + α1 R + α2 Y + ε
5. θ=λ M(λ) = α0 + α1 R(λ) + α2 Y(λ) + ε
6. θ, λ −> 0 ln(M) = α0 + α1 ln(R) + α2 ln(Y) + ε
7. θ, λ = 1 M = α0 + α1 R + α2 Y + ε
The following program considers a general case of Box-Cox variable transformation.

In addition to the regression parameters α, there are two power transformation
parameters: θ for the left-hand side variable and λ for the right-hand side. All these
parameters (α,λ,θ) are estimated jointly. For model estimation, it is typical to scale
the variables so that their power transformations do not under- or over-flow in
arithmetic computation. This will make the interpretation of the estimated parameters
more difficult. The sensible approach is to convert the estimated values of the
parameters to unit-free measurements of elasticity. Therefore, for the general model,
the estimated elasticity of the interest rate is
∂ln(M ) R ∂M Rλ
= = α1 θ
∂ln(R ) M ∂R M
104
Similarly, the elasticity of GNP is ∂ln(M ) = α Y . The elasticity at the means (of
λ
∂ln(Y )
2
Mθ
data variables) should be reported for model interpretation.
/*
** Lesson 7.2: Box-Cox Transformation
** U.S. Money Demand Equation
** Greene (1999), Chapter 10
*/
1 use gpe2;
3 load x[21,4]=gpe\money.txt;
@ scale and re-arrange data: m,r,y @
4 x=(x[2:21,3]/1000)~(x[2:21,2])~(x[2:21,4]/1000);
5 call reset;
6 _method=0;
7 _iter=200;
8 _step=1;
9 _conv=1;
10 _jacob=&jf;
@ starting linear model @

11 b=x[.,1]/(ones(rows(x),1)~x[.,2:3]);
12 _b=b|1.0|1.0;
/*
@ starting log model @
b=ln(x[.,1])/(ones(rows(x),1)~ln(x[.,2:3]));
_b=b|-0.01|0.01;
*/
13 _nlopt=1; @ MAXLIK @
14 call estimate(&rf,x);
15 end;
16 proc jf(data,b); @ jacobian @

17 local k;
18 k=rows(b); @ the last parameter @
19 retp(data[.,1]^(b[k]-1));
20 endp;
21 proc rf(data,b); @ residual: general model @

22 local r,m,y,e;
@ box-cox transformation @
23 m=(data[.,1]^b[5]-1)/b[5];
24 r=(data[.,2]^b[4]-1)/b[4];
25 y=(data[.,3]^b[4]-1)/b[4];
26 e=m-b[1]-b[2]*r-b[3]*y;
27 retp(e);
28 endp;
The residual function for the general model rf is defined in lines 21 through 28.
Notice that the residual function is written in such a way that the first column of the
data matrix is the dependent variable. The procedure jf, given in the block from
lines 16 to 20, defines the Jacobian terms for the likelihood function. Recall that the
Jacobian term is just Yθ-1 for the Box-Cox model, where Y is the left-hand side
dependent variable. As we have mentioned earlier in Lesson 7.1, the GPE input
variable _jacob controls the use of Jacobian transformation in defining the log-
likelihood and sum-of-squares functions. In the case of Box-Cox variable
transformation, the residuals are weighted with the inverse of geometric mean of
Jacobians: 1/[(∏i=1,2,…,N |Ji|)1/N] and Ji = Yiθ-1. When _jacob=0, the Jacobians are
not used (vanishing log-Jacobians is assumed). When _jacob=1, the numerical
105
Jacobians are computed from the residual function with the assumption that the first
column of the data matrix is the dependent variable under consideration. When
_jacob is set to the location (address) of a procedure defining the analytical
Jacobians, the result of the procedure called is used. Here, in line 10:
_jacob = &jf;
the input control variable _jacob is assigned the result of the procedure
jf(data,b). Therefore, the analytical Jacobian transformation is used for log-
likelihood maximization and sum-of-squares (weighted) minimization. By defining
and applying the Jacobian transformation, as we have done here, we guarantee that
our parameters will be efficiently estimated.
Although numerical Jacobians may be requested instead, by setting _jacob=1 in

line 10, we recommend against using it. The numerical approximation of Jacobian
transformation is typically slower and results in less accurate parameter estimates
than explicitly defining the analytical formula. As in the case of defining the residual
function, the Jacobian function is problem-dependent and it involves differentiation
of the residual function with respect to the dependent variable. For optimization,
another round of differentiation of the residual and Jacobian functions is needed.
Rounding errors due to excess numerical approximation may be too severe to yield
an accurate solution.
Two sets of starting values of the parameters may be tried: one from the linear model
estimates and the other from log model. In the program, we start with the linear
model estimates (lines 11 and 12). The alternative is to start with the log model
estimates as given in the comment block immediately below line 12. Just to make
sure that you achieve the same solution starting from several different initial values
of the parameters, run program lesson7.2, and check the following result:

-----------------------------

Tolerance = 0.001
Initial Result:
Sum of Squares = 7.0963
Log Likelihood = -18.017
Parameters = -3.1694 -0.014921 1.5881 1.0000 1.0000
Using Berndt-Hall-Hall-Hausman (BHHH) Algorithm
Final Result:
Log Likelihood = 22.978
Gradient of Log Likelihood = -8.2374e-005 -3.0382e-005 -2.8533e-005 -
0.00032734 -6.0704e-005
Asymptotic
X1 -14.503 12.127 -1.1959
X2 -14.067 50.134 -0.28058
X3 56.399 106.75 0.52831
106
X4 -2.6723 1.8562 -1.4396

X5 -0.96447 0.54949 -1.7552
Model interpretation of Box-Cox variable transformation is more difficult than that

of linear models. However, the elasticity at the mean of each variable is computed
and interpreted as below:
M (-0.96) = -14.5 - 14.07 R (−2.67) + 56.4 Y (-2.67)

s.e. (0.55) (12.13) (50.13) (1.86) (106.75) (1.86)
Elasticity - 0.087 4.252
The same model may be estimated with the technique of weighted least squares. As
long as the Jacobian is used, the sum-of-squares function is derived from the
residuals, weighted by the inverse of the geometric mean of the Jacobians. Just
replace line 13 with:
_nlopt=0;
This should produce the same result as that from the maximum likelihood
estimation. However, if we attempt to minimize the sum-of-squared unweighted
residuals, then the estimation result will not be efficient. It can even be biased.
Check it out by deleting line 10 or changing it to:
_jacob=0;
The program of Lesson 7.2 is readily modifiable to accommodate all of the special
cases of Box-Cox transformations. For example, for the case θ = λ, let’s define the
residual function rf1 as follows:
proc rf1(data,b);
local r,m,y,e;
@ box-cox transformation @
m=(data[.,1]^b[4]-1)/b[4];
r=(data[.,2]^b[4]-1)/b[4];
y=(data[.,3]^b[4]-1)/b[4];
e=m-b[1]-b[2]*r-b[3]*y;
retp(e);
endp;
To run this special case, modify the starting values for the parameters (note that the
number of parameters is changed as well) and call estimate with &rf1. That is,
the lines from 12 to 14 should read like this:
_b=b|1.0;
_nlopt=1;
call estimate(&rf1,x);
Other cases such as linear or log transformation on one side of the equation can be
estimated as long as the respective residual function is defined and used correctly.
You may have to experiment with different combinations of optimization options
and starting values to find all the solutions. Also, the last two cases of linear and log
models may be more conveniently estimated with linear least squares. We leave the
remainder of these special cases to you as exercises.
From Lesson 7.2, we have estimated the general Box-Cox transformation model,
together with many special cases. It is useful to tabulate and compare the estimation
results of all these models. Based on the sample data series in money.txt, what is the
107
most appropriate functional form for the U.S. money demand equation? To answer
this question, some knowledge of statistical inference will be necessary.
Statistical Inference in Nonlinear Models

A fundamental assumption of statistical inference is the normality of the model error:
ε = F(Z,β). In particular, ε is assumed to be identically independently normally
distributed with zero mean and constant variance σ2, or ε ~ nii(0,σ2). Thus the
estimated least squares or maximum likelihood parameter b of β is normally
distributed with the mean as β and the estimated variance-covariance matrix as:
[ ( )]
Var (b ) = − E ∂ 2 ll (b ) / ∂β ′∂β
−1
= s [½E (∂ S(b ) / ∂β′∂β )]

2 2 −1
where s2 = S(b)/N is the estimated asymptotic variance of the model. Note that S(b)
= e'e and e = F(Z,b).
In many situations, it is of interest to test the validity of a set of J non-sample

restrictions of the parameters, linear or nonlinear (continuous and differentiable),
expressed as the following vector-valued equation:
c(β) = 0
If there are J active parameter restrictions, let the restricted parameter estimator and
its variance-covariance matrix be b* and Var(b*), respectively. For example, the
simplest case of a linear restriction c(β) = β - β0 (possibly a vector) confines the
parameter vector β to be near β0. The following three tests are useful for inference
about the model restrictions.
Wald Test
Without estimating the constrained model, the unconstrained parameter estimator b
is expected to satisfy the constraint equation closely, if the hypothesis is true. That is,
c(b) = 0. The Wald test statistic:
′
W = c(b ) { Var[c(b )] } c(b )
−1
has a Chi-square distribution with J degrees of freedom (remember that J is the

number of restrictions). With the first-order linear approximation of the constraint
function c(β) at b,
′
{
W = c(b ) [∂c(b ) ∂β ] [Var(b )] [∂c(b ) ∂β ]
′
} c(b)
−1
Note that this test statistic does not require the computation of the constrained
parameters.
Lagrangian Multiplier (LM) Test

Given the J-element constraint equation c(β) = 0, let b* denote the maximum
likelihood estimator of the parameter vector β with the constraint in place. The
108
Lagrangian multiplier test is based on the score vector ∂ll(b*)/∂β of the original
parameterization of the log-likelihood function. If the constraints hold, then
∂ll(b*)/∂β should be close to ∂ll(b)/∂β for the unconstrained parameter estimator b,
which is of course zero. The Lagrangian multiplier test statistic is written as:
LM = (∂ll(b*)/∂β) [Var(b*)] (∂ll(b*)/∂β)'
The estimated variance-covariance matrix of the constrained estimator b* is

computed as follows:
Var(b*) = H-1 [I - G'(G H-1G')-1H-1]
where H = [-∂ll(b*)2/∂β'∂β] and G = [∂c(b*)/∂β]. In practice, the LM test statistic is

easily approximated with the following formula:
 ′
−1
′
LM =  [e * ′ (∂ e * ∂ β )]  (∂ e * ∂ β ) (∂ e * ∂ β ) [e * ′ (∂ e * ∂ β )]  / (e * ′ e * /N )
 
 
 
where e* = F(Z,b*) is the vector of residuals evaluated at the constrained maximum

likelihood solution, and ∂e*/∂β = ∂F(Z,b*)/∂β. Note that this test statistic is based on
the constrained parameters alone.
Likelihood Ratio (LR) Test

If both the constrained and unconstrained maximum likelihood solutions are
available, then the Likelihood Ratio (LR) test statistic defined by
LR = -2(ll(b*)-ll(b))
follows a Chi-square distribution with J degrees of freedom, in which there are J

constraints in the equation c(β) = 0. In terms of sum-of-squares, it is equivalent to
LR = N ln(S(b*)/S(b))
Lesson 7.3: Hypothesis Testing for Nonlinear Models

Returning to Lesson 7.1 (see also Lesson 6.5) on CES production function,
ln(Q) = β1 + β4 ln(β2Lβ3 + (1-β2)K β3) + ε
let’s verify the nonlinear equality constraint: β4 = 1/β3. The following program
implements the Wald, Lagrangian multiplier, and Likelihood Ratio tests, based on
constrained and unconstrained maximum likelihood estimates of the parameters. The
unconstrained model is the same as in Lesson 7.1 except that we are working with
maximum likelihood estimation (instead of sum-of-squares). The constrained
residual function rfc is defined in lines 38 through 45 in which the constraint β4 =
1/β3 is substituted into the function, eliminating the parameter β4. The single
constraint, expressed as β4β3 – 1 = 0, is given in lines 46 through 48 and is named as
the eqc procedure. In line 11 and 12, the constrained model is estimated, and the
estimated parameters and log-likelihood function value are saved for later use.
109
/*
** Lesson 7.3: Hypothesis Testing for Nonlinear Models
** CES Production Function: b[4]=1/b[3]
** Judge, et al. (1988), Chapter 12
*/
1 use gpe2;
4 call reset;
5 _nlopt=1; @ MAXLIK: log-likelihood maximization @
6 _method=5;
7 _iter=100;
8 _tol=1.0e-5;
9 _conv=1;
10 _jacob=0; @ vanishing log-jacobians @
/* Lagrangian Multiplier Test */

@ based on constrained estimation @
11 _b={1.0,0.5,-1.0};
12 call estimate(&rfc,x);
13 b1=__b|(1/__b[3]); @ original parameterization, b[4]=1/b[3] @
14 ll1=__ll; @ log-likelihood @
15 e=rf(x,b1); @ estimated errors @
16 s2=meanc(e^2); @ estimated error variance @
17 g=gradp2(&rf,x,b1); @ gradient of error function @
18 lm=(e'g)*invpd(g'g)*(g'e)/s2;
/* Wald Test */
@ based on unconstrained estimation @
19 _b={1.0,0.25,-1.0,-1.0};
20 call estimate(&rf,x);
21 b2=__b; @ estimated parameters @
22 vb2=__vb; @ estimated var-cov. of parameters @
23 ll2=__ll; @ log-likelihood @
24 w=eqc(b2)'*invpd(gradp(&eqc,b2)*vb2*gradp(&eqc,b2)')*eqc(b2);
/* Likelihood Ratio Test */

25 lr=-2*(ll1-ll2);
26 print "Wald Test = " w;

27 print "Lagrangian Multiplier Test = " lm;
28 print "Likelihood Ratio Test = " lr;
29 end;
30 proc rf(data,b); @ unconstrained residual function @

31 local l,k,q,e;
32 l=data[.,1];
33 k=data[.,2];
34 q=data[.,3];
35 e=ln(q)-b[1]-b[4]*ln(b[2]*l^b[3]+(1-b[2])*k^b[3]);
36 retp(e);
37 endp;
38 proc rfc(data,b); @ constrained residual function @

39 local l,k,q,n,e;
40 l=data[.,1];
41 k=data[.,2];
42 q=data[.,3];
43 e=ln(q)-b[1]-(1/b[3])*ln(b[2]*l^b[3]+(1-b[2])*k^b[3]);
44 retp(e);
45 endp;
46 proc eqc(b); @ constraint function @

47 retp(b[3]*b[4]-1);
48 endp;
110
Here is the estimation result of the constrained model as of line 12 (see also
output7.3 for more details):

-----------------------------

Tolerance = 1e-005
Initial Result:
Parameters = 1.0000 0.50000 -1.0000
Final Result:
Gradient of Log Likelihood = -6.2600e-006 3.4304e-006 -5.5635e-007
Asymptotic
X1 0.11849 0.070742 1.6749
X2 0.32238 0.10324 3.1225
X3 -3.4403 1.7791 -1.9338
We now return to the program of Lesson 7.3. To compute the Lagrangian multiplier
test statistic, the estimated errors are recalculated from the residual function rf (line
15). In addition, the variance (line 16) and the derivatives (line 17) of estimated
errors are needed for implementing the LM formula in line 18. Note that the gradient
computation of line 17 is for a function with two arguments, the data matrix as the
first and parameter vector as the second. The procedure gradp2 is built into GPE
with the consideration that user-defined functions are constructed from the
combination of a data matrix and a parameter vector. It serves the same purpose as
the GAUSS built-in procedure gradp to compute the gradient vector of a
continuous differentiable function with respect to the parameters. The result of
gradp2(&rf,x,b1)of line 17 is a 30 by 4 matrix of derivatives of the residual
function rf with respect to 4 parameters of b1 over a sample of 30 observations of
x.
The Wald test is based on the unconstrained model. The unrestricted regression
model is the same as reported in Lesson 7.1. Based on the maximum likelihood
estimation using the unconstrained residual function rf (lines 30-37), the Wald test
statistic is computed from the constraint function eqc (and its first derivatives
gradp(&eqc,b2)) evaluated at the estimated parameter b2 (lines 19-24). Finally,
log-likelihood function values of both constrained and unconstrained estimations are
used to compute the Likelihood Ratio test statistic in line 25. The following output
summarizes the result:
Wald Test = 0.078241
111
Lagrangian Multiplier Test = 0.092612

Likelihood Ratio Test = 0.082175
All three test statistics are small and close to 0. Comparing with the critical value of
the Chi-square distribution for 1 degree of freedom, we conclude that the restriction
β4 = 1/β3 should not be rejected. Therefore the CES production function should be
represented as follows:
ln(Q) = 0.12 - 1/3.44 ln (0.32 L-3.44 + 0.68 K-3.44)

s.e. (0.07) (0.10) (1.78)
Lesson 7.4: Likelihood Ratio Tests of Money Demand Equation

Now back to the question at the end of Lesson 7.2: what is the most appropriate
functional form for the U.S. money demand equation? If we have successfully
estimated all the models resulting from Box-Cox variable transformations, it is easy
to compute the LR test statistics from each pair of nested models. Note that the log
transformation is a limiting case where the power coefficient approaches zero, which
is not exactly the same as setting the relevant power coefficient to zero.
Nevertheless, LR tests are valid asymptotically. The critical value 3.84 is based on
the Chi-square distribution of 1 degree of freedom at the 5% level of significance. By
applying the general model to this specific example, we demonstrate the strategy for
hypothesis testing:
From the general model λ ≠ θ, does λ = 1? No (LR = 5.92)

does λ -> 0? Yes (LR = 3.37)
does θ = 1? No (LR = 7.09)
does θ -> 0? Yes (LR = 2.44)
Yes (LR = 2.70)
does λ = θ?
From the model λ = θ, does λ = θ = 1? No

does λ = θ -> 0? Yes
The conclusion is obvious that the linear equation of a log model (θ, λ −> 0) will be
the choice for this set of data. The model is linear in the parameters and can be
estimated more accurately using linear least squares. The log model is the limiting
case of Box-Cox transformation, and the estimates obtained from linear regression
are close to those of nonlinear method. The above calculation is based on the
following estimation result, which you should be able to duplicate:
Money Demand Equation Log-Likelihood

θ≠λ M (-0.96) = -14.5 – 14.07 R (−2.67) + 56.4 Y (-2.67) 22.978
1. θ −> 0 ln(M) = -2.23 + 0.0005 R (1.04) + 1.22 Y (1.04) 21.760
2. θ=1 M = -3.03 + 0.0000007 R (4.91) + 0.023 Y (4.91) 19.433
3. λ −> 0 M (-0.23) = -3.62 – 0.022 ln(R) + 3.58 ln(Y) 21.289
4. λ=1 M (-0.021) = -3.54 – 0.0002 R + 1.27 Y 20.019
5. θ=λ M (-0.35) = -4.34 –0.065 R (-0.35) + 5.17 Y (-0.35) 21.648
6. θ, λ −> 0 ln(M) = -3.64 – 0.03 ln(R) + 3.66 ln(Y) 21.833
7. θ, λ = 1 M = -3.17 – 0.015 R + 1.59 Y 8.022
112
VIII
Discrete and Limited Dependent Variables
There are many situations in which the dependent variable of a regression equation is
discrete or limited (truncated) rather than continuous. As we have seen in the
discussion of dummy variables in Chapter IV, some or all of the explanatory
variables in a regression model are qualitative in nature, and therefore only take on a
limited number of values. In the case of dummy variables, those values are 0 and 1.
In this chapter we will consider only the simplest form of qualitative choice models:
binary choice and tobit (censored regression) models. The binary choice (or the “yes
or no” decision) will take on one of two discrete values, 1 or 0. The censored
regression model allows for the dependent variable to follow a mix of discrete and
continuous distributions. Here we learn how to implement and estimate the binary
choice and tobit limited dependent variable models as applications of nonlinear
regression.
Binary Choice Models

Consider a linear regression model Y = Xβ + ε, where
Yi = 1 with probability Pi
0 with probability 1-Pi
It should be clear that Xi explains the probability of Yi equaling 1 or 0. If we let Pi =

Prob(Yi=1|Xi) = F(Xiβ), then 1-Pi = Prob(Yi=0|Xi) = 1-F(Xiβ).
Since E(Yi|Xi) = (1)F(Xiβ) + (0)(1-F(Xiβ)) = F(Xiβ), we may interpret the estimated

model using the following marginal effects:
∂ E (Y i | X i ) ∂ F (X i β )
= β = f (X i β )β
∂X i ∂ (X i β )
where f (X β ) = ∂ F (X i β ) . Given a sample of N independent observations, the

∂ (X i β )
i
likelihood function is
L(β) = ∏i=1,2,...,N PiYi (1-Pi)1-Yi = ∏i=1,2,...,N F(Xiβ)Yi (1-F(Xiβ))1-Yi
The log-likelihood function is then:
ll(β) = ln(L(β)) = ∑i=1,2,...,N [Yi lnF(Xiβ) + (1-Yi) ln(1-F(Xiβ))]
To maximize ll(β) with respect to β, we solve from the following first-order

condition:
∂ ll (β )  Yi 1 − Yi 
∂β
= ∑ 
i = 1 , 2 ,..., N
−  fiX i
 Fi 1 - Fi 
 Y − Fi 
= ∑ i =1 , 2 ,..., N  i  fiX i = 0
 Fi (1 − Fi ) 
where Fi = F (X i β ) and f = f (X β ) = ∂ F (X i β ) . Note that f X = ∂ F (X i β ) .

∂ (X i β )
i i i i
∂β
Finally, we need the hessian matrix  ∂ ll
2
(β )  to be negative definite, allowing us
 
 ∂β∂β ′ 
−1
to write the estimated variance-covariance matrix of β as Var(β ) = − E ∂ll (β )  .
2
  
  ∂β∂β ′ 
Linear Probability Model

Let Pi = F(Xiβ) = Xiβ. It is immediately clear that E(Yi|Xi) = Xiβ. In addition, with
the linear probability model, we are assured of a heterogeneous error structure:
E(εi) = (1-Xiβ)Pi + (-Xiβ)(1-Pi) = Pi - Xiβ

Var(εi) = E(εi2) = Pi(1-Xiβ)2 + (1-Pi)(-Xiβ)2
= Pi(1-Pi)2 + (1-Pi)(-Pi)2 = (1-Pi)Pi = (1-Xiβ)(Xiβ)
The range of Var(εi) is between 0 and 0.25. Furthermore, since E(Yi|Xi) = F(Xiβ) =
Xiβ, a linear function, there is no guarantee that the estimated probability Pi or 1-Pi
will lie within the unit interval. We can get around the problem of Pi taking values
outside the unit interval by considering a specific probability distribution or
functional transformation for Pi. A commonly used probability distribution is the
normal distribution giving rise to the probit model, while a commonly used
functional transformation is the logistic curve function giving rise to the logit model.
Probit Model
Let Pi = F(Xiβ) = ∫−∞Xiβ 1/(2π)1/2 exp(-z2/2) dz. Then we call Pi (based on the
cumulative normal distribution), the probit for the i-th observation. The model Yi =
F-1(Pi) + εi is called the probit model, where F-1(Pi) = Xiβ is the inverse of the
cumulative normal distribution F(Xiβ). For those concerned that we chosen the
above specification seemingly out of thin air, the probit model can be derived from a
model involving a continuous, unobserved, or latent, variable Yi* such that Yi* =
Xiβ + εi, where εi follows a standard normal density.6 Suppose the value of the
observed binary variable Yi depends on the sign of Yi* as follows:
Yi = 1 if Yi* > 0
0 if Yi* ≤ 0
6
If εi is a normal random variable with zero mean and standard error σ, then the probability of
Yi = 1 is written as Pi = F(Xiβ/σ). Since β/σ appears in the density function as a ratio, they are
not separately identified. Therefore, it is convenient to normalize σ to be one. The standard
normal distribution is sometimes referred to as the z-distribution, where the random variable
is zi = εi/σ = εi, given σ =1.
114
DISCRETE AND LIMITED DEPENDENT VARIABLES
Therefore,
Pi = Prob(Yi=1|Xi) = Prob(Yi*>0|Xi) = Prob(εi > -Xiβ)

= ∫ -Xiβ∞ 1/(2π)1/2 exp(-z2/2) dz
= ∫−∞Xiβ 1/(2π)1/2 exp(-z2/2) dz
For maximum likelihood estimation, we solve the following first-order condition:
 Yi − Fi 
∑  f i Xi = 0
 Fi (1 − Fi ) 
i =1, 2 ,..., N
where fi and Fi are, respectively, the probability density and cumulative density
functions of a standard normal random variable evaluated at Xiβ. That is,
Fi = F(X iβ ) = ∫
Xi β
−∞
1 ( )
2π exp − z 2 / 2 dz
and,
∂F(X iβ )
fi =
∂ (X iβ )
=1 [ 2
]
2π exp - (X iβ ) / 2 .
Furthermore, it can be shown that for the maximum likelihood estimates of β the
expected value of the (negative definite) hessian is
′
 ∂ll 2 (β )  f 2X X
E  = −∑i =1,2,..., N i i i .
 ∂β∂β′  Fi (1 − Fi )
The estimated variance-covariance matrix of β is computed as the inverse of

negative expected hessian.
In interpreting the model, the probability E(Yi|Xi) = F(Xiβ) = Pi will be of interest.

The marginal effects of the j-th explanatory variable Xij are written as:
∂E(Yi | X i ) ∂F(X iβ )
= β j = f (X iβ )β j = f i β j
∂X ij ∂ (X iβ )
Lesson 8.1: Probit Model of Economic Education

This example (see Greene 1999, Example 19.1; Spector and Mazzeo, 1980)
examines the effect of a new teaching method (known as PSI) on students’ grades.
The following variables are used:
GRADE An indicator of whether the student’s grade on an examination

improved after exposure to the new teaching method.
PSI An indicator of whether the student was exposed to the new teaching
method.
TUCE Score of a pretest that indicates entering knowledge of the material
115
prior to the introduction of PSI.

GPA Grade point average.
The qualitative model is formulated as follows:
GRADE = β0 + β1GPA + β2TUCE + β3PSI + ε
The data file grade.txt is used. The following program estimates the probit model
specification of the above equation. The log-likelihood function for each data
observation is defined in lines 20 through 26 with the procedure named probitf.
Since the first derivatives of the log-likelihood function are rather straightforward
analytically, we also write the procedure probitf1 to calculate these derivatives
(lines 27 to 34). The analytical first derivatives may be used in optimization to
improve the accuracy of the solution (see Chapter VI).
If the residual function is defined and called when estimating a nonlinear regression
model, the GPE default objective function is the log-likelihood function for
maximization and sum-of-squares for minimization. Whether the problem is
maximization or minimization is controlled by the value of the input variable
_nlopt. Setting _nlopt=1 specifies a maximization problem. Setting _nlopt=0
indicates a minimization problem. There are certain cases of maximum likelihood
estimation in which the log-likelihood function may be defined instead of the
residual function for each sample observation. All the qualitative choice models
discussed in this chapter fall in this second category. Setting _nlopt=2 (see line
11) informs GPE that the maximization is performed on the sum of the component
(log-likelihood) functions. We could write the total log-likelihood function to
estimate the model, but there may be a loss of numerical accuracy due to
compounding running errors in evaluating the function and its derivatives.
/*
** Lesson 8.1: Probit Model of Economic Education
** Greene (1999), Example 19.1
** See also Spector and Mazzeo (1980)
*/
1 use gpe2;
3 n=33;
4 load data[n,4]=gpe\grade.txt;
5 gpa=data[2:n,1];
6 tuce=data[2:n,2];
7 psi=data[2:n,3];
8 grade=data[2:n,4];
9 z=gpa~tuce~psi~ones(rows(grade),1);
10 call reset;
@ probit model: estimation @

11 _nlopt=2; @ using component log-likelihood @
12 _method=4;
13 _iter=50;
14 _b={0.5,0.0,0.5,0};
15 call estimate(&probitf,grade~z);
/*
_derive=&probitf1;
call estimate(&probitf,grade~z);
*/
@ probit model: interpretation @
16 b=__b;
116
17 print " Probability Slopes";;

18 print cdfn(z*b)~(pdfn(z*b).*b[1:rows(b)-1]');
19 end;
/* log-likelihood function of probit model */

20 proc probitf(x,b);
21 local k,z,f;
22 k=rows(b);
23 z=x[.,2:k+1]*b;
24 f=cdfn(z); @ normal cdf @
25 retp(x[.,1].*ln(f)+(1-x[.,1]).*ln(1-f));
26 endp;
/* 1st derivatives of log-likelihood function of probit model */

27 proc probitf1(x,b);
28 local z,k,f,g;
29 k=rows(b);
30 z=x[.,2:k+1]*b;
31 f=cdfn(z); @ normal cdf @
32 g=pdfn(z); @ normal pdf @
33 retp((x[.,1].*(g./f)-(1-x[.,1]).*(g./(1-f))).*x[.,2:k+1]);
34 endp;
Running the program, we obtain the estimated probit model as follows (see also
output8.1 for the detailed results of interim iterations):

-----------------------------

Tolerance = 0.001
Initial Result:
Parameters = 0.50000 0.00000 0.50000 0.00000
Final Result:
Parameters = 1.6258 0.051729 1.4263 -7.4523
Gradient Vector = -2.5489e-005 -0.00031045 -3.6749e-006 -5.7386e-006
X1 1.6258 0.69373 2.3436
X2 0.051729 0.083925 0.61637
X3 1.4263 0.59520 2.3964
X4 -7.4523 2.5467 -2.9263
Recall that, given the estimated parameters, we are mostly concerned with the
probability or the conditional expected value E(Y|X). With respect to the three
explanatory variables, GPA, TUCE, and PSI, the slopes (or marginal effects) are of
interest. Lines 16 through 18 of lesson8.1 calculate and report the probability and the
relevant marginal effects. The next part of the output shows these results for each
observation:
Probability Slopes
117
0.018171 0.072553 0.0023084 0.063651

0.053081 0.17584 0.0055949 0.15427
0.18993 0.44108 0.014034 0.38696
0.018571 0.073911 0.0023517 0.064843
0.55457 0.64253 0.020443 0.56369
0.027233 0.10207 0.0032476 0.089546
0.018503 0.073683 0.0023444 0.064642
0.044571 0.15292 0.0048656 0.13416
0.10881 0.30333 0.0096511 0.26611
0.66312 0.59360 0.018887 0.52077
0.016102 0.065434 0.0020819 0.057406
0.19356 0.44623 0.014198 0.39148
0.32333 0.58391 0.018579 0.51227
0.19518 0.44851 0.014270 0.39348
0.35634 0.60608 0.019284 0.53172
0.021965 0.085218 0.0027114 0.074762
0.045694 0.15601 0.0049640 0.13687
0.030851 0.11322 0.0036023 0.099327
0.59340 0.63075 0.020069 0.55336
0.65719 0.59758 0.019013 0.52426
0.061929 0.19852 0.0063164 0.17416
0.90454 0.27577 0.0087744 0.24194
0.27319 0.54072 0.017204 0.47438
0.84745 0.38335 0.012197 0.33631
0.83419 0.40485 0.012881 0.35518
0.48873 0.64834 0.020629 0.56880
0.64241 0.60683 0.019308 0.53237
0.32867 0.58783 0.018703 0.51571
0.84002 0.39555 0.012585 0.34702
0.95224 0.16164 0.0051428 0.14180
0.53996 0.64535 0.020533 0.56617
0.12354 0.33195 0.010562 0.29122
If you do not want to see the long list of E(Y|X) (probability) and ∂E(Y|X)/∂X
(marginal effects), they may be computed at the means of the explanatory variables.
To see what happens, insert the following statement after line 16:
z=meanc(z)';
Logit Model
Let Pi = F(X i β ) = 1
, where Pi as defined is the logistic curve. The
1 + exp(− X i β )
model Yi = Xiβ+ εi = F-1(Pi) + εi is called the logit model. We can easily derive the
logit model from the odd ratio model, in which we assume that the log of the ratio of
Pi
the probabilities (Pi and 1-Pi) is equal to Xiβ. Thus we assume ln  1-P  = Xiβ.
 i
Solving for Pi yields:
exp(X iβ ) 1
Pi = =
1 + exp(X iβ ) 1 + exp(− X iβ )
For maximum likelihood estimation, we solve the first-order condition:
 Yi − Fi 
∑   fi Xi = 0
 Fi (1 − Fi ) 
i =1, 2 ,..., N
Because of the logistic functional form in use,
118
1
Fi = F(X i β ) =
1 + exp(− X i β )
and
∂F(X iβ ) exp(− X iβ )
fi = = = Fi (1 − Fi ),
∂ (X iβ ) [1 + exp(− X iβ )]2
the first-order condition amounts to the following simpler expression:
∑ i =1, 2 ,..., N
(Yi − Fi )X i = 0.
The variance-covariance matrix of β is estimated with
−1
  ∂ll 2 (β ) 
Var (β ) = − E 
  ∂β∂β ′ 
where the hessian matrix  ∂ll (β )  = - ∑i=1,2,...,NFi(1-Fi)Xi'Xi is negative definite.

2
 ∂β∂β ′ 
 
To interpret the model, we define the marginal effect of the j-th explanatory variable
Xij as:
∂E(Yi | X i ) ∂F(X iβ )
= β j = f (X iβ )β j = f i β j = Fi (1 − Fi )β j
∂X ij ∂(X iβ )
As you can see, the logit model is similar in construction to the probit model. Only
the choice of transformation function is different.
Lesson 8.2: Logit Model of Economic Education

We will now rerun the model of Lesson 8.1, using the logit model instead of the
probit model. We will need only to change the transformation function from the
cumulative normal probability to the logistic curve. Specifically, lines 20-26 define
the component log-likelihood function logitf for the logit model, while the
corresponding first derivative function logitf1 is defined in lines 27-34.
/*
** Lesson 8.2: Logit Model of Economic Education
** See also Spector and Mazzeo (1980)
*/
1 use gpe2;
3 n=33;
4 load data[n,4]=gpe\grade.txt;
5 gpa=data[2:n,1];
6 tuce=data[2:n,2];
7 psi=data[2:n,3];
119
8 grade=data[2:n,4];
9 z=gpa~tuce~psi~ones(rows(grade),1);
10 call reset;
@ logit model: estimation @

11 _nlopt=2; @ using component log-likelihood @
12 _method=4;
13 _iter=50;
14 _b={0.5,0.0,0.5,0};
15 call estimate(&logitf,grade~z);
/*
_deriv=&logitf1;
call estimate(&logitf,grade~z);
*/
@ logit model: interpretation @
16 p=1./(1+exp(-z*__b));
17 print " Probability Slopes";;
18 print p~(p.*(1-p).*__b[1:rows(__b)-1]');
19 end;
/* log-likelihood function of logit model */

20 proc logitf(x,b);
21 local k,z,f;
22 k=rows(b);
23 z=x[.,2:k+1]*b;
24 f=1./(1+exp(-z)); @ same as: f=exp(z)./(1+exp(z)); @
@ logistic distribution function @
25 retp(x[.,1].*ln(f)+(1-x[.,1]).*ln(1-f));
26 endp;
/* 1st derivatives of log-likelihood function of logit model */

27 proc logitf1(x,b);
28 local z,k,f;
29 k=rows(b);
30 z=x[.,2:k+1]*b;
31 f=1./(1+exp(-z)); @ same as: f=exp(z)./(1+exp(z)); @
32 @ logistic distribution function @
33 retp((x[.,1]-f).*x[.,2:k+1]);
34 endp;
The estimated results are similar to those of the probit model. Instead of showing the
detailed output for comparison, we present the estimated probabilities and marginal
effects of the probit and logit models, evaluated at the means of three explanatory
variables:
Probit Logit
Probability 0.26581 0.25282
Marginal Effects GPA 0.53335 0.53386
TUCE 0.01697 0.01798
PSI 0.04679 0.04493
Extensions of binary choice models to the cases with more than two choices are
interesting, though the derivations are tedious. Multiple choice models include
unordered (independent or nested) and ordered (with a preference rank) choices.
Both the probit and logit model specifications for multiple choice are possible, but
they are beyond the scope of the current discussion.
120
Limited Dependent Variable Models

If the random decision variable follows a mixture of discrete and continuous
distributions, we have the limited dependent variable (or censored regression) model.
Recall the latent variable interpretation of the probit model,
Yi* = Xiβ + εi
where εi follows a normal probability distribution, and
Yi = 1 if Yi* > 0
0 if Yi* ≤ 0
Suppose, however, that Yi is censored—that is, we restrict the number (or kinds) of
values that Yi can take. As an example, consider the following:
Yi = Yi* if Yi* > 0

0 if Yi* ≤ 0
That is,
Yi = Xiβ + εi if Xiβ + εi > 0

0 otherwise
This model is called the tobit (or Tobin’s probit) model. Define fi and Fi to be the
probability density function and cumulative density function of a standard normal
random variable evaluated at Xiβ/σ. That is,
Fi = F(X iβ / σ ) = ∫
−∞
Xi β σ
1 (
2π exp − z 2 2 dz )
fi = f (Xiβ/σ) = 1 2π exp − (Xiβ σ) 2[ 2
]
For the observations such that Yi = 0 or Yi* = Xiβ + εi ≤ 0, the likelihood function is
Prob(Yi = 0) = Prob(εi ≤ -Xiβ) = Prob(εi/σ ≤ -Xiβ/σ) = 1 – F(Xiβ/σ) = 1-Fi
If Yi > 0, on the other hand, then the likelihood function is simply the normal density
function:
1/(2πσ2)1/2 exp [-(Yi-Xiβ)2/(2σ2)]
Therefore the likelihood function for the tobit model is a mixture of the above
discrete and continuous distributions depending on the values taken by the dependent
variable (i.e., zero or positive):
L = ∏{i |Yi = 0}(1-Fi) ∏{i |Yi > 0}1/(2πσ2)1/2 exp [-(Yi-Xiβ)2/(2σ2)]
The corresponding log-likelihood function is
[ ( )
ll = ∑{i|Y =0} ln(1 − Fi ) − 1 2 ∑{i|Y >0} ln(π ) + ln σ 2 + (Yi − Xiβ ) σ 2
i i
2
]
121
Then, for the maximum likelihood estimation, we solve from the following first-
order conditions:
i
(
∂ll ∂β = − (1 σ )∑{i|Y =0} f i Xi (1 − Fi ) + 1 σ 2 )∑{ i |Yi >0}
(Yi − Xiβ) Xi = 0
(
∂ll ∂σ 2 =1 2 1 σ 3 )∑{i |Yi = 0} i
(
f X iβ (1 − Fi ) − 1 2 1 σ 2 )∑{ i |Yi >0}
[1 − (Y − X β)
i i
2
]
σ2 = 0
To interpret the estimated coefficients of the model, we may use three conditional
expected values:
E(Yi*|Xi) = Xiβ
E(Yi |Xi,Yi>0) = Xiβ + E(εi |Yi > 0)

= Xiβ + E(εi |εi > -Xiβ)
= Xiβ + σ fi/Fi > E(Yi*|Xi)
E(Yi|Xi) = Fi E(Yi |Xi,Yi > 0)

= Fi Xiβ + σ fi
The first expected value (corresponding to the “uncensored” case) is easy to obtain.
The last expected value will be of particular interest if our sample contains many
censored observations. Accordingly, for the j-th explanatory variable, the
corresponding marginal effects are:
∂E(Yi*|Xi)/∂Xij = βj
∂E(Yi|Xi,Yi>0)/∂Xij = βj [1- (Xiβ/σ)(fi/Fi) –(fi/Fi)2]
∂E(Yi|Xi)/∂Xij = Fi ∂E(Yi|Xi,Yi > 0)/∂Xij + E(Yi |Xi ,Yi > 0) ∂Fi/∂Xij

= Fi βj
We note that the last censored marginal effect differs from the first uncensored one
by a scale factor equal to the probability of that observation not being censored. In
other words, the scale factor is equal to Fi (recall that Fi is 1-Prob(Yi = 0)).
The tobit model is often estimated for comparison with the alternative probit or count
model specifications. The model can be easily extended to consider more than one
censoring point. For instance, we could censor both tails of the distribution. This is
an example of a doubly censored regression.
Lesson 8.3: Tobit Analysis of Extramarital Affairs

This example is taken from Greene (1999, Example 20.12), which is based on Fair
(1978). The study examines the qualitative responses to a question about extramarital
affairs from a sample of 601 men and women married for the first time. The
dependent variable is:
Y Number of affairs in the past year: 0, 1, 2, 3, 4-10 (coded as 7), 11-365 (coded
as 12).
122
The preponderance of zeros (no affairs) may not render the tobit model to be the best
for the study. The complete data set used in the article is available from the text file
fair.txt, but we present only the restricted model using a subset of five explanatory
variables as follows:
Z2 Age.
Z3 Number of years married.
Z5 Degree of religiousness: 1 (anti-religious), … , 5 (very religious).
Z7 Hollingshead scale of occupation: 1, … , 7.
Z8 Self-rating of marriage satisfaction: 1 (very unhappy), … , 5 (very happy).
The regression equation is:
Y = β0 + β2 Z2 + β3 Z3 + β5 Z5 + β7 Z7 + β8 Z8 + ε
The conclusion and interpretation of the estimated model are left to the interested
reader. Our emphasis here is the implementation of tobit analysis using GPE and
GAUSS. To do so, we need to briefly explain the maximum likelihood estimation
procedure. Recall that the likelihood function of the tobit model is a mixture of
discrete and continuous normal likelihoods, depending on the censored point (zero)
of the dependent variable. Unlike in the probit model, the standard error is an explicit
unknown parameter which must be estimated together with the regression
parameters. In lines 22 to 28 of the following program, the procedure
tobitf defines the log-likelihood function for each sample observation. For
maximum likelihood estimation, we need to set _nlopt=2 (line 10), which
instructs GPE to maximize the sum of the individual log-likelihood functions.7
/*
** Lesson 8.3: Tobit Analysis of Extramarital Affairs
** See also R. Fair, JPE, 86, 1978, 45-61
*/
1 use gpe2;
3 n=602;
4 load data[n,15]=gpe\fair.txt;
5 y=data[2:n,13];
6 z=data[2:n,5 6 8 11 12]; @ use z2, z3, z5, z7, z8 @
7 call reset;
@ Uncensored OLS estimation @

8 call estimate(y,z);
@ Tobit model estimation @

9 z=z~ones(rows(z),1); @ RHS variables inc. constant @
10 _nlopt=2; @ using component likelihood function @
11 _method=5;
12 _iter=100;
13 _b=__b|5.0;
14 call estimate(&tobitf,y~z);
@ Tobit model interpretation based on E(y) @

15 b=__b[1:6];
16 s=__b[7];
17 ey=cdfn(z*b/s).*(z*b)+s.*pdfn(z*b/s);
7
Because the size of this nonlinear optimization is beyond the limits of GAUSS Light, the
professional version of GAUSS should be used for Lesson 8.3.
123
18 em=cdfn(z*b/s).*b';
19 print "Expected Value Marginal Effects";;
20 print ey~em[.,1:5];
21 end;
/* Log-likelihood function: tobit model */

22 proc tobitf(x,b);
23 local y,z,k;
24 k=rows(b);
25 y=x[.,1];
26 z=(y-x[.,2:k]*b[1:k-1])/b[k];
27 retp((y.==0).*ln(cdfn(z))+(y.>0).*ln(pdfn(z)/b[k]));
28 endp;
Remember that tobit is a nonlinear model. First, the uncensored model is estimated
by ordinary least squares (line 8). The estimated parameters are then used as the
initial values in the tobit model estimation (line 14). Here is the result of the
estimated tobit model, which converges after 60 iterations:

-----------------------------

Tolerance = 0.001
Initial Result:
Parameters = -0.050347 0.16185 -0.47632 0.10601 -0.71224
5.6082 5.0000
Final Result:
Parameters = -0.17933 0.55414 -1.6862 0.32605 -2.2850
8.1742 8.2471
Gradient Vector = 3.0579e-005 3.0303e-006 2.2826e-006 1.4280e-006
2.3271e-006 4.5474e-007 -2.7294e-006
X1 -0.17933 0.079136 -2.2661
X2 0.55414 0.13459 4.1174
X3 -1.6862 0.40378 -4.1761
X4 0.32605 0.25443 1.2815
X5 -2.2850 0.40787 -5.6022
X6 8.1742 2.7419 2.9812
X7 8.2471 0.55363 14.896
For interpretation of the estimated parameters, we compute the expected values of

censored observations and the corresponding marginal effects of the explanatory
variables (see lines 14-19). The computation is for each sample observation.
To save space we do not list the 601 observations of expected values and marginal
effects. One alternative is to calculate the average of the estimated expected values
124
and marginal effects:

print meanc(ey~em[.,1:5])';
A second alternative is to evaluate the expected values and marginal effects at the
means of the explanatory variables by inserting the following statement before line
16:
z=meanc(z)';
We conclude with an important remark. In Lesson 8.3, the dependent variable is

actually more like count data than a continuous variable, for it records the number of
occurrences of extramarital affairs within a year. As a matter of fact, this model is
well-suited for probit analysis. If we define the dependent variable Y as a binary
variable (0 for no affairs and 1 otherwise), then the same data set can be used for
either probit or logit analysis. However, the estimated parameters, expected values,
and marginal effects of the probit (or logit) model are not directly comparable to
those of the tobit model. The dependent variable being “explained” or “predicted” is
different under tobit estimation than under probit or logit estimation. With a tobit
model, we are predicting the mean number of affairs. With a probit or logit model,
we are predicting the probability of an affair occurring within a year. The probit
analysis is not concerned with the actual number of affairs. We leave the model
comparison to interested readers.
125
IX
Heteroscedasticity
Heteroscedasticity is a common problem with cross-sectional data, in which unequal
model variance is observed. Ordinary least squares estimation with a
heterogeneously distributed error structure leads to inefficient estimates of the
regression parameters. In order to correct for this inefficiency, the source of
heteroscedasticity in relation to one or more variables must be identified.
To illustrate how to test and correct for the problem of heteroscedasticity, the
following relationship of public school spending (SPENDING) and income
(INCOME) across 50 states in the U.S. is considered:
SPENDING = β0 + β1 INCOME + β2 INCOME2 + ε
To estimate this equation, which is used for all the lessons in this chapter, a cross-
sectional data file greene.txt is used.8 It gives per capita public school expenditure
and per capita income by state in 1979. Let’s take a look at the data file greene.txt
we will be dealing with. The data file contains three columns. The first column is the
state identifier (STATE), the second column is per capita expenditure on public
schools (SPENDING), and the third column is per capita income (INCOME).
Viewing greene.txt in the Edit window reveals a problem with the data. Notice that
WI (Wisconsin) is missing a data observation. The row WI has “NA” for the
corresponding value in the SPENDING column. GAUSS sees “NA” as a character
string, not suitable for numerical computation. GAUSS has commands that convert
character strings, such as “NA,” to a symbol that it can interpret as a missing value.
The first part of each lesson in this chapter walks you through the process of
converting greene.txt with its missing values to useable data. Several new GAUSS
commands are introduced for this purpose.
Heteroscedasticity-Consistent Covariance Matrix

Assuming heteroscedasticity, the ordinary least squares estimator is unbiased but
inefficient. We can keep the unbiased least squares estimator, but correct for
inefficiency with an estimated heteroscedasticity-consistent covariance matrix.
Lesson 9.1 below demonstrates the use of the input control variable _hacv to
correct the covariance matrix.
Lesson 9.1: Heteroscedasticity-Consistent Covariance Matrix

In this lesson, we estimate the regression equation of public school spending
(SPENDING) with regressors income (INCOME) and squared income (INCOME2)
using the data file greene.txt.
8
This example was used in Greene (1997, Chapter 12), but it has been removed from the
updated fourth edition (Greene, 1999).
For a heteroscedastic regression model, a consistent estimate of the variance-

covariance matrix is obtained by approximating the original heteroscedastic
variances with the squares of estimated residuals from ordinary least squares
estimation. That is,
() ( )
Var β̂ = (X′X ) X′Σ̂X (X′X )
−1 −1
where X is the data matrix of regressors, β̂ is the ordinary least squares estimator of
the parameters vector β, and Σ̂ is a diagonal variance-covariance matrix (i.e., the
estimator of E(εε') with the elements being the squares of the estimated regression
residuals.
From two least squares estimations, one with the ordinary variance-covariance
matrix and the other with the heteroscedasticity-consistent covariance matrix, we can
directly compare the results of these regressions. In GPE, by setting the input
variable _vcov=1 (see line 11 of Lesson 9.1), the details of the variance-covariance
matrix are presented. The second regression estimation with the newly introduced
input variable _hacv=1 (see line 13 of Lesson 9.1) computes the heteroscedasticity-
consistent estimates of the variance-covariance matrix instead of the inefficient one
from the ordinary least squares.
/*
** Lesson 9.1: Heteroscedasticity-Consistent
** Variance-Covariance Matrix
*/
1 use gpe2;
3 load greene[52,3]= gpe\greene.txt;
4 data = greene[2:52,.];
5 data = miss(data,"NA"); @ NA to missing value @
6 data = packr(data); @ deletes row w/miss value@
7 spending = data[.,2];
8 income = data[.,3]/10000;
9 call reset;
/* Ordinary Least Squares */

10 _names = {"spending","income","income^2"};
11 _vcov = 1;
12 call estimate(spending,income~income^2);

13 _hacv = 1; @ with hetero. consistent var-cov @
15 end;
Line 5 introduces the miss command of GAUSS. It modifies the matrix data by
comparing each element in the matrix data to the character string “NA.” If an
element in data equals “NA,” it is replaced with a dot (.), GAUSS’s symbol for a
missing value. In the next line, packr(data) deletes any rows that contain any
missing values in the matrix data. After data has been packed (line 6), the number
of rows in data is reduced to 50. Refer to the GAUSS manual or on-line help to
find more information on the commands miss and packr.
128
HETEROSCEDASTICITY
Try combining lines 4, 5, and 6 for efficiency:

data=packr(miss(greene[2:52,.],"NA"));
The result of the first least squares estimation with the option to print out the
estimated variance-covariance matrix (lines 10 to 12) is:

------------------------
Dependent Variable = SPENDING

Log-Likelihood Function Value = -271.27

Explained 2.8709E+005 2 1.4355E+005 44.684 1.3445E-011
Residual 1.5099E+005 47 3212.5
Total 4.3808E+005 49 8940.3

INCOME -1834.2 828.99 -2.2126 0.031820 0.094335
INCOME^2 1587.0 519.08 3.0574 0.0036770 0.16590
CONSTANT 832.91 327.29 2.5449 0.014275 0.12111

INCOME 6.8722E+005
INCOME^2 -4.2844E+005 2.6944E+005
CONSTANT -2.7021E+005 1.6709E+005 1.0712E+005
INCOME INCOME^2 CONSTANT

INCOME 1.0000
INCOME^2 -0.99567 1.0000
CONSTANT -0.99591 0.98352 1.0000
In order to compare the ordinary least squares estimates of standard errors and the
variance-covariance matrix with the heteroscedasticity-consistent variance-
covariance matrix, we look at a portion of the output from the second regression
estimation (lines 13 and 14):

INCOME -1834.2 1243.0 -1.4756 0.14673 0.044275
INCOME^2 1587.0 829.99 1.9121 0.061968 0.072177
CONSTANT 832.91 460.89 1.8072 0.077137 0.064972

INCOME 1.5452E+006
INCOME^2 -1.0296E+006 6.8889E+005
CONSTANT -5.7170E+005 3.7941E+005 2.1242E+005
129
INCOME 1.0000
INCOME^2 -0.99796 1.0000
CONSTANT -0.99789 0.99182 1.0000
In this example, the estimated heteroscedasticity-consistent standard errors and the

variance-covariance matrix of the regression parameters are both larger than their
ordinary least squares counterparts. The results indicate that there are smaller t-
ratios, leading to the conclusion of statistically insignificant parameter estimates.
Such correction to the model is useful when the problem of heteroscedasticity is of
unknown form.
Weighted Least Squares

As mentioned earlier, if we can identify one or more of the regressors in the model as
the source of heterogeneity in the variances, then a variation of least squares called
weighted least squares is recommended. This estimation method adjusts all data
series across the regression equation in accordance with the origin of the
heteroscedasticity. Lesson 9.2 details the Goldfeld-Quandt approach, while Lesson
9.3 presents more general testing procedures devised by Breusch-Pagan and White.
Lesson 9.2: Goldfeld-Quandt Test and Correction for Heteroscedasticity

Continuing on with the spending-income regression equation of Lesson 9.1, this
lesson introduces the Goldfeld-Quandt method to test for heteroscedasticity. It then
corrects for this condition by using the method of weighted least squares.
The Goldfeld-Quandt test requires data to be sorted according to the size of the
independent variable suspected to be the source of heteroscedasticity. The entire data
set is then divided into three parts. The middle group is dropped, and the regression
is run using only the groups containing the smallest and largest values. Separate
regressions are estimated on each of the groups of smallest and largest values.
Residual sum-of-squares (RSS) from both groups are then compared in the Goldfeld-
Quandt test statistic.
As in the previous Lesson 9.1, the data are read from greene.txt and corrected for
missing values. The Goldfeld-Quandt test requires a sorted data series in accordance
with the suspected source of heteroscedasticity. Sorting the rows in the matrix,
data, by the information in the third column (that is, the variable INCOME) is done
in line 7. INCOME is sorted from its smallest value to its largest. The GAUSS Help
menu gives greater details about the data sorting commands such as sortc used
here.
/*
** Lesson 9.2: Goldfeld-Quandt Test and
** Correction for Heteroscedasticity
*/
1 use gpe2;
6 data = packr(data); @ deletes row w/miss value @
7 data = sortc(data,3); @ sort data (income), in ascending order @
130
HETEROSCEDASTICITY
9 income = data[.,3]/10000;
10 call reset;
/* Goldfeld-Quandt Test */
12 _begin = 1;
13 _end = 17;
15 mss1 =__rss/14; @ N-K = 17-3 = 14 @
16 _begin = 34;
17 _end = 50;
19 mss2 =__rss/14; @ N-K = 17-3 = 14 @
20 print "Goldfeld-Quandt Test Statistic = " mss2/mss1;
/* Weighted Least Squares */

21 call reset;
22 _weight = 1/income;
24 end;
Selecting the first group of the 17 smallest observations to regress for the Goldfeld-
Quandt test is done by restricting the data matrix to only include observations from 1
to 17 (lines 12 and 13). The use of the output control variable, __rss, is introduced
in line 15. __rss stores the sum-of-squared residuals from the latest regression
estimation. Each time when the procedure estimate is called, output control
variables are assigned new values. To save the value of __rss/14, the mean sum-
of-squares of residuals, for later use in the Goldfeld-Quandt test statistic, it is
assigned to variable mss1. Similarly, lines 16, 17, and 18 select the 17 largest
observations and run the regression, assigning the resulting __rss/14 to the
variable mss2.
Since we are only interested in the RSS from the regressions, the outputs from the
first two estimations are not printed here. Instead, the result of Goldfeld-Quandt test
statistic from line 20 is given as follows:
Goldfeld-Quandt Test Statistic = 1.9444
This statistic is computed according to the formula:
RSS2
N2 − K
RSS1
N1 − K
which follows the F-distribution with N2-K and N1-K degrees of freedom, where N1
and N2 are the number of observations corresponding to the two separate samples,
and K is the number of regressor parameters, including the constant term. Since the
Goldfeld-Quandt method requires that the test statistic be greater than 1, the largest
RSS (RSS2) must be in the numerator. The computed value of 1.94 is smaller than
the critical value of F(14,14) at a 5% level of significance (that is, 2.40), so we could
not reject the hypothesis of homoscedasticity.
131
Although there seems to be no problem of heteroscedasticity in the spending-income

equation based on the Goldfeld-Quandt test, we warn the reader to never totally rely
on statistical results from just one single test. We continue on to show you the
weighted least squares method for correcting heteroscedasticity.
The problem of heteroscedasticity is corrected using weighted least squares. In line

22, the statement:
_weight = 1/income;
tells GPE to estimate the model using 1/INCOME as the weight incorporated in the
regression estimation called on line 23. All the variables, both dependent and
independent, including the constant term, are weighted (or divided by INCOME).
The rationale is that the variable INCOME2 may be used to approximate the
heterogeneous variances of the model.
Because input variables _begin and _end had been set to non-default values
earlier in the program, calling reset again (line 21) is the simplest way to insure
that all control variables are reset to their default values for the new estimation to
come.
Delete line 21 or change it to

@ call reset; @
See what happens. Why?
The regression output corrected for heteroscedasticity is available in the output file
output9.2. Compare the earlier ordinary least squares estimation with the
heteroscedasticity-consistent covariance matrix from Lesson 9.1,
SPENDING = 832.91 - 1834.2 INCOME + 1587.0 INCOME2

s.e. (460.89) (1243.0) (829.99)
and with the weighted least squares estimation using 1/INCOME as the weighting
variable,

s.e. (333.61) (872.07) (563.71)
Notice that the numbers in parentheses are the estimated standard errors of the
coefficients. The results of the two regressions are similar, but in theory, weighted
least squares estimation is more efficient.
Lesson 9.3: Breusch-Pagan and White Tests for Heteroscedasticity

In this lesson, we will briefly explain two other methods to test for
heteroscedasticity: the Breusch-Pagan and White tests for general heteroscedasticity.
Many econometrics references cover the operation of these tests in more detail.
In an auxiliary regression, the Breusch-Pagan test for general heteroscedasticity uses

all explanatory variables including the constant term as the regressors to compute the
test statistic. The test procedure does not rely on any particular variable or functional
form to specify heteroscedasticity. In addition, it does not require the data to be
132
HETEROSCEDASTICITY
sorted. However, the Breusch-Pagan test does assume residual normality for accurate
results. The alternative Koenker-Basset test is more forgiving in regards to the
assumption of normality.
The White test is based on the computation of the heteroscedasticity-consistent

covariance matrix as described in Lesson 9.1, in which the explanatory variables and
their squares and cross products are used to fit the squared residuals in a regression
setting. The resulting goodness of fit is the test statistic.
We note that both the Breusch-Pagan and White tests for general heteroscedasticity
do not offer information about the source and the form of heteroscedasticity. To
correct for this problem, a more specific heteroscedastic structure of the model may
be required.
It is quite easy to implement the Breusch-Pagan and White tests for

heteroscedasticity in GPE. What you need is to set a positive value to the input
control variable _bptest.
/*
** Lesson 9.3: Breusch-Pagan and White Tests
** for Heteroscedasticity
*/
1 use gpe2;
6 data = packr(data); @ deletes row w/mis value @
8 income = data[.,3]/10000;
9 call reset;
/* Breusch-Pagan and White Tests */

11 _bjtest = 1;
12 _bptest = 1;
14 end;
Lines 1 through 10 are similar to lesson9.1. Keeping in mind that the working of the
Breusch-Pagan test assumes that residuals are normally distributed, so we have
included the Bera-Jarque test for residual normality on line 11 (_bjtest=1). Line
12 (_bptest=1) performs the Breusch-Pagan and White tests for general
heteroscedasticity.
Let’s examine the output now. The regression result from the first estimation (line
13) is the same as the result discussed in the previous Lesson 9.2, with additional
pieces of information: the Bera-Jarque test for normality, and the Breusch-Pagan and
White tests for heteroscedasticity:
Bera-Jarque Wald Test for Normality

Asymptotic Standard Error of Residuals = 54.952
Skewness of Residuals = -0.083322
Kurtosis of Residuals = 3.3877
Chi-Sq( 2) Prob>Chi-Sq
0.37107 0.83066
133
Breusch-Pagan and White LM Tests for Heteroscedasticity

Breusch-Pagan Test 18.903 2 7.8553e-005
Koenkar-Basset Test 15.834 2 0.00036454
White Test 21.159 4 0.00029443
The result of the Bera-Jarque test reveals normality in the residuals (refer to Lesson
3.6 for details on the Bera-Jarque test for residual normality). The last section of the
regression output is what we are interested in: the Breusch-Pagan and White tests for
heteroscedasticity. It is set up to test the null hypothesis of homoscedasticity. That is,
for the Breusch-Pagan test, if the computed test value is less than the critical value of
the Chi-square distribution with two degrees of freedom, we fail to reject the
hypothesis that the model error is homogeneously distributed. A similar conclusion
is obtained from the White test, which is based on the R2 statistic of the auxiliary
regression with 4 degrees of freedom. Note the low P-values for both the Breusch-
Pagan and White test statistics, leading us to reject the hypothesis of
homoscedasticity and conclude that heteroscedasticity exits in the model.
Remember the requirement of residual normality for the Breusch-Pagan test? If the
residuals are not normally distributed, we need to use a more general version of the
Breusch-Pagan test, called the Koenkar-Basset test. The closer to normal the
residuals are, the more similar these two test statistics. If absolute normality exists,
the computed values of the two tests will be identical. Since the estimated residuals
are indeed normally distributed for this example as shown earlier, both tests return
rather close values, 18.9 and 15.8, respectively. Our conclusion of heteroscedasticity
is the same from both the Breusch-Pagan and Koenkar-Basset test statistics.
Nonlinear Maximum Likelihood Estimation

A more general approach is to consider the regression model, linear or nonlinear,
with the heteroscedastic variance-covariance matrix
∑ = σ2Ω
where Ω is a positive diagonal matrix representing the proportional variance weight

of each data observation. It is impossible to estimate the general matrix Ω. Usually Ω
is re-parameterized with a vector α of only a few parameters. By assuming a normal
probability distribution for the model’s error term, maximum likelihood estimation is
implemented to estimate the regression parameters and heteroscedastic variances.
Consider a sample of N data observations Z = [Y,X] in fitting the regression model

F(Z,β) = ε, where Y is the left-hand side dependent variable, X is the right-hand side
independent variable, and β consists of the unknown parameters. For each data
observation i, let εi(β) = F(Zi,β) and assume εi(β) is normally independently
distributed with zero mean and positive variance σi2. The log-likelihood function,
ll(β,σ2) = ll(β,σ12,σ22,…,σN2) for brevity, is
ll(β,σ2) = -N/2 ln(2π) - ½ ∑i=1,2,…,N ln(σi2) - ½ ∑i=1,2,…,N (εi(β)2 / σi2)
Given the general form of heteroscedasticity, there are too many unknown
parameters. For practical purposes, some hypothesis of heteroscedasticity must be
assumed:
134
HETEROSCEDASTICITY
σi2 = σ2 h(Xi,α)
where σ2 > 0 and the heteroscedastic function h depends on part or all of the
regressors and a vector of parameters α. Given a specific formulation of
heteroscedasticity, hi(α) = h(Xi,α) for brevity, the log-likelihood function is written
as:
ll(β,α,σ2) = -N/2 (ln(2π) + ln(σ2))

1
- ½ ∑i=1,2,…,N ln(hi(α)) - ∑ (ε (β)2 / hi(α))
2σ2 i=1,2,…,N i
Let εi*(β,α) = εi(β) / hi(α) and substitute out σ2 with ε*(β,α)'ε*(β,α)/N, then
the concentrated log-likelihood function is
ll*(β,α) = -N/2 (1 + ln(2π) + ln(ε*(β,α)'ε*(β,α)/N)) - ½ ∑i=1,2,…,N ln(hi(α))
As the variances must be explicitly estimated, σi2 = σ2 hi(α), the objective log-
likelihood function is inevitably complicated. To maximize the log-likelihood
function, the techniques of nonlinear optimization of Chapter VI are applicable.
Consider the following examples of heteroscedasticity specifications, in which X is

assumed to be a single variable for simplicity. The corresponding functional forms of
h(α) and ε* are defined for the concentrated log-likelihood function:
ll*(β,α) = -N/2 [1 + ln(2π) + ln(ε*'ε*/N)] - ½ ∑i=1,2,…,N ln(hi(α))
σi2 hi(α) εi*

1. σ2(Xiα) Xiα εi / (Xiα)½ (Note: Xiα > 0)
2. σ2(Xiα)2 (Xiα)2 εi / Xiα
3. σ2exp(Xiα) exp(Xiα) εi / exp(Xiα)½
4. σ2(Xiα) Xiα εi / Xiα/2
The last two cases may be expressed in log form:
3. ln(σi2)= ln(σ2) + α Xi
4. ln(σi2)= ln(σ2) + α ln(Xi)
Depending on whether the variable X is log-transformed or not, both cases of

multiplicative heteroscedasticity are essentially the same. For (3) and (4), if α = 0 the
model is homoscedastic. If α = 2, we have case (2).
Lesson 9.4: Multiplicative Heteroscedasticity

Let’s reexamine the heteroscedastic relationship of public school spending
(SPENDING) and income (INCOME) of the previous three lessons:
135
The data set is given in greene.txt. This time we will find and compare the
maximum likelihood estimates based on the following hypothesis of multiplicative
heteroscedasticity:
σi2 = σ2 INCOMEiα
Lesson 9.2 has demonstrated the weighted least squares estimation for the case of α
= 2. The alternative expression of multiplicative heteroscedasticity is
σi2 = σ2 exp(α INCOMEi)
If the variable INCOME is in log form for the alternative expression of

heteroscedasticity, the two forms of multiplicative heteroscedasticity are the same.
/*
** Lesson 9.4: Multiplicative Heteroscedasticity
** Greene (1997), Chap. 12.5
*/
1 use gpe2;
3 load data[52,3]=gpe\greene.txt;
4 data=data[2:52,2]~(data[2:52,3]/10000); @ scale data @
5 data=packr(miss(data,"NA")); @ take care of missing obs @
6 b=data[.,1]/(ones(rows(data),1)~data[.,2]~(data[.,2]^2));
7 call reset;
8 _method=4;
9 _iter=100;
10 _restart=10;
11 _b=b|2.0;
12 _nlopt=1;
13 call estimate(&llf,data);
14 end;
15 proc llf(data,b);
16 local n,y,x,e,h,ll;
17 y=data[.,1]; @ public school spending @
18 x=data[.,2]; @ income @
19 n=rows(y);
20 h=x^b[4]; @ multiplicative hetero @
/*
h=exp(b[4]*x);
*/
21 e=(y-b[1]-b[2]*x-b[3]*(x^2))./sqrt(h);
22 ll=-0.5*n*(1+ln(2*pi)+ln(e'e/n))-0.5*sumc(ln(h));
23 retp(ll);
24 endp;
The first part of the program loads and scales the data, which are the same as in
previous lessons. Line 6 computes the linear model estimates as the starting values of
parameters for nonlinear maximum likelihood estimation (see line 11). The objective
log-likelihood function llf is defined in lines 15 through 23. The specific form of
multiplicative heteroscedasticity is given in line 20. Since the estimation has
experienced some difficulty in improving the function value in its final iterations, we
set _restart=10 in line 10 to restart the iteration in case of failure. If you have
trouble understanding what this program is doing, a review of Chapter VI and the
program lessons on nonlinear models there is recommended.
136
HETEROSCEDASTICITY
The result of maximum likelihood estimation is given in the following:

---------------------------------------------

Tolerance = 0.001
Initial Result:
Parameters = 832.91 -1834.2 1587.0 2.0000

Parameters = 563.08 -1141.3 1150.6 3.2550
Parameters = 560.95 -1124.8 1132.8 3.2986
Parameters = 560.69 -1124.1 1132.4 3.2985
Parameters = 560.69 -1124.1 1132.4 3.2984
Parameters = 560.69 -1124.1 1132.4 3.2985
Final Result:
Parameters = 560.69 -1124.1 1132.4 3.2985
Gradient Vector = 2.0276e-008 0.00000 0.00000 1.7233e-006
Hessian Matrix =
-0.020623 -0.014837 -0.010848 -0.00024046
-0.014837 -0.010848 -0.0080655 9.0987e-005
-0.010848 -0.0080655 -0.0061037 -0.00013901
-0.00024046 9.0987e-005 -0.00013901 -0.58515
Asymptotic
X1 560.69 354.11 1.5834
X2 -1124.1 943.28 -1.1917
X3 1132.4 621.04 1.8233
X4 3.2985 1.3790 2.3920
The estimated public school spending-income relationship is summarized as:

s.e. (354.11) (943.28) (621.04)
In addition, the heteroscedastic variance is related to the variable INCOME as

follows:
σi2 = σ2 INCOMEi3.3
It is easy to modify lesson9.4 for the alternative (exponential) form of multiplicative

heteroscedasticity. Line 20 in the definition of log-likelihood function would be
replaced by:
h=exp(b[4]*x);
137
which is currently ignored by GAUSS within the comment notations (/* */). Make
the change and run the modified program. The resulting regression equation looks
like this,

s.e. (364.97) (974.17) (643.10)
and the variance equation is:
σi2 = σ2 exp(4.23 INCOMEi)
Refer to the output file generated from the program and verify the above regression
results. To summarize the discussion of heteroscedastic regression models, we put
together and compare the estimation results of the public school spending-income
relationship:
(1) (2) (3) (4)

Lesson 9.1 Lesson 9.2 Lesson 9.4 Lesson 9.4
β1 -1834.2 -1399.3 -1124.1 -1076.2
(1243.0) (872.07) (943.28) (974.17)
β2 1587.0 1311.3 1132.4 1097.7
(829.99) (563.71) (621.04) (643.10)
β0 832.91 664.58 560.69 544.53
(460.89) (333.61) (354.11) (364.97)
α 0 2 3.2985 4.2344
(1.3790) (1.7364)
The numbers in parentheses are the estimated standard errors of the parameters. To
recap the essence of each model: (1) Lesson 9.1 is an ordinary least squares with
heteroscedasticity-consistent variance-covariance matrix; (2) Lesson 9.2 is a
weighted least squares using 1/INCOME as the weight (i.e., σi2 = σ2 INCOMEi2); (3)
and (4) are the maximum likelihood estimators with multiplicative heteroscedasticity
(i.e., σi2 = σ2 INCOMEiα and σi2 = σ2 exp(α INCOMEi), respectively) as presented
in Lesson 9.4.
Which one is most accurate? We leave that decision to you.
138
X
Autocorrelation
Autocorrelation is a problem most likely associated with time series data. It concerns
the relationship between previous and current error terms in a regression model. In
the simplest case, the serial correlation is of first order where the correlation between
current and immediate previous errors is nonzero. A more complicated error
structure can include autoregressive (AR) and moving average (MA) terms.
OLS (ordinary least squares) estimation with autocorrelated error structure results in
a loss of efficiency. Therefore, statistical inference using t and F test statistics cannot
be trusted.
In this chapter, we revisit the multiple regression model of U.S. production function,
using the labor (L), capital (K), and output (X) data series of cjx.txt:
ln(X) = β0 + β1 ln(L) + β2 ln(K) + ε
Lesson 10.1 below demonstrates the use of the input control variable _hacv to
obtain a consistent estimator of the variance-covariance matrix, when ordinary least
squares is used. Several tests for the existence of autocorrelation are given in Lesson
10.2. Correction methods for first-order autocorrelation to improve the efficiency of
parameter estimates are presented in Lessons 10.3 and 10.4. Since a more
complicated structure of autocorrelation may be identified, the estimation of
autoregressive and moving average error structures is considered in the last three
lessons. Lesson 10.5 is a model with higher-order autocorrelation. The technique of
maximum likelihood is introduced in Lesson 10.6, while Lesson 10.7 covers the
nonlinear method.
Autocorrelation-Consistent Covariance Matrix

Given the existence of autocorrelation in a regression equation, the ordinary least
squares estimator is unbiased but inefficient. Following from the treatment of the
heteroscedasticity-consistent covariance matrix introduced in the previous chapter,
we can keep the unbiased parameter estimators but correct for the variance-
covariance estimator with an autocorrelation-consistent covariance matrix.
Combining both problems of heteroscedasticity and autocorrelation, the Newey-West
estimator of the heteroscedasticity-autocorrelation-consistent covariance matrix is a
simple approach to deal with an unspecified structure of heteroscedasticity and
autocorrelation. The drawback is that the order of autocorrelation must be
predetermined for the computation of Newey-West estimators.
For a regression model with an unspecified structure of heteroscedasticity and

autocorrelation, the consistent estimator of the variance-covariance matrix is
() ( )
Var β̂ = (X′X ) X′Σ̂X (X′X )
−1 −1
where X is the data matrix of regressors, β̂ is the ordinary least squares estimator of
the parameter vector β , and Σ̂ is the Newey-West covariance estimator for
autocorrelated and possibly heterogeneous disturbances. Refer back to Lesson 9.1 for
more details on the consistent covariance matrix in the context of heteroscedasticity.
Lesson 10.1: Heteroscedasticity-Autocorrelation-Consistent Covariance Matrix

Based on the Cobb-Douglas production model for U.S. manufacturing introduced in
Lesson 3.4, we first repeat the ordinary least squares estimation, printing out the
OLS variance-covariance matrix. Then two more least squares estimations are
performed, one with an estimated autocorrelation-consistent covariance matrix and
the other with a heteroscedasticity-autocorrelation-consistent covariance matrix.
Thus, we can directly compare the results of these regression estimations.
/*
** Lesson 10.1: Heteroscedasticity Autocorrelation
** Consistent Variance-Covariance Matrix
*/
1 use gpe2;
3 load data[40,6]= gpe\cjx.txt;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 call reset;

9 _names = {"X","L","K"};
10 _vcov = 1;

12 _hacv = {0,4};@ with auto consistent var-cov @

14 _hacv = {1,4};@ w/hetero auto consist var-cov @
16 end;
Recall that by setting the GPE input variable _vcov = 1 (see line 10), the
estimated variance-covariance matrix is presented. Instead of using the inefficient
variance-covariance matrix from ordinary least squares, computation of the
consistent covariance matrix is controlled by the input variable _hacv. _hacv is
either a scalar or a two-element vector. The first element of _hacv is reserved for
heteroscedasticity correction as shown earlier in Lesson 9.1, while the second
element is the order of autocorrelation to be considered for the estimator of an
autocorrelation-consistent variance-covariance matrix. Therefore, line 12 of the
program:
_hacv = {0,4};
144
AUTOCORRELATION
will compute the fourth-order autocorrelation-consistent variance-covariance matrix.

The mixture of heteroscedasticity and a fourth-order autocorrelation-consistent
covariance matrix is obtained by setting line 14 to:
_hacv = {1,4};
Why the fourth-order autocorrelation correction? There is no particular reason for

this choice. As a matter of fact, we should try different numbers of orders to
compute all the consistent covariance matrices, to find the proper order that
stabilizes the resulting covariance matrix.
We now analyze the output of three regression estimations. The first least squares
estimation with the option to print out the estimated variance-covariance matrix
(lines 10 to 11) is as follows:

------------------------

Log-Likelihood Function Value = 77.286

Explained 8.0310 2 4.0155 3332.2 1.3921E-041
Residual 0.043382 36 0.0012051
Total 8.0744 38 0.21248

L 1.4508 0.083228 17.431 3.9260E-019 0.89407
K 0.38381 0.048018 7.9930 1.7130E-009 0.63960
CONSTANT -3.9377 0.23700 -16.615 1.8332E-018 0.88464

L 0.0069270
K -0.0038020 0.0023057
CONSTANT -0.018806 0.0092666 0.056169
L K CONSTANT

L 1.0000
K -0.95134 1.0000
CONSTANT -0.95338 0.81428 1.0000
L K CONSTANT
Since autocorrelation is suspected for most time series data, the second regression
estimation (lines 12 and 13) is carried out with the fourth-order autocorrelation-
consistent standard errors and the variance-covariance matrix:

L 1.4508 0.10980 13.213 2.2423E-015 0.82905
145
K 0.38381 0.058368 6.5756 1.1907E-007 0.54568

CONSTANT -3.9377 0.33421 -11.782 6.5278E-014 0.79407

L 0.012055
K -0.0060940 0.0034069
CONSTANT -0.035362 0.016267 0.11170
L K CONSTANT

L 1.0000
K -0.95089 1.0000
CONSTANT -0.96366 0.83388 1.0000
L K CONSTANT
Finally, heteroscedasticity and the fourth-order autocorrelation-consistent covariance

matrix is the outcome of the last regression estimation (lines 13 and 14):

L 1.4508 0.11561 12.549 1.0409E-014 0.81392
K 0.38381 0.062253 6.1653 4.1818E-007 0.51358
CONSTANT -3.9377 0.33788 -11.654 8.9304E-014 0.79048

L 0.013366
K -0.0069673 0.0038755
CONSTANT -0.037946 0.018529 0.11416
L K CONSTANT

L 1.0000
K -0.96804 1.0000
CONSTANT -0.97140 0.88092 1.0000
L K CONSTANT
Putting together the estimated regression equations with the three sets of estimated
standard errors, we have
ln(X) = -3.94 + 1.45 ln(L) + 0.38 ln(K)

s.e. (ols) (0.24) (0.08) (0.05)
s.e. (ac) (0.33) (0.11) (0.06)
s.e. (hac) (0.34) (0.12) (0.06)
In general, the consistent estimators of the covariance matrix are larger than their
ordinary least squares counterparts. The consequence is a higher probability of type
II error (incorrectly accepting the null hypothesis) for the estimators. In this example,
all three estimated variance-covariance matrices of the coefficients are quite similar
in spite of the consistency correction for autocorrelation and heteroscedasticity.
Detection of Autocorrelation
Given each observation of a linear regression model
Yi = Xi β + εi
the linear form of autocorrelated errors is written as:
146
AUTOCORRELATION
ε i = ρ1ε i-1 + ρ2ε i-2 + … + ρpεi-p + υi
where Y is the dependent variable, X is a vector of explanatory independent

variables, and β is the associated parameter vector. The simplest case of
autocorrelated error structure is autocorrelation of order 1 (or AR(1)), where p = 1. It
is always worthwhile to check for the existence of autocorrelation before considering
correcting the problem.
Lesson 10.2: Tests for Autocorrelation

This lesson explores several methods to test for autocorrelation. The most popular
test is the Durbin-Watson bounds test, which is designed to check for first-order
serial correlation. The convenience of the Durbin-Watson bounds test has its limits:
it tests only for first-order autocorrelation and the regression must include a constant
term. Additionally, there cannot be lagged dependent variables in the regression
equation. The Breusch-Godfrey LM test is more forgiving: lagged dependent
variables may be included and it can be used to test for higher orders of
autocorrelation. Checking the autocorrelation and partial autocorrelation coefficients
can reveal a more complicated autoregressive and moving average structure of
autocorrelated errors. This is accomplished with the Box-Pierce and Ljung-Box Q
test statistics. However, except for the Durbin-Watson bounds test, all tests for
autocorrelation need a large sample size to be useful.
The popular Durbin-Watson test statistic is a part of the residual statistics output
reported with the regression estimation. That is, it is available with the statement:
_rstat = 1;
To call the Breusch-Godfrey LM test statistic, we need to specify the order of

autocorrelation to be tested. Indeed, this is a cumulative test for no serial correlation
up to the order specified. Therefore, we need only to find the autocorrelation at a
rather low order to confirm the problem. For example, to test for autocorrelation up
to the fourth order, use the statement:
_bgtest = 4;
Autocorrelation and partial autocorrelation coefficients are typically computed and

plotted for a long period of lags. By examining these coefficients and their
distribution, a pattern of autoregressive and moving average structures of
autocorrelated residuals may be identified. The Chi-square-based Box-Pierce and
Ljung-Box Q test statistics work the same way as the Breusch-Godfrey LM test. For
example, 12-lag autocorrelation and partial autocorrelation functions are called with
the statement:
_acf = 12;
The selection of the fourth-order autocorrelation for the Breusch-Godfrey LM test is

only a suggestion. What we need is the lowest number of orders to test for
autocorrelation. Again, the use of 12 lags for calculating and plotting autocorrelation
and partial autocorrelation coefficients is arbitrary. As a rule of thumb, about one
quarter of the sample size should offer sufficient information concerning the sample
autocorrelation.
147
/*
** Lesson 10.2: Tests for Autocorrelation
*/
1 use gpe2;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 call reset;
9 _names = {"X","L","K"};
10 _rstat = 1;
11 _rplot = 2;
12 _bgtest = 4; @ Breusch-Godfrey, 4th order @
13 _acf = 12; @ auto & partial auto,12 lags @
15 end;
Looking at the program, the use of input control variables _bgtest (line 12) and
_acf (line 13) are new. You may set the value of _bgtest to any order of
autocorrelation for the Breusch-Godfrey LM test, in this case up to 4 orders are
tested. Similarly, for calculating and plotting autocorrelation and partial
autocorrelation functions, the number of lags must be given to the variable _acf. In
this example, 12 lags seem sufficient for a data size of about 40 observations.
To see the plotting of more than one graphs in multiple windows, the following
statement must be included anywhere before calling estimate:
pqgwin many;
Let’s analyze the first part of the estimation result (which is the same as that of
Lesson 10.1), paying attention to the Durbin-Watson test statistic:

The first-order Rho is the estimated first-order serial correlation coefficient

(ρ) which ranges from –1 to 1. At a value of 0.57, we can see that autocorrelation is a
problem. To test the statistical significance of ρ = 0, the Durbin-Watson bounds test
is used here. The computed Durbin-Watson statistic of 0.858 lies below the lower
bound critical value of 1.382 for a regression using 39 observations and 2
explanatory variables (not including constant term) at a 5% level of significance.
The second part of the result concerns the Breusch-Godfrey LM test of

autocorrelation, up to the fourth order:
Breusch-Godfrey LM Test for Autocorrelation

AR( 1) 13.205 1 0.00027923
AR( 2) 20.331 2 3.8471e-005
AR( 3) 22.221 3 5.8666e-005
AR( 4) 22.445 4 0.00016339
148
AUTOCORRELATION
The Breusch-Godfrey LM test is compared with the critical value of the Chi-square
distribution with degrees of freedom equal to the number of orders of
autocorrelation. P-values of each order tested are given to simplify the analysis. It
becomes clear that autocorrelation exists from the very first order upward. Since the
first order exhibits the problem of serial correlation as the Durbin-Watson bounds
test suggests, all LM tests for cumulative higher orders will certainly also identify
autocorrelation.
The last part of the output lists and displays autocorrelation and partial
autocorrelation coefficients for regression residuals up to 12 lags. Standard errors of
these coefficients are useful to spot the significance of lags for autoregressive and
moving average structures of autocorrelated residuals. In addition, both Box-Pierce
and Ljung-Box Q test statistics are computed for each lag. Similar to the Breusch-
Godfrey LM test, these accumulative tests follow a Chi-square distribution with
degrees of freedom corresponding to each individual number of lags, adjusted for the
number of regression coefficients whenever necessary.
Autocorrelation and Partial Autocorrelation Functions

ARMA Model Specification
Mean = -2.8701e-012 Standard Error = 0.033788
Lag AR S.E.(AR) PAR S.E.(PAR) Box-Pierce Ljung-Box
1 0.55892 0.20411 0.55892 0.16013 12.183 13.145
2 -0.046422 0.20438 -0.52183 0.16013 12.267 13.238
3 -0.19341 0.20902 0.27527 0.16013 13.726 14.900
4 0.054712 0.20939 0.12555 0.16013 13.843 15.037
5 0.20105 0.21428 -0.093386 0.16013 15.419 16.937
6 0.047409 0.21455 -0.073459 0.16013 15.507 17.046
7 -0.20382 0.21946 -0.12344 0.16013 17.127 19.122
8 -0.32496 0.23147 -0.18444 0.16013 21.246 24.569
9 -0.20559 0.23610 0.043822 0.16013 22.894 26.822
10 -0.15817 0.23881 -0.36648 0.16013 23.870 28.202
11 -0.23694 0.24476 -0.047891 0.16013 26.059 31.407
12 -0.24987 0.25121 0.010437 0.16013 28.494 35.105
Both moving average and autoregressive processes of lower orders are visibly
identifiable from the significant values of autocorrelation and partial autocorrelation
coefficients, respectively. These include the first lag of autocorrelation as well as the
first and second lags of partial autocorrelation coefficients. Moreover, Box-Pierce
and Ljung-Box Q test statistics confirm the problem of autocorrelation starting from
the first lag.
In summary, all these tests for autocorrelation suggest that our model may need to be
re-specified. Moreover, the correct specification may not involve just the simple
first-order correction. Nevertheless, the next two lessons will explain the correction
mechanisms of autocorrelation for the first-order model. For higher-order
autocorrelation and even the mixture of autoregressive and moving average error
structure, a proper model specification must first be identified.
Correction for Autocorrelation

The GPE package offers several different methods to correct for autocorrelation. The
default is the Prais-Winsten modified Cochrane-Orcutt iterative method, which
applies a scaled data transformation 1-ρ2 to the first observation consistent with an
AR(1) error structure. Due to the use of this transformation, only the estimation of an
AR(1) process is applicable.
149
In GPE, use the input control variable _ar to specify the order of autocorrelation for
estimation. Since the correction mechanism may require several iterations to
complete, another control variable _iter must be set to a large enough number to
achieve convergence. Therefore the following statements are minimal for estimating
a regression model, which corrects for the first-order serial correlation within the
limit of 50 iterations:
_ar = 1;
_iter = 50;
Lesson 10.3 below illustrates the default method of Cochrane-Orcutt iterative

estimation of the AR(1) model.
As an alternative to the Prais-Winsten modification, the original Cochrane-Orcutt

iterative method does not include the first observation or its transformation. It is
simply dropped from the data set used in the regression estimation. Therefore,
adding the statement:
_drop = 1;
will use the traditional Cochrane-Orcutt method to estimate and to correct for an
AR(1) error process. As a matter of fact, this method applies to the autocorrelated
error structure of both first and higher orders. It just drops more observations in the
beginning of the sample, with a cost of inefficiency, partially due to the loss of
degrees of freedom. For estimation with higher-order autocorrelation, it is not
necessary to specify the variable _drop. For example,
_ar = 4;
will estimate and correct for the AR(4) process using the traditional Cochrane-Orcutt
iterative method in which the initial four observations are dropped automatically.
The Cochrane-Orcutt method only converges to a local solution. In rare cases, there
may exist more than one solution for an autocorrelated error structure. The Hildreth-
Lu grid search method guarantees that the global solution will be found for an AR(1)
model. Similar to the Prais-Winsten modification to the original Cochrane-Orcutt
iterative method, the Hildreth-Lu grid search method may include the first
observation with proper transformation. Alternatively, dropping the first observation
is an option with the cost of a decrease in efficiency. Again, we note that the
Hildreth-Lu method applies to an AR(1) model only. The Hildreth-Lu method is
activated by letting:
_method = 2;
Based on the GAUSS program of Lesson 10.3, the Hildreth-Lu grid search method is
introduced in Lesson 10.4, in which the global solution for the estimated AR(1)
model is ensured.
Both the Cochrane-Orcutt iterative and the Hildreth-Lu grid search methods offer the
option of using least squares or maximum likelihood criterion for optimization. They
are the same if the first observation is dropped. However, with the transformed first
observation included, the use of different optimization criteria may result in finding
different solutions, although they are usually close. We have noted the use of the
input control variable _method to select different methods of estimation for an
autocorrelated error structure. _method can be either a scalar or a 2-element vector.
150
AUTOCORRELATION
When employing a 2-element vector, the first element of _method selects the
estimation method, while the second element selects either least squares or
maximum likelihood criterion for optimization. For example,
_method = {0,1};
calls for the estimation of autocorrelation coefficients using the Cochrane-Orcutt

iterative procedure based on maximum likelihood criterion. The default method is
the Cochrane-Orcutt iterative method and the default optimization criterion is the
least squares. See Appendix A for more information.
There is another method implemented for estimating the first-order autocorrelation,

which includes the transformed first observation of data series using maximum
likelihood optimization criterion. This method is the Beach-MacKinnon iterative
maximum likelihood algorithm, specified by:
_method = 1;
Lesson 10.3: Cochrane-Orcutt Iterative Procedure

This lesson walks through the Prais-Winsten modified Cochrane-Orcutt iterative
procedure based on the least square criterion. Unless otherwise specified, GPE
defaults to this method for estimation and correction for a first-order autoregressive
error structure. The most basic model is presented here.
/*
** Lesson 10.3: Cochrane-Orcutt Iterative Procedure
*/
1 use gpe2;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 call reset;
9 _names = {"X","L","K"};
10 _rstat = 1;
11 _ar = 1; @ AR(1) error structure @
12 _iter = 50; @ 50 iter for default C-O @
14 end;
As shown below, the regression output of this program is more complicated than
previous ones without the autocorrelation correction.

------------------------
Order of Autoregressive Errors = 1

Convergence Tolerance Level = 0.00100
151
Using Cochrane-Orcutt Iterative Least Squares Algorithm

Iteration = 0 Sum of Squares = 0.04338 Log Likelihood = 77.28607
Rho = 0.00000
Rho = 0.57053
Rho = 0.66944
Rho = 0.72411
Rho = 0.75755
Rho = 0.77780
Rho = 0.78976
Rho = 0.79670
Rho = 0.80069
Rho = 0.80297
Rho = 0.80427
Rho = 0.80501
Rho Std.Error t-Ratio

0.80501 0.095517 8.4279
NOTE: R-Square, AOV are computed from original series.


Explained 8.0200 2 4.0100 5460.0 1.9940E-045
Residual 0.026440 36 0.00073443
Total 8.0744 38 0.21248

L 1.0680 0.15125 7.0612 2.7289E-008 0.58072
K 0.55812 0.095595 5.8384 1.1428E-006 0.48635
CONSTANT -2.7298 0.45549 -5.9931 7.0980E-007 0.49943

Notice that the second block of output reports the Cochrane-Orcutt iterative results
set in lines 11, 12, and 13 of the program. Every iteration is listed until convergence
is reached. At the end of the iterations, we have the following results: the estimated
Rho, standard error, and t-ratio associated with the first-order serial coefficient. In
this example, the significant non-zero Rho value of 0.805 is used to correct the least
squares regression. In summary,
ln(X) = -2.73 + 1.07 ln(L) + 0.56 ln(K) + ε
152
AUTOCORRELATION
s.e. (0.46) (0.15) (0.10)
ε = 0.805 ε-1
s.e. (0.096)
To call for the estimation of autocorrelation coefficients using Cochrane-Orcutt

iterative procedure based on the maximum likelihood criterion, add the following
line before calling estimate in line 13:
_method = {0,1};
The verbose listing of iterations may be suppressed by setting the control variable
_print = 0. See Appendix A for details.
If the iterative estimation terminates prematurely due to exceeding the maximum

iteration limit, the estimation result may not be reliable. A larger value of _iter
must be given and the model re-estimated.
There is another optional input control variable _tol, which adjusts the
convergence tolerance level. Its default value is set to 0.001.
The remaining regression output is familiar, including the Durbin-Watson test

statistic computed for the model corrected for an AR(1) error structure. Recall from
the test result of Lesson 10.2 that the extent of autocorrelation specification may be
more complicated than the estimated AR(1). Indeed, the first-order Rho of 0.38 for
the corrected model supports the notion that we may need to specify and correct for
higher orders of autocorrelation.
Lesson 10.4: Hildreth-Lu Grid Search Procedure

To be sure that the estimated coefficient of the first-order serial correlation is indeed
a global solution, we need only to specify the estimation method in the previous
program. Modify the program to call and run the Hildreth-Lu grid search procedure
by adding the statement
_method = 2;
in line 12 as follows:
/*
** Lesson 10.4: Hildreth-Lu Grid Search Procedure
*/
1 use gpe2;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 call reset;
9 _names = {"X","L","K"};
10 _rstat = 1;
12 _method = 2; @ H-L method @
153
13 _iter = 50; @ 50 iterations @

15 end;
To call for the estimation of autocorrelation coefficients using the Hildreth-Lu grid
search procedure based on the maximum likelihood criterion, try this in place of line
12:
_method = {2,1};
Running the above program returns many iterations and lengthy output. Remember
that these iterations are controlled by the global variables _iter and _tol. For
viewing the estimation result, we refer readers to the output file output10.4.
The estimated model with AR(1) error structure is summarized as follows:
ln(X) = -2.66 + 1.05 ln(L) + 0.56 ln(K) + ε

s.e. (0.47) (0.15) (0.10)
ε = 0.826 ε-1
s.e. (0.090)
The results are basically the same as those obtained by using the Cochrane-Orcutt
method. Although the Hildreth-Lu grid search is costly in terms of computer
resources, the global nature of the estimated autocorrelation coefficients is superior
to the local solution found with either of the Cochrane-Orcutt methods.
Lesson 10.5: Higher-Order Autocorrelation

As we have seen, many methods and options are available for estimating and
correcting an AR(1) model. For higher-order autocorrelation, the traditional
Cochrane-Orcutt iterative method is used, with a truncated sample in which initial
unusable observations are dropped. Based on the Cochrane-Orcutt method
introduced in Lesson 10.3, the first part of this lesson reexamines the estimated
AR(1) model and tests for possible higher-order autocorrelation. We recall the use of
the Breusch-Godfrey test (_bgtest) and autocorrelation functions (_acf) in
addition to the Durbin-Watson bounds test to check for higher-order problems after
the AR(1) model is estimated. Higher-order problems are identified and corrected in
the second part of this lesson. Further tests for autocorrelation reveal no more
information can be extracted from the estimated residuals.
/*
** Lesson 10.5: Higher-Order Autocorrelation
*/
1 use gpe2;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 call reset;
9 _names = {"X","L","K"};
10 _rstat = 1;
154
AUTOCORRELATION
11 _rplot = 2;
12 _bgtest = 4;
13 _acf = 12;


19 end;
Lines 12 and 13 add the options to perform tests for higher-order autocorrelation.
These include the Breusch-Godfrey LM test up to the fourth order of autocorrelation
and a plot of 12-lag autocorrelation functions. The AR(1) model is re-estimated (line
16) with the following test results:
Breusch-Godfrey LM Tests for Autocorrelation

AR( 1) 5.5093 1 0.018916
AR( 2) 14.853 2 0.00059525
AR( 3) 14.883 3 0.0019193
AR( 4) 15.021 4 0.0046572

Mean = 0.00084871 Standard Error = 0.026364
1 0.36888 0.18061 0.36888 0.16013 5.3068 5.7257
2 -0.29377 0.19247 -0.49754 0.16013 8.6724 9.4552
3 -0.32527 0.20608 0.031588 0.16013 12.799 14.155
4 0.058462 0.20651 0.12368 0.16013 12.932 14.311
5 0.22715 0.21282 -0.015040 0.16013 14.944 16.737
6 -0.026579 0.21290 -0.15509 0.16013 14.972 16.772
7 -0.17667 0.21663 0.061130 0.16013 16.189 18.331
8 -0.15212 0.21935 -0.15998 0.16013 17.092 19.525
9 0.11512 0.22090 0.19394 0.16013 17.608 20.231
10 0.10311 0.22213 -0.19689 0.16013 18.023 20.817
11 -0.14423 0.22451 -0.090591 0.16013 18.834 22.005
12 -0.21271 0.22962 -0.026934 0.16013 20.599 24.685
From the visual display of autocorrelation functions as well as the results of several
test statistics (Breusch-Godfrey LM test, Box-Pierce and Ljung-Box Q tests), higher
orders of autocorrelation, or a mixture of autocorrelation and moving average
processes is suspected. In particular, the coefficients for the first lag of
autocorrelation and the first two lags of partial autocorrelation are still statistically
significantly different from zero. The second part of Lesson 10.5 goes ahead to
estimate the AR(3) model using the traditional Cochrane-Orcutt iterative method.
The possibility of a mixed error structure with a moving average process is discussed
in the next lesson.
Since the option to test for higher orders of autocorrelation is still included in the
program (see lines 12 and 13), the estimated AR(3) model is also tested for problems
of autocorrelation. Here are the estimation and test results with the AR(3) error
structure:

------------------------
155

NOTE: Estimation Range Has Been Adjusted.


Rho = 0.00000 0.00000 0.00000
Rho = 1.0537 -0.81992 0.31827
…
Rho = 1.3540 -0.92057 0.39938

1.3540 0.14121 9.5891
-0.92057 0.21656 -4.2508
0.39938 0.14174 2.8178


Explained 7.0896 2 3.5448 8157.7 3.4129E-045
Residual 0.014340 33 0.00043453
Total 7.1357 35 0.20388

L 1.0433 0.11306 9.2277 1.1644E-010 0.72070
K 0.54676 0.074920 7.2979 2.2460E-008 0.61743
CONSTANT -2.5236 0.38943 -6.4802 2.3638E-007 0.55996

First-Order Rho = -0.0040825

AR( 1) 0.41856 1 0.51766
AR( 2) 0.86990 2 0.64730
AR( 3) 3.3268 3 0.34393
AR( 4) 3.4588 4 0.48416

Mean = 3.4232e-015 Standard Error = 0.020241
1 -0.0039777 0.16667 -0.0039777 0.16667 0.00056961 0.00061843
2 -0.086337 0.16791 -0.086355 0.16667 0.26892 0.30054
3 -0.20318 0.17460 -0.20543 0.16667 1.7551 2.0119
4 -0.0027129 0.17460 -0.016352 0.16667 1.7554 2.0122
5 0.10710 0.17642 0.074897 0.16667 2.1683 2.5184
6 -0.18193 0.18156 -0.23380 0.16667 3.3599 4.0277
156
AUTOCORRELATION
7 0.060756 0.18212 0.073555 0.16667 3.4928 4.2018

8 -0.14740 0.18540 -0.15926 0.16667 4.2749 5.2633
9 0.15411 0.18893 0.092475 0.16667 5.1299 6.4667
10 0.019960 0.18899 0.0049687 0.16667 5.1443 6.4876
11 -0.20234 0.19491 -0.24119 0.16667 6.6182 8.7281
12 -0.24101 0.20302 -0.28134 0.16667 8.7093 12.039
Based on the Breusch-Godfrey LM test up to the fourth order as well as the plot of
the 12-lag autocorrelation function and the associated Box-Pierce and Ljung-Box Q
test statistics, the estimated model with AR(3) error structure is now free of
autocorrelation. The estimated model is superior to that of correcting only for first-
order autocorrelation.
In summary,
ln(X) = -2.52 + 1.04 ln(L) + 0.55 ln(K) + ε

s.e. (0.39) (0.11) (0.08)
ε = 1.35 ε-1 - 0.93 ε-2 + 0.40 ε-3

s.e. (0.14) (0.22) (0.14)
Autoregressive and Moving Average (ARMA) Models: An Introduction

GPE can handle the estimation of a more complicated regression model involving
autoregressive and moving average autocorrelated errors. In addition to the p-th
order autoregressive structure AR(p) discussed earlier such as
ε i = ρ1ε i-1 + ρ2ε i-2 + … + ρpε i-p + υi
the q-th order moving average error structure MA(q) is specified as
ε i = υi - θ1υ i-1 - θ2υ i-2 - … - θqυ i-q
Or in combination with the autoregressive structure, that is ARMA(p,q), we have
ε i = ρ1ε i-1 + ρ2ε i-2 + … + ρpεi-p - θ1υ i-1 - θ2υ i-2 - … - θqυ i-q + υi
where the filtered error term υi is assumed to be normally independently distributed.

Given a mixed error process of AR(p) and MA(q), or ARMA(p,q), the estimation of
the p-element AR and q-element MA parameters is typically carried out by a
nonlinear optimization algorithm. To be more specific, nonlinear least squares or
maximum likelihood is called for in the estimation of an error structure with a
moving average component. Without the moving average specification, the estimated
autoregressive parameters are computed by the methods described earlier in previous
lessons.
With GPE, the following input control variables are relevant to the estimation of an
error structure with autoregressive and moving average components:
• _arma Autoregressive moving average orders

• _nlopt Nonlinear least squares or maximum likelihood
• _method Nonlinear optimization method
• _iter Iteration limit
157
• _conv Convergence criteria

• _tol Convergence tolerance level
Among these control variables, only _arma is new. The other variables are related
to nonlinear model estimation discussed in Chapters VI and VII. The variable
_arma is used to specify the orders of the autoregressive and moving average
components of the error structure. It is a column vector consisting of at least two
elements: the first is the order of the autoregressive component, followed by the
order of the moving average. For example,
_arma = {1,1};
specifies an error process of the first-order autoregressive and first-order moving

average. A pure moving average of q-th order is set by _arma = {0,q}.
Obviously, _arma = {p,0} is identical to _ar = p where p is the order number
of the autocorrelation. For parameter estimation, providing the initial guess values of
autoregressive and moving average parameters will be helpful for the convergence.
These values could be appended to the variable _arma, in which the first two
elements are always the order of the respective process. We note that the estimated
nonlinear ARMA error structure is conditional upon data initialization for the
beginning observations necessary to compute the process. For convenience, in GPE,
we initialize the data series with the sample mean of model errors.
Lesson 10.6: ARMA(1,1) Error Structure

In normal situations, a higher order of autoregressive structure can be specified with
a lower order moving average structure, and vice versa. In order to keep the
estimation simple, we recommend the use of a lower order model. Of course, this is a
matter of taste and also depends on the theory and computational experience. Earlier
lessons on estimating the U.S. production function suggest that the model is better
described with an AR(3) error structure (see Lesson 10.5). An alternative would be
to estimate the error process using a MA(1) structure or a mixed structure of
autoregressive and moving average at a lower order, say, ARMA(1,1). Lesson 10.6
demonstrates the estimation of an ARMA(1,1) model.
/*
** Lesson 10.6: ARMA(1,1) Error Structure
*/
1 use gpe2;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 call reset;
9 _names = {"X","L","K"};
10 _rstat = 1;
11 _rplot = 2;
12 _bgtest = 4;
13 _acf = 12;
14 _arma = {1,1}; @ ARMA(1,1) error structure @

15 _nlopt = 1; @ maximum likelihood estimation @
16 _method = 5; @ QHC optimization method @
158
AUTOCORRELATION

19 end;
To double check for autocorrelation after the specification of ARMA(1,1) is

estimated, we keep _bgtest and _acf (lines 12 and 13) in the program.
The ARMA(1,1) structure is estimated with the nonlinear maximum likelihood

method using the quadratic hill-climbing algorithm. It takes 34 iterations to achieve
the convergence. Here is the abridged regression output of the estimated Cobb-
Douglas production function with an ARMA(1,1) error structure:

------------------------
Maximum Likelihood Estimation for Nonlinear Error Structure

ARMA(1 ,1 ) Autoregressive Moving Average Process

Tolerance = 0.001
Initial Result:
Parameters = 1.4508 0.38381 -3.9377 0.00000 0.00000

Iteration = 1 Step Size = 2.3579 Log Likelihood = 89.926
Parameters = 1.4410 0.38615 -3.8967 0.45558 -0.46020
Iteration = 34 Step Size = 1.0000 Log Likelihood = 93.016
Parameters = 1.1051 0.54833 -2.8796 0.62424 -0.67141
Final Result:
Parameters = 1.1051 0.54833 -2.8796 0.62424 -0.67141
Gradient of Log Likelihood = 0.022545 0.019469 0.0044089 -
0.00020935 0.00015670
Parameter Std.Error t-Ratio

AR(1 ) 0.62424 0.22900 2.7259
MA(1 ) -0.67141 0.17318 -3.8770


Explained 8.0216 2 4.0108 7456.8 7.4158E-048
Residual 0.019363 36 0.00053787
Total 8.0744 38 0.21248
159

L 1.1051 0.16031 6.8934 3.0026E-008 0.54923
K 0.54833 0.083877 6.5373 9.3096E-008 0.52285
CONSTANT -2.8796 0.54958 -5.2397 5.8625E-006 0.41314


AR( 1) 0.50222 1 0.47852
AR( 2) 0.86082 2 0.65024
AR( 3) 5.3236 3 0.14958
AR( 4) 5.4880 4 0.24079

1 -0.0037201 0.16013 -0.0037201 0.16013 0.00053974 0.00058235
2 -0.049522 0.16052 -0.049537 0.16013 0.096186 0.10657
3 -0.23055 0.16880 -0.23150 0.16013 2.1691 2.4674
4 0.064036 0.16942 0.060518 0.16013 2.3290 2.6547
5 0.24181 0.17805 0.23366 0.16013 4.6095 5.4047
6 -0.13491 0.18065 -0.20050 0.16013 5.3193 6.2865
7 0.032581 0.18080 0.086890 0.16013 5.3607 6.3396
8 -0.26120 0.19023 -0.18931 0.16013 8.0215 9.8587
9 0.18387 0.19474 0.11395 0.16013 9.3400 11.661
10 0.0024843 0.19474 -0.030052 0.16013 9.3403 11.661
11 -0.10407 0.19616 -0.16014 0.16013 9.7627 12.280
12 -0.13637 0.19857 -0.093640 0.16013 10.488 13.381
If you prefer to use nonlinear least squares instead of maximum likelihood to

estimate the ARMA model, just delete line 15 or change it to:
_nlopt = 0;
The result should be the same. Why?
Summarizing, the estimated model with ARMA(1,1) error structure is:
ln(X) = -2.88 + 1.11 ln(L) + 0.55 ln(K) + ε

s.e. (0.55) (0.16) (0.08)
ε = 0.62 ε-1 + 0.67 υ-1

s.e. (0.23) (0.17)
Both the parameter estimates of AR(1) and MA(1) are statistically significant and
useful for correcting autocorrelation. Based on the Breusch-Godfrey LM test up to
the fourth order, as well as the 12-lag autocorrelation function plot and the
associated Box-Pierce and Ljung-Box Q test statistics, the estimated ARMA(1,1)
model is as good as that of the AR(3) specification presented in the Lesson 10.5.
Both models are essentially equivalent.
160
AUTOCORRELATION

It is clear that a regression model with autocorrelation is intrinsically a nonlinear
model. Even with the basic linear regression equation, ε = Y – Xβ, the functional
form of model error for estimation is nonlinear in the parameters. Consider the
AR(1), MA(1), and ARMA(1,1):
Residual Residual Function

AR(1) υi = εi – ρ εi-1 (Yi – ρYi-1) – (Xi – ρXi-1)β
MA(1) υi = εi + θ υi-1 (Yi – Xiβ) + θ υi-1
ARMA(1,1) υi = εi – ρ εi-1 + θ υi-1 (Yi – ρYi-1) – (Xi – ρXi-1)β + θ υi-1
The nonlinearity of AR(1) is clearly due to the product of parameters β and ρ, while
MA(1) is recursively weighted by θ. ARMA(1,1) is a mixed process of AR(1) and
MA(1), and therefore contains both of aforementioned nonlinearities. For model
estimation, the beginning observation of data series may be lost if not properly
initialized. The built-in ARMA estimation of GPE is conditional upon the simple
data initialization with the sample mean. We have seen the Prais-Winsten
transformation for the first observation of AR(1): 1-ρ2 Y1 and 1-ρ2 X1. This adds
more nonlinear complexity into the model and makes maximum likelihood the
preferred method for estimation.
Given N sample data observations of Z = [Y,X], the concentrated log-likelihood

function for the AR(1) model is
ll*(β,ρ) = -N/2 (1+2π-ln(N)) –N/2 ln(υ'υ) + ln(1-ρ2)
For MA(1), the recursive process starts with the initial residual υ0 which is typically
set to its expected value (i.e., zero) or the sample mean. An alternative is to estimate
υ0 directly. The concentrated log-likelihood function is simpler but conditional to the
initialization of υ0 as follows:
ll*(β,θ,υ0) = -N/2 (1+2π-ln(N)) –N/2 ln(υ'υ)
The concentrated log-likelihood function of the mixed process ARMA(1,1) is similar

to that of AR(1) in which the residual function depends on both ρ and θ (in addition
to β) and is subject to the initialization of υ0 and transformation for the first data
observation.
Lesson 10.7: Nonlinear ARMA Model Estimation

We continue with the previous example of U.S. Cobb-Douglas production function
and estimate the three autoregressive error structures: AR(1), MA(1), and
ARMA(1,1). Using the method of nonlinear maximum likelihood, the regression
parameter β and autoregressive coefficients ρ and/or θ are estimated jointly. For
MA(1) and ARMA(1,1), the initialization of residuals with zero expected value is
applied. As we have mentioned earlier, the nonlinear method may produce different
results compared with the linear iterative approximations as employed in the
Cochrane-Orcutt approach (Lesson 10.3) or the GPE built-in method of conditional
nonlinear maximum likelihood (Lesson 10.6). However, the results are not
drastically different.
161
This lesson differs from the previous ones in that the program setup is for nonlinear
maximum likelihood estimation. Nonlinear maximum likelihood was covered in
Chapter VII, and it is helpful to go back for a thorough review of the estimation
technique.
For each of three autoregressive models, the residual function must be defined in
order to maximize the objective log-likelihood function. The block from line 34 to
line 41 defines the residual function for an AR(1) model. The MA(1) specification is
given in lines 42 through 48, while the ARMA(1,1) is specified from line 49 to line
57. Notice that AR(1) and ARMA(1,1) require the use of Jacobians in the likelihood
function. The Jacobian function is defined in the block from line 29 to line 33.
/*
** Lesson 10.7: Maximum Likelihood Estimation
** AR(1), MA(1), ARMA(1,1)
*/
1 use gpe2;
4 year = data[2:40,1];
5 X = ln(data[2:40,2]);
6 L = ln(data[2:40,3]);
7 K = ln(data[2:40,5]);
8 data=X~L~K;
@ OLS estimates as initial values @

9 b=data[.,1]/(ones(rows(data),1)~data[.,2:3]);
10 call reset;
11 _nlopt=1;
12 _method=0;
13 _iter=100;
14 _conv=1;
15 _b=b|0.5;
16 _jacob=&jcb;
17 _names = {"CONSTANT","LN(L)","LN(K)","AR(1)"};
18 call estimate(&ar,data);
19 _b=b|0;
20 _jacob=0;
21 _names = {"CONSTANT","LN(L)","LN(K)","MA(1)"};
22 call estimate(&ma,data);
23 _b=b|0.5|0;
24 _jacob=&jcb;
25 _names = {"CONSTANT","LN(L)","LN(K)","AR(1)","MA(1)"};
26 call estimate(&arma,data);
27 end;
28 proc jcb(x,b); @ jacobian for AR(1) and ARMA(1,1) @

29 local j;
30 j=ones(rows(x),1);
31 j[1]=sqrt(1-b[4]^2);
32 retp(j);
33 endp;
34 proc ar(x,b);
35 local n,e,u;
36 n=rows(x);
37 e=x[.,1]-b[1]-b[2]*x[.,2]-b[3]*x[.,3];
162
AUTOCORRELATION
38 u=e-b[4]*lagn(e,1);
@ first obs transformation @
39 u[1]=sqrt(1-b[4]^2)*e[1];
40 retp(u);
41 endp;
42 proc ma(x,b);
43 local n,e,u;
44 n=rows(x);
45 e=x[.,1]-b[1]-b[2]*x[.,2]-b[3]*x[.,3];
46 u=recserar(e,e[1],b[4]); @ u[1]=e[1] since u[0]=0 @
/*
@ recursive computation of errors using @
@ built-in RECSERAR is the same as below: @
u=e; @ initialize: u[1]=e[1] @
i=2;
do until i>n;
u[i]=e[i]+b[4]*u[i-1];
i=i+1;
endo;
*/
47 retp(u);
48 endp;
49 proc arma(x,b);
50 local n,e,u,v;
51 n=rows(x);
52 e=x[.,1]-b[1]-b[2]*x[.,2]-b[3]*x[.,3];
53 u=e-b[4]*lagn(e,1);
@ first obs transformation @
54 u[1]=sqrt(1-b[4]^2)*e[1];
55 v=recserar(u,u[1],b[5]);
56 retp(v);
57 endp;
Using the linear least squares estimates as initial values of parameters (line 9), lines
15-18 carry out the estimation for the AR(1) model. Here is the result:

-----------------------------

Tolerance = 0.001
Initial Result:
Parameters = -3.9377 1.4508 0.38381 0.50000
Final Result:
Gradient of Log Likelihood = 4.9312e-005 0.00023054 0.00020964 2.6217e-
005
Asymptotic
CONSTANT -2.8220 0.56933 -4.9567
LN(L) 1.0926 0.17753 6.1546
LN(K) 0.54995 0.096866 5.6775
163
AR(1) 0.77832 0.13059 5.9599
The BHHH method is used for log-likelihood function maximization (see line 12),
and it converges in 17 iterations. As seen in line 15, the initial value of the AR(1)
coefficient is set to 0.5. The solution is close to that of the Cochrane-Orcutt (Lesson
10.3) and Hildreth-Lu (Lesson 10.4) procedures. The crucial point of this model is
the use of first observation transformation (line 39) and the resulting Jacobian
function must be incorporated for the exact maximum likelihood estimation.
Similarly, the MA(1) model is estimated in lines 19-22 with the starting value of the
MA(1) coefficient at 0. Here is the estimation result:

-----------------------------

Tolerance = 0.001
Initial Result:
Parameters = -3.9377 1.4508 0.38381 0.00000
Final Result:
Gradient of Log Likelihood = -0.00016182 -0.00094511 -0.00082061 4.1086e-
005
Asymptotic
CONSTANT -3.6176 0.27477 -13.166
LN(L) 1.3379 0.094876 14.102
LN(K) 0.44264 0.054316 8.1495
MA(1) -0.81620 0.095539 -8.5431
As the MA(1) model does not use the first-observation transformation of AR(1), the
Jacobian function should not be called (see line 20). The residual function is defined
with an autoregressive recursive series using the GAUSS built-in function
recserar (line 46). The initialization of the recursive series is the expected value
of the series, which is zero. Line 46 shows the use of recserar with initialization.
Check the GAUSS manual or online help for more information about the procedure
recserar. The computation of autoregressive recursive series is also explained in
the comment block immediately below line 46.
Conditional to the initialization of the recursive moving average series
υi = (Yi – Xiβ) + θ υi-1
we have obtained the maximum likelihood estimates of β and θ as shown above. For
i = 1, υ0 = 0 is assumed. The alternative is to estimate υ0 together with β and θ.
Simply replace line 46 with the following:
164
AUTOCORRELATION
u=recserar(e,e[1]+b[5]*b[4],b[4]);
where b[5] is the unknown υ0 to be estimated with the rest of the parameters.
Starting from the initial guess of (θ,υ0) at (0,0), in addition to the linear least squares
estimator of β, the model is estimated exactly the same way as before except that
line 19 should be:
_b=b|0|0;
How do the estimation results differ from those of Lesson 10.7, which assumed υ0 =
0? We leave this question to you.
For ARMA(1,1), both the first observation transformation of AR(1) and the
autoregressive recursive series with initialization of MA(1) are required. The model
is estimated in lines 23-26. Here is the estimation result:

-----------------------------

Tolerance = 0.001
Initial Result:
Parameters = -3.9377 1.4508 0.38381 0.50000 0.00000
Final Result:
Gradient of Log Likelihood = -0.00012631 -0.00065739 -0.00055447
0.00015394 -1.6021e-006
Asymptotic
CONSTANT -2.6041 0.42941 -6.0644
LN(L) 1.0321 0.12537 8.2320
LN(K) 0.57271 0.071019 8.0641
AR(1) 0.66145 0.14519 4.5559
MA(1) -0.71077 0.12402 -5.7309
Again, for estimating the above ARMA(1,1) model:
υi = (Yi – ρYi-1) – (Xi – ρXi-1)β + θ υi-1
υ0 = 0 is assumed for i = 1. The alternative is to estimate υ0 together with β, ρ, and

θ. Simply replace line 55 with the following:
v=recserar(u,u[1]+b[6]*b[5],b[5]);
where b[6] is the unknown υ0 to be estimated with the rest of the parameters.
Starting from the initial guess of (ρ,θ,υ0) at (0.5,0,0), in addition to the linear least
squares estimator of β, the model is estimated exactly the same way as before except
165
that line 23 should be changed to:

_b=b|0.5|0|0;
To summarize the discussion of autoregressive regression models, we put together

and compare the estimation results of the U.S. Cobb-Douglas production function:
ln(X) = β0 + β1 ln(L) + β2 ln(K) + ε
where X is output, and the two factor inputs are labor L and capital K. The following
table presents the parameter estimates (numbers in parentheses are the estimated
standard errors of the parameters) and the corresponding log-likelihood function
value ll for each model.
(1) (2) (3) (4) (5) (6) (7)

Lesson Lesson Lesson Lesson Lesson Lesson Lesson
10.1 10.3 10.4 10.7 10.7 10.7 10.6
β1 1.451 1.068 1.050 1.093 1.338 1.032 1.105
(0.083) (0.151) (0.153) (0.178) (0.095) (0.125) (0.160)
β2 0.384 0.558 0.563 0.550 0.443 0.573 0.548
(0.048) (0.096) (0.098) (0.097) (0.054) (0.071) (0.084)
β0 -3.94 -2.73 -2.66 -2.82 -3.62 -2.60 -2.88
(0.237) (0.456) (0.465) (0.569) (0.275) (0.429) (0.550)
ρ 0 0.805 0.826 0.778 0 0.661 0.624
(0.096) (0.090) (0.131) (0.145) (0.229)
θ 0 0 0 0 -0.816 -0.711 -0.671
(0.096) (0.124) (0.173)
ll 77.286 86.420 86.379 86.438 90.226 93.879 93.016
The top row of the table identifies the model and its corresponding lesson: (1)
Lesson 10.1 is the ordinary least squares estimates without autocorrelation
correction; (2) Lesson 10.3 is the AR(1) model using the Cochrane-Orcutt iterative
procedure; (3) Lesson 10.4 is the same AR(1) model using Hildreth-Lu grid search
method; (4), (5), and (6) are based on Lesson 10.7, using nonlinear maximum
likelihood estimation for the model AR(1), MA(1), and ARMA(1,1), respectively.
The last column (7) is the ARMA(1,1) model estimated with the GPE built-in
conditional maximum likelihood method in Lesson 10.6. All the methods use the
entire sample of 39 observations from 1929 to 1967. For model comparison, the
statistics of pair-wise Likelihood Ratio will be useful. It is immediately clear that the
model must be corrected for autocorrelation. The plain OLS model (1) is rejected
based on LR tests with all the other models. Finally, the structure of the
autoregressive moving average ARMA(1,1) of both Lessons 10.6 and 10.7 cannot be
rejected.
166
XI
Distributed Lag Models
With the proper use of distributed lags, regression models can be expanded to
include dynamic features such as long-run and short-run elasticities and multipliers
for analysis. In this chapter we will consider two popular setups of distributed lags:
geometric, or Koyck lags, and polynomial, or Almon lags. The former is an infinite
distributed lags model with a geometric declining pattern, which in turn can be
transformed into a lagged dependent variable model. The latter is a finite distributed
lags model with polynomial functional restrictions. The combination of the two is the
so-called autoregressive distributed lag (ARDL) model.
Lagged Dependent Variable Models

Applications of lagged dependent variable models include partial adjustment and
adaptive expectation estimations. These models relate the long-run and short-run
behavior of influential variables. However, regression models with lagged dependent
variables may possess some undesirable characteristics. Possible problems include
correlated error terms and random regressors. The resulting least squares estimation
is biased, inconsistent, and inefficient. The lagged dependent variable model is
considered in Lesson 11.1.
For estimating such a model with lagged dependent variables, instrumental variable
(IV) estimation is suggested. IV weighs the trade-off between “biasedness” and
“inefficiency” and obtains a “consistent” parameter estimator which minimizes the
ill effects of using lagged dependent variables. Instrumental variable estimation for a
lagged dependent variable model is the focus of Lesson 11.2.
Lesson 11.1: Testing for Autocorrelation with Lagged Dependent Variable

A classical consumption-income relationship based on the Permanent Income
Hypothesis is a good example to demonstrate the construction, testing, and
estimation of the lagged dependent variable model:
C = β0 + β1Y + β2 C-1 + ε
where C is consumption and Y is income. Assuming a simple partial adjustment of

short-run consumption towards its long-run equilibrium, the lagged dependent
variable C-1 is included in the model specification.
To estimate the above consumption-income equation, we introduce a new data file

usyc87.txt. This data file consists of three variables: YEAR, Y, and C. YEAR is just
the time indicator ranging from 1929 to1994. Y is personal disposable income, and C
is personal consumption expenditure. Both income and consumption time series are
expressed in billions of 1987 dollars. In total there are 66 observations.
In this lesson we will estimate a lagged dependent variable model. The only new
input control variable is _dlags, which is used to specify the number of lags
desired. For example, by setting _dlags = 1, GPE inserts the first lag of the
dependent variable into the regression equation for least squares estimation. The
lagged dependent variable is added in front of the other explanatory variables. The
Durbin-H test statistic is automatically included in the output of _rstat when
_dlags is set to a value greater than zero.
/*
** Lesson 11.1: Lagged Dependent Variable Model
** Estimation and Testing for Autocorrelation
*/
1 use gpe2;
3 load z[67,3] = gpe\usyc87.txt;
4 y = z[2:67,2];
5 c = z[2:67,3];
6 call reset;
7 _names = {"c","y"};
8 _rstat = 1;
9 _dlags = 1;
10 call estimate(c,y);
11 _ar = 1;
12 _iter = 50;
14 end;
If more than one lag is needed, just change the value of _dlags to the desired
positive number of lags.
To estimate the model is simple, but to evaluate and interpret the effect of a lagged
dependent variables is not. Line 9 specifies that the model to be estimated includes
the first lag of the dependent variable. The following call to estimate (line 10)
proceeds to carry out least squares estimation of the model. Since _rstat is set to 1
in line 8, a summary of residual statistics including the new Durbin-H test statistic is
presented.
Alternatively, you can create the lagged dependent variable and then include it in
estimate as an independent variable. In GAUSS, the lagged variable is
constructed with the command lag1 or lagn. This method requires the extra step
of handling the initial observation lost from lagging the dependent variable explicitly
so that the program will run. Setting the GPE control variable _begin to the
beginning of the usable sample for estimation may be necessary. In addition, GPE
will not treat the variable you created differently from the rest of the explanatory
variables. Therefore, the Durbin-H test statistic, unique to the lagged dependent
variable model, is not computed. In passing, we note that testing linear restrictions
involving lagged dependent variables requires specifying restrictions on those
variables explicitly.
The result of the first least squares estimation (line 10) is given below:

------------------------
Dependent Variable = C
168
DISTRIBUTED LAG MODELS

Lagged Dependent Variables Used = 1


Explained 5.8344E+007 2 2.9172E+007 42648. 4.9624E-098
Residual 42410. 62 684.03
Total 5.8387E+007 64 9.1229E+005

C-1 0.69465 0.057390 12.104 5.6098E-018 0.70265
Y 0.29660 0.051205 5.7925 2.4894E-007 0.35114
CONSTANT -0.56350 6.4092 -0.087920 0.93022 0.00012466

Sum of Squared Residuals = 42410.
Durbin-H Statistic = 4.1974
Notice that the estimation range given in the first block of the output is from 2 to 66,
using 65 observations. This is because of the use of the first lag of the dependent
variable on the right-hand side of the regression equation. Next is a statement giving
the number of lags included in the estimation.
The last line of the first block of output is the Durbin-H statistic. Given the first-
order Rho at 0.45 with the Durbin-H test statistic as high as 4.2 (comparing with the
critical values of a standardized normal distribution), the problem of autocorrelation
is readily apparent. Methods of correction for autocorrelation discussed in the
previous chapter should be used to improve the results of the model estimation.
Lines 11 to 13 of the program correct and then re-estimate the model with a first-
order autocorrelated error structure. The default Cochrane-Orcutt iterative method is
used. We refer the reader to the output file output11.1 for details of the
regression results. In summary, here is our estimated lagged dependent variable
model with AR(1) error structure:
C = -6.284 + 0.487 Y + 0.484 C-1 + ε

s.e. (14.61) (0.065) (0.073)
ε = 0.648 ε-1
s.e. (0.098)
The correction for first-order autocorrelation is certainly a right step to improve the
model. It may not be a bad idea to continue to carry out testing for higher orders of
autocorrelation. Remember to use the following statements before the last call of
169
estimate in the above program:

_bgtest = 4;
_acf = 12;
With the results not shown here, we did not find a significant autocorrelation
problem at higher orders.
Lesson 11.2: Instrumental Variable Estimation

The more serious specification problem related to the lagged dependent variable
model is random regressors. In Lesson 11.1, although the problem of autocorrelation
seems easy to resolve by correcting the first-order serial correlation, the estimated
parameters are still biased and inefficient. These problems are due to the ill effects of
random regressors that may be involved through the use of lagged dependent
variables.
To handle a lagged dependent variable model estimation with instrumental variables,

GPE implements an estimation technique which uses the current and lagged
explanatory variables as instruments for the lagged dependent variable. If longer lags
of the dependent variable are adopted or an autocorrelated error structure is
identified, the instrumental variable estimation procedure may need to include more
lags of the explanatory variables as well. In the context of a lagged dependent
variable model, instrumental variable estimation is activated by setting the following
input control variable:
_ivar = 1;
The alternative is to specify the entire data matrix for the variable _ivar. This is
useful for applying instrumental variable estimation in other contexts such as
measurement error in the regression model. We note that the matrix specification of
_ivar requires that its size (rows and columns) to be at least as large as that of the
data matrix of explanatory variables.
We now continue on from the end of Lesson 11.1, adding the option to estimate the
model using instrumental variables in the following program:
/*
** Lesson 11.2: Lagged Dependent Variable Model
** Instrumental Variable Estimation
*/
1 use gpe2;
3 load z[67,3]=gpe\usyc87.txt;
4 y=z[2:67,2];
5 c=z[2:67,3];
6 call reset;
7 _names={"c","y"};
8 _rstat=1;
9 _dlags=1;
10 _ar=1;
11 _iter=50;
13 _ivar=1;
15 end;
170
The only new addition to the program is line 13:

_ivar = 1;
which calls for the use of internal instrumental variables that GPE will construct for
the lagged dependent variables. In addition, the autocorrelation correction is
requested for the first order (line 10), hence one additional lag of explanatory
variables is needed as part of the instrumental variables.
Alternatively, line 13 can be replaced with the explicitly defined instrumental

variables as follows:
_ivar = y~lagn(y,1)~lagn(y,2)~ones(rows(y),1);
_begin = 3;
The advantage is that you have more control over the addition of relevant
instrumental variables in order to improve the small-sample properties of the
estimator. In contexts other than the lagged dependent variable model, instrumental
variable estimation may be requested with the variable _ivar explicitly assigned to
a data matrix no smaller than that of explanatory variables.
We note that the scalar definition of _ivar = 1 will only work when specifying a
positive number of _dlags. _ivar = 1 without _dlags (or _dlags = 0)
will result in a program error.
Looking at the output file output11.2, the results of the first regression in this
lesson are the same as the results of the second regression of Lesson 11.1. The
second estimation of this lesson performs instrumental variable estimation while at
the same time correcting for first-order serial correlation. We will show you only a
partial result of the second regression estimation:

------------------------

NOTE: Lagged Dependent Variables Used = 1
NOTE: Instrumental Variables Used = 4


Iteration = 0 Sum of Squares = 50104.92594 Log Likelihood = -304.02779
Rho = 0.00000
Rho = 0.63232
…
Rho = 0.66806

0.66806 0.097245 6.8699
171


Explained 5.7131E+007 2 2.8566E+007 56110. 2.6192E-100
Residual 31055. 61 509.10
Total 5.7228E+007 63 9.0838E+005

C-1 0.45966 0.094109 4.8843 7.8449E-006 0.28114
Y 0.50728 0.084052 6.0353 1.0194E-007 0.37388
CONSTANT -3.8035 16.185 -0.23501 0.81499 0.00090457

Sum of Residuals = -1.67559E+001
Because instrumental variables are used essentially to replace lagged dependent

variables in the model estimation, testing and correction for autocorrelation is now
the same as in the classical model. The Durbin-Watson test statistic can be applied
as usual.
The model uses four instrumental variables: the original, the first and second lags of
the explanatory independent variable Y, and the constant term. The second lag is
included due to the first-order serial correlation being specified for model estimation.
Comparing the estimation results obtained when instrumental variables are not used,
C = -6.284 + 0.487 Y + 0.484 C-1 + ε

s.e. (14.61) (0.065) (0.073)
ε = 0.648 ε-1
s.e. (0.098)
with our estimated model using instrumental variables,
C = -3.804 + 0.507 Y + 0.460 C-1 + ε

s.e. (16.185) (0.084) (0.094)
ε = 0.668 ε-1
s.e. (0.097)
we see that their parameter estimates are similar. But the current estimated standard
errors of the parameters are slightly larger than the standard errors resulting from not
using the instrumental variables. Nevertheless, the conclusions of statistical
inferences are not affected in this example.
172
We will keep the interpretation of the estimated model using instrumental variables
brief. First, the short-run marginal propensity to consume is 0.51. With the estimated
coefficient 0.46 for the lagged dependent variable, the long-run consumption change
is about 0.94 for each dollar increase of income. To realize 50% of the total effect
(that is, half of 0.94 or 0.47) will take 0.89 years. This is the concept of median lag
frequently used in dynamic analysis. The other measurement is the lag-weighted
average or the mean lag, which is computed at about 0.85 years.
Remember the formula for computing the median lag and mean lag? Let λ be the
estimated parameter of the lagged dependent variable. Then the median lag is
computed as ln(0.5) , and the mean lag is λ .
ln(λ ) 1− λ
Polynomial Lag Models

By imposing polynomial functional restrictions on the finite distributed lags of some
or all explanatory variables, the dynamic model can be estimated with traditional
restricted least squares methodology. In this model we specify the polynomial
distributed lag structure for each explanatory variable. Given the number of
distributed lags q, a polynomial function of order p is used to describe p-1 number of
turning points in the lag structure. In addition, we can add end-point restrictions to
“tie down” the effects of distributed lags at either or both ends. For a polynomial lag
model to be meaningful, p must be greater than 1, and the total number of lags (q)
must be greater than the polynomial orders (p). If q equals p, the distributed lag
model is without polynomial restrictions.
In GPE, a polynomial lag model is defined with the input control variable _pdl. The
variable _pdl is a 3-column matrix with the number of the rows corresponding to
the number of explanatory variables. A polynomial lag structure must be defined for
each variable (row) in the _pdl matrix. The 3-column entry for each explanatory
variable must be separately called out for the lags q, polynomial orders p, and end-
point restrictions r in that order. End-point restrictions “tie down” one or both ends
of the polynomial’s curve, enforcing a theoretical or empirical justification of the lag
structure. For variables that do not have the polynomial lag structure, the 3-column
entry 0 0 0 should be used. Normally the constant term is not included in the
_pdl matrix, unless otherwise specified.
Estimating a model with a polynomial lag structure defined for each explanatory
variable is essentially the same as restricted least squares. The number of restrictions
is (q-p) polynomial restrictions plus the number of end-point restrictions. Any
additional linear restrictions imposed with the variable _restr must take into
account the correct structure of right-hand side variables that _pdl may add to the
original equation.
Lesson 11.3: Almon Lag Model Revisited

In this lesson, we will revisit the relationship between appropriations (CAPP) and
capital expenditure (CEXP) first introduced in Lesson 4.1. The data file almon.txt
provides quarterly data series containing the values of CEXP and CAPP in the period
from 1953 to 1967. In the following, we try to duplicate the original study of S.
Almon published in 1965. Almon’s regressions were computed using a fourth-order
polynomial and lags extending in seven periods. She also included seasonal dummy
173
variables whose four coefficients were constrained to sum to zero. Also, both end-
points are restricted on the polynomial lags.
/*
** Lesson 11.3: Polynomial Distributed Lag Model
** Almon Lag Model Revisited
*/
1 use gpe2;
4 cexp = almon[2:61,2];
5 capp = almon[2:61,3];
6 qt = almon[2:61,1];
7 q1=(qt%10).==1; @ quarterly seasonal dummies @

8 q2=(qt%10).==2;
9 q3=(qt%10).==3;
10 q4=(qt%10).==4;
11 call reset;
12 _rstat = 1;
13 _end = 36;
/* restrictions on all dummy variables */

14 _const = 0;
15 _restr = {0 0 0 0 0 0 0 0 1 1 1 1 0};
16 _pdl = {7 4 2,
0 0 0,
0 0 0,
0 0 0,
0 0 0};
17 _names={"cexp","capp","q1","q2","q3","q4"};
18 call estimate(cexp,capp~q1~q2~q3~q4);
19 end;
In her original discussion, Almon used only 36 observations, therefore we end our
estimation at 36 (line 13). As you will remember, suppressing the constant term (line
14) is necessary to avoid the dummy variable trap when using all four seasonal
dummy variables. The reason for not dropping one dummy variable is so that we can
impose a linear restriction, summing the coefficients of these four dummy variables
to zero (line 15). Line 16 defines the polynomial lag structures with _pdl. Each row
of _pdl controls a different explanatory variable and rows are separated by
commas. Remember that carriage returns are not “seen” by GAUSS. Three columns
of _pdl specify the following: q = lags, p = orders, and r = end-point restrictions,
respectively. There are four possible settings for end-point restrictions: -1
(beginning), 1 (ending), 2 (both), or 0 (no restriction). The first row of _pdl in line
16 assigns 7 lags, to the fourth order, with both endpoints restricted to the variable
CAPP. The four dummy variables are not affected since each entry is set to 0 from
the second to the last rows of _pdl.
Run the program to see how _pdl works.9
9
As pointed out by Greene (1997, Chapter 17), it was not possible to reproduce Almon’s
original results. Our regression results match with Greene’s results.
174
------------------------

Distributed Lags Variables Used = 7
WARNING: Constant Term Suppressed.


F( 6, 17) Prob>F
1.1338 0.38486


Explained 6.5671E+006 6 1.0945E+006 45.590 1.2384E-011
Residual 5.5218E+005 23 24008.
Total 6.1508E+006 29 2.1210E+005

CAPP 0.086812 0.020858 4.1620 0.00037595 0.42960
CAPP-1 0.12315 0.014384 8.5619 1.3128E-008 0.76118
CAPP-2 0.13424 0.011553 11.620 4.1900E-011 0.85445
CAPP-3 0.13671 0.020442 6.6879 8.0157E-007 0.66041
CAPP-4 0.13859 0.020447 6.7778 6.5118E-007 0.66637
CAPP-5 0.13931 0.011649 11.959 2.3667E-011 0.86146
CAPP-6 0.12972 0.014578 8.8986 6.5882E-009 0.77492
CAPP-7 0.092066 0.020974 4.3894 0.00021327 0.45584
Q1 -13.302 50.453 -0.26365 0.79440 0.0030131
Q2 -7.0170 50.455 -0.13907 0.89060 0.00084023
Q3 -7.6275 50.450 -0.15119 0.88114 0.00099286
Q4 27.946 48.255 0.57915 0.56812 0.014373

Sum of Squared Residuals = 5.5218E+005
Sum of Residuals = 1.88769E+002
In the output, seven lags of CAPP are estimated with the adjusted sample size. Look
at the results of hypothesis testing for linear restrictions. Although we have explicitly
defined only one restriction to sum all the seasonal effects across four quarters to
zero, there are six restrictions. How are the other five restrictions entering the model?
Remember the 7 lags and 4 orders of the polynomial lag structure for the variable
CAPP? Equivalently, there are 3 restrictions (that is, 7-4). On top of them, there
175
were two end-point restrictions. The computed Wald F-test statistic of 1.13 (with P-
value 0.38) suggests these 6 linear restrictions cannot be jointly rejected. Similar
results are obtained from the other test statistics. We notice the insignificant seasonal
effect for all four quarters. The model may be re-estimated without quarterly dummy
variables but with a constant.
Of course, there are problems of model misspecification. As you can see from the
Durbin-Watson statistic, autocorrelation is a serious problem that was not addressed
in the original Almon study. We leave for an exercise the challenge of refining this
model further.
Autoregressive Distributed Lag Models

A polynomial (finite) lag structure may be combined with geometric (infinite) lag to
form an ARDL (autoregressive distributed lag) model. The result is a powerful
dynamic specification which captures the essences of two types of distributed lag
models. The benefit is a better fit of the model with improved (or “whiter”) error
structure. The cost is the potential problem of random regressors and
multicollinearity. Autocorrelation must be evaluated and tested with the presence of
lagged dependent variables. By including more lags of the dependent variable, the
estimated model is assured of being free of misspecification. Further, the model
should be checked for dynamic stability.
To implement an ARDL model using GPE, we need to specify _pdl for the
appropriate polynomial lag structure (restricted or unrestricted) of the explanatory
variables and _dlags for the number of lags of dependent variable. Additional
parameter restrictions may be imposed on these lag variables as well. We have seen
the input control variables _pdl and _dlags used separately in previous lessons.
For specifying an ARDL model, both _pdl and _dlags are required.
Lesson 11.4: Almon Lag Model Once More

As an example of ARDL model, we will improve upon the Almon Lag Model of
Lesson 11.3 by including lags of the dependent variables. First, we need to fix up a
few issues from the regression output of Lesson 11.3. The quarterly seasonality is not
statistically significant, and the parameter restriction on the seasonal dummy
variables is not necessary. We eliminate four quarterly seasonal dummy variables,
and in their place insert a constant term (constant is included in the regression by
default). In the program lesson11.4, line 14, we keep the same polynomial lag
structure for the single explanatory variable. That is, we include seven lags with
fourth-order polynomials, and with end-point restrictions on both sides. Because
quarterly dummy variables are deleted from the model, _pdl consists of only one
row (for the explanatory variable). Now we will deal with the autocorrelation
problem associated with the polynomial lag model of Lesson 11.3. By adding lags of
the dependent variable in the regression equation, serial dependence of the model
errors may be reduced. Through trial and error, and testing for no serial correlation,
we decide on two augmenting lags of the dependent variable.10 Line 15 of lesson11.4
does exactly that.
10
To correct for autocorrelation, we could continue on the model of Lesson 11.3 and assume
that the error structure is AR(1) or even higher-order. Such derivation will result in
complicated non-linear restrictions involving lagged variables (dependent and independent).
176
/*
** Lesson 11.4: Autoregressive Distributed Lag Model
** Almon Lag Model Once More
*/
1 use gpe2;
4 cexp = almon[2:61,2];
5 capp = almon[2:61,3];
6 qt = almon[2:61,1];
7 q1=(qt%10).==1; @ quarterly seasonal dummies @

8 q2=(qt%10).==2;
9 q3=(qt%10).==3;
10 q4=(qt%10).==4;
11 call reset;
12 _rstat = 1;
13 _end = 36;
14 _pdl = {7 4 2};
15 _dlags = 2;
16 _names={"cexp","capp"};
17 call estimate(cexp,capp);
18 end;
Running the program, we have the following:

------------------------

NOTE: Lagged Dependent Variables Used = 2
NOTE: Distributed Lags Variables Used = 7

F( 5, 18) Prob>F
0.77945 0.57735

We can view it as a restricted version of ARDL model. For example, assuming AR(1)
correlation ε = ρ ε-1 + u for the long-run relation Y = α + β X + ε is the same as assuming the
short-run dynamics Y = a + b X + c X-1 + ρY-1 + u with the non-linear restriction b = -c/ρ. In
other words, given ε = ρ ε-1 + u, we must have a = α/(1-ρ), b = β, c = -βρ.
177

Explained 6.0288E+006 5 1.2058E+006 227.20 8.8193E-019
Residual 1.2206E+005 23 5307.1
Total 6.1508E+006 28 2.1967E+005

CEXP-1 1.2541 0.16540 7.5824 1.0625E-007 0.71426
CEXP-2 -0.62565 0.16853 -3.7124 0.0011458 0.37469
CAPP 0.033586 0.018926 1.7746 0.089205 0.12043
CAPP-1 0.041515 0.016618 2.4982 0.020075 0.21343
CAPP-2 0.038166 0.012788 2.9845 0.0066274 0.27916
CAPP-3 0.033262 0.017913 1.8568 0.076188 0.13036
CAPP-4 0.031873 0.021607 1.4751 0.15373 0.086433
CAPP-5 0.034417 0.021396 1.6086 0.12135 0.10113
CAPP-6 0.036656 0.021233 1.7263 0.097699 0.11471
CAPP-7 0.029698 0.018826 1.5775 0.12833 0.097636
CONSTANT 228.22 160.09 1.4256 0.16742 0.081186

Sum of Squared Residuals = 1.2206E+005
Durbin-H Statistic = -0.24902
The estimated model is a restricted ARDL model. The restrictions come in the form
of the fourth-order (of 7 lags) polynomial and end-point restrictions. There are 5
restrictions because of the polynomial lag structure assumed, and these restrictions
are statistically significant based on all the tests. The first two lags of the dependent
variable, with coefficients 1.25 and –0.63, are statistically significant. The stability
of a dynamic equation hinges on the characteristic equation for the autoregressive
part of the model. It is easy to show that the model is stable.11 By augmenting two
lags of the dependent variables, the model is free of autocorrelation as required.
11
Solving the characteristic function 1-1.25z + 0.63z2 = 0, z = 0.9921 ± 0.7766i. It is obvious
that the solutions are greater than 1 in absolute value. Thus two complex solutions of z lie
outside of unit circle.
178
XII
Generalized Method of Moments
Recall from the study of maximum likelihood estimation that assumptions regarding
the underlying probability density or likelihood function of a model structure are
rather strong, typically including the assumption that the model error is normally
distributed. The alternative to the maximum likelihood approach, known as
generalized method of moments (GMM), does away with assumptions regarding the
probability density or likelihood function. Instead, GMM estimation begins by
specifying a set of identities, or moment functions, involving the model variables and
parameters, and then finds the set of parameters that best satisfies those identities
according to a quadratic criterion function. As a result, the GMM estimator is
consistent. For some ideal cases, it can be shown to be as efficient as a maximum
likelihood estimator. In addition to the classical least squares and maximum
likelihood methods, GMM serves as an alternative for regression parameter
estimation. Even for the purpose of estimating the parameters for a probability
distribution of a random variable, GMM is a viable alternative to maximum
likelihood estimation.
GMM estimation is nonlinear in nature. In the following, we shall revisit the problem
of estimating a probability distribution first seen in Lesson 6.3. Instead of using the
maximum likelihood method, GMM is introduced to estimate the parameters of a
gamma probability distribution. It is generalized to study a nonlinear regression
model of rational expectations as done by Hansen and Singleton (1982), where a set
of moment equations or orthogonality conditions are estimated. Finally, the special
cases of linear regression models are considered. For linear models, GMM is more
general than the least squares estimation. It is analogous to an instrumental variable
estimator which accounts for heteroscedasticity and autocorrelation. One of the
advantages of GMM estimation is that it is less dependent on the underlying
probability density or likelihood function. Classical least squares and maximum
likelihood estimation methods are special cases of GMM.
GMM Estimation of Probability Distributions

GMM estimation relies on the specification of a set of identities, known as moment
functions, involving variables and parameters. A moment function is defined as the
expectation of some continuous function m of a random variable X with a parameter
vector θ:
E[m(X,θ)] = 0
Let’s consider a simple example. Suppose X is a random variable for which the
population mean is defined as θ = E(X). Then, E(X) - θ = E(X-θ) = 0. The moment
function is m(X,θ) = X - θ = 0 so that E[m(X,θ)] = 0. In the now familiar maximum
likelihood case, m(X,θ) = ∂ll(θ)/∂θ and E[∂ll(θ)/∂θ] = 0, where ll(θ) is the log-
likelihood function with unknown parameters θ. Moments are used to describe the
characteristics of a distribution, and much of the statistical estimation focuses on the
problem of moment functions (or in the case above, orthogonality conditions). In

addition, a function of moments is also a moment. A more general class of estimators
based on moment functions has been shown to exhibit desirable asymptotic (or large-
sample) properties. For empirical estimation, GMM estimators are based on the
sample moment functions:
m(θ) = 1/N ∑i=1,2,...,N m(Xi,θ)' = 0
where Xi is a sample observation of the random variable and θ is the vector of

unknown parameters. If there are K parameters (i.e., θ = (θ1, θ2, ..., θK)'), we will
need at least K moment equations in order to successfully estimate the parameters
(i.e., m(θ) = (m1(θ), m2(θ), ..., mL(θ))', L ≥ K). In other words, the number of sample
moment equations must be at least as large as the number of parameters to be
estimated. This is just the classical identification problem. If L = K, this model is
exactly identified. If L > K, the model is over-identified with L-K functional
restrictions. The optimization problem is to minimize the quadratic criterion
function:
Q(θ) = m(θ)'W m(θ)
where W is a positive definite weighting matrix. Optimally, W is chosen to be the

inverse of the estimated consistent covariance matrix of m(θ). That is, W = W(θ) =
[Var(m(θ))]-1 and Var(m(θ)) = 1/N2 ∑∑i,j=1,2,...,N m(Xi,θ)'m(Xj,θ).
To ensure that W is positive definite, we need to say some things about its
autocovariance structure. For example,
Var(m(θ)) = S0(θ) + ∑j=1,2,...,p(1 - j/(p+1)) (Sj(θ) + Sj(θ)')

S0(θ) = m(θ)m(θ)' = 1/N2 ∑i=1,2,...,N m(Xi,θ)'m(Xi,θ)
Sj = m(θ)m-j(θ)' = 1/N2 ∑i=j+1,...,N m(Xi,θ)'m(Xi-j,θ)
j = 1, ..., p < N
where p is the degree of autocovariance assumed in the model. This is the White-
Newey-West estimator of Var(m(θ)), which guarantees positive definiteness by
down-weighting higher-order autocovariances.
For an exactly identified model, the optimal value of Q is zero and therefore the
choice of weighting matrix W is irrelevant. For an over-identified case, there are L-K
moment restrictions which must be satisfied with a minimal positive value (or
penalty) of Q. The function of weighting matrix W as constructed is to place the
importance of each individual moment function. Typically, the first iteration of
GMM estimation starts with the special case of W = I (the identity matrix). In other
words, we find the estimator θ0 of θ that minimizes the quadratic function, Q(θ) =
m(θ)'m(θ), with the associated asymptotic covariance matrix:
Var(θ0) = [G(θ0)'G(θ0)]-1G(θ0)'[Var(m(θ0))] G(θ0) [G(θ0)'G(θ0)]-1
where G(θ0) = ∂m(θ0)/∂θ is the L by K matrix of derivatives. With the initial

parameter estimates θ0, let W = W(θ0) = [Var(m(θ0))]-1 and then minimize the
quadratic function:
Q(θ) = m(θ)'W m(θ)
180
GENERALIZED METHOD OF MOMENTS
The asymptotic covariance matrix for the resulting GMM estimator θ1 of θ is:
Var(θ1) = [G(θ1)'W G(θ1)]-1G(θ1)'W [Var(m(θ1))] WG(θ1) [G(θ1)'W G(θ1)]-1
Updating the weighting matrix W = W(θ1) = [Var(m(θ1))]-1 and reiterating the

optimization process until convergence, the final GMM estimator θ* of θ is
obtained, with the following asymptotic covariance matrix:
Var(θ*) = [G(θ*)'W(θ*)G(θ*)]-1
We note that convergence is not necessary for a consistent GMM estimator of θ. If

our estimate θ* of θ is to be asymptotically efficient, the optimal weighting matrix
W = W(θ*) must have been used. In other words, the iterative estimation process
must converge on the solution θ*. With the optimal weighting matrix, θ* is
asymptotically normally distributed with mean θ and covariance Var(θ*). The value
of quadratic function Q at the optimal solution θ* is:
Q* = Q(θ*) = m(θ*)'W(θ*)m(θ*)
Q* serves as the basis for hypothesis testing of moment restrictions. If there are L
moment equations with K parameters (L > K), the Hansen test statistic Q* follows a
Chi-square distribution with L-K degrees of freedom. Justification for including L-K
additional moment functions is made based on the value of Q*.
Lesson 12.1 Gamma Probability Distribution

Recall the classical maximum likelihood method of estimating a probability
distribution discussed in Chapter VI. We now return to Lesson 6.3, estimating the
two parameters of the gamma distribution of the INCOME variable. Consider four
sample moment functions of the gamma probability distribution function with
unknown parameters λ and ρ:
m1(λ,ρ) = 1/N ∑i=1,2,...,N Xi - ρ/λ

m2(λ,ρ) = 1/N ∑i=1,2,...,N Xi2 - ρ(ρ+1)/λ2
m3(λ,ρ) = 1/N ∑i=1,2,...,N ln(Xi) - dlnΓ(ρ)/dρ + ln(λ)
m4(λ,ρ) = 1/N ∑i=1,2,...,N 1/Xi - λ/(ρ-1)
The GMM estimator of θ = (λ,ρ) is obtained from minimizing the weighted sum-of-
squares:
Q(θ) = m(θ)'W m(θ)
where m(θ) = (m1(θ), m2(θ), m3(θ), m4(θ))' and W is a positive definite symmetric
matrix. Conditional to the weighting scheme W, the variance-covariance matrix of θ
is estimated by:
Var(θ) = [G(θ)'W G(θ)]-1 G(θ)'W [Var(m(θ))] WG(θ) [G(θ)'W G(θ)]-1
If we let W equal the inverse of the covariance matrix of m(θ), or [Var(m(θ))]-1, then
Var(θ) = [G(θ)'W G(θ)]-1.
181
From here, we can show that the GMM class of estimators includes the maximum
likelihood estimator as a special case. Solving from the score of the log-likelihood
function based on the gamma distribution:
∂ll(X,θ)/∂λ = N (ρ/λ) - ∑i=1,2,...,N Xi = 0
∂ll(X,θ)/∂ρ = N ln(λ) – N dlnΓ(ρ)/dρ + ∑i=1,2,...,N ln(Xi) = 0
where ll(X,θ) = N [ρln(λ) - lnΓ(ρ)] - λ ∑i=1,2,...,N Xi + (ρ-1) ∑i=1,2,...,N ln(Xi). Thus, the
maximum likelihood estimate of θ = (λ,ρ) is an exactly identified GMM with m(θ) =
(m1(θ), m3(θ)). For this exactly identified case, the weighting matrix W is irrelevant,
and the optimal criterion Q is zero.
GMM is a nonlinear estimation method. Nonlinear optimization methods were

discussed in Chapter VI. Chapter VII presented the applications of nonlinear least
squares and maximum likelihood estimation in econometrics. GMM estimation is a
straightforward example of a nonlinear minimization problem. The first part of the
program lesson12.1 sets up a standard nonlinear optimization problem. The data
series INCOME is read from the text file yed20.txt. As in Lesson 6.3, the INCOME
variable is scaled by 10 and renamed x (line 4).12 The next few lines specify the
minimization problem (_nlopt=0 in line 6) to be estimated by the QHC method
(_method=5 in line 7) with no more than 100 iterations (_iter=100 in line 8).
Line 9 specifies the initial values of the two parameters ρ and λ, respectively.
/*
** Lesson 12.1: GMM Estimation of a Gamma Distribution
** See Greene (1999), Example 4.26 and Example 11.4
*/
1 use gpe2;
4 x=data[2:21,1]/10; @ income: data scaling may help @
5 call reset;
6 _nlopt=0; @ it is a minimization problem @
7 _method=5;
8 _iter=100;
9 _b={3,1}; @ initial values of parameters @

10 _hacv=1; @ hetero consistent covariance @
@ assuming serially uncorrelated @
11 call estimate(&gmmqw,x);
@ using the results of previous estimator @
12 _b=__b; @ for initial value of parameters and @
13 gmmw=invpd(gmmv(x,_b)); @ for computing the weight matrix @
15 call gmmout(x,__b); @ print GMM output @
/*
_b=__b;
call estimate(&gmmq,x);
call gmmout(x,__b); @ print GMM output @
*/
12
Our implementation of this example is slightly different from that of Example 11.4 in
Greene (1999). Instead of using the scaled moment equations as Greene did, we scale the data
series first and then estimate the original moment equations as described. The numerical
results are more stable and easier to evaluate.
182
16 end;
/*
User-defined moments equations, must be named mf
based on gamma distribution: b[1]=rho, b[2]=lambda
*/
17 proc mf(x,b);
18 local n,m;
19 n=rows(x);
20 m=zeros(n,4);
21 m[.,1]=x-b[1]/b[2];
22 m[.,2]=x^2-b[1]*(b[1]+1)/(b[2]^2);
23 m[.,3]=ln(x)-gradp(&lngamma,b[1])+ln(b[2]);
24 m[.,4]=1/x-b[2]/(b[1]-1);
25 retp(m);
26 endp;
/*
Log of gamma distribution function
*/
27 fn lngamma(x)=ln(gamma(x));
28 #include gpe\gmm.gpe;
Most of the computation details and algorithms of GMM estimation in GAUSS are
grouped in a module named GMM.GPE. There are many ways to include a module
in your program. The simplest is to use the GAUSS compiler directive #include.
It will include the specified module during the compilation of your program. We
suggest including the module GMM.GPE at the end of your program. If you have
properly installed the GPE package with your version of GAUSS, GMM.GPE is
located in the GPE subdirectory. Putting source codes in a separate file hides their
implementation “secrets.” If you are interested in the programming details, you can
examine the program listing of GMM.GPE available in Appendix B-1.
The module GMM.GPE defines two objective functions, gmmqw and gmmq. The
former uses a predefined weighting matrix, while the latter computes the weighting
matrix together with the unknown parameters. In addition, the procedure gmmout
prints the regression output. Since GMM is a nonlinear optimization method, it
requires a user-defined moment function with the name mf which, like the other
functions (e.g. residual function for nonlinear least squares or maximum likelihood
estimation), depends on a sample data matrix x and a parameter vector b.
Based on the gamma probability distribution, two parameters λ and ρ will be

estimated from the four moment functions we defined earlier (see lines 17 through
26 which define the procedure mf(x,b)). One tricky part is computing the
derivative of the log of the gamma function for the third moment equation (line 23).
The logarithm of the gamma function is defined as a separate one-line function in
line 27, so that the GAUSS built-in gradient function gradp is applied in line 23.
One of the key advantages of GMM is that it allows for a flexible specification of
covariance structure, including heteroscedasticity and autocorrelation. We have seen
the use of the GPE control variable _hacv to compute the heteroscedasticity-
autocorrelation-consistent covariance matrix in the context of heteroscedasticity
(Chapter IX) and autocorrelation (Chapter X). _hacv is used similarly for nonlinear
GMM estimation, except that heteroscedasticity-consistent covariance is the default
option here. If we assume first-order autocovariance, then line 10 needs to be
modified as follows:
183
_hacv = {1,1};
The first element of _havc directs the computation of the heteroscedasticity-

consistent covariance matrix, while the second element specifies the order of
autocovariance requested. See Appendix A for more information about the global
control variable _hacv.
We begin the GMM estimation by minimizing the objective function gmmqw (line
11) with the default weighting matrix I (identity matrix). The estimated parameters
are used as starting point for the next iteration. First, we start from the estimated
parameters (line 12) and compute the corresponding weighting matrix (line 13).
Then the improved consistent parameter estimates are obtained (line 14) and printed
(line 15). We could continue on updating the weighting matrix and estimating the
parameters until convergence. Equivalently, we could estimate the parameters
together with the associated weighting matrix. However, finding a convergent
solution is not guaranteed due to a high degree of nonlinearity in the objective
function.
Running lesson12.1, the first set of GMM estimation results, based on the identity
weighting matrix, is as follows:

---------------------------------------------
Assuming Generalized Method of Moments

Tolerance = 0.001
Initial Result:
Parameters = 3.0000 1.0000

Iteration = 1 Step Size = 1.0000 Value = 1.9867
Parameters = 3.0010 0.86896
…
Parameters = 2.3691 0.74112
Final Result:
Parameters = 2.3691 0.74112
Gradient Vector = -0.21423 0.44266
X1 2.3691 0.010345 229.02
X2 0.74112 0.0028050 264.21
The second set of GMM estimation results is based on the previous estimates of the
parameters and the associated weighting matrix (see lines 12-14 of the program
lesson12.1). As is standard practice in nonlinear optimization, the estimated standard
errors and t-ratios of the parameters are computed from the inverse of the hessian
matrix calculated during minimization of the quadratic criterion function. Note that
184
these are not the GMM estimates of the standard errors and t-ratios. The correct
estimates of the standard errors and t-statistics of the parameters are computed at the
end. In addition, the Hansen test statistic of the moment restrictions is presented.
Calling the procedure gmmout (see line 15 of the program lesson12.1) gives us the
following result:

---------------------------------------------

Tolerance = 0.001
Initial Result:
Parameters = 2.3691 0.74112

Parameters = 3.3687 0.74667
…
Parameters = 2.8971 0.84839
Final Result:
Parameters = 2.8971 0.84839
Gradient Vector = -487.16 -6.2162
X1 2.8971 0.0079292 365.37
X2 0.84839 0.062556 13.562
GMM Estimation Result

=====================
2.8971 0.0044004 658.37
0.84839 0.11996 7.0721
Hansen Test Statistic of the Moment Restrictions

Chi-Sq( 2) = 4.4604
For the two parameters of the gamma distribution, ρ and λ, we now compare their
GMM estimators with the maximum likelihood (ML) estimators obtained earlier
from Lesson 6.3 in Chapter VI. The standard errors are in parentheses.
ML GMM
ρ 2.4106 (0.7161) 2.8971 (0.0044)
λ 0.7707 (0.2544) 0.8484 (0.120)
GMM Estimation of Econometric Models

The GMM estimation of econometric models can be considered as an extension of
the IV (instrumental variable) estimation method. IV estimation is widely used for
models with random regressors (e.g. lagged dependent variables) which exhibit
185
contemporaneous correlation with the model’s errors. The advantage of GMM over
IV is that the model need not be homoscedastic and serially independent. The
covariance matrix of the averages of sample moments is taken into account for
minimizing the GMM criterion function.
For notational convenience, let X be a combined data matrix of endogenous

(dependent) and predetermined (independent or explanatory) variables in the model.
β is a K-element vector of unknown parameters. Suppose there are L moment
equations, m(X,β) = (m1(X,β), ..., mL(X,β)), where L ≥ K. The model formulation is
not limited to the case of a single equation. Generalization to a system of linear or
nonlinear equations is straightforward.
Corresponding to the moment conditions E(m(X,β)) = 0, we write the sample

moment equations as follows:
m(β) = 1/N ∑i=1,2,...,N m(Xi,β)' = 0
Recall that the objective function, known as the GMM criterion function, to be
minimized is Q(β) = m(β)'W m(β). We would optimally choose W to be equal to
[Var(m(β))]-1. To find the β* which minimizes Q(β), we solve the zero-gradient
conditions: ∂Q(β*)/∂β = 0. Our estimator β* of β will be asymptotically efficient and
normally distributed with mean β and covariance matrix:
Var(β*) = {G(β*)'[Var(m(β*))]-1G(β*)}-1
where G(β*) = ∂m(β*)/∂β and [Var(m(β*))] is the White-Newey-West estimator of

the covariance matrix of the sample moments.
Nonlinear IV Estimation
Now we consider the regression model ε = ε(β) = F(Y,X,β) (or Y- f(X,β)), where Y
is the endogenous or dependent variable, and X consists of predetermined (or
independent) variables. β is a K-element parameter vector. Let Z be a set of L
instrumental variables, for which we assume L ≥ K. Under the general assumption
that E(ε) = 0 and Var(ε) = E(εε') = Σ = σ2Ω, we can write the model as E(Z'ε) = 0.
The sample moment functions are defined by m(β) = 1/N Z'ε(β) with covariance
matrix:
Var(m(β)) = 1/N2 Z' E[ε(β)ε(β)'] Z = 1/N2 Z'Σ(β)Z
Therefore, GMM estimation amounts to finding the β* which minimizes:
Q(β) = ε(β)'Z [Z'Σ(β)Z]-1Z'ε(β)
The resulting GMM estimator β* is asymptotically normally distributed with mean β

and covariance matrix:
Var(β*) = {(∂ε(β*)/∂β)'Z [Z'Σ(β*)Z]-1Z'(∂ε(β*)/∂β)}-1
where Σ(β∗) is the White-Newey-West estimator of Σ = σ2Ω.
186
Linear IV Estimation
If the model is linear, or ε = ε(β) = Y - Xβ, then the GMM estimator of β is
equivalent to the IV estimator:
β* = (X'Z[Z'Σ(β∗)Z]-1Z'X)-1 X'Z[Z'Σ(β∗)Z]-1Z'Y
Var(β*) = {X'Z[Z'Σ(β∗)Z]-1Z'X }-1
If the instrumental variables Z = X, then
β* = (X'X)-1X'Y
Var(β*) = (X'X)-1[X'Σ(β∗)X](X'X)-1
Special Cases
If the IV model is homoscedastic and serially uncorrelated, that is Σ = σ2I, then
β* = (X'Z[Z'Z]-1Z'X)-1X'Z[Z'Z]-1Z'Y
Var(β*) = σ2(β∗) {X'Z[Z'Z]-1Z'X}-1
where σ2(β∗) = 1/N ε(β∗)'ε(β∗). If the instrumental variables Z = X, this further

reduces to the ordinary least squares estimator:
β* = (X'X)-1X'Y
Var(β*) = σ2(β∗) (X'X)-1
Hypothesis Testing
Based on the statistical inference for nonlinear regression models (see Chapter VII,
Lesson 7.3 in particular), there are three corresponding test statistics for testing linear
or nonlinear restrictions on the GMM estimate of β. Suppose there are J constraint
equations written in the form c(β) = 0. Let β* be the unconstrained GMM estimator
of β, and let b* be the constrained estimator of β. All the three test statistics
discussed below will follow a Chi-square distribution with J degrees of freedom.
Wald Test
The Wald test statistic, based on the unconstrained estimator β*, is defined as:
W = c(β*)'[Var(c(β*)]-1c(β*)
= c(β*)' {(∂c(β*)/∂β) [Var(β*)] (∂c(β*)/∂β)'}-1 c(β*)
Lagrangian Multiplier (LM) Test

Let α = ∂Q(b*)/∂β = 2 m(b*)'W G(b*), where G(b*) = ∂m(b*)/∂β. If the constraints
hold, then α approaches to 0. The LM statistic is:
LM = α[Var(α)]-1α'
= m(b*)'W G(b*)[G(b*)'W G(b*)]-1G(b*)'W m(b*)
187
Likelihood Ratio (LR) Test

If we estimate both β* and b*, the LR test statistic is computed as:
LR = Q(b*) - Q(β*)
We note that both β* and b* are computed using the same consistent estimator of the
weighting matrix W.
Lesson 12.2 A Nonlinear Rational Expectations Model

An important application of GMM estimation is to estimate the first-order conditions
(or Euler equations) of a dynamic optimization problem. As an example, we will use
a model common in the finance literature. Suppose a representative stockholder at
current time t, who tries to maximize a concave expected utility function of
consumption over an indefinite future horizon. This can be modeled as:
∑τ=0,...,∞ βτ E{u(Ct+τ) | Zt}
where Zt is the information available to the consumer at time t and Ct+τ is the
consumption τ periods from t. 0 < β <1 is known as the discount factor of time
preference. Given N different stocks, the optimal consumption-investment allocation
must satisfy the following condition:
u'(Ct) = β E{u'(Ct+1) [(Pi,t+1+Di,t+1)/Pt] | Zt}
for i = 1,...,N. u'(Ct) = ∂u/∂Ct is the marginal utility of consumption at time t. Pi,t+1 is
the price of stock i at time t+1 and Di,t+1 is the dividend per share of stock i at t+1.
The ratio (Pi,t+1+Di,t+1)/Pi,t represents the returns of investment in stock i between
periods t and t+1. In other words, this merely defines the equilibrium condition that
the marginal utility of consumption in the current period must equal the expected
return next period from investing in stock i. Assume that the utility function exhibits
constant relative risk aversion as:
u(Ct) = Ctα/α for α<1
where 1-α is known as the coefficient of relative risk aversion, and 1-α > 0. Then,
for each i = 1, ..., N, the optimal decision-rule is
Ctα-1 = β E{Ct+1α-1 [(Pi,t+1+Di,t+1)/Pt] | Zt}
Equivalently, for each stock i = 1, ..., N, we must have
βE{[(Ct+1/Ct)α-1] [(Pi,t+1+Di,t+1)/Pt] | Zt} = 1
The hypothesis of rational expectations assumes that the intertemporal decision-

making should be independent from the historical information available at the time at
which the decision is made. Therefore, the derived orthogonality condition for each
stock i = 1, ..., N is:
E [Zt {β[(Ct+1/Ct)α-1] [(Pi,t+1+Di,t+1)/Pt] -1}] = 0
188
For more detailed description of the model, see Hansen and Singleton (1982).13 In
terms of our econometric estimation, the model may be expressed with the
orthogonality condition: E[Z ε(X,θ)] = 0, where X = [X1,X2,X3], θ = (β,α), and
ε(X,θ) = [βX1α−1X2 – 1, βX1α−1X3 – 1]
The data file gmmq.txt installed in the GPE subdirectory consists of three variables
(335 observations from January 1959 to December 1978, though not the original
Hansen-Singleton data):
X1 Ratio of two-period consumption, Ct+1/Ct.

X2 Value-weighted returns of NYSE stock market, (Pt+1+Dt+1)/Pt where Pt+1 is the
price and Dt+1 is the dividend payoff of stock at t+1.
X3 Risk-free rate of returns (T-Bill rate).
We note that this model consists of a system of two nonlinear equations. The
instrumental variables Z consist of one or several lags of X and a constant. The
following program lesson12.2 implements and estimates the Hansen-Singleton
rational expectations model. The program structure looks similar to that of Lesson
12.1. The main difference is the model specification described in the block from line
19 to line 27 for the moment function procedure mf(x,b), which reflects exactly
the model described above. We use one lag of each of the three variables X1, X2, X3,
and a constant as the instrumental variables (see line 22). More lags may be included
for the additional instrumental variables. The two orthogonality equations for stock
and risk-free returns are concatenated to form the moment functions system (see
lines 23 and 24). The rest of the program performs the GMM estimation. First, the
initial estimates of (β,α) are obtained by assuming the default identity weighting
matrix in the objective function for minimization (line 10). Then, with the resulting
covariance matrix of the moment functions, the consistent parameter estimates are
calculated (see lines 11-13). An efficient solution is obtained from simultaneously
estimating the parameters and the corresponding covariance matrix in the second
iteration (see lines 15 and 16).
/*
** Lesson 12.2: A Nonlinear Rational Expectation Model
** GMM Estimation of Hansen-Singleton Model (Ea, 1982)
*/
1 use gpe2;
3 load x[335,3]=gpe\gmmq.txt; @ data columns: @
@ (1) c(t+1)/c(t) (2)vwr (3)rfr @
4 call reset;
5 _nlopt=0;
6 _method=5;
7 _tol=1.0e-5;
8 _iter=100;
9 _b={1,0}; @ GMM estimation with initial @

10 call estimate(&gmmqw,x); @ identity weighting matrix @
11 _b=__b; @ GMM estimation with external @
13
For computational implementation of the model and the data file gmmq.txt used in this
lesson example, see also the Hasen-Heaton-Ogaki GMM package from the American
University GAUSS archive at http://www.american.edu/academic.depts/cas/econ/gaussres/
GMM/GMM.HTM.
189
12 gmmw=invpd(gmmv(x,__b)); @ covariance weighting matrix @

15 _b=__b; @ may be a convergent efficient solution @

16 call estimate(&gmmq,x); @ with internal covariance weighting matrix @
18 end;
/*
User-defined moments functions, must be named mf
*/
19 proc mf(x,b);
20 local z,n,m;
21 n=rows(x);
22 z=ones(n,1)~lagn(x,1); @ IV @
23 m=z.*(b[1]*(x[.,1]^(b[2]-1)).*x[.,2]-1);
24 m=m~(z.*(b[1]*(x[.,1]^(b[2]-1)).*x[.,3]-1));
25 @ nonlinear multiple equations system @
26 retp(packr(m));
27 endp;
28 #include gpe\gmm.gpe;
The first part of the output (the first iteration, using the identity matrix as the
weighting matrix) is only preparation for computing the consistent parameter
estimates in the second iteration of the GMM estimation. The result of second
consistent estimation is shown below:

---------------------------------------------

Tolerance = 1e-005
Initial Result:
Parameters = 0.99977 -0.00012883

Parameters = 1.0014 0.00045142
…
Parameters = 0.99919 0.85517
Final Result:
Parameters = 0.99919 0.85517
Gradient Vector = -0.00026951 0.00017303
X1 0.99919 0.00012573 7946.9
X2 0.85517 0.044497 19.219
190

=====================
0.99919 0.00047887 2086.6
0.85517 0.17853 4.7901

Chi-Sq( 6) = 13.306
The consistent estimates (0.9992, 0.8552) of the two parameters are obtained using 8
instrument variables (4 for each of the two equations). The Hansen test statistic of
the extra 6 moment restrictions is barely statistically significant at 5% level (with the
critical value 12.59).

---------------------------------------------

Tolerance = 1e-005
Initial Result:
Parameters = 0.99919 0.85517

Parameters = 0.99934 0.85482
…
Parameters = 0.99950 0.78735
Final Result:
Parameters = 0.99950 0.78735
Gradient Vector = 0.0069545 0.0056644
X1 0.99950 9.8686e-005 10128.
X2 0.78735 0.037531 20.979

=====================
0.99950 0.00047957 2084.2
0.78735 0.17736 4.4393

Chi-Sq( 6) = 12.474
The final GMM estimate of (β,α) at (0.9995, 0.7874) is consistent with the result of
Hansen and Singleton (1982). Since it is the convergent solution, it is also efficient.
However, the 6 moment restrictions are significant only at the 10 percent level.
191
It may be of interest to perform hypothesis tests on the coefficient of relative risk

aversion 1-α. Two cases that are often of practical interest are α approaching 0
(indicating a logarithmic utility function) and α = 1 (indicating a linear utility
function), we leave applying hypothesis testing in the GMM framework to interested
readers.
Linear GMM
GMM estimation of a linear regression model is essentially the same as IV
estimation with a general covariance matrix. In Chapters IX and X, we discussed the
computation of White-Newey-West estimates of heteroscedasticity-autocorrelation-
consistent covariance by using the GPE control variable _hacv. In the case of the
IV approach, Lesson 11.2 introduced the use of another GPE control variable,
_ivar, to specify a set of instruments for linear model estimation. _ivar may be
internally determined in the case of autocorrelation correction for a lagged dependent
variable model. Alternatively, it may be externally specified as a data matrix
consisting of a list of instrumental variables. Refer to your econometrics textbook for
the requirements and proper use of instrumental variables in a regression model.
In GPE, _ivar may be combined with _hacv to carry out GMM estimation. In
addition, as in a nonlinear model estimation, we can set the control variable _iter
to allow for iterated computation of the parameter estimates. If the solution
converges, it is efficient. Even if it does not converge, the GMM estimator is still
consistent.
Lesson 12.3 GMM Estimation of U.S. Consumption Function

In the following, we present a simple example of a consumption function derived
from Hall’s Life Cycle-Permanent Income Hypothesis (Hall, 1978) with a quadratic
utility function. The intertemporal model formulation is similar to that of Lesson
12.2 on the consumption-investment decision-making, with the exception that we
consider only the consumption plan with a quadratic utility function. The use of a
quadratic functional form leads to a linear consumption function:
Ct+1 = β0 + β1Ct + εt+1
Ct+1 and Ct are expected and current consumption, respectively. Let Zt be historical
information available to the consumer at time t or earlier. Then the orthogonality
condition becomes:
E(Ztεt+1) = 0
From other theories of consumption, the instrumental variables Z may include levels
of income Y and consumption C in addition to a constant. That is,
Zt = [ 1 Ct Yt]
Further lags of C and Y may be added to the model as needed. The consumption-
income relationship was studied in Chapter XI with income as the explanatory
variable. Based on an Euler equation of a constrained expected utility maximization,
in this example, the future consumption is affected indirectly by the current and past
income as the instrumental variables. Using the U.S. time series data of usyc87.txt,
192
Lesson 12.3 demonstrates GMM estimation of the linear consumption function as

follows:
/*
** Lesson 12.3: GMM Estimation of U.S. Consumption Function
** GMM Estimation of a Linear Regression Model
*/
1 use gpe2;
3 load z[67,3]=gpe\usyc87.txt;
4 y = z[2:67,2];
5 c = z[2:67,3];
6 call reset;
7 _names={"c","c1"};
8 _rstat=1;
9 _rplot=2;
10 _dlags=1;
11 _ivar=ones(rows(y),1)~lagn(c~y,1);
12 _hacv={1,1};
13 _iter=100;
14 call estimate(c,0);
15 end;
Two variables, named C and Y, are read from the data file usyc87.txt. Line 10,
_dlags = 1, specifies the lagged dependent variable model. GMM estimation for
the linear autoregressive consumption function is given in line 14, with the
instrumental variables _ivar specified in line 11. We use only the first lag of
income and consumption variables in addition to a constant as the instruments. The
first-order autocovariance structure is specified in line 12, in which the White-
Newey-West estimate will be computed. The computation of GMM estimates will be
iterated until convergence or until the limit set in line 13, _iter = 100. The
empirical results in greater detail can be found in the output file output12.3.
Of course the estimated model is not good enough to be free of misspecification. It

serves the purpose for demonstrating GMM estimation of a linear model. To improve
the model, either more lags of consumption should be included for a comprehensive
autoregressive specification or the model should explicitly correct for serial
correlation. We leave the rest of the model improvement task to the interested reader.
193
XIII
Systems of Simultaneous Equations
GPE can estimate systems of linear and nonlinear equations. For a system of linear
equations, you need to define the endogenous and predetermined (including lagged
endogenous, current and lagged exogenous) variables. By selecting and identifying
the relevant variables for each equation, the procedure estimate carries out the
system model estimation as in the case of a single equation model. For a system of
nonlinear equations, it becomes more involved to define the functional form for the
model equations. In this chapter, the classic example of Klein Model I (1950) is used
to demonstrate the estimation of a system of linear regression equations. The special
case of seemingly unrelated regressions (SUR) is considered with the Berndt-Wood
model of energy demand. Finally, re-examining the Klein Model, nonlinear
maximum likelihood estimation is shown to be useful for estimating a system of
nonlinear equations.
Linear Regression Equations System

We follow the conventional matrix representation of a system of linear equations as:
YB + XΓ = U
Let N be the number of observations, G be the number of endogenous variables

(therefore, the number of equations), and K be the number of predetermined
variables. Then Y, a N×G matrix, and X, a N×K matrix, are the respective data
matrices of endogenous and predetermined variables. The corresponding G×G
matrix B associated with the endogenous variable matrix Y, and the G×K matrix Γ
associated with predetermined variable matrix X are the sparse parameter matrices in
which the unknown nonzero elements need to be estimated. Finally, U is the N×G
stochastic error matrix.
Given the data matrices Y and X, the unknown parameters in B and Γ can be
estimated using a variety of methods. GPE implements both single equation (limited
information) methods and simultaneous equations (full information) methods. Before
the parameter estimation, the model of multiple equations must be properly specified
to account for the relevant variables and restrictions. In GPE, this is done by
specifying the structure of the parameter matrices B and Γ. It uses a couple of
specification matrices to define the stochastic equations and fixed identities of the
system by representing the parameter matrices of the system as arrays of 1’s, 0’s, and
–1’s, signifying the inclusion or exclusion of variables from particular equations. In
the following, we discuss the three most important input variables that control the
specification and estimation of a simultaneous linear equations model.
• _eq
• _id
• _method
First, the variable _eq specifies the stochastic equation matrix for system model
estimation. This is a Gs by (G+K) matrix with elements -1, 0, and 1 arranged in
accordance with the order of endogenous variables Y followed by the predetermined
variables X. Note that Gs is the number of stochastic equations and G ≥ Gs. For each
stochastic equation, there is exactly one row of _eq to define it. In the stochastic
equation specification matrix _eq, an element -1 indicates the left-hand side
endogenous variable. Only one -1 entry is allowed in each equation. An element 1
indicates the use of an endogenous and/or a predetermined variable on the right-hand
side of an equation. The zeros indicate the corresponding unused or excluded
variables in the equation. Constant terms are not normally included in the equation
specification. If _eq is not specified, or _eq=0 by default, a SUR equations system
is assumed. In this case, Gs = G, and the _eq matrix consists of a G×G sub-matrix
with -1 in the diagonals and zeros elsewhere (endogenous variables portion of _eq),
and a G×K sub-matrix consisting entirety of ones (predetermined variables portion of
_eq).
The second input variable _id specifies the identity equation specification matrix
for a system model. _id is similar in size and setup to _eq, except that its entries
can be any value as required by the model. If _id is not specified, or _id=0 by
default, there is no identity. To ensure system compatibility, the number of rows in
two specification matrices _eq and _id must sum to G, the total number of
endogenous variables or equations in the system.
The input variable _method controls the use of the specific method of estimation.
In the context of simultaneous linear equations, the available estimation methods are:
_method=0 Ordinary least squares (OLS, the default method)

_method=1 Limited information maximum likelihood (LIML)
_method=2 Two-stage least squares (2SLS)
_method=3 Three-stage least squares (3SLS, may be iterative)
_method=4 Full information maximum likelihood (FIML, may be iterative)
Note that LIML and FIML are not true nonlinear maximum likelihood estimation
methods. Instead they are types of instrumental variables estimation. In GPE, three
variants of the FIML method are available:
_method={4,0} (or 4) FIML instrumental variable method (Hausman, 1975)

_method={4,1} FIML linearized method (Dhrymes, 1970)
_method={4,2} FIML Newton method
2SLS and 3SLS are flavors of the instrumental variables estimation method, where
the instrumental variables used are all the predetermined variables in the system. For
estimation of a linear system model, external instrumental variables may be
requested and specified in the matrix _ivar. The data matrix _ivar will be
combined with all the predetermined variables to form the basis for instrumental
variable estimation. A constant term is automatically included in _ivar. For
technical details and comparisons of different estimation methods for a linear
equations system, refer to your econometrics textbook.
196
SYSTEM OF SIMULTANEOUS EQUATIONS
Klein Model I (1950)

The pedagogical example of Klein Model I is typically used to demonstrate the
model estimation of a linear regression equations system. It was the first U.S.
economy-wide econometric model consisting of three stochastic equations with
about 10 variables covering the prewar period from 1920 to 1941. klein.txt is a text
data file containing these variables which will be used in the Klein Model I. There
are several variations of the Klein Model in the literature. For the convenience of
illustration, we adopt the following specification of the model:
C = α0 + α1 P + α2 P-1 + α3 (W1 + W2) + ε1

I = β0 + β1 P + β2 P-1 + β3 K + ε2
W1 = γ0 + γ1 X + γ2 X-1 + γ3 A + ε3
X=C+I+G
P = X – T – W1
K = K-1 + I
The variables used are:
C Consumption in billions of 1934 dollars.

I Investment.
W1 Private wage bill.
X Total private income before taxes, or X = Y + T - W2 where Y is after-tax
income.
P Private profits.
K Capital stock in the beginning of year, or capital stock lagged one year.
W2 Government wage bill.
G Government non-wage spending.
T Indirect taxes plus net exports.
A Year – 1931 (a time trend).
The first three equations are stochastic with unknown parameters α’s, β’s, and γ’s,
respectively. The remaining three equations are accounting identities. Since the sum
of private and public wage bills (W1+W2) appears in the first equation, it is more
convenient to introduce one additional endogenous variable W, total wage bill, with
the accounting identity:
W = W1 + W2
The alternative is to impose a linear parameter restriction of identical parameter

value for W1 and W2 in the first equation. The resulting model consists of 7
endogenous variables (C, I, W1, X, P, K, W) and 8 predetermined variables (X-1, P-1,
K-1, W2, G, T, A, and Constant). Lesson 13.1 below implements the model as
described.
Lesson 13.1: Klein Model I

In the program, from line 3 to line 15, the data file klein.txt is loaded and each
variable used in the model is defined. Data matrix yvar of endogenous variables is
defined in line 16. Line 17 defines the data matrix xvar of predetermined variables
which includes three lagged endogenous variables and four exogenous variables. By
default, a constant term is automatically included in each stochastic equation. Then,
in the next two lines, are the two important control variables: _eq and _id. We
197
explain the construction of the first line of these two matrices, and leave you to
puzzle out the rest.
First, the _eq matrix (line 21) specifies which variables are to be included in which
stochastic equations. Recall the first stochastic equation in the Klein Model I:
C = α0 + α1 P + α2 P-1 + α3 (W1 + W2) + ε1
Then the first line of the _eq matrix is:

@ C I W1 X P K W XL PL KL W2 A G T 1 @
_eq = {-1 0 0 0 1 0 1 0 1 0 0 0 0 0, …
Note that the column under the variable C contains a –1. This means that C is the
left-hand side variable of the first stochastic equation. Since I and W1 are not in the
first stochastic equation, we place 0’s in their respective columns. Looking again at
this equation, we see that the right-hand side variables are: constant, P, P-1, and W
(remember that W = W1+ W2). To let GPE know that P, P-1, and W are in the first
equation, we place 1’s in their respective places in the _eq matrix. This does not
mean that their coefficients are restricted to be equal to one, it merely tells GPE to
include those particular variables in the equation. GPE includes the constant
automatically. All other variables (namely X, K, X-1, K-1, W2, A, G, T) are not in the
first equation. Putting 0’s in the respective places of these variables in the _eq
matrix lets GPE know that it should not include these variables in the first stochastic
equation of the system.
/*
** Lesson 13.1: Klein’s Model I
** Simultaneous Equation System of Klein’s Model I
*/
1 use gpe2;
3 load data[23,10] = gpe\klein.txt;
4 a=data[2:23,1]-1931; @ time trend: 1931 = 0 @

5 c=data[2:23,2]; @ consumption @
6 p=data[2:23,3]; @ profit income @
7 w1=data[2:23,4]; @ private wage income @
8 i=data[2:23,5]; @ investment @
9 k1=data[2:23,6]; @ lagged capital stock @
10 x=data[2:23,7]; @ private total income @
11 w2=data[2:23,8]; @ public wage income @
12 g=data[2:23,9]; @ government spending @
13 t=data[2:23,10]; @ tax @
14 k=k1[2:22]|209.4; @ capital stock @

15 w=w1+w2; @ total wage income @
16 yvar=c~i~w1~x~p~k~w;
17 xvar=lag1(x~p~k)~w2~a~g~t;
18 call reset;
19 _names={"c","i","w1","x","p","k","w",
"x-1","p-1","k-1","w2","a","g","t"};
20 _vcov=1;
21 _eq = {-1 0 0 0 1 0 1 0 1 0 0 0 0 0,
0 -1 0 0 1 0 0 0 1 1 0 0 0 0,
0 0 -1 1 0 0 0 1 0 0 0 1 0 0};
22 _id = { 1 1 0 -1 0 0 0 0 0 0 0 0 1 0,
198
0 0 -1 1 -1 0 0 0 0 0 0 0 0 -1,
0 1 0 0 0 -1 0 0 0 1 0 0 0 0,
0 0 1 0 0 0 -1 0 0 0 1 0 0 0};
23 _begin=2;
24 _method=0; @ OLS estimation @

25 call estimate(yvar,xvar);
26 _method=1; @ LIML estimation @
28 _method=2; @ 2SLS estimation @
30 _iter=100;
31 _method=3; @ 3SLS estimation (iterative) @
33 _method=4; @ FIML estimation @
35 end;
Similarly, line 22 specifies the identity equations of the model. Take the first identity
equation as an example:
X=C+I+G
It involves the variables X, C, I, and G. Therefore, in the first row of the _id
matrix, only the relevant columns have non-zero values. Typically these entries are 1
or -1, but they could be any other values as required. Variables not used in the
definition of an identity have zeros in the corresponding places of the _id matrix.
The first row of the _id matrix looks like this:
_id = { 1 1 0 -1 0 0 0 0 0 0 0 0 1 0, …
The easiest way to understand the construction of the system model described so far
is to relate it with the matrix representation:
YB + XΓ = U
where Y corresponds to yvar and X corresponds to xvar in the program. The

matrix _eq|_id (vertical concatenation of _eq and _id) is the specification
matrix corresponding to the transpose of B|Γ (vertical concatenation of B and Γ).
Because of the use of lag variables, the first observation of the data is lost and
estimation must start from the second observation (see line 23). In lines 24 through
34, five estimation methods are carried out. They are OLS, LIML, 2SLS, 3SLS, and
FIML. It is of interest to see the covariance matrices of the equations, thus line 20
sets the option _vcov=1 to show the variance-covariance matrix across equations
and across parameters as well. Note that 3SLS and FIML are iterative methods, and
it is wise to set an iteration limit for the solution to converge. Line 30, _iter=100,
does the job.
Running the program of Lesson 13.1 will generate about 20 pages of output. To save
space, we will only present the results of 2SLS because of its popularity in the
literature. You should run the program in its entirety and check the output file to see
the complete results. In a summary table, the parameter estimates of these methods
are listed and compared. You need to check your econometrics textbook for the
199
evaluation of the pros and cons of these estimation methods, in particular the
differences between limited information and full information estimation methods.
The regression results of a typical linear system model are divided into two parts: the
results of the system of equations as a whole, and the results of each separate
equation. Here we present only the first part of 2SLS for the estimated parameters,
including the variance-covariance matrices across equations. For the rest of
regression results by equation, we refer to the output file output13.1.
Simultaneous Linear Equations Estimation

----------------------------------------
Number of Endogenous Variables = 7
Number of Predetermined Variables = 8
Number of Stochastic Equations = 3
Two Stages Least Squares Estimation
System R-Square = 0.97711

Equation Variable Estimated Asymptotic

Name Name Coefficient Std Error t-Ratio
C P 0.017302 0.11805 0.14657
W 0.81018 0.040250 20.129
P-1 0.21623 0.10727 2.0158
CONSTANT 16.555 1.3208 12.534
I P 0.15022 0.17323 0.86718
P-1 0.61594 0.16279 3.7838
K-1 -0.15779 0.036126 -4.3677
CONSTANT 20.278 7.5427 2.6885
W1 X 0.43886 0.035632 12.316
X-1 0.14667 0.038836 3.7767
A 0.13040 0.029141 4.4746
CONSTANT 1.5003 1.1478 1.3071
Asymptotic Variance-Covariance Matrix of Equations

C 1.0441
I 0.43785 1.3832
W1 -0.38523 0.19261 0.47643
C I W1
Asymptotic Variance-Covariance Matrix of Coefficients

P 0.013936
W -0.0015260 0.0016200
P-1 -0.0095710 -0.00053085 0.011506
CONSTANT -0.015344 -0.032733 -0.0047520 1.7445
P 0.0040375 0.0012781 -0.0046423 -0.045188 0.030008
P-1 -0.0031110 -0.0011810 0.0051395 0.017372 -0.025772
K-1 -0.00037500 0.00039811 -0.00013045 -0.0080439 0.0041898
CONSTANT 0.057936 -0.082066 0.020400 2.1124 -0.92485
X -0.0018149 5.1138E-005 0.0011799 0.0092111 0.00059433
X-1 0.0018888 -8.0164E-005 -0.0017884 0.00071044 -0.00058617
A 0.00066034 -0.00052372 0.00047623 0.0027719 0.00047661
CONSTANT -0.00052229 0.0015772 0.032840 -0.61273 -0.0017044
P W P-1 CONSTANT P
P-1 0.026499
K-1 -0.0038717 0.0013051
CONSTANT 0.77760 -0.26903 56.892
X -0.00033125 -9.0028E-005 0.013436 0.0012696
X-1 0.00061121 9.9812E-005 -0.020120 -0.0011997 0.0015082
A -0.00066991 0.00017448 -0.032062 -0.00030915 5.1023E-005
200
CONSTANT -0.015548 -0.00038089 0.36894 -0.0066842 -0.015405

P-1 K-1 CONSTANT X X-1
A 0.00084920
CONSTANT 0.015608 1.3174
A CONSTANT
The system estimation methods such as 3SLS and FIML may be iterated until the
convergent solution is found. To control the iteration, as in the case of nonlinear
iteration methods, the following input control variables may be applied: _iter,
_tol, and _restart. We refer readers to Appendix A for more details of these
control variables. For example, the statement _iter=100 of line 30 sets the
number of iterations for 3SLS and FIML estimation. This is to ensure the near
efficiency (theoretically speaking) of the parameter estimates, if they are found
before exhausting the limit of iterations. It will be warned that the results may be
unreliable when iterations exceed the limit.
In Lesson 13.1, line 33,

_method = 4;
indicates that the instrumental variables method of FIML is used for estimation. It is
equivalent to state the scalar 4 as a vector {4,0}. The alternatives are either setting
_method to {4,1} for a linearized FIML or setting it to {4,2} for the Newton
method. It is mind-boggling that, for the same problem, not all the FIML methods
will converge to the same solution (provided there is convergence of a solution). It is
now common wisdom that different methods may produce different results due to
different algorithms in use for nonlinear model estimation.
We now present the summary table of the parameter estimates obtained from five
methods we use in estimating the Klein Model I. Numbers in parentheses are
asymptotic standard errors.
Eq. Variable OLS LIML 2SLS 3SLS FIML/IV

C P 0.19293 -0.22251 0.017302 0.16451 -0.23190
(0.082065) (0.20175) (0.11805) (0.096198) (0.23178)
W 0.79622 0.82256 0.81018 0.76580 0.80182
(0.035939) (0.055378) (0.040250) (0.034760) (0.037137)
P-1 0.089885 0.39603 0.21623 0.17656 0.38545
(0.081559) (0.17360) (0.10727) (0.090100) (0.18362)
Const. 16.237 17.148 16.555 16.559 18.340
(1.1721) (1.8403) (1.3208) (1.2244) (1.8637)
I P 0.47964 0.075185 0.15022 -0.35651 -0.80082
(0.087377) (0.20218) (0.17323) (0.26015) (0.35761)
P-1 0.33304 0.68039 0.61594 1.0113 1.0516
(0.090747) (0.18817) (0.16279) (0.24876) (0.30823)
K-1 -0.11179 -0.16826 -0.15779 -0.26019 -0.14811
(0.024048) (0.040798) (0.036126) (0.050868) (0.033826)
Const. 10.126 22.591 20.278 42.895 27.267
(4.9175) (8.5458) (7.5427) (10.593) (7.7850)
W1 X 0.43948 0.43394 0.43886 0.37478 0.23412
(0.029158) (0.067937) (0.035632) (0.031103) (0.045546)
X-1 0.14609 0.15132 0.14667 0.19365 0.28464
(0.033671) (0.067054) (0.038836) (0.032402) (0.042736)
201
A 0.13025 0.13159 0.13040 0.16792 0.23487

(0.028711) (0.032386) (0.029141) (0.028929) (0.033488)
Const. 1.4970 1.5262 1.5003 2.6247 5.7963
(1.1427) (1.1884) (1.1478) (1.1956) (1.7621)
Log-Likelihood -141.92156 -122.81636 -121.56134 -86.14210 -83.32381
Lesson 13.2: Klein Model I Reformulated

In lesson13.1, we saw the use of the input variable _begin (line 23) to control the
starting data observation for model estimation, and of _vcov (line 20) to show the
estimated variance-covariance matrix of the parameters. There are many other input
variables useful for controlling single equation model estimation that are applicable
to system model estimation. These include, but are not limited to: _const,
_dlags, and _restr. In addition, for presenting the estimation results, control
variables such as _rstat and _rplot are useful. For details of using these input
control variables, we refer the reader to Appendix A: GPE Control Variables.
lesson13.2 below is a reformulation of Klein Model I. There are no new things

added. Instead, it demonstrates the use of a few of the input control variables
mentioned above in order to offer an alternative formulation of the model. This may
be a more useful representation for the purpose of forecasting and simulation. In the
following, we will list the program without running it.
/*
** Lesson 13.2: Klein’s Model I Reformulated
** Using _dlags and _restr
*/
1 use gpe2;
3 load data[23,10] = gpe\klein.txt;
4 a=data[2:23,1]-1931; @ time trend: 1931 = 0 @

13 t=data[2:23,10]; @ tax @
14 k=k1[2:22]|209.4; @ capital stock @

@ do not include lagged endog. var. in xvar and _names @

16 yvar=c~i~w1~x~p~k;
17 xvar=w2~a~g~t;
18 call reset;
19 _names={"c","i","w1","x","p","k",
"w2","a","g","t"};
20 _vcov=1;
@ do not include lagged endog. var. in _eq and _id @

@ C I W1 X P K W2 A G T 1 @
21 _eq = {-1 0 1 0 1 0 1 0 0 0,
0 -1 0 0 1 0 0 0 0 0,
0 0 -1 1 0 0 0 1 0 0};
202
22 _id = { 1 1 0 -1 0 0 0 0 1 0,
0 0 -1 1 -1 0 0 0 0 -1,
0 1 0 0 0 -1 0 0 0 0};
@ using _dlags option to add the specified lagged endog. var. @

@ after the entire list of endog. var. in _eq and _id @
@ C I W1 X P K @
23 _dlags = {0 0 0 0 1 0,
0 0 0 0 1 1,
0 0 0 1 0 0,
0 0 0 0 0 0,
0 0 0 0 0 0,
0 0 0 0 0 1};
@ restriction: W1 and W2 share the same coef. in C eq. @

@ EQ: C | I | W1 @
@ VAR: W1 P P1 W2|P P1 K1|X X1 A|q **incl. lagged endog. @
24 _restr = { 1 0 0 -1 0 0 0 0 0 0 0};
25 _method=2; @ 2SLS estimation @

27 _iter=100;
28 _method=3; @ 3SLS estimation (iterative) @
30 _method=4; @ FIML estimation @

32 end;
We notice the use of _dlags in line 23 to specify the lagged endogenous variables
in each equation. _dlags is a G×G matrix whose entries indicate the number of
lags for each endogenous variable (column) in each equation (row). If there is no lag
of a particular endogenous variable in a particular equation, the corresponding entry
in _dlags is set to 0. The resulting lag variables are appended to the list of
endogenous variables to form the complete system. Instead of hard-coding the
relevant lags as in Lesson 13.1, the advantage of using _dlags to specify the
model’s lag structure is to let GPE control the dynamic specification of the model.
This feature will be useful for forecasting and simulation.
Considering linear restrictions of the model’s parameters is a cumbersome task in a

simultaneous equations system. Whenever possible, it is recommended to substitute
out the restrictions (as we did in Lesson 13.1 by defining W = W1+W2).
Nevertheless, we recall that linear restrictions are expressed in the form Rβ=q, where
β is the parameter vector and _restr=[R|q] is the restriction matrix. The number of
rows in _restr is the number of restrictions imposed. We first introduced the use
of variable _restr in Lesson 3.4, to specify the linear restrictions in a single
regression equation. For a system model, in the matrix R, restrictions are stacked
horizontally in accordance with the order of equations. q is a column vector of the
restricted values. Cross equation restrictions, if any, can be coded accordingly. In
general, restrictions on the constant terms are not required. In cases with restrictions
involving the constant term, we have to explicitly treat the constant term as one of
the exogenous variables. The input control variable _const is used for this purpose.
That is, coding the restrictions with the constant term (as one of the exogenous
variables) requires setting _const=0 first. Then the restrictions involving the
constant term can be specified correctly. Normally, a constant term will be added for
each equation, unless otherwise specified. This variable can be a column vector with
0 (no constant) or 1 (with constant) associated with each equation.
203
Leaving Klein Model I for the moment, let’s move on to consider a special class of
simultaneous linear regression equations which has broad application. More
examples involving the usage of the above mentioned input control variables will
follow.
Seemingly Unrelated Regression Equations System (SUR)

Recall the matrix representation of a simultaneous linear equations system: YB + XΓ
= U. Let B = -I (minus identity matrix) and E = -U. Then, the resulting system is a
seemingly unrelated regression system of equations:
Y = XΓ + E
The system is seemingly unrelated because of the correlated error structure of the
model due to the embedded parameter restrictions or data constraints. In other words,
errors in one equation may be correlated with errors in other equations.
The estimation of a cost-minimizing factor demand system, developed by Berndt and

Wood (1975), is an application of seemingly unrelated regression. The system is
derived from an indirect translog cost function of four factors: capital (K), labor (L),
energy materials (E), and non-energy materials (M). Assuming constant returns to
scale and price normalization, the real unit cost function is:
ln(c) = β0 + ∑i=K,L,E βi ln(pi) + ½ ∑i=K,L,E ∑j=K,L,E βij ln(pi) ln(pj)
where c = (C/PM)/Q is the normalized unit cost (C is total cost and Q is total output),
and the normalized factor price is pi = Pi/PM for i = K, L, E. All the βs are unknown
parameters of the cost function. Invoking the Shepard Lemma, we can derive the
factor shares as Si = PiXi/C = ∂ln(c)/∂ln(pi), where Xi is the quantity demanded of the
i-th factor (i = K, L, E). Therefore, adding the error terms, the system of factor
demand equations for model estimation is written as:
Si = βi + ∑j=K,L,E βij ln(pj) + εi

i = K, L, E
The symmetry condition, βij = βji, must be imposed as parameter restrictions. We

note that the factor M is treated as a numeraire, and it is clear that
SM = 1- ∑i=K,L,E Si
Lesson 13.3: Berndt-Wood Model

The price and quantity data of the Berndt-Wood Model come in two files: bwp.txt
and bwq.txt. The time series from 1947 to 1971 covers the period before oil
embargo. Quantity series are in billions of dollars, and the price series are the 1947-
based indexes.
/*
** Lesson 13.3: Berndt-Wood Model
** Seemingly Unrelated Regression Estimation
** Factor Shares System with Symmetry Restrictions
*/
1 use gpe2;
204
3 n=26;
4 load x[n,6] = gpe\bwq.txt;
5 year=x[2:n,1];
6 qy=x[2:n,2];
7 qk=x[2:n,3];
8 ql=x[2:n,4];
9 qe=x[2:n,5];
10 qm=x[2:n,6];
11 load x[n,6] = gpe\bwp.txt;
12 py=x[2:n,2];
13 pk=x[2:n,3];
14 pl=x[2:n,4];
15 pe=x[2:n,5];
16 pm=x[2:n,6];
17 tc=pk.*qk + pl.*ql + pe.*qe + pm.*qm;

18 sk=(pk.*qk)./tc;
19 sl=(pl.*ql)./tc;
20 se=(pe.*qe)./tc;
21 sm=(pm.*qm)./tc;
22 pk=ln(pk./pm);
23 pl=ln(pl./pm);
24 pe=ln(pe./pm);
25 yv=sk~sl~se;
26 xv=pk~pl~pe;
27 call reset;
28 _names={"sk","sl","se","pk","pl","pe"};
@ PK PL PE|PK PL PE|PK PL PE| q @

29 _restr = { 0 1 0 -1 0 0 0 0 0 0,
0 0 1 0 0 0 -1 0 0 0,
0 0 0 0 0 1 0 -1 0 0};
30 _method=0;
31 call estimate(yv,xv);
32 _method=1;
34 _method=2;
36 _iter=50;
37 _method=3;
39 _method=4;
41 _method={4,1};
43 _method={4,2};
45 end;
The program is rather straightforward. It loads the two data files and calculates the
necessary variables such as factor shares and normalized prices needed for model
estimation. We do not use either _eq or _id to define the equations system. First,
there are no identity equations. All the equations in the model are stochastic. The
model is in the form Y = XΓ + E where Y corresponds to yv and X corresponds to
xv in the program, and all variables in the xv matrix appear in each equation. Recall
that this is exactly the default structure of the _eq matrix.
205
For the Berndt-Wood example of three equations or endogenous variables (factor

shares) and three exogenous variable (factor prices), the _eq matrix could be
specified as follows:
_eq = {-1 0 0 1 1 1,
0 –1 0 1 1 1,
0 0 –1 1 1 1};
The more challenging task is to specify the symmetry condition for the parameters.
As shown in lesson13.2, linear restrictions are expressed in Rβ= q and _restr =
[R|q]. Recall that the number of rows in _restr is the number of restrictions
imposed. As explained before, the restrictions are stacked horizontally in matrix R,
and q is a column vector of the restricted values. There are three symmetry
restrictions for the Berndt-Wood Model across three equations: βKL = βLK, βKE = βEK,
and βLE = βEL. Line 29 does exactly that by setting the correct entries in the _restr
matrix. No restrictions on the constant terms are needed.
The first row of _restr corresponds to the first restriction βKL = βLK. The entry for
the variable PL of SK equation is set to 1, while the entry for the variable PK of SL
equation is –1. Since there is a zero for the q’s column, it amounts to βKL - βLK = 0,
or βKL = βLK. By the same token, the other two restrictions, βKE = βEK and βLE = βEL,
are expressed in the second and third rows of _restr, respectively:
@ PK PL PE|PK PL PE|PK PL PE| q @
_restr = { 0 1 0 -1 0 0 0 0 0 0,
0 0 1 0 0 0 -1 0 0 0,
0 0 0 0 0 1 0 -1 0 0};
Although we estimate the model with all the available methods, there are only two
sets of solutions. One is from the limited information method, and the other from full
information. Of course, the single equation method is not appropriate for a seemingly
unrelated equations system in which parameter restrictions bind the equations
together. Cross equation covariance is of interest in the multiple equations system.
Iterative 3SLS and FIML are the methods of choice for this particular class of model.
In the literature, the estimated parameters of a production function are rich in terms
of elasticity interpretation. Elasticities are simple to compute once the parameters
have been estimated according to the formula:
Elasticities of Substitution Price Elasticities

β ij + S i S j
ζ ij = η ij = S j ζ ij
Si S j
β ii + Si (Si − 1)
ζ ii = i, j = K, L, E, M
Si2
We leave the elasticity interpretation of the production function to the reader. Here is
the summary output of 3SLS estimation of the Berndt-Wood Model (see
output13.2 for the estimated results by equation):

----------------------------------------
206
Three Stages Least Squares Estimation

Tolerance = 0.001
Iteration = 1 Log Likelihood = 344.55

Parameters = 0.030634 -0.00035814 -0.0097343 0.057091 -0.00035814
0.075102 -0.0044159 0.25349 -0.0097343 -0.0044159 0.018921
0.044330
Iteration = 2 Log Likelihood = 344.57
Parameters = 0.029791 -0.00038017 -0.010208 0.057029 -0.00038017
0.075403 -0.0044256 0.25340 -0.010208 -0.0044256 0.018761
0.044291


SK PK 0.029791 0.0059443 5.0117
PL -0.00038017 0.0038638 -0.098392
PE -0.010208 0.0034011 -3.0014
CONSTANT 0.057029 0.0013574 42.013
SL PK -0.00038017 0.0038638 -0.098392
PL 0.075403 0.0068108 11.071
PE -0.0044256 0.0024401 -1.8137
CONSTANT 0.25340 0.0021210 119.47
SE PK -0.010208 0.0034011 -3.0014
PL -0.0044256 0.0024401 -1.8137
PE 0.018761 0.0053539 3.5042
CONSTANT 0.044291 0.00088399 50.103

SK 9.9232E-006
SL 8.0000E-006 2.8720E-005
SE 4.6387E-006 4.7966E-006 3.1884E-006
SK SL SE
Lesson 13.4: Berndt-Wood Model Extended

Extending from the basic system model of factor demand equations, the stochastic
unit cost function:
ln(c) = β0 + ∑i=K,L,E βi ln(pi) + ½ ∑i=K,L,E ∑j=K,L,E βij ln(pi) ln(pj) + εc
may be added to the model of Lesson 13.3 to form a four-equation system. The idea
is to explicitly estimate the cost function from which the factor share equations are
derived. In particular, this model allows us to estimate the scale parameter β0 of the
cost function. In addition, both first-order βi and second-order βij parameters are
constrained to equal the corresponding parameters of the factor share equations.
The parameter restrictions are certainly more involved in the extended model. Since
the restrictions involve constant terms of each equation, we need to address the issue
of regression intercept explicitly. In lesson13.4 below, we first define the constant
vector one in line 32, and include it in the list of exogenous variables xv in line 34.
The model is then estimated without the constant term or _const=0 (line 38). Line
39 specifies 13 linear restrictions among 23 variables (including constant terms for
each equation). Identifying and restricting the parameters of the unit cost function
with those of derived factor share equations is accomplished in the first 10 rows of
207
_restr. The last 3 rows are the familiar three symmetry conditions across factor
demand equations, as specified in Lesson 13.3.
/*
** Lesson 13.4: Berndt-Wood Model Extended
** Seemingly Unrelated Regression Estimation
** Full System with Restrictions
*/
1 use gpe2;
3 n=26;
4 load x[n,6] = gpe\bwq.txt;
5 year=x[2:n,1];
6 qy=x[2:n,2];
7 qk=x[2:n,3];
8 ql=x[2:n,4];
9 qe=x[2:n,5];
10 qm=x[2:n,6];
11 load x[n,6] = gpe\bwp.txt;
12 py=x[2:n,2];
13 pk=x[2:n,3];
14 pl=x[2:n,4];
15 pe=x[2:n,5];
16 pm=x[2:n,6];
17 tc=pk.*qk + pl.*ql + pe.*qe + pm.*qm;

18 sk=(pk.*qk)./tc;
19 sl=(pl.*ql)./tc;
20 se=(pe.*qe)./tc;
21 sm=(pm.*qm)./tc;
22 c=ln(tc./pm./qy);
23 pk=ln(pk./pm);
24 pl=ln(pl./pm);
25 pe=ln(pe./pm);
26 pkpk=0.5*pk.*pk;
27 pkpl=pk.*pl;
28 pkpe=pk.*pe;
29 plpl=0.5*pl.*pl;
30 plpe=pl.*pe;
31 pepe=0.5*pe.*pe;
32 one=ones(rows(c),1);
33 yv=sk~sl~se~c;
34 xv=pk~pl~pe~pkpk~pkpl~pkpe~plpl~plpe~pepe~one;
35 call reset;
36 _names={"sk","sl","se","c","pk","pl","pe",
"pkpk","pkpl","pkpe","plpl","plpe","pepe","one"};
37 _iter=50;
38 _const=0;
@ |----yv----|------------xv--------------| @
@ SK SL SE C PK PL PE KK KL KE LL LE EE 1 @
39 _eq[4,14] = { -1 0 0 0 1 1 1 0 0 0 0 0 0 1,
0 -1 0 0 1 1 1 0 0 0 0 0 0 1,
0 0 -1 0 1 1 1 0 0 0 0 0 0 1,
0 0 0 -1 1 1 1 1 1 1 1 1 1 1};
40 _restr[12,23] =
@ P P P K K K L L E @
@PK PL PE 1|PK PL PE 1|PK PL PE 1|K L E K L E L E E 1|q @
{0 0 0 -1 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 -1 0 0 0 0 0 1 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 0 -1 0 0 1 0 0 0 0 0 0 0 0,
-1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0,
0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0,
0 0 0 0 0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 1 0 0,
208
0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0,
0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0,
0 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0,
0 1 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 1 0 0 0 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0 0,
0 0 0 0 0 0 1 0 0 -1 0 0 0 0 0 0 0 0 0 0 0 0 0};
41 _method=3;
43 end;
The extended Berndt-Wood Model is estimated with 3SLS. To save space, we leave
out the lengthy results of the model estimation. You should run the program and
check out the results yourself. Instead, we compare the parameter estimates for two
versions of the Berndt-Wood Model. We do not expect the parameter estimates to be
the same, or close, for the two models, even though the same 3SLS method is used
for model estimation. Numbers in the parentheses are standard errors.
Eq. Variable Basic Model Extended Model

SK ln(PK) 0.029791 (0.0059443) 0.040482 (0.0048854)
ln(PL) -0.00038017 (0.0038638) 0.029914 (0.0031598)
ln(PE) -0.010208 (0.0034011) -0.0043608 (0.0026280)
Constant 0.057029 (0.0013574) 0.049289 (0.0015800)
SL ln(PK) -0.00038017 (0.0038638) 0.029914 (0.0031598)
ln(PL) 0.075403 (0.0068108) 0.096927 (0.0065949)
ln(PE) -0.0044256 (0.0024401) 0.014914 (0.0019469)
Constant 0.25340 (0.0021210) 0.24955 (0.0022281)
SE ln(PK) -0.010208 (0.0034011) -0.0043608 (0.0026280)
ln(PL) -0.0044256 (0.0024401) 0.014914 (0.0019469)
ln(PE) 0.018761 (0.0053539) 0.019229 (0.0047603)
Constant 0.044291 (0.00088399) 0.039379 (0.0010016)
C ln(PK) 0.049289 (0.0015800)
ln(PL) 0.24955 (0.0022281)
ln(PE) 0.039379 (0.0010016)
ln(PK)ln(PK) 0.040482 (0.0048854)
½ ln(PK)ln(PL) 0.029914 (0.0031598)
½ ln(PK)ln(PE) -0.0043608 (0.0026280)
ln(PL)ln(PL) 0.096927 (0.0065949)
½ ln(PL)ln(PE) 0.014914 (0.0019469)
ln(PE)ln(PE) 0.019229 (0.0047603)
Constant -0.16689 (0.010360)
Log-Likelihood 344.56744 390.87112
Based on the Likelihood Ratio statistic: -2(344.56744-390.87112) = 92.60736,

compared with the critical value 18.31 (Chi-square distribution of 10 degrees of
freedom at 95% significance level), the gain in efficiency of the parameter estimates
using the full system, including the unit cost function and share equations, is
obvious.

The method of full information maximum likelihood is intrinsically a nonlinear
optimization method. We can write a general representation of a system of nonlinear
equations as follows:
209
F(Y,X,β) = U
Assuming N data observations, Y is an N by G data matrix of G endogenous

variables, X is an N by K data matrix of K predetermined variables (including
exogenous and lagged variables), β is the parameter matrix, and U is an N by G error
matrix of G stochastic equations. Let
F = [F1, F2, ..., FG]'

Y = [Y1, Y2, ..., YG]
X = [X1, X2, ..., XK]
U = [U1, U2, ..., UG]'
β = [β1, β2, ..., βG]'
Then, the model can be rewritten in G separate stochastic equations as:
F1(Y,X,β1) = U1
F2(Y,X,β2) = U2
…
FG(Y,X,βG) = UG
Note that identity equations are substituted out in order to avoid the complication of
using constrained optimization. Also, not all the columns of data matrices Y and X
are used in each equation of the system. However, there must be at least one distinct
endogenous variable appearing in each equation. The parameter vector βj effectively
selects the variables included in the equation j.
Log-Likelihood Function of Nonlinear Equations System

We now briefly describe the methodology of nonlinear FIML. Chapter VI covered
the general framework of nonlinear optimization, and must be reviewed. Many
econometrics textbooks and journal articles discuss the implementation in more
detail.
Assume U is normally independently distributed with mean 0 (a G-element zero

vector) and covariance Σ⊗I. That is, Cov(Uj,Uk) = σjk, where σjk is an element of the
G by G cross-equation covariance matrix Σ. For each equation j, Uj is zero in mean
and it has homogeneous variance σjj or σj2.
Constructing from the joint normal probability density of Ui and the Jacobian factor
Ji = Ji(β) = det(∂Ui./∂Yi.) for each observation i, the likelihood function is
(2π)-G/2 |det(Σ)|-1/2 exp(-½UiΣ-1Ui') |det(Ji)|
It follows immediately that the log-likelihood function ll(β,Σ|Y,X) for a sample of N

observations is
-NG/2 ln(2π) -N/2 ln|det(Σ)| -½ Σi=1,2,...,N UiΣ-1Ui' + Σi=1,2,...,N ln|det(Ji)|
The concentrated log-likelihood function is obtained by substituting the estimated

covariance matrix Σ = U'U/N as follows:
210
ll*(β|Y,X) = -NG/2 (1+ln(2π)) -N/2 ln(det(U'U/N)) + Σi=1,2,...,N ln|det(Ji)|
The FIML estimator of the parameters vector β = [β1, β2, ..., βG]' is obtained by
maximizing the above concentrated log-likelihood function.
Special Case: Linear Equations System

A linear equations system is typically represented with the matrix form YB + XΓ =
U, where the sparse parameter matrices B and Γ are used to identify the variables
included in the respective equations (and the identity restrictions, if any). We adopt
the notation β to indicate the combined elements of estimated parameters in B and Γ.
To estimate the linear model YB + XΓ = U, the Jacobian term is the same for each
data observation. That is, Ji = det(∂Ui/∂Yi) = det(B) for all i = 1,2,…,N. Thus the
corresponding concentrated log-likelihood function is
ll*(β|Y,X) = -NG/2 (1+ln(2π)) -N/2 ln(det(U'U/N)) + N ln|det(B)|
Lesson 13.5: Klein Model I Revisited

We are now ready to estimate the Klein Model I using nonlinear FIML. First, we
need to eliminate the three identities by substituting them into the stochastic
equations so that there are three endogenous variables corresponding to the three
stochastic equations in the original presentation of the model. One representation of
the model (see Goldfeld and Quandt, 1972, p.34) is
P = a0 + a1 (W1+W2) + a2 (K-K-1+G+W2-T) + a3 P-1 + u1

W1 = b0 + b1 (P+T) + b2 X-1 + b3 A + u2
K = r0 + r1 P + r2 P-1 + r3 K-1 + u3
Note that the original Klein model (Klein, 1950) used a variable G' = G + W2, and
many parameter restrictions have to be built into the system for correct estimation.
To represent the model in the form YB + XΓ = U, we have
Y = [P W1 K]
X = [P1 K1 X1 W2 (G+W2) T A 1]
-1 b1 r1 a3 0 r2
B =  a1 -1 0
 a1 0 -1
 -a2 0 r3 
 
0 b2 0
a1 0 0
Γ=
 
a2 0 0
-a2 b1 0
 0 b3 0
1 1 1 
/*
** Lesson 13.5: Klein Model I Revisited
** Nonlinear FIML Estimation, Goldfeld-Quandt (1972), p.34
*/
1 use gpe2;
3 load data[23,10]=gpe\klein.txt;
211
4 a=data[2:23,1]-1931; @ time trend: 1931 = 0 @

13 t=data[2:23,10]; @ tax @
14 k=k1[2:22]|209.4; @ capital stock @

@ use klein original data @

16 data=packr(p~w1~k~lag1(p~k~x)~w2~(g+w2)~t~a);
17 data=data-meanc(data)'; @ change data format to: @
@ deviation-from-mean form @
18 call reset; @ reset default control variables @
19 _nlopt=1;
20 _method=6; @ modified quadratic hill climbing @
21 _iter=100; @ set 50 maximal iterations @
22 _step=1;
23 _tol=1.0e-4; @ set tolerance level @
24 _conv=2; @ set convergence
25 criteria @
26 _vcov=1;
27 _b={0.2041,0.1025,0.22967, @ initial parameter values @
0.72465,0.23273,0.28341, @ for nonlinear FIML @
0.23116,0.541,0.854};
28 _names={"a1","a2","a3","b1","b2","b3","r1","r2","r3"};
29 call estimate(&klein,data);
30 end;
31 proc klein(x,c); @ klein model 1 @

32 local n,u,beta,gama,a,b,r,ll;
33 a=c[1 2 3]; b=c[4 5 6]; r=c[7 8 9];
34 n=rows(x); @ number of observations @
35 beta=-eye(3); @ initialize beta @
36 gama=zeros(7,3); @ initialize gama @
37 beta[2,1]=a[1]; @ assign values to beta and gama @
38 beta[3,1]=a[2];
39 gama[1,1]=a[3];
40 beta[1,2]=b[1];
41 gama[3,2]=b[2];
42 gama[7,2]=b[3];
43 beta[1,3]=r[1];
44 gama[1,3]=r[2];
45 gama[2,3]=r[3];
46 gama[4,1]=a[1]; @ parameter restrictions @
47 gama[5,1]=a[2];
48 gama[6,1]=-a[2];
49 gama[2,1]=-a[2];
50 gama[6,2]=b[1];
51 u=x[.,1:3]*beta+x[.,4:10]*gama; @ stochastic errors @
52 ll=-0.5*n*3*(1+ln(2*pi))+ @ log-likelihood value @
-0.5*n*ln(det(u'u/n))+n*ln(abs(det(beta)));
53 retp(ll);
54 endp;
Gamma is a built-in function in GAUSS to compute the gamma function, therefore

we use gama for the variable name in Lesson 13.5 above. It is not a typo.
212
The first part of the data manipulation is the same as in the linear system of Lesson
13.1. The relevant input variables controlling nonlinear optimization are discussed in
chapters VI and VII. The objective log-likelihood function is defined in lines 30
through 53. For nonlinear maximum likelihood estimation, we reduce the size of the
problem by using the deviation from the mean of the data series so that the constant
terms of each equation are eliminated (see line 17). The following is the result of
running lesson13.5:

---------------------------------------------

Tolerance = 0.0001
Initial Result:
Parameters = 0.20410 0.10250 0.22967 0.72465 0.23273
0.28341 0.23116 0.54100 0.85400
Using Modified Quadratic Hill-Climbing Algorithm

Parameters = 0.076508 0.36162 0.45511 0.68255 0.31692
0.22009 0.095885 0.41955 0.96932
…
Parameters = -0.16079 0.81143 0.31295 0.30568 0.37170
0.30662 -0.80101 1.0519 0.85190
Final Result:
Parameters = -0.16079 0.81143 0.31295 0.30568 0.37170
0.30662 -0.80101 1.0519 0.85190
A1 -0.16079 0.098663 -1.6297
A2 0.81143 0.38368 2.1149
A3 0.31295 0.11847 2.6417
B1 0.30568 0.16223 1.8843
B2 0.37170 0.049169 7.5596
B3 0.30662 0.047628 6.4378
R1 -0.80101 0.84311 -0.95007
R2 1.0519 0.42533 2.4730
R3 0.85190 0.046840 18.187

A1 0.0097344
A2 -0.035851 0.14721
A3 0.0064035 -0.031598 0.014034
B1 -0.011472 0.041600 -0.0049678 0.026317
B2 0.0017000 -0.0066728 0.0015841 -0.0038648 0.0024176
B3 0.00042010 -0.0040841 0.00081371 -0.0036929 -0.00025486
R1 -0.071452 0.30359 -0.066130 0.084909 -0.0091796
R2 0.027921 -0.12886 0.043653 -0.020684 0.0067868
R3 0.0029682 -0.010417 0.0014621 -0.0059803 0.0014133
A1 A2 A3 B1 B2
B3 0.0022684
R1 -0.0095885 0.71083
213
R2 -0.00072454 -0.30500 0.18091

R3 0.00081869 -0.015241 0.0034790 0.0021940
B3 R1 R2 R3
As a mathematical exercise, you may want to verify the relationships of parameters

between the two representations of Klein Model I we discussed in Lesson 13.1 and
13.5:
a0 = α0/(1-α1) b0 = γ0/(1-γ1) r 0 = β0
a1 = (α2-1)/(1-α1) b1 = γ1/(1-γ1) r 1 = β1
a2 = 1/(1-α1) b2 = γ2/(1-γ1) r 2 = β2
a3 = α3/(1-α1) b3 = γ3/(1-γ1) r3 = β3+1
The advantage of nonlinear FIML is to allow for nonlinear equations in a

simultaneous system, but the computation becomes more involved in defining the
nonlinear function for numerical optimization. Because the Klein Model includes
only the linear equations, it is no surprise that the parameter estimates obtained from
the nonlinear FIML method and from a variant of linear instrumental variable FIML
method (using the option _method=4, as shown in the table at the end of Lesson
13.1) are very close, if not numerically identical.
Nonlinear FIML FIML/IV

α1 -0.2324 -0.23190
α2 0.80184 0.80182
α3 0.38568 0.38545
α0 18.340
β1 -0.801 -0.80082
β2 1.0518 1.0516
β3 -0.1481 -0.14811
β0 27.267
γ1 0.23412 0.23412
γ2 0.28468 0.28464
γ3 0.23483 0.23487
γ0 5.7963
Log-Likelihood -83.324 -83.32381
214
XIV
Unit Roots and Cointegration
So far the econometric models we have constructed are mostly based on economic
theory or empirical evidence. In many situations involving time series data, we will
have to rely on information drawn from the data generating process (DGP) . An
example of this would be a time series with an autocorrelated error structure.
Considering a time series as a DGP, the data may possess a trend, cycle, or
seasonality (or any combination). By removing these deterministic patterns, we
would hope that the remaining DGP is stationary. However, most nonstationary data
series are stochastic. “Spurious” regressions with a high R-square but a near-two
Durbin-Watson statistic, often found in time series literature, are mainly due to the
use of stochastic nonstationary data series.
Given a time series DGP, testing for a random walk is a test of stationarity. It is also
called a unit roots test. Testing for the problem of unit roots for each time series is
more of a process than it is a step. This chapter will chart the procedure to test for
unit roots. If a problem is identified, the original data are differenced and tested
again. In this way, we are able to identify the order of the integrated process for each
data series. Once all data series have completed this process, they are regressed
together and tested for a cointegrating relationship.
Since the tests we use, Dickey-Fuller (DF) and augmented Dickey-Fuller (ADF),
require the model’s error structure to be individually independent and
homogeneously distributed, anomalous serial correlation in time series must be
treated before these tests can be applied. Therefore, serial correlation is tested and
corrected as the pretest step of each unit root test. Instead of directly correcting the
error structure through the integration process, we will modify the dynamics of the
data generating process with lagged dependent variables.
We follow the “top down” approach to carry out both the DF and ADF tests for unit
roots, by testing for the most complicated problems first and then simplifying our
model if problems are absent. We will formulate and test a hierarchy of three
models. First, we estimate the Random Walk Model with trend and drift, or the
Model III, as follows:
∆Xt = α + β t + (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
where the dependent variable ∆Xt = Xt-Xt-1 is the first difference of the data series
Xt. Using augmented lags of dependent variable Σi=1,2,… ρi ∆Xt-i ensures a white noise
εt for the unit root test. The optimal lag may be selected based on criteria such as AIC
(Akaike Information Criterion) and BIC (Schwartz Baysian Information Criterion).
Testing the hypothesis that ρ = 1 (so that the coefficient of Xt-1 is equal to zero) is the
focus of the unit root tests.
If the unit root is not found in the Model III, we continue the process by estimating
the Random Walk Model with Drift, or the Model II, as follows:
∆Xt = α + (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
And finally, if the unit root is not found in the Model II, we estimate the Random
Walk Model, or Model I:
∆Xt = (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
Testing for unit roots is the first step of time series model building. For a univariate
case, several versions of the DF and ADF tests are available. For multivariate time
series, after unit root tests for each variable, a cointegration test should be carried out
to ensure that the multiple regression model is not spurious. For testing cointegration
of a set of variables, the necessary causal relationship among variables may not be
available for the single equation ADF-type testing due to Engle and Granger (1987).
Johansen’s vector autoregression (VAR) representation of the model and the relevant
Likelihood Ratio tests are suggested for the multivariate case.
Testing for Unit Roots

The process of the augmented Dickey-Fuller (ADF) test starts from estimating
Model III in which autocorrelation has been removed from the data series. The ADF
test for unit roots then steps through three models, testing each model’s estimated
coefficients to see if they are statistically significantly different from zero. Computed
t- and F-statistics are compared against critical values from various Dickey-Fuller τ
and φ distributions. Critical values of ADF τ and φ distributions for all three models
are given in Appendix C, Statistical Table C-1 and C-2, respectively. We note that
unit root tests have low power to reject the null hypothesis. Hence, if the null
hypothesis of a unit root is rejected, there is no need to proceed further.
Estimate and Test Model III

∆Xt = α + β t + (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
1. Test ρ = 1, using the ADF τρ distribution (t-statistic) for Model III. If the null
hypothesis is rejected, we conclude that there are no unit roots in X. Otherwise,
continue on to Step 2.
2. Test β = 0 given ρ = 1, using the ADF τβ distribution (t-statistic) or the ADF φ3

distribution (F-statistic) for Model III. If the null hypothesis is rejected, we need to
test ρ = 1 again using the normal distribution as follows (see Step 3). Otherwise, go
to Estimate and Test Model II.
3. Test ρ = 1 using the normal distribution. If the null hypothesis is rejected, we

conclude that there are no unit roots. Otherwise, we conclude that the data series is
nonstationary, and restart the test process using the differenced data series.
Estimate and Test Model II

∆Xt = α + (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
216
UNIT ROOTS AND COINTEGRATION
1. Test ρ = 1, using the ADF τρ distribution (t-statistic) for Model II. If the null
hypothesis is rejected, we conclude that there are no unit roots in X. Otherwise,
continue on to Step 2.
2. Test α = 0 given ρ = 1, using the ADF τα distribution (t-statistic) or the ADF φ1

distribution (F-statistic) for Model II. If the null hypothesis is rejected, we need to
test ρ = 1 again using the normal distribution as follows (see Step 3). Otherwise, go
to Estimate and Test Model I.
3. Test ρ = 1 using normal distribution. If the null hypothesis is rejected, we

conclude that there are no unit roots. Otherwise, we conclude that data series is
nonstationary, and restart the test process using the differenced data series.
Estimate and Test Model I

∆Xt = (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
Test ρ = 1, using the ADF τρ distribution (t-statistic) for Model I. If the null
hypothesis is rejected, we conclude that there are no unit roots in X. Otherwise, we
conclude that the data series is nonstationary, and restart the test process using the
differenced data series.
Many macroeconomic time series have been scrutinized for unit roots and
cointegration. In this chapter, two economic time series, Y (real personal disposable
income) and C (real personal consumption expenditure), from usyc87.txt are used to
illustrate the process of ADF tests for unit roots and cointegration. The same data
series were used in the example of U.S. income-consumption relationship studied in
Chapter XI.
Lesson 14.1: Augmented Dickey-Fuller Test for Unit Roots

Based on time series of personal consumption and personal disposable income from
usyc87.txt, this lesson performs the ADF unit root test procedure. The program is
written to allow straightforward testing of different variables or different
transformations of the same variable.
Personal Consumption Expenditure

This program is designed to easily allow the testing of more than one data series.
Starting at line 7, the variable X is tested for a unit root. To change data series to be
tested, just assign a different data series in line 6 to the variable X. The level (not
differenced) series of C, personal consumption, is examined first.
/*
** Lesson 14.1: Unit Root Tests
*/
1 use gpe2;
4 y = z[2:67,2];
5 c = z[2:67,3];
/* select one variable to work on */
217
6 x = c;
/* difference the data if needed */

7 diff = 0;
8 j = 1;
9 do until j > diff;
10 x = x-lagn(x,1);
11 j = j+1;
12 endo;
13 x1 = packr(lagn(x,1)); @ sample truncated @

14 dx = packr(x-lagn(x,1));
15 trend = seqa(1,1,rows(dx));
16 call reset;
17 _names={"dx","trend","x1"};
18 _rstat = 1;
19 _dlags = 3; @ augmented terms if needed @
/* Model III */
20 call estimate(dx,trend~x1);
21 _restr = {0 0 0 1 0 0,
0 0 0 0 1 0}; @ DF joint test @
22 call estimate(dx,trend~x1);
23 end;
Let’s walk through the program. Lines 7 through 12 introduce a Do Loop to simplify
taking the difference of our data, if needed. Line 7 specifies the number of
differences diff on the data series. Then, from line 8 to 12, a Do Loop is used to
transform the necessary differences for the data series when the variable diff is
greater than 0. In line 7 we begin with the original data series in level:
diff = 0;
The next two lines (lines 13 and 14) work on the selected variable to compute the
lagged and the differenced values necessary for the test specification. A GAUSS
command packr is used to eliminate the initial observations of data which are lost
due to the lag operation. Next, a trend variable is generated (line 15) and included for
the estimation of Model III.
Line 19 is the result of a pretest of the model to ensure a classical or white noise
error structure, which is necessary for the ADF test of unit roots. Through a process
of trial and error, we found that for the consumption series C, the addition of three
lags of the dependent variable to the test equation is enough to remove
autocorrelation and maintain the classical assumption for the model error.
Model III is now estimated and tested for unit roots (line 20). Keep in mind that most
computed t-statistics will fall in the left tail of the ADF τ distribution and will be
negative. The second restricted least squares estimation (lines 21 to 22) is there to
carry out the ADF φ-test (based on the F-statistic) for the joint hypotheses of unit
root and no trend, provided that the first regression equation reveals a unit root. We
note that the definition of the restriction matrix of _restr must take into account
the three lags of dependent variables included in front of the explanatory
independent variables.
The following is the result of the estimated Model III in which three lags of the
dependent variable are augmented for analyzing personal consumption.
218

------------------------
Dependent Variable = DX



Explained 43171. 5 8634.1 11.821 7.9646E-008
Residual 40904. 56 730.42
Total 84074. 61 1378.3

DX1 0.32414 0.12529 2.5872 0.012301 0.10676
DX2 -0.16381 0.13150 -1.2457 0.21807 0.026963
DX3 -0.25278 0.12477 -2.0260 0.047541 0.068291
TREND 2.3924 0.98917 2.4186 0.018859 0.094575
X1 -0.020042 0.018215 -1.1003 0.27590 0.021162
CONSTANT 2.7427 7.8871 0.34775 0.72933 0.0021548

Durbin-H Statistic = -1.4164
Because of the use of lagged dependent variables, the sample range is adjusted. As a
pretest, we see that the errors for the test model are not autocorrelated, therefore
various ADF tests for unit roots are applicable. Starting at Step 1, with the estimated
t-statistic of –1.10 for the coefficient of the lagged variable X1 in the test equation
(vs. the τρ critical value of –3.5 at 5% level of significance, see Table C-1), the unit
root problem is clearly shown. Given the unit root, we continue on to Step 2, testing
the zero-value coefficient of the trend variable. Based on the ADF t-statistic for the
variable TREND, the hypothesis of no trend is barely rejected at a 10% level of
significance. Notice that, from Table C-1, the τβ critical value is 2.81 and 2.38 at 5%
and 10% levels of significance, respectively. However, the joint hypotheses of unit
root and no trend may be better served with the ADF φ-test based on the F-statistic.
The following result of hypothesis testing is due to the restrictions specified in line
21:

F( 2, 56) Prob>F
10.836 0.00010513
219
Wald Chi-Square Test 21.673 2 1.9669e-005

Likelihood Ratio Test 20.284 2 3.9398e-005
With the Wald F-test statistic of 10.836, compared with the critical value of the ADF
φ3 distribution for Model III (6.73 at 5% significance, see Table C-2), the conclusion
of unit root and no trend leads to the confirmation of unit root with a traditional
normal test. Unit root for the variable C is confirmed, so the level series of C is
nonstationary.
Since the level data series is nonstationary, it must be differenced then estimated and
tested again for unit roots. Based on the above program, it is easy to make changes to
carry out the unit root test for the first differenced consumption series. First, line 7
should read as:
diff = 1;
The Do Loop of lines 8 through 12 translates the original level series to the first
difference series. From this point on, the program will evaluate the data series in the
first difference. We also found that there is no need to augment the model with
lagged dependent variables, since the model error is already free of correlation.
Therefore, line 19 is changed to:
_dlags = 0;
Model III is estimated and tested for unit roots for the first difference of the
consumption series. The ADF test for the joint hypotheses of unit root and no trend
in lines 21 to 22 must be modified or deleted. A simple way to remove these lines
from being “seen” by GAUSS is to comment them out in between “/*” and “*/”.
Here is the output of the modified program running the unit root test on the first
difference of the consumption series:

------------------------


Explained 25248. 2 12624. 15.401 3.8510E-006
Residual 50001. 61 819.69
Total 75250. 63 1194.4

TREND 0.91047 0.25521 3.5675 0.00070864 0.17262
X1 -0.66639 0.12008 -5.5496 6.5763E-007 0.33550
CONSTANT 3.0604 7.2425 0.42256 0.67410 0.0029186

220

Based on the ADF t-statistic –5.55 for the lagged variable X1, the conclusion of no
unit root in the first difference data is immediate and obvious. Therefore, we
conclude that the consumption series is an integrated series of order one. We know
this because taking the first difference makes the data stationary.
Personal Disposable Income

We continue on to test the second data series, personal disposable income Y, for unit
root. The original program for analyzing consumption level data is used, except the
variable of interest now selected in line 6 is Y.
x = y;
Also, from the pretest for the classical error structure, it appears that augmenting the
first lag of the dependent variable is necessary for “whitening” the error term.
Therefore, in line 19:
_dlags = 1;
Accordingly, for computing the F-statistic from the second restricted least squares
estimation, we also modify the restriction matrix in line 21:
_restr = {0 1 0 0,
0 0 1 0};
Here is the estimation result of Model III for personal income level series:

------------------------



Explained 44744. 3 14915. 10.012 1.9204E-005
Residual 89380. 60 1489.7
Total 1.3412E+005 63 2128.9

DX1 0.13228 0.12474 1.0605 0.29318 0.018399
TREND 3.2200 1.2776 2.5203 0.014404 0.095731
X1 -0.038953 0.022835 -1.7059 0.093203 0.046257
CONSTANT 3.8784 10.136 0.38263 0.70335 0.0024341
221

We see that by comparing the t-statistic of X1 (-1.71) with the corresponding ADF
critical values (-3.50 at 5% significance, see Table C-1), there is a unit root. Based
on the joint test for unit root and no trend hypotheses, a trend is also presented. This
is the purpose of the second regression estimation. The following Wald F-test result
should be checked with the critical values of the ADF φ3 distribution for Model III
(6.73 at 5% significance, see Table C-2):

F( 2, 60) Prob>F
7.8478 0.00093843
The level series of personal disposable income is clearly nonstationary. By

modifying the program again in line 7,
diff = 1;
and deleting the last part of the ADF joint F-test, the first difference of the income
series augmented with one lag of dependent variable is reexamined as follows:

------------------------



Explained 69607. 3 23202. 14.790 2.7151E-007
Residual 92557. 59 1568.8
Total 1.6216E+005 62 2615.5Variable Estimated
Standard t-Ratio Prob Partial
DX1 -0.073578 0.12989 -0.56648 0.57322 0.0054095
TREND 0.99944 0.35282 2.8327 0.0063057 0.11972
X1 -0.79417 0.16721 -4.7496 1.3442E-005 0.27660
CONSTANT 9.4345 10.400 0.90713 0.36803 0.013755

Durbin-H Statistic = NA
222
Based on the ADF t-test statistic, -4.75, for the lagged variable X1, the first
differenced income series is stationary and free of unit roots. As with consumption,
personal income is an integrated series of order one.
The above example of testing for unit roots in the personal consumption and income
data series is carried out based on Model III. We did not go down the hierarchy
further to test Model II or Model I since most of the test results are clear-cut at the
level of Model III. For other macroeconomic time series, you may be required to test
Model II or Model I as well.
Below is the summarized result of the unit root tests:
Series N Lags ρ-1 (τρ) β (τβ) φ3

C 62 3 -0.02 (-1.10) 2.39 (2.42) 10.84*
Y 64 1 -0.04 (-1.71) 3.22 (2.52)* 7.85*
∆C 64 0 -0.67 (-5.55)*
∆Y 63 1 -0.79 (-4.75)*
All tests are based on Model III. The following annual data series from 1929 to 1994
are tested (in rows): C = Personal consumption expenditure in billions of 1987
dollars; Y = Personal disposable income in billions of 1987 dollars; ∆C = Annual
change in personal consumption expenditure; ∆Y = Annual change in personal
disposable income. Also the following notations are used (in columns): N = number
of observations; Lags = augmented lag terms in the test equation; ρ-1 = estimated
coefficient of the lag variable; β = estimated coefficient of the trend variable; τρ = t-
statistic hypothesis of unit root; τβ = t-statistic hypothesis of no trend, given unit
root; and φ3 = F-statistic hypotheses of unit root and no trend. The asterisk (*)
indicates rejection of the null hypothesis at a 5% statistical significance level based
on ADF distributions (see Table C-1 for critical values of t-statistics and Table C-2
for critical values of F-statistics).
As many previous studies have suggested, income and consumption data are
nonstationary in level, but their first difference or change series are stationary. In
summary, both income and consumption are of the first-order integrated series.
Testing for Cointegrating Regression

The next interesting question is, statistically, can we find a meaningful nonspurious
income-consumption relationship as the classical Permanent Income Hypothesis
claims? To answer this question, we need to look at the problem in a more general
framework of multivariate regression.
Suppose there are M variables, Z1, … , ZM. Let Yt = Zt1 and Xt = [Zt2, ..., ZtM].
Consider the following regression equation:
Yt = α + Xtβ + εt
In general, if Yt, Xt ~ I(1), then εt ~ I(1). But, if εt can be shown to be I(0), then the
set of variables [Yt, Xt] is said to be cointegrated, and the vector [1 -β]' (or any
multiple of it) is called a cointegrating vector. Depending on the number of variables
M, there are up to M-1 linearly independent cointegrating vectors. The number of
223
linearly independent cointegrating vectors that exists in [Yt, Xt] is called the
cointegrating rank.
To test for the cointegration of the set of variables [Yt, Xt], two approaches are used.
If the causality of Y on X is clear, then the Engle-Granger or ADF test based on the
regression residuals may be applied. The alternative is to work with the VAR system
of all variables under consideration. This is the Johansen approach to the
cointegration test, to be discussed later.
Cointegration Test: The Engle-Granger Approach

Based on the regression model
Yt = α + Xtβ + εt
the Engle-Granger test for cointegration is to test for unit root for the residuals of the
above regression model. That is, based on Model I, the auxiliary test equation is
written as:
∆εt = (ρ-1)εt-1 + ut
where εt = Yt - α - Xtβ, and ∆εt is defined as εt - εt-1. The rationale is that if the
residual εt has unit root, regressing Y on X may not completely capture the
underlying (nonstationary) trends of all these variables. The estimated model does
not reveal the meaningful relationship, although it may fit the data well. This is the
crux of the spurious regression problem. However, if a cointegrating vector can be
found among the variables that causes the error term εt to be stationary or I(0), then
we can attach meaning to the estimated regression parameters.
We note that the above unit root test equation on the regression residuals does not
have a drift or trend. In order to apply ADF-type testing for a unit root, the model
may be augmented with lagged dependent variables as needed:
∆εt = (ρ-1)εt-1 + ∑j=1,2,... ρt-j∆εt-j + ut
Alternatively, the cointegrating test regression may be expressed as the following

Error Correction Model:
∆Yt = ∆Xtβ + (ρ-1)(Yt-1 - α - Xt-1β) + ∑j=1,2,... ρt-j (∆Yt-j - ∆Xt-jβ) + ut
If we can reject the null hypothesis of unit root on the residuals εt, we can say that
variables [Yt, Xt] in the regression equation are cointegrated. The cointegrating
regression model may be generalized to include trend as follows:
Yt = α + γ t + Xtβ + εt
Notice that the trend in the cointegrating regression equation may be the result of
combined drifts in X and/or Y. Critical values of the ADF τρ distribution for spurious
cointegrating regression are given in Table C-3 of Appendix C. These values are
based on the work of Phillip and Ouliaris (1990), and depend on the number of
cointegrating variables and their trending behaviors for large samples.
224
Furthermore, MacKinnon’s table of critical values of cointegration tests for

cointegrating regression with and without trend (named Model 2 and Model 3,
respectively) is given in Appendix C, Table C-4. It is based on simulation
experiments by means of response surface regression in which critical values depend
on the type of model, number of variables, and are adjusted for sample size.
MacKinnon’s table is easier and more flexible to use than that of Phillip and
Ouliaris. We note that the univariate case (K=1) of MacKinnon’s table (top portion)
corresponds to the critical values of ADF distributions (testing unit roots for Models
I, II, and III).
Lesson 14.2: Cointegration Test: Engle-Granger Approach

Given that both the income (Y) and consumption (C) series are integrated of order
one (that is, I(1)), the long-run relationship:
Ct = β0 + β1 Yt + εt
will be meaningful only if the error εt is free of unit roots. The test for cointegration
between C and Y thus becomes a unit root test on the regression residuals:
∆εt = (ρ-1) εt-1 + Σj=1,2,… ρj ∆εt-j + ut
/*
** Lesson 14.2: Cointegration Test
** Engle-Granger Approach
*/
1 use gpe2;

4 y = z[2:67,2];
5 c = z[2:67,3];
6 call reset;
7 _names = {"c","y"};
/* Unit Roots Test on Residuals */

9 x = __e; @ set x to regression residuals @
10 x1 = packr(lagn(x,1)); @ sample truncated @
11 dx = packr(x-lagn(x,1));
12 _names = {"dx","x1"};
13 _rstat = 1;
14 _dlags = 2; @ augmented terms if needed @
15 _const = 0; @ no intercept term @
16 call estimate(dx,x1);
17 end;
The program reads in and uses both income (Y) and consumption (C) data series, and
runs a regression of the consumption-income relationship. Here, we are not
interested in investigating or refining the error structure of the regression equation
(though we must make sure that no autocorrelated structure exists in the error term of
the cointegrating regression). Instead, we want to test the residuals for the presence
of unit roots. In GPE, residuals are available as the output variable __e immediately
after the regression equation is estimated. Line 9 sets the variable X to the vector of
residuals:
225
x = __e;
and prepares for unit root testing on this variable in the rest of the program. This later
portion of codes (lines 10 through 16) is the same as that in lesson14.1 for testing
unit roots of a single variable. Again, line 14 is the result of a pretest to ensure the
white noise error structure for unit root test:
_dlags = 2;
It turns out that we need to have two lags of the dependent variable augmented to the
test equation. We recall that both income (Y) and consumption (C) variables include
linear trend from our earlier unit roots tests on the respective variable. This fact must
be considered when we use the appropriate ADF τρ distribution for cointegration
tests (using Model 2a or Model 3 of Table C-3). The alternative is to use
MacKinnon’s table (Table C-4) for testing the cointegrating regression model. We
present only the results relevant to the cointegration test in the following (see the
generated output file output14.2 for more details):

------------------------

WARNING: Constant Term Suppressed.



Explained 10786. 3 3595.5 5.2123 0.0028932
Residual 41388. 60 689.80
Total 52190. 63 828.41

DX1 0.33306 0.12268 2.7149 0.0086440 0.10941
DX2 0.17315 0.13228 1.3089 0.19555 0.027762
X1 -0.29001 0.082515 -3.5146 0.00084459 0.17073

Testing for the cointegration of two variables, C and Y with trend, the computed t-
statistic for the lagged variable X1 in the test equation is –3.52, which is right on the
borderline of rejecting the null hypothesis of unit root at a 5% level of significance
(looking at Table C-3, the critical value of ADF cointegration t-statistic τρ for K=2 at
226
5% is –3.42 for Model 2a). A similar conclusion is obtained by using the critical
values of MacKinnon (Table C-4). Although these results do not give us
overwhelming confidence that the long-run income-consumption relationship is
legitimate, empirical studies based on the Permanent Income Hypothesis still stand.
Single equation cointegration tests can only be valid when the specific causal
relation of the underlying multiple regression is correct. If the causal relationship of
C and Y is not as clean-cut as the Permanent Income Hypothesis suggests, we need
to run and test the reverse regression equation.
Cointegration Test: The Johansen Approach

The Engle-Granger cointegration test discussed in the previous section is only
appropriate when the direction of causality involved in the regression equation is
clear. If there are more than two variables involved in a regression model, the
direction of causality may not be clear, or one-sided. In this case, we turn to
Johansen’s multivariate cointegration test.
Given a set of M variables Zt=[Zt1, Zt2, ..., ZtM], and considering their feedback
simultaneity, Johansen’s cointegration test based on FIML (full information
maximum likelihood) is derived from the following:
• VAR (vector autoregression) System Model Representation

• FIML Estimation of the Linear Equations System
• Canonical Correlations Analysis
Similar to the random walk (unit roots) hypothesis testing for a single variable with
augmented lags, we write a VAR(p) linear system for the M variables Zt:
Zt = Zt-1Π1 + Zt-2Π2 + ... + Zt-pΠp + Π0 + Ut
where Πj, j=1,2,...M, are the MxM parameter matrices, Π0 is a 1xM vector of
deterministic factors (drifts and trends). Moreover, we assume the 1xM error vector
Ut is independently normally distributed with a zero mean and a constant covariance
matrix Σ = Var(Ut) = E(Ut'Ut) across M variables.
The VAR(p) system can be transformed using the difference series of the variables,
resembling the error correction model, as follows:
∆Zt = ∆Zt-1Γ1 + ∆Zt-2Γ2 + ... + ∆Zt-(p-1)Γp-1 + Zt-1Π + Γ0 + Ut
where I denotes the identity matrix, Π = ∑j=1,2,...,pΠj - I, Γ1 = Π1 - Π - I , Γ2 = Π2 + Γ1,
Γ3 = Π3 + Γ2, … , and Γ0 = Π0 for notational convenience. Recall that Γ0 is a vector
of deterministic factors including drifts and trends. If both drift and trend (µ0 + µ1t)
exist in Zt, then Γ0 = -µ0 Π + µ1(Γ+Π) - µ1Π t where Γ = I - ∑j=1,2,...,p-1Γj.
A few words about the vector Γ0 (or Π0) of the deterministic factors. We consider
only the case of constant vector Γ0 that is restricted such that µ1Π = 0 (no trend),
then Γ0 = -µ0 Π + µ1 Γ. It is easy to see that (1) if µ1 = 0, µ0 is the only deterministic
factor (drift) for Zt, or Γ0 = -µ0 Π; (2) if µ1 ≠ 0, then the VAR(p) model consists of
drift and linear trend components, or Γ0 = -µ0 Π + µ1 Γ.
227
If Zt ~ I(1), then ∆Zt ~ I(0). In order for the variables in Zt to be cointegrated, we

must have Ut ~ I(0). That is, we must show the last term in the error correction
equation: Zt-1Π ~ I(0). By definition of cointegration, the parameter matrix Π must
contain r (0<r<M) linearly independent cointegrating vectors such that ZtΠ ~ I(0).
Therefore, the cointegration test for Zt amounts to checking the rank of matrix Π,
denoted Rank(Π). If Rank(Π) = r > 0, we may impose parameter restrictions Π = -
BA' where A and B are Mxr matrices. Given the existence of the constant vector Γ0,
there can be up to M-r random walks or drift trends. Such common trends in the
variables may be removed in the case of Model II below. We consider the following
three models:
• Model I: VAR(p) representation without constant vector, i.e., Γ0 = 0.

• Model II: VAR(p) representation with constant vector but the trend removed
(drift only, i.e., Γ0 = -µ0 Π).
• Model III: VAR(p) representation with constant vector (drift trend, i.e.,
Γ0 = -µ0Π + µ1Γ).
For model estimation of the above VAR(p) system, where Ut is independently

normally distributed with zero mean and constant covariance matrix Σ, we derive the
log-likelihood function for Model III:
ll(Γ1,Γ2,..., Γp-1,Γ0,Π,Σ) = - MN/2 ln(2π) - N/2 ln|det(Σ)| - ½ ∑t=1,2,...,NUtΣ-1Ut'
Since the maximum likelihood estimate of Σ is U'U/N, the concentrated log-

likelihood function is written as:
ll*(Γ1,Γ2,..., Γp-1,Γ0,Π) = - NM/2 (1+ln(2π)-ln(N)) - N/2 ln|det(U'U)|
The actual maximum likelihood estimation can be simplified by considering the

following two auxiliary regressions:
1. ∆Zt = ∆Zt-1Φ1 + ∆Zt-2Φ2 + ... + ∆Zt-(p-1)Φp-1 + Φ0 + Wt

2. Zt-1 = ∆Zt-1Ψ1 + ∆Zt-2Ψ2 + ... + ∆Zt-(p-1)Ψp-1 + Ψ0 + Vt
We see that Γj = Φj-ΨjΠ, for j=0,1,2,...,p-1, and Ut = Wt - VtΠ. If Φ0 = Ψ0 = 0, then

Γ0 = 0, implying no drift in the VAR(p) model. However, Γ0 = 0 needs only the
restriction that Φ0 = Ψ0Π.
Plugging in the auxiliary regressions, we can now write the concentrated log-
likelihood function as
ll*(W(Φ1,Φ2,...,Φp-1,Φ0), V(Ψ1,Ψ2,...,Ψp-1,Ψ0),Π)
= - NM/2 (1+ln(2π)-ln(N)) - N/2 ln|det((W-VΠ)'(W-VΠ))|
Maximizing the above concentrated log-likelihood function is equivalent to

minimizing the sum-of-squares term det((W-VΠ)'(W-VΠ)). Conditional on
W(Φ1,Φ2,...,Φp-1,Φ0) and V(Ψ1,Ψ2,...,Ψp-1,Ψ0), the least squares estimate of Π is
(V'V)-1V'W. Thus,
det((W-VΠ)'(W-VΠ)) = det(W(I-V(V'V)-1V')W')
= det((W'W)(I-(W'W)-1(W'V)(V'V)-1(V'W))
228
= det(W'W) det(I-(W'W)-1(W'V)(V'V)-1(V'W))
= det(W'W) (∏i=1,2,...,M(1-λi))
where λ1, λ2, ..., λM are the ascending ordered eigenvalues of the matrix
(W'W)-1(W'V)(V'V)-1(V'W). Therefore the resulting double concentrated log-
likelihood function (concentrating on both Σ = U'U/N and Π = (V'V)-1V'W) is
ll**(W(Φ1,Φ2,...,Φp-1,Φ0), V(Ψ1,Ψ2,...,Ψp-1,Ψ0))
= - NM/2 (1+ln(2π)-ln(N)) - N/2 ln|det(W'W)| - N/2 ∑i=1,2,...,M ln(1-λi)
Given the parameter constraints that there are 0<r<M cointegrating vectors, that is Π
= -BA' where A and B are Mxr matrices, the restricted concentrated log-likelihood
function is similarly derived as follows:
llr**(W(Φ1,Φ2,...,Φp-1,Φ0), V(Ψ1,Ψ2,...,Ψp-1,Ψ0))
= - NM/2 (1+ln(2π)-ln(N)) - N/2 ln|det(W'W)| - N/2 ∑i=1,2,...,rln(1-λi)
Therefore, with the degree of freedom M-r, the Likelihood Ratio test statistic for at
least r cointegrating vectors is
-2(llr** - ll**) = -N ∑i=r+1,r+2,...,Mln(1-λi)
Similarly the Likelihood Ratio test statistic for r cointegrating vectors against r+1
vectors is
-2(llr** - llr+1**) = -N ln(1-λr+1)
A more general form of the Likelihood Ratio test statistic for r1 cointegrating vectors
against r2 vectors (0 ≤ r1 ≤ r2 ≤ M) is
-2(llr1** - llr2**) = -N ∑i=r1+1,r1+2,...,r2ln(1-λi)
The following table summarizes the two popular cointegration test statistics: the
maximal eigenvalue test statistic λmax(r) and the trace test statistic λtrace(r). By
definition, λtrace(r) = ∑r1=r,r+1,…,Mλmax(r1). For the case of r = 0, they are the tests for no
cointegration. If M=r+1, the two tests are identical.
Cointegrating H0: r1 = r H0: r1 = r

Rank (r) H1: r2 = r+1 H1: r2 = M
0 -N ln(1-λ1) -N ∑i=1,2,...,Mln(1-λi)
1 -N ln(1-λ2) -N ∑i=2,3,...,Mln(1-λi)
... ... ...
M-1 -N ln(1-λM) -N ln(1-λM)
Test Statistics λmax(r) λtrace(r)
The critical values of λmax(r) and λtrace(r) for testing the specific number of
cointegrating vectors or rank r are given in Statistical Table C-5. Three models (no
constant, drift only, and trend drift) are presented.
The procedure of Johansen’s cointegration Likelihood Ratio tests is implemented as

the GPE module program JOHANSEN.GPE. The module is located in the GPE
229
subdirectory (see also Appendix B-2). Interested readers can study the details of
implementation of the Likelihood Ratio test we outline above. To perform the
Johansen’s cointegration test, in lesson14.3, we include the module at the end of the
program (line 17). The test is done by calling the procedure johansen with three
input arguments: z = data matrix, p = lags of VAR structure, and c is the model or
constant (0=no, 1=drift only, 2=trend drift):
call johansen(z,p,c);
The lengthy implementation closely follows the theoretical discussion above. We

will concentrate on the application of cointegration test statistics, and leave the
programming details of the procedure to you.
Lesson 14.3: Cointegration Test: Johansen Approach

Returning to the data series of income (Y) and consumption (C) we have studied so
far, the Johansen cointegration test considers a VAR representation of the data
matrix, and estimates the model as a simultaneous linear equations system. In
lesson14.3, line 7 defines the data matrix consisting of two variables: y (income) and
c (consumption). A VAR model with 3 lags (i.e., _dlags=3 of line 11) has been
shown to best describe the data series under consideration (that is, to “whiten” the
residuals). This can be verified by checking the estimation result of the VAR system
as specified from line 9 through line 12. Although the estimation is carried out with
the system method of 3SLS (line 10), it should be the same as any limited
information estimation method (why?).
/*
** Lesson 14.3: Cointegration Test
** Johansen Approach
*/
1 use gpe2;

4 y = z[2:67,2];
5 c = z[2:67,3];
6 ns = {"c","y"};
7 data = y~c; @ data matrix for cointegration test @
8 call reset;
9 _rstat=1;
10 _method=3;
11 _dlags=3; @ find the proper order p @
12 call estimate(data,0); @ for VAR(p) model estimation @
@ Johansen cointegration test based on VAR(3) model @

13 call johansen(data,3,2); @ model with trend drift @
14 call johansen(data,3,1); @ model with drift only @
15 call johansen(data,3,0); @ model with no drift @
16 end;
17 #include gpe\johansen.gpe;
We present only the summary of the estimation results, and leave out the details of
each equation. We note that the model estimation is used to determine the lag
structure of the VAR system. In this example, VAR(3) has been shown to be the
appropriate model.
230
----------------------------------------

Lagged Endogenous Variables Used = 6


Y1 Y1-1 0.88597 0.16655 5.3196
Y1-2 0.29313 0.24559 1.1935
Y1-3 -0.36656 0.18016 -2.0346
Y2-1 0.59551 0.23153 2.5720
Y2-2 -0.74215 0.35751 -2.0759
Y2-3 0.36302 0.23135 1.5691
CONSTANT 13.691 9.4196 1.4535
Y2 Y1-1 -0.12601 0.12090 -1.0423
Y1-2 0.36104 0.17828 2.0251
Y1-3 -0.11330 0.13079 -0.86627
Y2-1 1.5421 0.16808 9.1748
Y2-2 -0.93942 0.25952 -3.6198
Y2-3 0.28254 0.16795 1.6823
CONSTANT 7.4753 6.8380 1.0932

Y1 1367.7
Y2 683.05 720.74
Y1 Y2
Cointegration Test (Model 3):

Cointegrating Eigv. Test Trace Test
Rank DF Statistic Statistic
0 2 18.958 25.103
1 1 6.1456 6.1456

0 2 23.890 35.150
1 1 11.260 11.260

0 2 23.547 33.056
1 1 9.5092 9.5092
Most importantly, cointegration tests based on eigenvalue and trace statistics are
given for each of the three models: trend drift (Model 3), drift only (Model 2), and
no constant (Model 1), in that order. These computed test statistics are compared
with the critical values of Statistical Table C-5 in Appendix C. Consider the case of
no cointegration (that is, cointegrating rank equals 0 with 2 degrees of freedom):
both λmax(0) and λtrace(0) statistics are statistically greater than the corresponding
critical values at a 5% level significance. We reject the null hypothesis of no
cointegration for the data series under consideration. Therefore, the time series
231
income (Y) and consumption (C) are cointegrated, confirming the previous Engle-
Granger or ADF test result based on the cointegrating regression residuals.
232
XV
Time Series Analysis
Continuing from the previous chapter in which we discussed a stationary vs.
nonstationary data generating process, in this chapter, we focus on the modeling of
stationary time series data. If the data series under consideration is nonstationary, we
assume that it is an integrated process and can be made stationary with the proper
amount of differencing. A random data generating process which is difference
stationary is the subject of modern time series analysis.
Autocorrelation occurs when previous and current observations of a random variable

are correlated. Chapter X discussed autocorrelation in detail. Serial correlation in the
mean is common in many economic time series, with the simplest case being first-
order correlation. More complicated model structures can include autoregressive and
moving average terms, known as ARMA processes. However, serial correlation in
the mean is not the only problem of autocorrelation in time series. Conditional to the
information available, the variance of a data generating process may not be constant
for all observations. Nonconstant variance, or heteroscedasticity, was studied in
Chapter IX. Even worse is serial correlation in the conditional variance. The
phenomenon of conditional variance correlation is often found in high-frequency
observations such as those studied in financial economics. Autoregressive
conditional heteroscedasticity, or the ARCH process, is another important time series
model structure to be studied in this chapter.
Typically, time series analysis is carried out in several steps: model identification,
estimation, diagnostic checking, and prediction. In this chapter we emphasize model
identification and estimation. Diagnostic checking is the repetition of the
identification step on the estimated model. Prediction is taken up later in Chapter
XVII. In many circumstances, economic theory offers no a priori data generating
process for a given variable, so model identification is often a trial and error process.
To extract structural information from a random variable, the process of model
identification consists of testing and estimation for the mean and variance of the
variable under consideration. In Chapter X, we used several procedures to test for
autocorrelation in an ARMA model. These tests include the Durbin-Watson bounds
test for first-order serial correlation, the Breusch-Godfrey LM test for higher-order
autocorrelation, and Box-Pierce and Ljung-Box Q test statistics based on different
lags of autocorrelation coefficients. In addition, the autocorrelation function (ACF)
and partial autocorrelation function (PACF) gave us useful clues as to the model’s
structure. Many examples shown in Chapter X demonstrated the use of a
combination of the above-mentioned testing procedures and statistics to study the
time series.
For ARCH modeling, the idea of variance correlation is new but the mechanics are
similar to ARMA analysis. Working on the squares of mean-deviation (or
regression) residuals, the corresponding ACF and PACF can assist in detecting the
autocorrelation in the variance. The associated Box-Pierce and Ljung-Box statistics
are useful to test the potential ARCH process. Analogous to the Breusch-Godfrey
LM test for autocorrelation in the mean, the Engle-Bollerslev LM test statistic is

used for testing autocorrelation in the variance.
In the following, we present the basic formulation of ARMA and ARCH models.
GPE implementation of model identification and estimation for regression models
with ARMA and ARCH effects are illustrated by examples.
Autoregressive and Moving Average Models

Consider a stationary data generating process ARMA(p,q) for a random variable Y:
Yt = δ + ρ1Yt-1 + ρ2Yt-2 + ... + ρpYt-p - θ1εt-1 - θ2εt-2 - ... - θqεt-q + εt
where εt is independently distributed with zero mean and constant variance σ2, or εt ~
ii(0,σ2), t = 1,2,...,N. As described in Chapter X, ARMA(p,q) is a mixed process of
AR(p) and MA(q), where p and q represent the highest order of autoregressive and
moving average parameters in the model, respectively. The model may also be
written as a general linear stochastic process:
Yt = µ + εt + ψ1εt-1 + ψ2εt-2 + ...
Recall that stationarity requirements for the process imply that the mean, variance,
and autocovariances of the variable must be finite constants:
Mean µ = E(Yt) < ∞

Variance γ0 = σ2∑i=0,...,∞ψi2 < ∞ ψ0 = 1
Autocovariance γj = σ2∑i=0,...,∞ψiψj+i < ∞
The coefficient of autocorrelation defined by φj = γj / γ0 serves as the foundation for

model identification. We have seen examples in chapters X and XI of using
autocorrelation and partial autocorrelation coefficients to model the ARMA error
structure. In particular, the Box-Pierce and Ljung-Box Q test statistics derived from
the autocorrelation coefficients are useful in identifying the autoregressive and
moving average time series. For details of model identification, we refer readers to
standard econometrics textbooks on time series analysis.
For parameter estimation, the ARMA(p,q) model may be written in the “inverted”
form as follows:
ρ(B)Yt = δ + θ(B)εt
or,
θ(B)-1[-δ+ρ(B)Yt] = εt
where B is the backshift operator, ρ(B) = 1 - ρ1B - ρ2B2 - ... - ρpBp, and θ(B) = 1 -
θ1B - θ2B2 - ... - θqBq. Conditional to the historical information (YN, ..., Y1), and data
initialization (Y0, ..., Y-p+1) and (ε0, ..., ε-q+1), the error sum-of-squares is defined by
S = ∑t=1,2,...,Nεt2
234
TIME SERIES ANALYSIS
In order to utilize all N data observations, data initialization may be needed for the
observations Y0, Y-1, ..., Y-p+1 with E(Yt) = δ / (1-ρ1-...-ρp), and ε0, ε-1, ..., ε-q+1 with
E(εt) = 0.14 In GPE, the data initialization used for the pre-sample observations is
simply the sample mean of the series.
Techniques of nonlinear optimization may be applied directly to minimize the sum-

of-squares objective function. The alternative is the maximum likelihood method, for
which we need to make additional assumptions about the probability distribution of
the error term. For each independent observation t, we assume the model error εt is
normally distributed with zero mean and constant variance σ2, that is εt ~ nii(0,σ2).
Then the concentrated log-likelihood objective function is
ll = -N/2 [1+ln(2π)-ln(N)+ln(∑t=1,2,...,Nεt2)]
Using nonlinear optimization methods, maximizing the above function with respect
to the parameters ρs, θs, and δ is straightforward (see Chapter VII for more details
on maximum likelihood estimation of a nonlinear regression model). The GPE
package implements the nonlinear maximum likelihood estimation for the ARMA
error structure in a linear regression framework.
Note that the model specification posited above is only tentative, pending diagnostic
checking on the estimated residuals. We do not know whether or not we have
included sufficiently high orders of AR and MA terms in the model specification. In
other words, we do not know whether our choice of orders, p for AR and q for MA,
were adequate. The “correct” p and q are usually determined through an iterative
process. We choose an initial number of AR and MA terms (usually at low values,
zero or one) and estimate the model. We then use the diagnostic tests on the
estimated residuals (e.g., Durbin-Watson, Breusch-Godfrey, Box-Pierce, and Ljung-
Box) to determine if serial correlation is still present in the model. If we still have
problems with serial correlation, we add AR or MA terms (i.e., increase the values of
p and q), re-estimate the model and rerun the diagnostic tests on the “new” residuals.
This process continues until the error term has been sufficiently “whitened.” In so
doing, we find the combination of AR and MA terms that removes the serial
correlation from the model. Note that when performing the diagnostic checking on
the estimated residuals, the degrees of freedom used to choose the critical value for
each test statistic is N-(K+p+q), where K is the number of regression parameters.
Lesson 15.1: ARMA Analysis of Bond Yields

This example demonstrates univariate time series analysis. bonds.txt is a data file
consisting of 5 years of monthly average yields on a Moody's Aaa rated corporate
bond (see also Greene, 1999, Example 18.1). The original level series is
nonstationary, but it can be shown to be an integrated process of the first order (or
I(1)) with no augmented lags (we leave this as an exercise, see Chapter XIV). Since
the first difference of an I(1) process is stationary, deriving from the unit roots test
equation, we will estimate the following second order autoregressive model:
Yt = ρ0 + ρ1 Yt-1 + ρ2 Yt-2 + ut
14
The alternative to data initialization is to estimate the unknown pre-sample observations of
ε0, ε-1, ..., ε-q+1 together with the model parameters. The problem becomes highly nonlinear
and complicated.
235
where ut ~ ii(0,σ2) or nii(0,σ2). We may examine the data series Yt by plotting the
correlogram of its autocorrelation and partial autocorrelation coefficients. For
univariate analysis, the ACF and PACF of the time series will be identical to those of
the residuals obtained from the mean-deviation regression. Up to the maximum
number of lags specified for the ACF and PACF, Box-Pierce and Ljung-Box test
statistics are useful for identifying the proper order of AR(p), MA(q), or ARMA(p,q)
process. In addition, the Breusch-Godfrey LM test may be used to verify higher
orders of autocorrelation, if any exist.
Although the entire diagnostic checking procedure is not shown here, by examining
the autocorrelation and partial autocorrelation coefficients as well as the relevant
diagnostic test statistics, we can show that an AR(2) specification is sufficient for the
bond yield model. The result is consistent with that of the stationarity test. Using a
time series of bond yields, a regression model with two (first and second) lagged
dependent variables is estimated in the following program lesson15.1.
/*
** Lesson 15.1: ARMA Analysis of Bond Yields
*/
1 use gpe2;
3 n=61; @ 1990.01 - 1994.12 @

4 load bonds[n,2]=gpe\bonds.txt;
5 y=bonds[2:n,2];
6 call reset;
7 _names={"yields"};
8 _rstat=1;
9 _rplot=2;
10 _dlags=2;
/*
_ar=2;
_iter=50;
*/
11 _bgtest=4;
12 _acf=12;
13 call estimate(y,0);
14 end;
The estimated model is summarized as follows (standard errors are in parentheses):
Yt = 0.4068 + 1.1566 Yt-1 – 0.2083 Yt-2

s.e. (0.2107) (0.1107) (0.1102)
Further study of the regression residuals using the ACF and PACF up to 12 lags does
not reveal higher-order autocorrelated structure in this model. The other tests
(Durbin-Watson test for the first lag and Breusch-Godfrey LM test up to the fourth
lags) suggest that a structure beyond AR(2) may be presented. But if such structure
does exist in this model, it is not reflected in the ACF and PACF. Conflicting results
from the use of different diagnostic tests are not unusual in empirical analysis. Run
this program, and see the output file output15.1 for details.
An alternative to including two lagged dependent variables in the model is to express

residuals of the mean-deviation model as an AR(2) process:
236
Yt = µ + εt
εt = φ1 εt-1 + φ2 εt-2 + ut
Or, equivalently
Yt = µ + φ1 εt-1 + φ2 εt-2 + ut
We note that µ = ρ0 /(1- ρ1 - ρ2) from the earlier specification with lagged dependent
variables. We now modify Lesson 15.1 by replacing line 10 of the program
(_dlags=2) with the following two statements:
_ar=2;
_iter=50;
Estimation of autocorrelated error structures is inherently nonlinear, so we will need

to ensure that the number of iterations is sufficient to converge on a solution. We
note that the control variable _ar is used to specify the autocorrelation order of the
model error. For a more general autoregressive and moving average model, the GPE
control variable _arma should be used instead. _arma is a column vector with the
first element being the autoregressive order, and the second being the moving
average order of the model structure. In this example, _ar=2 is equivalent to
_arma={2,0}. For a pure moving average model, you would set the first element
of _arma to zero. For example, _arma={0,1} defines the first-order moving
average process. See Appendix A for more information about the use of the GPE
control variable _arma.
Running the revised lesson15.1, we obtain the following result (standard errors are in
parentheses):
Yt = 7.877 + 1.1566 εt-1 - 0.2083 εt-2

s.e. (0.3882) (0.08597) (0.08654)
Comparing the estimated autoregressive parameters with those of Lesson 15.1, we

find the two sets of coefficients are very similar. In addition, diagnostic checking
indicates that AR(2) is a sufficient specification for this model. We note that the
divisor N (N=58 in this example) is used in calculating the standard errors for a
nonlinear model.15 Therefore, the resulting standard errors of the parameters are
smaller in this model, as compared with the lagged dependent variables specification
of Lesson 15.1.
The classical univariate ARMA analysis is easily extended to a more general

regression model with multiple regressors. There are two approaches, as shown
below.
ARMA Analysis for Regression Residuals

The full model consists of the following two equations:
15
With the model of Lesson 15.1, the divisor used is N-K where K is 3.
237
Yt = Xtβ + εt
φ(B)εt = θ(B)ut
Or, equivalently
Yt = Xtβ + φ1εt-1 + φ2εt-2 + ... + φpεt-p - θ1ut-1 - θ2ut-2 - ... - θqut-q + ut
where ut ~ ii(0,σ2) or nii(0,σ2), t = 1,2,...,N. The model identification and estimation

proceeds in the same way as univariate ARMA analysis. Regression parameters (βs)
and ARMA parameters (φs and θs) must be simultaneously estimated through
iterations of nonlinear functional (sum-of-squares or log-likelihood) optimization.
For statistical inference, the degrees of freedom must account for all the unknown
parameters in the model.
ARMAX Regression Model: Transfer Function Approach

The model to be estimated is
ρ(B)Yt = Xtβ + εt
εt = θ(B)ut
Or, equivalently
Yt = Xtβ + ρ1Yt-1 + ρ2Yt-2 + ... + ρpYt-p - θ1ut-1 - θ2ut-2 - ... - θqut-q + ut
where ut ~ ii(0,σ2) or nii(0,σ2), t = 1,2,...,N. We identify the proper autoregressive

and moving average orders and estimate the model in a similar fashion to the
classical univariate ARMA analysis. The difference between ARMA and ARMAX
model specification lies in the treatment of autoregressive components of the error
structure. For the former, the model error is specified solely by an ARMA
representation. On the other hand, in an ARMAX model, lagged dependent variables
are used in conjunction with only the moving averages of errors. In GPE, however, it
is possible to include the autoregressive error terms in an ARMAX model for
estimating non-stationary time series with ARMA error structure.
As mentioned earlier, in GPE, ARMA analysis is called with the input control
variable _arma. _arma is a column vector containing at least two elements
specifying the type of ARMA model to be estimated. The first element of _arma
denotes autoregressive order of the ARMA process, while the second element
denotes the moving average order. Specifying only the autoregressive portion and
including a zero for the moving average portion yields a pure AR specification (vice
versa for a pure MA specification). Optional initial values of the autoregressive and
moving average coefficients may be appended to the vector _arma along with their
respective orders. Supplying the initial values is useful for starting the iterative
estimation process from non-zero values of ARMA coefficients. For example,
_arma = {1,1,0.5,0.1};
The model ARMA(1,1) is specified, with the initial values 0.5 for the AR(1)
coefficient and 0.1 for the MA(1) coefficient. Nonlinear estimation of this model will
begin from the specified set of starting values.
238
Lesson 15.2: ARMA Analysis of U.S. Inflation

In this example we will identify and estimate a time series regression model of U.S.
inflation. The inflation rate in this example is measured as the quarterly rate of
percent change in price:
∆Pt = 100 [ln(Pt) - ln(Pt-1)]
Inflation is believed to be affected by excess monetary growth (i.e., monetary growth

that is faster than the growth of real output) and by external economic shocks. Excess
monetary growth is defined as ∆Mt - ∆Yt, where
∆Mt = 100 [ln(M1t) - ln(M1t-1)]

∆Yt = 100 [ln(GNPt) - ln(GNPt-1)]
The basic regression model of inflation is presented as follows:
∆Pt = β0 + β1(∆Mt-1-∆Yt-1) + εt
The lagged values of the inflation rate (or the disturbance term) will serve to model
the effects of external shocks to the economy. The data file usinf.txt consists of 136
quarterly observations (from 1950 Q1 to 1984 Q4) of data for price (implicit GNP
deflator) Pt, money stock M1t, and output (GNP) Yt.
To keep the model simple, we include the first lag of the dependent variable in the
regression and examine the patterns of ACF and PACF. Significant spikes (or non-
zero values of autocorrelation and partial autocorrelation coefficients) appear up to
the third or fourth lags for both functions, indicating a complicated structure in the
model’s error term. We will not go through the entire identification process here.
Interested readers can “comment out” lines 15 to 18 in the program lesson15.2
below, and decide the proper ARMA or ARMAX specification for themselves based
on their observations of the behavior of the ACF and PACF.
/*
** Lesson 15.2: ARMA Analysis of U.S. Inflation
*/
1 use gpe2;
3 n=137;
4 load data[n,4]=gpe\usinf.txt;
5 y=ln(data[2:n,2]);
6 m=ln(data[2:n,3]);
7 p=ln(data[2:n,4]);
8 dp=packr(100*(p-lagn(p,1)));
9 dm=packr(100*(m-lagn(m,1)));
10 dy=packr(100*(y-lagn(y,1)));
11 call reset;
12 _rstat=1;
13 _rplot=2;
14 _acf=12;
15 _dlags=1;
16 _arma={0,3};
17 _method=5;
239
18 _iter=100;
19 call estimate(dp,lagn(dm-dy,1));
20 end;
The final model for estimation is a lagged dependent variable model (line 15) with a
third-order moving average specification (line 16). Maximum likelihood estimation
of the model is carried out using the QHC optimization method (line 17). The output
of running lesson15.2 is stored in the file output15.2.
In summary, the estimated model (with standard errors in parentheses) is
∆Pt = 0.1008 + 0.0146 (∆Mt-∆Yt) + 0.9151∆Pt-1

s.e. (0.0516) (0.0199) (0.0429)
- 0.505εt-1 + 0.0232εt-2 + 0.204εt-3
(0.093) (0.0986) (0.0884)
The lag of the dependent variable (∆Pt-1) plays an important role in the regression
equation. Although the second lag of the moving average is insignificant, the first
and third are significant. Further analysis of the ACF and PACF does not show
autocorrelation in the regression residuals.
Autoregressive Conditional Heteroscedasticity

We have thus far concentrated on the classical time series modeling, which focuses
on the expected value (mean) of the variable. In many financial and monetary
economic applications, serial correlation over time is characterized not only in the
mean but also in the conditional variance. The latter is the so-called autoregressive
conditional heteroscedasticity or ARCH model. It is possible that the variance is
unconditionally homogeneous in spite of the presence of conditional
heteroscedasticity. Using GPE, analysis of ARCH effects is no more complicated
than setting a few input control variables.
Consider the time series linear regression model:
Yt = Xtβ + εt
At time t, conditional to the available historical information Ht, we assume that the
error structure follows a normal distribution: εt|Ht ~ nii(0,σ2t) where the variance is
written as:
σ2t = α0 + δ1σ2t-1 + ... + δpσ2t-p + α1ε2t-1 + ... + αqε2t-q

= α0 + Σj=1,2,...pδjσ2t-j + Σi=1,2,...qαiε2t-i
Let υt = ε2t-σ2t, αi = 0 for i > q, δj = 0 for j > p, and m = max(p,q). Then, the above
serially correlated variance process may be conveniently rewritten as an
ARMA(m,p) model for ε2t. That is,
ε2t = α0 + Σi=1,2,...m(αi+δi)ε2t-i - Σj=1,2,...pδjυt-j + υt
By assuming E(υt) =0, E(ε2t) is the estimated variance of σ2t. This is the general
specification of autoregressive conditional heteroscedasticity, or GARCH(p,q),
240
according to Bollerslev (1986). If p = 0, this GARCH(0,q) process simply reduces to

an ARCH(q) process:
σ2t = α0 + Σi=1,2,...qαiε2t-i
ARCH(1) Process
The simplest case, pioneered by Engle (1982) sets q = 1 (while p = 0). This
ARCH(1) process can be written as:
σ2t = α0 + α1ε2t-1
The ARCH(1) model can be summarized as follows:
Yt = Xtβ + εt
εt = ut(α0 + α1ε2t-1)½ where ut ~ nii(0,1)
This specification gives us the conditional mean and variance, E(εt|εt-1) = 0 and σ2t =
E(ε2t|εt-1) = α0 + α1ε2t-1, respectively. Note that the unconditional variance of εt is
E(ε2t) = E[E(ε2t|εt-1)] = α0 + α1E(ε2t-1). If σ2 = E(ε2t) = E(ε2t-1), then σ2 = α0/(1-α1),
provided that |α1| < 1. In other words, the model may be free of general
heteroscedasticity even when we assume that conditional heteroscedasticity is
present.
ARCH-M(1) Model
An extension of the ARCH(1) model is ARCH(1) in mean, or ARCH-M(1) model,
which adds the heterogeneous variance term directly into the regression equation
(assuming a linear model):
Yt = Xtβ + γσ2t + εt
σ2t = α0 + α1 ε2t-1
The last variance term of the regression may be expressed in log form ln(σ2t) or in
standard error σt. For example, Yt = Xtβ + γln(σ2t) + εt. Moreover, to ensure the
model stability and positive values of variances, we will need to constrain σ2t by
forcing α0 > 0 and 0 ≤ α1 < 1.
Hypothesis Testing for ARCH and GARCH Processes

As with ARMA modeling, specifying the correct order of the model is an important
step in estimating the ARCH and GARCH processes. Luckily, the close connection
between GARCH and ARMA allows us to compute the ARMA autocorrelation and
partial autocorrelation coefficients based on the squares of standardized regression
residuals (εt/σt)2 . The GPE control input variable _acf2 calculates the ACF, PACF,
and the associated diagnostic test statistics on (εt/σt)2 up to the number of lags we
specified. For example, to examine 12 lags of ACF and PACF of (εt/σt)2, use the
statement:
_acf2 = 12;
241
The other testing procedure involves checking ARCH(q) against ARCH(0), or

GARCH (p,q) against GARCH(p,0) for a given p. The Engle-Bollerslev LM Test of
ARCH effects (Bollerslev, 1986) is carried out using the test statistic NR2, where N
is the sample size and R2 (R-square statistic) is obtained from the ARCH(q)
regression. The Engle-Bollerslev test statistic is distributed as a Chi-square with q
degrees of freedom. The Engle-Bollerslev LM test of ARCH effects resembles the
Breusch-Godfrey LM test for AR effects. This is a cumulative test for no ARCH
effects up to the order specified. Therefore, we need only to test for ARCH effects at
a low order to confirm their existence. In GPE, the Engle-Bollerslev LM test
procedure is called using the control variable _ebtest in a similar fashion to
_acf2. For example, to test for ARCH effects up to the 6th order, use the statement:
_ebtest = 6;
For more information about the use of _acf2 and _ebtest, see Appendix A.
ARCH Model Estimation

Recall the normal log-likelihood function of a heteroscedastic regression model
ll = -½ N ln(2π) – ½ Σt=1,2,...,Nln(σ2t) – ½ Σt=1,2,...,N(ε2t/σ2t)
with the general conditional heteroscedastic variance GARCH(p,q) process:
σ2t = α0 + δ1σ2t-1 + δ2σ2t-2 + ... + δpσ2t-p + α1ε2t-1 + α2ε2t-2 + ... + αqε2t-q
The parameter vector (α,δ) is estimated together with the regression parameters (e.g.,
ε = Y - Xβ) by maximizing the log-likelihood function, conditional to the data
initialization ε20, ε2-1, ..., ε2-q, σ20, σ2-1, ..., σ2-p. In GPE, the data initialization used for
the pre-sample observations is simply the sample variance of the error series E(ε2t) =
Σt=1,2,...,N ε2t/N.
In estimating a GARCH model, the estimated variance for each observation must be
positive. We could assume the following parameter restrictions:
α0 > 0; αi≥0, i=1,2,...q; δj≥0, j=1,2,...,p
However, this set of restrictions is sufficient but not necessary (see Nelson and Cao,
1992) for the positive values of variances.
To estimate a model with ARCH or GARCH effects, we introduce the input control
variable _garch. _garch is a column vector with at least two elements. The first
element defines the autoregressive order of the GARCH model, while the second
element is the moving average order of the model. The rest of components in the
vector _garch, if given, specify the initial values of the GARCH parameters in
accordance with the orders given, as well as the initial value of the constant term.
The constant term in the variance equation is important because it may indicate a
homoscedastic structure if all the other slope parameters are insignificant. For
example,
_garch = {1,1};
242
specifies a GARCH(1,1) model. If we write instead,
_garch = {1,1,0.1,0.1,0.5};
then the initial values of the GARCH(1,1) are also given. The first two values of 0.1
are the initial values of autoregressive and moving average components of the
GARCH (1,1) process, respectively. The last element, 0.5, is the constant. The
nonlinear model estimation will begin from this set of starting values. Finally, we
remark that GPE implementation of the GARCH model estimation includes another
input variable _garchx to allow for external effects in the variance equation:
σ2t = α0 + Σj=1,2,...pδjσ2t-j + Σi=1,2,...qαiε2t-i + Xtγ
where Xt is a data matrix of external variables which may influence the variances. γ
is the corresponding parameter vector. Setting the data matrix to the variable
_garchx will do the trick. For example,
_garchx = x;
where x is the data matrix of the external variables already in place, which must have
the same number of observations as the residual variances. For more information
about the use of _garch and _garchx, see Appendix A.
Lesson 15.3 ARCH Model of U.S. Inflation

In this example, we focus on univariate ARCH analysis of U.S. inflation. We have
seen the ARMA regression analysis of U.S. inflation rate in Lesson 15.2. The data
are read from the file usinf.txt as in Lesson 15.2, but we use only the price variable.
Our study is based on the example given in Bollerslev (1986) and Greene (1999,
Example 18.2).
We will test, identify, and estimate the appropriate GARCH variance structure for
the variable ∆Pt, defined as the percentage change of implicit price GNP deflator. We
specify 12 lags of ACF and PACF for the squared mean-deviation residuals and
compute Engle-Bollerslev LM test statistics up to the sixth lag:
_acf2 = 12;
_ebtest = 6;
Just to be sure, the ARMA structure in the mean is also investigated by examining 12
lags of ACF and PACF for the mean-deviation residuals and 6 lags for Breusch-
Godfrey LM test statistics:
_acf = 12;
_bgtest = 6;
We consider a three-lag autoregressive model of ∆Pt in conjunction with a

GARCH(1,1) error process. If you would like to go through the model identification
process, simply comment out lines 14 through 17 in the following lesson15.3.
/*
** Lesson 15.3: ARCH Analysis of U.S. Inflation
*/
1 use gpe2;
243
3 n=137;
4 load data[n,4]=gpe\usinf.txt;
5 p=ln(data[2:n,4]);
6 dp=packr(100*(p-lagn(p,1)));
7 call reset;
8 _rstat=1;
9 _rplot=2;
10 _acf=12;
11 _bgtest=6;
12 _acf2=12;
13 _ebtest=6;
14 _dlags=3;
15 _garch={1,1,0.5,0.5,0.1};
16 _method=5;
17 _iter=100;
18 call estimate(dp,0);
19 end;
We note that the initial values of the GARCH(1,1) parameters are used in order to
successfully estimate the model (see line 15). Running lesson15.3, we obtain the
following result (see the generated output file output15.3 for details):16
∆Pt = 0.119 + 0.341 ∆Pt-1 + 0.214 ∆Pt-2 + 0.325 ∆Pt-3

s.e. (0.056) (0.088) (0.088) (0.087)
σ2t = 0.00573 + 0.882 σ2t-1 + 0.0799 ε2t-1

s.e. (0.0066) (0.056) (0.0496)
Based on the standard normal test, we see that σ2t-1 is statistically different from zero,
but the constant term and ε2t-1 are not. The model may be re-estimated with
GARCH(1,0) specification.
To be sure that the estimated GARCH(1,1) model does not have higher-order
structures in either the ARMA or GARCH specifications, the following extract of
output on diagnostic checking of the estimated model consists of: (1) ACF and
PACF for the estimated residuals and Breusch-Godfrey LM test for ARMA
specification; (2) ACF and PACF for the squared estimated standardized residuals
and Engle-Bollerslev LM test for ARCH specification. With an exception at the
twelfth lag of ACF and PACF for GARCH specification (possibly an outlier), the
estimated GARCH(1,1) model is acceptable for describing the U.S. inflation rate.

AR( 1) 0.66868 1 0.41351
AR( 2) 0.79698 2 0.67133
AR( 3) 0.82301 3 0.84396
AR( 4) 1.5805 4 0.81229
AR( 5) 1.7334 5 0.88467
AR( 6) 2.3775 6 0.88192
Engle-Bollerslev LM Test for Autoregressive Conditional Heteroscedasticity
16
Because we use a different set of U.S. inflation rate data, the estimated model does not
match with the Example 18.2 of Greene (1999).
244
Based on Squared Standardized Residuals

ARCH( 1) 1.3974 1 0.23716
ARCH( 2) 2.0844 2 0.35269
ARCH( 3) 2.4708 3 0.48060
ARCH( 4) 3.5246 4 0.47415
ARCH( 5) 3.5422 5 0.61702
ARCH( 6) 3.8524 6 0.69665

1 0.065389 0.087410 0.065389 0.087039 0.56439 0.57732
2 0.034597 0.087514 0.030451 0.087039 0.72238 0.74017
3 0.031536 0.087600 0.027494 0.087039 0.85366 0.87653
4 0.041261 0.087747 0.036716 0.087039 1.0784 1.1118
5 0.012524 0.087761 0.0058822 0.087039 1.0991 1.1336
6 0.0021592 0.087761 -0.0021918 0.087039 1.0997 1.1343
7 -0.12321 0.089062 -0.12686 0.087039 3.1034 3.2823
8 -0.080462 0.089611 -0.068807 0.087039 3.9580 4.2058
9 -0.019052 0.089641 -0.0039795 0.087039 4.0059 4.2580
10 -0.011847 0.089653 0.0011377 0.087039 4.0245 4.2784
11 -0.095777 0.090425 -0.082072 0.087039 5.2353 5.6193
12 -0.067064 0.090801 -0.049412 0.087039 5.8290 6.2823
GARCH Model Specification based on Squared Standardized Residuals

1 -0.10274 0.087953 -0.10274 0.087039 1.3933 1.4252
2 0.082417 0.088536 0.072628 0.087039 2.2899 2.3494
3 -0.069017 0.088943 -0.054563 0.087039 2.9186 3.0025
4 0.10580 0.089891 0.090016 0.087039 4.3961 4.5492
5 -0.015788 0.089912 0.010947 0.087039 4.4290 4.5839
6 -0.030121 0.089988 -0.048974 0.087039 4.5488 4.7113
7 -0.050698 0.090204 -0.047526 0.087039 4.8880 5.0750
8 0.095577 0.090968 0.085506 0.087039 6.0939 6.3781
9 0.00024805 0.090968 0.019248 0.087039 6.0939 6.3781
10 -0.056418 0.091233 -0.067360 0.087039 6.5140 6.8396
11 -0.12419 0.092505 -0.12321 0.087039 8.5499 9.0942
12 0.21915 0.096358 0.20097 0.087039 14.889 16.173
At this point, you may be wondering whether there exist ARCH effects for the
inflation rate model we considered earlier in Lesson 15.2. The mixture of ARMA
and ARCH effects may be identified and estimated for the model. We leave the
validation of ARCH effects in Lesson 15.2 to interested readers.
Lesson 15.4 ARCH Model of Deutschemark-British Pound Exchange Rate

This example investigates the “long-run volatility” persistence of the Deutschemark-
British pound exchange rate (see Bollerslev and Ghysels, 1986). Daily exchange rate
data from January 3, 1984 to December 31, 1991 (1974 observations) are used (see
data text file dmbp.txt). The model of interest is
Yt = 100 [ln(Pt) - ln(Pt-1)] = µ + εt
where Pt is the bilateral spot Deutschemark-British pound exchange rate. Thus Yt is

the daily nominal percentage returns from exchanging the two currencies. Similar to
the testing procedure carried out in Lesson 15.3, we will identify and estimate the
appropriate GARCH variance structure for the variable Yt. Because of the large data
245
sample, longer lags may be used for the tests with ACF and PACF.17 We leave out
the details of identification and report only the chosen model for estimation.
We find that the mean returns of Deutschemark-British pound exchange are

essentially zero, and there is no evidence of ARMA structure. However, a high order
ARCH or a mixed GARCH process is suggested. Therefore, in lesson15.4, the model
is estimated with GARCH(1,1) effects (see line 14). We keep the code section of
model identification (lines 10 through 13) for the purpose of performing diagnostic
tests on the estimated model.
/*
** Lesson 15.4: GARCH(1,1) Model of DM/BP Exchange Rate
** Bollerslev and Ghysels (1996), JBES, 307-327.
*/
1 use gpe2;
@ Deutschemark/British Pound Exchange Rate @

3 n=1974; @ 1-3-1984 to 12-31-1991 @
4 load data[n,2]=gpe\dmbp.txt;
5 x=data[.,1];
6 call reset;
7 _names={"xrate"};
8 _rstat=1;
9 _rplot=2;
@ model identification @
10 _acf2=12;
11 _ebtest=6;
12 _acf=12;
13 _bgtest=6;
@ model estimation @
14 _garch={1,1};
15 _method=6;
16 _iter=100;
17 call estimate(x,0);
18 end;
Using the modified QHC method (line 15), the result of maximum likelihood
estimation of the GARCH(1,1) model is given below:

------------------------
Dependent Variable = XRATE
Mean of Dependent Variable = -0.016427

GARCH( 1, 1) Autoregressive Conditional Heteroscedasticity Process

17
As the size of the data is beyond the limit of GAUSS Light, the professional version of
GAUSS should be used.
246
Tolerance = 0.001
Initial Result:
Parameters = -0.016427 0.00000 0.00000 0.10000
Using Modified Quadratic Hill-Climbing Algorithm

Iteration = 1 Step Size = 2.3579 Log Likelihood = -1305.1
Parameters = -0.013777 0.072340 0.015314 0.16692
…
Parameters = -0.0061905 0.80598 0.15313 0.010761
Final Result:
Parameters = -0.0061905 0.80598 0.15313 0.010761
Gradient Vector = 0.067109 -3.2813 -2.7642 -17.896

HAR( 1) 0.80598 0.073406 10.980
HMA( 1) 0.15313 0.054232 2.8236
CONSTANT 0.010761 0.0065879 1.6334
R-Square = -0.00047409 R-Square Adjusted = -0.00047409


Explained 3.7079E-029 0
Residual 436.50 1973 0.22123
Total 436.29 1973 0.22113

CONSTANT -0.0061905 0.0091932 -0.67338 0.50079 0.00022977

CONSTANT 8.4515E-005
CONSTANT

CONSTANT 1.0000
CONSTANT
Squared Correlation of Observed and Predicted = 5.3640E-006

The GARCH(1,1) model is summarized as follows:
σt2 = 0.01076 + 0.80598 σ2t-1 + 0.15313 εt-12

s.e. (0.0066) (0.0734) (0.0542)
247
With the exception of the constant term, all other parameters are significantly
different from zero based on the standard normal test.
If the underlying assumption of normal distribution for the model is questionable,

the estimated variance-covariance matrix may be adjusted. This is easily done by
setting the control variable _vcov=3 before calling the estimate statement (line
17). As the result of quasi-maximum likelihood estimation, the robust standard
errors for the parameters are computed. See Appendix A for more details.
Diagnostic checking on the estimated GARCH(1,1) model does not suggest a higher-
order ARMA or GARCH specification. All the statistical tests presented below
confirm that the estimated GARCH(1,1) model describes the volatility of the returns
of the Deutschemark-British pound exchange rate reasonably well.

AR( 1) 0.17342 1 0.67709
AR( 2) 1.4489 2 0.48458
AR( 3) 3.8242 3 0.28109
AR( 4) 4.5111 4 0.34124
AR( 5) 5.2147 5 0.39025
AR( 6) 5.2316 6 0.51447
Engle-Bollerslev LM Test for Autoregressive Conditional Heteroscedasticity

Based on Squared Standardized Residuals
ARCH( 1) 2.5119 1 0.11299
ARCH( 2) 2.6312 2 0.26832
ARCH( 3) 4.2403 3 0.23666
ARCH( 4) 4.2406 4 0.37442
ARCH( 5) 4.2422 5 0.51510
ARCH( 6) 6.7980 6 0.33993

Mean = -0.010236 Standard Error = 0.47024
1 0.0093663 0.022509 0.0093663 0.022507 0.17318 0.17344
2 -0.025323 0.022524 -0.025413 0.022507 1.4390 1.4418
3 0.034169 0.022550 0.034675 0.022507 3.7436 3.7523
4 0.019958 0.022559 0.018659 0.022507 4.5299 4.5409
5 0.017487 0.022566 0.018896 0.022507 5.1335 5.1468
6 -0.0023945 0.022566 -0.0029546 0.022507 5.1449 5.1581
7 -0.016242 0.022572 -0.016632 0.022507 5.6656 5.6812
8 0.016314 0.022578 0.014906 0.022507 6.1910 6.2093
9 0.016177 0.022584 0.014567 0.022507 6.7076 6.7288
10 0.011128 0.022587 0.012590 0.022507 6.9520 6.9747
11 -0.037358 0.022618 -0.037292 0.022507 9.7069 9.7478
12 -0.0013434 0.022618 -0.0011303 0.022507 9.7105 9.7514
GARCH Model Specification based on Squared Standardized Residuals

1 0.035668 0.022536 0.035668 0.022507 2.5113 2.5152
2 -0.0064933 0.022537 -0.0077754 0.022507 2.5946 2.5986
3 -0.029035 0.022556 -0.028562 0.022507 4.2587 4.2669
4 -0.0016120 0.022556 0.00040097 0.022507 4.2638 4.2720
5 -0.00056223 0.022556 -0.00090229 0.022507 4.2644 4.2726
6 -0.035159 0.022584 -0.036022 0.022507 6.7046 6.7228
7 -0.025093 0.022598 -0.022674 0.022507 7.9475 7.9714
8 -0.013861 0.022602 -0.012704 0.022507 8.3268 8.3526
9 -0.0019132 0.022602 -0.0033856 0.022507 8.3340 8.3598
10 0.018812 0.022610 0.017422 0.022507 9.0326 9.0627
248
11 -0.013855 0.022615 -0.016063 0.022507 9.4116 9.4441

12 -0.016590 0.022621 -0.016835 0.022507 9.9548 9.9913
249
XVI
Panel Data Analysis
We have seen two popular types of data used in econometric analysis: time-series
and cross-sectional data. However, in some circumstances, the economic data may
be a composition of time series and cross sections (i.e., the observations of several
individuals over time). International statistics, company surveys, and longitudinal
data sets are common examples. Modeling these panel data sets calls for some quite
complex stochastic specifications. In this chapter, we introduce the basic
programming techniques for panel data analysis.
For each cross section (individual) i=1,2,...N and each time period (time) t=1,2,...T,
we write the regression equation as follows:
Yit = Xitβit + εit
Suppose that the regressors Xit include a constant term. Let βit = β and assume εit = ui
+ vt + eit. Note that we assume the identical β for all i and t, and consider their
differences in the components of the error term εit. Here ui represents the individual
difference in intercept (so that the individual effect is β0+ui, where β0 is the intercept
parameter in β) and vt is the time difference in intercept (so that the time effect is
β0+vt). Two-way analysis includes both time and individual effects. Throughout
much of this chapter, however, we will assume vt = 0. That is, there is no time effect
and only the one-way individual effects will be analyzed.
We further assume that eit is a classical error term, with zero mean, homogeneous
variance, and there is neither serial correlation nor contemporaneous correlation.
That is, the error term is not correlated across individuals or time periods. Also, eit is
assumed to be uncorrelated with the regressors Xit. That is,
E(eit) = 0
E(eit2) = σ2e
E(eitejt) = 0, for i≠j
E(eiteiτ) = 0, for t≠τ
E(Xiteit) = 0
Fixed Effects Model

Assume that the error component ui, the individual difference, is fixed (or
nonstochastic), but varies across individuals. In this case, the model error simply
reduces to εit = eit. The model is expressed as:
Yit = (Xitβ + ui) + eit
where ui is interpreted to be the change in the intercept from individual to individual.

As defined earlier, the individual effect is ui plus the intercept, and this model is
known as the fixed effects model. To estimate a model with individual fixed effects,
consider the following equation:
Yit = (Xitβ + ui) + εit (i=1,2,...,N; t=1,2,...,T)
Let Yi = [Yi1, Yi2, ..., YiT]', Xi = [Xi1, Xi2, ..., XiT]', εi = [εi1, εi2, ..., εiT]', and υi = [ui,
ui, ..., ui]' (a column vector of T elements of ui). The pooled (stacked) model is
υ1 ε1
Y2  = X2  β + υ2  + ε2 , or
Y1  X1 
 …  …  …  …
YN  XN  υN  εN 
Y = Xβ + υ + ε
Dummy Variables Approach

For each i, define NT×1 vector Di with the element:
Dit = 1 if (i-1)×T+1 ≤ i×t ≤ i×T

0 otherwise
Then D = [D1, D2, ..., DN-1] is NT×(N-1) matrix of N-1 dummy variables.
Mathematically, D is the first N-1 columns of the matrix I⊗ι where I is NxN identity
matrix and ι is Tx1 column vector of ones.
Ordinary least squares can be used to estimate the model with dummy variables as
follows:
Y = Xβ + Dδ +ε
Since X includes a constant term, we will only need N-1 dummy variables for
estimation and the estimated δ measures the individual change from the intercept.
The individual effects are then computed as the sum of the intercept coefficient and
the estimated dummy variable parameter for each individual.
Deviation Approach
Although the dummy variable approach is simple, the size of the problem may
become difficult to handle if the number of cross sections (individuals) is large. An
alternative is the deviation approach.
Let Ymi = (Σt=1,2,...,TYit)/T, Xmi = (Σt=1,2,...,TXit)/T, and emi = (Σt=1,2,...,Teit)/T. By

estimating the following mean deviation model, we can obtain within-estimates of
the parameters:
(Yit - Ymi) = (Xit - Xmi)β + (eit - emi)
Or, equivalently
Yit = Xitβ + (Ymi - Xmiβ) + (eit - emi)
252
PANEL DATA ANALYSIS
Note that the constant term drops out due to the deviation transformation. As a result,
we can conclude the individual effects as ui = Ymi - Xmiβ. The variance-covariance
matrix of individual effects can be estimated as follows:
Var(ui) = v/T + Xmi [Var(β)] Xmi'
where v is the estimated variance of the mean deviation regression with NT-N-K
degrees of freedom. Note that K is the number of explanatory variables not counting
the constant term.
We can also estimate the model by using only the calculated individual means (as
opposed to the deviations from the mean):
Ymi = Xmiβ + ui + emi
The parameter estimates produced from this specification are referred to as the
between-estimates, and are related to the within-estimates of the parameters.
Hypothesis Testing for Fixed Effects

With the dummy variable model, we can test the null hypothesis that δ = 0 (i.e., that
there are no fixed effects) using the standard Wald F-test. The equivalent test statistic
for the deviation model is computed from the restricted (pooled model) and
unrestricted (mean deviation model) sum-of-squared residuals. That is, the statistic
RSS R − RSS U
N −1
RSS U
NT - N - K
follows an F distribution with N-1 and NT-N-K degrees of freedom.
Lesson 16.1: One-Way Panel Data Analysis: Dummy Variable Approach

As an example of one-way panel data analysis, we will duplicate a study of
efficiency in production of airline services presented in Greene (1999), Chapter 14.
The data file airline.txt consists of 6 firms for 15 years (1970 to 1984) with the
following variables:
I Cross section index: 6 airline firms

T Time index: 15 years from 1970 to 1984
C Cost (total cost of services)
Q Output (revenue passenger miles)
PF Fuel price
LF Load factor (rate of capacity utilization, measured as the average rate at
which seats on the airline’s planes are filled)
For panel data analysis, allowing for individual effects, the model for the total cost of
production is:
ln(Cit) = αi + β1 ln(Qit) + β2 ln(PFit) + β3 ln(LFit) + εit
253
We notice that the intercept αi is taken to be constant over time t and specific to the
individual firm i. The interpretation of slope parameters is straightforward in that
1
β1 > 0, β2 > 0, and β3 < 0. Moreover, the economies of scale defined as  -1 ,
 β1 
measures the efficiency of production.
The following program implements the fixed effects analysis using dummy
variables. For a typical regression, we need only to include five dummy variables for
the case of six firms. The estimated parameters associated with dummy variables
represent the change from the intercept (or the base case). If you are interested in the
fixed effects for each individual firm, you may use the full set of 6 dummy variables
in the regression equation without including the intercept. Since the use of dummy
variables in the regression was explained earlier in Chapter IV, this program is easy
to follow. In passing, note the use of a GAUSS built-in command dummybr to
generate the necessary dummy variables (see line 18).
/*
Lesson 16.1: One-Way Panel Data Analysis, Dummy Variable Approach
Cost of Production for Airline Services I
*/
1 use gpe2;
3 load data[91,6] = gpe\airline.txt;
4 panel=data[2:91,1:2]; @ panel definition @
5 n=6;
6 t=15;
@ stacked data series, by sections @

7 cs=ln(data[2:91,3]); @ log cost (stacked) @
8 qs=ln(data[2:91,4]); @ log output (stacked) @
9 pfs=ln(data[2:91,5]); @ log fuel price (stacked) @
10 lfs=data[2:91,6]; @ load factor (stacked) @
11 call reset;
12 _names = {"c","q","pf","lf","d1","d2","d3","d4","d5","d6"};
/* pooled estimates */
13 ys=cs;
14 xs=qs~pfs~lfs;
15 call estimate(ys,xs);
16 rssr=__rss;
17 dfr=__df;
@ use one less dummy variables with intercept @

18 d=dummybr(panel[.,1],seqa(1,1,n-1));
19 call estimate(ys,xs~d);
20 rssur=__rss;
21 dfur=__df;
22 f=((rssr-rssur)/(dfr-dfur))/(rssur/dfur);
23 print "Wald F Test Statistic";
24 print "for No Fixed Individual Effects = " f;
25 end;
The estimation results include the pooled regression and the dummy variable
regression. The Wald F-test statistic for fixed effects is computed from the estimated
sum-of-squares of the restricted (pooled) and unrestricted (dummy variables)
regressions (see line 22 in the program). Here is the output of running lesson16.1:
254
PANEL DATA ANALYSIS

------------------------


Explained 112.71 3 37.568 2419.3 6.5875E-083
Residual 1.3354 86 0.015528
Total 114.04 89 1.2814

Q 0.88274 0.013255 66.599 8.7911E-076 0.98098
PF 0.45398 0.020304 22.359 1.3601E-037 0.85322
LF -1.6275 0.34530 -4.7133 9.3090E-006 0.20529
CONSTANT 9.5169 0.22924 41.514 1.1294E-058 0.95247

------------------------


Explained 113.75 8 14.219 3935.8 1.5066E-101
Residual 0.29262 81 0.0036126
Total 114.04 89 1.2814

Q 0.91928 0.029890 30.756 1.9519E-046 0.92112
PF 0.41749 0.015199 27.468 8.3708E-043 0.90305
LF -1.0704 0.20169 -5.3071 9.5003E-007 0.25801
D1 -0.087062 0.084199 -1.0340 0.30421 0.013027
D2 -0.12830 0.075728 -1.6942 0.094071 0.034223
D3 -0.29598 0.050023 -5.9169 7.5281E-008 0.30179
D4 0.097494 0.033009 2.9535 0.0041106 0.097225
D5 -0.063007 0.023892 -2.6372 0.010020 0.079071
CONSTANT 9.7930 0.26366 37.142 1.2279E-052 0.94454
Wald F Test Statistic

for No Fixed Individual Effects = 57.732
Given the critical value of the distribution F(5, 81) at 5% level of significance, it is
clear that the cost structures among the six airline firms are somewhat different. In
other words, we reject the null hypothesis that there are no fixed effects. The fixed
255
effects are calculated by adding the parameters of the dummy variables to the
intercept.
Remember that an alternative is to include all six dummy variables and estimate the
model without an intercept. That is, replace line 18 with the following two
statements:
_const=0;
d=dummybr(panel[.,1],seqa(1,1,n));
The individual fixed effects are summarized in the following table (numbers in
parentheses are the estimated standard errors):
Firm Individual Effect

1 9.7059 (0.19312)
2 9.6647 (0.19898)
3 9.4970 (0.22496)
4 9.8905 (0.24176)
5 9.7300 (0.26094)
6 9.7930 (0.26366)
Random Effects Model

Consider the model with individual effects: Yit = Xitβit + ui + eit. We now assume that
the error component ui, the individual difference, is random (or stochastic) and
satisfies the following assumptions:
E(ui) = 0 (zero mean)

E(ui2) = σ2u (homoscedasticity)
E(uiuj) = 0 for i≠j (no cross-section correlation)
E(uieit) = E(uiejt) = 0 (independent from each eit or ejt)
Then, the model error is εit = ui + eit, which has the following structure:
E(εit) = E(ui + eit) = 0

E(εit2) = E((ui + eit)2) = σ2u + σ2e
E(εitεiτ) = E((ui + eit)(ui + eiτ)) = σ2u, for t≠τ
E(εitεjt) = E((ui + eit)(uj + ejt)) = 0, for i≠j
In other words, for each cross section i, the variance-covariance matrix of the model
error εi = [εi1, εi2, ...,εiT]' is the following T×T matrix:
2
+ σ2u σ2u … σ2u
σ e
2 2 2
σ uσ e+σ u… σ2u 
Σ=  =σ 2
eI + σ2u
… …… …
 σ2u σ2u … σ2e + σ2u 
If we let ε be an NT-element vector of the stacked errors ε1, ε2, ..., εN, then E(ε) = 0
and E(εε') = Σ⊗I, where I is an N×N identity matrix and Σ is the T×T variance-
covariance matrix defined above.
256
PANEL DATA ANALYSIS
Recall the pooled model for estimation, Y = Xβ + ε, where ε = [ε1,ε2,...,εN], εi =

[εi1,εi2,...,εiT]', and the random error has two components: εit = ui + eit. By
assumption, E(ε) = 0, and E(εε') = Σ⊗I. The generalized least squares estimate of β
is
β = [X'(Σ−1⊗I)X]-1X'(Σ−1⊗I)y
Since Σ-1 can be derived from the estimated variance components σ2e and σ2u, in
practice the model is estimated using the following partial deviation approach.
1. Estimate the model Y = Xβ + ε as a fixed effects model, using the dummy

variable approach, to obtain the estimated variance σ2e.
2. Assuming the randomness of ui, estimate the between parameters of the model:
Ymi = Xmiβ + (ui + emi)
where the error structure of ui + emi satisfies:
E(ui + emi) = 0
E((ui + emi)2) = σ2u + (σ2e/T)
E((ui + emi)(uj + emj)) = 0, for i≠j
Let v = σ2e and v1 = T σ2u + σ2e. Define w = 1 - (v/v1)½.
3. Use w to transform (partial deviations) the data as follows:
Y*it = Yit - w Ymi

X*it = Xit - w Xmi
Then the model for estimation becomes:
Y*it = X*itβ + ε*it
where ε*it = (1-w) ui + eit - w emi. Or, equivalently
Yit = Xitβ + w (ymi - Xmiβ) + ε*it
It is easy to validate that
E(ε*it) = 0
E(ε*2it) = σ2e
E(ε*itε*iτ) = 0 for t≠τ
E(ε*itε*jt) = 0 for i≠j
The least squares estimate of [w (Ymi - Xmiβ)] is interpreted as the change in the
individual effects.
257
Hypothesis Testing for Random Effects

To test the null hypothesis of no correlation between the error terms ui + eit and ui +
eiτ, we will use the following Breusch-Pagan LM test statistic based on the estimated
residuals of the restricted (pooled) model, εit (i=1,2,...N, t=1,2,...,T). The LM test
statistic is distributed as a Chi-square with one degree of freedom (note that εmi =
Σt=1,2,...,Tεit/T):
NT
2(T - 1)
[(∑ i =1,2,..., N
(∑ t =1,2,...T
ε it ) ) (∑
2
i =1,2,..., N ∑ t =1,2,..., T
) ]
ε it 2 − 1
2
or,
NT
2(T - 1)
[(∑ i =1,2,..., N
(Tε ) ) (∑
i
m 2
i =1,2,..., N ∑ t =1,2,..., T
) ]
ε it 2 − 1
2
Hausman Specification Test for Fixed or Random Effects

We have discussed the fixed and random effects models, and now you may be
wondering exactly what the difference is between the two models. How does one
decide to estimate a fixed vs. random effects model? The Hausman specification test
answers this question. It tests the null hypothesis that there is no difference between
the fixed and random effects models. Failure to reject the null hypothesis usually
implies that the fixed effects model is safe to use.
The Hausman test begins by noting that the difference between the fixed and random
effects models is in their respective covariance matrices. Let bfixed be the estimated
slope parameters of the fixed effects model (using the dummy variable approach),
and let brandom be the estimated slope parameters of the random effects model.
Similarly, let Var(bfixed) and Var(brandom) be the estimated covariance matrices for the
fixed and random effects models, respectively. The Hausman specification test
statistic is:
(brandom-bfixed)'[Var(brandom)-Var(bfixed)]-1(brandom-bfixed)
The Hausman test statistic is distributed as a Chi-square with degrees of freedom

equal to the number of slope parameters.
Lesson 16.2: One-Way Panel Data Analysis: Deviation Approach

We continue the previous example on the cost of airline services production. Instead
of using the dummy variable approach, we apply the deviation approach to
investigate the fixed effects and random effects. Recall that the main difference
between fixed and random effects lies in the assumption of the covariance structure
of the model. For the fixed effects model, total deviations of the data series from the
group means are used. For the random effects model, on the other hand, partial
deviations are employed. The deviation approach for one-way panel data analysis is
implemented in a GPE module program: PANEL1.GPE. In order to compute the
fixed effects and random effects, we run four regressions: pooled regression,
between-groups or means regression, within-groups full deviations regression, and
within-groups partial deviations regression. Three sets of hypothesis testing are
performed: Wald F-test for fixed effects, LM test for random effects, and Hausman
specification test comparing fixed and random effects. At the end, a summary of the
258
PANEL DATA ANALYSIS
panel data analysis is presented, including the estimated individual intercept

parameters for both the fixed and random effects models. PANEL1.GPE is installed
in the GPE subdirectory. The interested reader can examine the code to make sense
of the implementation (see also Appendix B-3). PANEL1.GPE can be included in
any part of your program with a compiler directive #include such as:
#include gpe\panel1.gpe;
We put the include directive at the end of program (see line 15 of lesson16.2). Then
one-way panel data analysis is called with the statement:
call panel1(y,x,n,t);
where y is the dependent variable and x is the data matrix of explanatory variables.
Both y and x are stacked according to the panel definition of n blocks (cross
sections) of t observations (time periods). To analyze fixed and random effects for
the airline services example, the program of Lesson 16.2 is given below:
/*
Lesson 16.2: One-Way Panel Data Analysis, Deviation Approach
Cost of Production for Airline Services II
*/
1 use gpe2;
4 panel=data[2:91,1:2]; @ panel definition @
5 n=6;
6 t=15;
/* stacked data series, by sections */

7 cs=ln(data[2:91,3]); @ log cost (stacked) @
8 qs=ln(data[2:91,4]); @ log output (stacked) @
9 pfs=ln(data[2:91,5]); @ log fuel price (stacked) @
10 lfs=data[2:91,6]; @ load factor (stacked) @
11 call reset;
12 _names = {"c","q","pf","lf"};
13 call panel1(cs,qs~pfs~lfs,n,t);
14 end;
15 #include gpe\panel1.gpe;
There are four sets of regression output, but we will present only the important
results of fixed effects and random effects models here:

------------------------
Mean of Dependent Variable = -6.3159E-016


Explained 39.068 3 13.023 3827.3 2.1614E-091
259
Residual 0.29262 86 0.0034026

Total 39.361 89 0.44226

Q 0.91928 0.029008 31.691 3.3110E-049 0.92112
PF 0.41749 0.014751 28.303 2.3780E-045 0.90305
LF -1.0704 0.19574 -5.4685 4.3807E-007 0.25801
CONSTANT -6.1586E-016 0.0061487 -1.0016E-013 1.0000 1.1666E-028

------------------------


Explained 40.185 3 13.395 3697.1 9.4659E-091
Residual 0.31159 86 0.0036231
Total 40.497 89 0.45502

Q 0.90668 0.025625 35.383 4.9455E-053 0.93572
PF 0.42278 0.014025 30.145 1.7169E-047 0.91354
LF -1.0645 0.20007 -5.3206 8.1016E-007 0.24765
CONSTANT 1.1873 0.025916 45.811 3.4012E-062 0.96064
The end of the estimation output produces a summary of the panel data analysis.
Three sets of hypothesis testing for fixed and random effects are given. Based on the
Wald F-test and the Breusch-Pagan LM test, it is clear that there exist both fixed
effects and random effects for this model. Based on the Hausman specification test,
however, there is no significant difference between the fixed and random effects.
Panel Data Model Estimation Procedure:

(1) Pooled Regression
(2) Between-Groups Regression
(3) Fixed Effects (Within-Groups) Regression
(4) Random Effects (Weighted Within-Groups) Regression
Wald F Test Statistic for Fixed Effects

F( 5, 81) = 57.732
Breusch-Pagan LM Test Statistic for Random Effects

Chi-Sq( 1) = 334.85
Hausman’s Test for Fixed and Random Effects

Chi-Sq( 3) = 0.75471
Within-Groups Estimates:
Fixed S.E. Random S.E.
0.91928 0.029890 0.90668 0.026404
0.41749 0.015199 0.42278 0.014451
-1.0704 0.20169 -1.0645 0.20615
-6.1586e-016 0.0063356 1.1873 0.026704
260
PANEL DATA ANALYSIS
One-Way Effects:
Section/Period Fixed S.E. Random S.E.
1.0000 9.7059 0.19323 9.6378 0.18313
2.0000 9.6647 0.19908 9.5979 0.18716
3.0000 9.4970 0.22505 9.4408 0.20686
4.0000 9.8905 0.24185 9.7780 0.21918
5.0000 9.7300 0.26102 9.6299 0.23371
6.0000 9.7930 0.26374 9.6831 0.23544
Finally, within-groups estimates of the slope parameters and the individual intercept
parameters are presented for the fixed effects and random effects models,
respectively. Note that the estimated fixed effects, derived from the deviation
approach, are the same as those of dummy variables approach. Furthermore, the
random effects are similar to the fixed effects, reinforcing the result of the Hausman
specification test that there is no significant difference between the two models.
Notice that the procedure panel1 is designed for study of individual (cross-
section) effects. To study the time effects, swap the panel definition n and t and
rearrange the stacked data series accordingly. For example, in lesson16.2, you can
insert the following statements (with comments for clarity) between lines 10 and 11:
@ re-arrange data, then swap n and t @
cs=vec(reshape(cs,n,t));
qs=vec(reshape(qs,n,t));
pfs=vec(reshape(pfs,n,t));
lfs=vec(reshape(lfs,n,t));
n=15;
t=6;
We leave the estimation and interpretation of the time period effects as an exercise.
Once you understand and master the idea of one-way panel data analysis, it is
straightforward to extend it to two-way analysis. Both cross-section and time period
effects are analyzed simultaneously under the respective assumptions of fixed effects
and random effects. Greene (1999) presented such an extension as two exercises in
Chapter 14. We implement the two-way analysis in the module program
PANEL2.GPE, which extends the module PANEL1.GPE for one-way analysis
used in Lesson 16.2. You may want to examine the code of PANEL2.GPE in
comparison with the outlined formula of Greene (1999), pp. 587-589. In essence, the
two-way analysis runs five regressions: a pooled regression, two between-groups
(time periods and cross sections) regressions, and two within-groups (full deviations
and partial deviations) regressions. From these regression estimations, we calculate
overall, cross section, and time period effects. As with one-way analysis, statistics
for testing fixed effects, random effects, and for comparing fixed and random effects
are computed. The module program PANEL2.GPE hides the details of
implementation from all but the most curious eyes. PANEL2.GPE can be found in
Appendix B-4 and it is installed in the GPE subdirectory.
Lesson 16.3: Two-Way Panel Data Analysis

Extending the analysis of one-way effects, in this example we re-estimate the airline
services production model to consider the two-way effects. We include
PANEL2.GPE at the end of the program. Similar to the one-way analysis, two-way
analysis is done with the statement (see line 14 of lesson16.3 below):
call panel2(y,x,n,t);
261
where y, the dependent variable, and x, the independent variables, are stacked
according to the panel definition of n blocks (cross sections) of t observations (time
periods). The rest of the program for two-way analysis is identical to the previous
lesson for one-way analysis.
/*
Lesson 16.3: Two-Way Panel Data Analysis
Cost of Production for Airline Services III
*/
1 use gpe2;
4 panel = data[2:91,1:2]; @ panel definition @
5 n=6;
6 t=15;
/* stacked data series, by sections */

7 cs = ln(data[2:91,3]); @ log cost (stacked) @
8 qs = ln(data[2:91,4]); @ log output (stacked) @
9 pfs = ln(data[2:91,5]); @ log fuel price (stacked) @
10 lfs = data[2:91,6]; @ load factor (stacked) @
11 xs = qs~pfs~lfs;
12 call reset;
13 _names = {"c","q","pf","lf"};
14 call panel2(cs,xs,n,t);
15 end;
It takes five regression estimations to carry out two-way panel data analysis. To save
space, we will report only the summary information as follows:

(2) Between-Groups (Cross Sections) Regression
(3) Between-Groups (Time Periods) Regression

F( 19, 67) = 23.102

Chi-Sq( 2) = 336.40

Chi-Sq( 3) = 183.54
0.81725 0.031851 0.90237 0.029742
0.16861 0.16348 0.42418 0.016306
-0.88281 0.26174 -1.0531 0.22948
6.1829e-016 0.0054155 1.0109 0.025968
Two-Way Effects:
Fixed Random
Overall 12.667 1.6784
Cross Sections Effects:

Sections Fixed Random
262
PANEL DATA ANALYSIS
1.0000 0.12833 7.9348

2.0000 0.065495 7.8933
3.0000 -0.18947 7.7292
4.0000 0.13425 8.0709
5.0000 -0.092650 7.9171
6.0000 -0.045956 7.9710
Time Periods Effects:

Periods Fixed Random
1.0000 -0.37402 -0.0023032
2.0000 -0.31932 0.00074765
3.0000 -0.27669 0.0030529
4.0000 -0.22304 0.0049901
5.0000 -0.15393 0.00044843
6.0000 -0.10809 -0.0013027
7.0000 -0.076864 -0.0011691
8.0000 -0.020733 -0.00015766
9.0000 0.047220 0.0025912
10.000 0.091728 -0.0018190
11.000 0.20731 -0.0018378
12.000 0.28547 0.00047461
13.000 0.30138 0.0022213
14.000 0.30047 0.0027990
15.000 0.31911 0.0043389
From the two-way analysis, we can see that the model exhibits significant fixed
effects and random effects. The magnitude and the pattern of the two effects are
different. From examining the “Time Periods Effects” in the output, we see that the
fixed effects are larger than the random effects. On the other hand, we see that for
the “Cross Sections Effects,” the magnitude of the random effects is greater than that
of the fixed effects.
Remember that to analyze one-way (time or individual) effects, PANEL1.GPE

should be included. PANEL2.GPE is used for analyzing two-way (time and
individual) effects. Furthermore, both modules are capable of dealing with missing
observations or unbalanced panels. The missing data observations must be
identified with GAUSS internal notation of missing value. See example lessons of
Chapter 9 for more details.
Seemingly Unrelated Regression System

The classical panel data analysis investigates only the intercept difference across
individuals or time periods. Consider a more general specification of the model:
Yit = Xitβi + εit (i=1,2,...,N; t=1,2,...,T)
Let Yi = [Yi1,Yi2,...,YiT]', Xi = [Xi1,Xi2,...,XiT]', and εi = [εi1,εi2,...,εiT]'. The stacked N

equations (T observations each) system is Y = Xβ + ε, or
ε1
Y2  =  0 X2 … 0 β + ε2 
Y1   X1 0 … 0
 …  … … … …  …
YN   0 0 … XN εN 
Notice that not only the intercept but also the slope terms of the estimated parameters
are different across individuals. Of course, the restrictions of identical slope terms
263
across individuals may be imposed for direct comparison with the classical methods.
The error structure of the model is summarized as follows:
E(ε) = 0
E(Xε) = 0
E(εε') = Σ⊗I
where Σ = [σij, i,j=1,2,...N] is the NxN variance-covariance matrix and I is a T×T

identity matrix. Notice that contemporaneous correlation across individuals is
assumed, although the assumption of no serial correlation is implied. The error
structure of this model is different than that of the random effects model described
above.
System estimation techniques such as 3SLS and FIML should be used for parameter
estimation in this kind of model, which is seemingly unrelated regression estimation
in the current context. The SUR estimation method was discussed in Chapter XIII.
Denote b and S as the estimated β and Σ, respectively. Then,
b = [X'(S-1⊗I)X]-1X'(S-1⊗I)y
Var(b) = [X'(S-1⊗I)X]-1
and S = ee'/T, where e = y-Xb is the estimated error ε.
The advantage of the SUR estimation method for panel data analysis is that it not
only allows the intercept difference between individuals (as in the fixed and random
effects models), but also allows the slope to vary among individuals. If the slope
parameters are assumed to be constant across individuals, the method differs from
the random effects model in the fundamental assumption of the covariance structure.
By allowing cross-section correlation, the restricted SUR method is more general
than the classical random effects model.
Lesson 16.4: Panel Data Analysis for Investment Demand: Deviation Approach
To demonstrate the different approaches for panel data analysis, we consider the
following classical example of investment demand (Greene, 1999, Chap. 15;
Grunfeld and Griliches, 1960; Boot and deWitt, 1960):
Iit = αi + β1i Fit + β2i Cit + εit
Where i = 5 firms: General Motors, Chrysler, General Electric,

Westinghouse, and United Steel.
t = 20 years: 1935-1954.
Iit = Gross investment.
Fit = Market value.
Cit = Value of the stock of plant and equipment.
The panel data of 20 years for 5 companies are available in 5 separate files, one for
each company. The data files used are: ifcgm.txt (General Motor), ifcch.txt
(Chrysler), ifcge.txt (General Electric), ifcwe.txt (Westinghouse), ifcus.txt (United
Steel).
264
PANEL DATA ANALYSIS
First we assume that β1i = β1 and β2i = β2 for all firms. In other words, we are
estimating the restricted SUR model by assuming that the slope parameters do not
vary across firms. To estimate and compare the fixed effects and random effects for
the model, we use the following program which is essentially the same as that of
lesson16.2. Since the five company data sets are read in separately as time series,
some manipulation is necessary to convert them into a stacked vector of dependent
variables and a stacked matrix of independent variables (see lines 8 through 14 in
lesson16.4 below). The stacked data format is required in order to use the
PANEL1.GPE module program.
/*
Lesson 16.4: Panel Data Analysis for Investment Demand
Deviation Approach
*/
1 use gpe2;
3 load gmc[21,4] = gpe\ifcgm.txt;

4 load chc[21,4] = gpe\ifcch.txt;
5 load gec[21,4] = gpe\ifcge.txt;
6 load wec[21,4] = gpe\ifcwe.txt;
7 load usc[21,4] = gpe\ifcus.txt;
8 i=gmc[2:21,2]~chc[2:21,2]~gec[2:21,2]~wec[2:21,2]~usc[2:21,2];
9 f=gmc[2:21,3]~chc[2:21,3]~gec[2:21,3]~wec[2:21,3]~usc[2:21,3];
10 c=gmc[2:21,4]~chc[2:21,4]~gec[2:21,4]~wec[2:21,4]~usc[2:21,4];
11 n=5; @ 5 cross sections (firms) @

12 t=20; @ 20 time periods (years) @
@ stacked data series, by firms @

13 ys = vec(i);
14 xs = vec(f)~vec(c);
15 call reset;
16 _names={"i","f","c"};
17 call panel1(ys,xs,n,t);
18 end;
As described earlier, using the module PANEL1.GPE to estimate the one-way fixed
and random effects gives us four sets of regression output: the pooled regression,
between-groups means regression, within-groups full deviations regression, and
within-groups partial deviations regression. You should check the details of each
regression output. We present only the summary results of the analysis.

(2) Between-Groups Regression

F( 4, 93) = 58.956

Chi-Sq( 1) = 453.82

Chi-Sq( 2) = 0.033043
265
0.10598 0.015891 0.10489 0.015112
0.34666 0.024161 0.34602 0.024770
-1.6507e-014 6.9118 -8.8082 8.1293
One-Way Effects:
Section/Period Fixed S.E. Random S.E.
1.0000 -76.067 66.886 -69.356 58.234
2.0000 -29.374 19.814 -33.176 19.376
3.0000 -242.17 33.321 -213.56 31.028
4.0000 -57.899 19.703 -57.575 19.263
5.0000 92.539 33.947 72.218 31.535
It is interesting to find the classical estimates of fixed effects and random effects are
similar. This is consistent with the very small Hausman specification test statistic
shown in the output.
Lesson 16.5: Panel Data Analysis for Investment Demand: SUR Method
By restricting β1i = β1 and β2i = β2 for all firms, the restricted SUR estimation
method is used in direct comparison with the classical methods of panel data
analysis. In Chapter XIII we implemented and estimated a system of linear demand
equations using the SUR estimation method. The use of the input control variable
_eq in estimating the simultaneous linear equations system was discussed in detail
in Chapter XIII. In Chapter III we introduced the implementation of restricted least
squares with the use of input control variable _restr. Parameter restrictions across
equations in a linear equations system were again discussed in Chapter XIII. You
may want to review these chapters and the relevant examples before working on this
lesson.
In Lesson 16.5, the restricted SUR method is estimated using iterative three-stage
least squares (_method=3). The result is the same as full information maximum
likelihood.
/*
Lesson 16.5: Panel Data Analysis for Investment Demand Function
Seemingly Unrelated Regression Estimation
*/
1 use gpe2;
3 load gmc[21,4] = gpe\ifcgm.txt;

4 load chc[21,4] = gpe\ifcch.txt;
5 load gec[21,4] = gpe\ifcge.txt;
6 load wec[21,4] = gpe\ifcwe.txt;
7 load usc[21,4] = gpe\ifcus.txt;
8 i=gmc[2:21,2]~chc[2:21,2]~gec[2:21,2]~wec[2:21,2]~usc[2:21,2];
9 f=gmc[2:21,3]~chc[2:21,3]~gec[2:21,3]~wec[2:21,3]~usc[2:21,3];
10 c=gmc[2:21,4]~chc[2:21,4]~gec[2:21,4]~wec[2:21,4]~usc[2:21,4];
11 yvar=i;
12 xvar=f~c;
13 call reset;
14 _names={"i-gm","i-ch","i-ge","i-we","i-us",
"f-gm","f-ch","f-ge","f-we","f-us",
"c-gm","c-ch","c-ge","c-we","c-us"};
@ I I I I I F F F F F C C C C C 1@
266
PANEL DATA ANALYSIS
15 _eq = {-1 0 0 0 0 1 0 0 0 0 1 0 0 0 0,
0 -1 0 0 0 0 1 0 0 0 0 1 0 0 0,
0 0 -1 0 0 0 0 1 0 0 0 0 1 0 0,
0 0 0 -1 0 0 0 0 1 0 0 0 0 1 0,
0 0 0 0 -1 0 0 0 0 1 0 0 0 0 1};
@ F C|F C|F C|F C|F C|q @

16 _restr = {-1 0 1 0 0 0 0 0 0 0 0,
-1 0 0 0 1 0 0 0 0 0 0,
-1 0 0 0 0 0 1 0 0 0 0,
-1 0 0 0 0 0 0 0 1 0 0,
0 -1 0 1 0 0 0 0 0 0 0,
0 -1 0 0 0 1 0 0 0 0 0,
0 -1 0 0 0 0 0 1 0 0 0,
0 -1 0 0 0 0 0 0 0 1 0};
17 _iter=200;
18 _method=3;
20 end;
You should run the program to get the full report of the estimation results. The
output of the restricted SUR estimation is lengthy, but can be summarized as
follows:

----------------------------------------

Tolerance = 0.001
…

I-GM F-GM 0.033825 0.0063427 5.3330
C-GM 0.15536 0.016608 9.3541
CONSTANT 360.69 51.871 6.9536
I-CH F-CH 0.033825 0.0063427 5.3330
C-CH 0.15536 0.016608 9.3541
CONSTANT 43.839 6.9344 6.3220
I-GE F-GE 0.033825 0.0063427 5.3330
C-GE 0.15536 0.016608 9.3541
CONSTANT -25.543 13.842 -1.8454
I-WE F-WE 0.033825 0.0063427 5.3330
C-WE 0.15536 0.016608 9.3541
CONSTANT 6.8931 4.7102 1.4634
I-US F-US 0.033825 0.0063427 5.3330
C-US 0.15536 0.016608 9.3541
CONSTANT 292.18 27.650 10.567

I-GM 39110.
I-CH 3359.9 575.64
I-GE 312.36 -49.662 670.86
I-WE 317.11 30.538 178.24 96.919
I-US 9742.5 1695.0 919.47 622.51 12240.
I-GM I-CH I-GE I-WE I-US
267
To compare the fixed effects, random effects, and SUR method, the estimated
parameters of the investment function are tabled together. The individual effects for
three methods (under different covariance assumptions) are shown in the rows of
intercept terms for each firm. Numbers in parentheses are the estimated standard
errors.
Fixed Effects Random Effects SUR Method

Slope F 0.10598 (0.01589) 0.10489 (0.01511) 0.033825 (0.006343)
C 0.34666 (0.02416) 0.34602 (0.02477) 0.15536 (0.01661)
Intercept GM -76.067 (66.886) -69.356 (58.234) 360.69 (51.871)
CH -29.374 (19.814) -33.176 (19.376) 43.839 (6.9344)
GE -242.17 (33.321) -213.56 (31.028) -25.543 (13.842)
WE -57.899 (19.703) -57.575 (19.263) 6.8931 (4.7102)
US 92.539 (33.947) 72.218 (31.535) 292.18 (27.650)
Although the estimates from the models of fixed effects and random effects are
similar, the parameter estimates obtained from the SUR estimation method are quite
different. The impact of different covariance assumptions when estimating the model
is obvious. Since the SUR method is typically applied to estimating a model with
varying slope as well as intercept terms, we can easily estimate the unrestricted
model by removing (or commenting out) the restriction statement in line 16 of
lesson16.5 above. By comparing the results to those of the restricted model, the
validity of the assumption of constant slopes may be tested. The following table
presents the comparison results of restricted and unrestricted estimates (standard
errors are in parentheses). The large Likelihood Ratio statistic of the two models,
calculated as 2 × [-459.092 -(-490.753)] = 63.322, leads us to the conclusion that the
slope parameters are not the same across the five firms under consideration.
Eq. Variable Unrestricted Model Restricted Model

I-GM F 0.12195 (0.020243) 0.033825 (0.006343)
C 0.38945 (0.031852) 0.15536 (0.01661)
Constant -173.04 (84.280) 360.69 (51.871)
I-CH F 0.067451 (0.017102) 0.033825 (0.006343)
C 0.30507 (0.026067) 0.15536 (0.01661)
Constant 2.3783 (11.631) 43.839 (6.9344)
I-GE F 0.037019 (0.01177) 0.033825 (0.006343)
C 0.11695 (0.021731) 0.15536 (0.01661)
Constant -16.376 (24.961) -25.543 (13.842)
I-WE F 0.053861 (0.010294) 0.033825 (0.006343)
C 0.026469 (0.037038) 0.15536 (0.01661)
Constant 4.4891 (6.0221) 6.8931 (4.7102)
I-US F 0.0886 (0.045278) 0.033825 (0.006343)
C 0.3093 (0.11783) 0.15536 (0.01661)
Constant 138.01 (94.608) 292.18 (27.650)
Log-Likelihood -459.092 -490.753
In summary, we have presented the classical methods of panel data analysis: fixed
effects and random effects. A more general SUR approach was introduced, which
allowed us to consider contemporaneous correlation across individuals, which the
classical methods ignore. Misspecification issues such as autocorrelation and
heteroscedasticity in panel data are important. In Chapter X we discussed the
268
PANEL DATA ANALYSIS
problem of autocorrelation associated with time series, while heteroscedasticity in

cross-sectional models was covered in Chapter IX. The combination of
autocorrelation and heteroscedasticity is common in models for panel data. The
treatment of autocorrelation for time-period effects and heteroscedasticity for cross-
section effects would be an integrated and complicated application of this and the
two aforementioned chapters, which we will not discuss here.
269
XVII
Least Squares Prediction
The art of forecasting lies in building a practical model for real world application,
and in the preceding chapters, we have presented all the tools necessary to do so in
GPE. This chapter introduces the few remaining GPE control variables dedicated
solely to least squares prediction and time series forecasting.
Least squares prediction is nothing more than the extrapolation of the estimated
regression model from a set of historical observations into the unknown future. It is
assumed that given the stable model structure, the future state is predictable from the
replication of history.
Predicting Economic Growth

In this chapter, we will consider a “conventional wisdom” that the future state of the
economy (measured in terms of real GDP growth) is predictable by an index called
the Composite Economic Leading Indicator, which is assembled and updated
monthly by the Conference Board and U.S. Department of Commerce. The Indicator
is a weighted average of 10 short-run economic factors, such as stock prices and
average hours worked. It is often reported in the media that this Leading Indicator
can predict the direction of the economy 3 to 9 months into the future.
The lessons in this chapter use the data file gdp96.txt. It consists of four variables:
QUARTER (quarterly index), GDP (Gross Domestic Product in billions of dollars),
PGDP (Implicit Price Deflator of GDP, 2000 = 100), and LEADING (Composite
Economic Leading Indicator, 1996 = 100). We note that the quarterly series
LEADING is the last month of the quarter.
The target variable is the annual growth rate of real GDP. The following GAUSS
statements generate the required data series of GDP growth:
rgdp = 100*gdp./pgdp;
growth = 100*(rgdp-lagn(rgdp,4))./lagn(rgdp,4);
First, Real Gross Domestic Product is expressed in billions of 2000 dollars. Then,
GDP growth is measured as the annual percentage rate of change in real GDP from
the same quarter last year. Although the causal relationship of the variables
LEADING and GROWTH is well grounded, we have to make sure that these two
variables are cointegrated. It turns out that both variables are stationary or I(0)
processes and thus do not have unit roots. Moreover, the two variables are
cointegrated. We leave the details of the unit roots and cointegration tests of
LEADING and GROWTH as exercises. See also Chapter XVI for a thorough
review.
We are now ready to construct a working model suitable for short-run structural
estimation and prediction. Since forecasting is a time-sensitive business, we reserve
the last two years of data for ex-post forecast evaluation. In other words, we are
going to estimate the model using data through 2001, and see how well the model
predicts real GDP growth in 2002 and 2003. We need to construct a working model
not only for historical estimation but also for forecasting18.
If the variable LEADING can actually predict GROWTH several quarters ahead,
then a distributed lag structure must be specified. As a result of trial and error, we
have found that both the first and fifth quarters lag of LEADING are useful in
explaining historical GROWTH. In addition, the momentum effect of GDP growth is
captured with a lagged dependent variable. The model error is identified to be a
MA(4) process. By construction, the dependent variable GROWTH is the annual rate
of GDP growth based on the same quarter in the previous year. The specification of
fourth-order moving average for the model error term should not, therefore, be
surprising. Of course, this may not be the only working specification of the model
you can construct. Throughout this book we have given examples of model building.
We leave the process of finding the best model for forecasting to you. However, we
emphasize the importance of using GPE variables such as _pdl and _dlags to
determine the short-run dynamics of the model structure.
We now turn to new forecasting features of the GPE package. In GPE, the work of
least squares prediction is done by a procedure called forecast. Forecasts are
usually computed after the estimation of the model. Calling forecast is similar to
calling estimate. In calling forecast, you need to specify only the dependent
and independent variables (in that order). The estimated parameters and the
associated covariance matrix of the regression model in the immediately preceeding
estimate statement are used to compute the forecasts for the same model. The
forecasting period defaults to begin one observation after the estimation period ends
and continues to the end of the data series. If future observations of the dependent
variable become available, ex-post forecast error statistics based on the results of
least squares prediction can be used for model evaluation.
If there are longer series of right-hand side explanatory variables, ex-ante forecasts
can be computed upon request. The GPE control variables _fbegin and _fend are
used to specify the beginning and ending of the multiple steps ahead of forecasts. In
most cases, ex-ante forecasting depends on the forecasts or scenario assumptions
made regarding the explanatory independent variables. If the Composite Economic
Leading Indicator can predict the economy three to nine months ahead as claimed,
our model certainly can point out the direction of GROWTH about one year in
advance of current Leading Indicator. Furthermore, by making scenario assumptions
about the variable LEADING (for example assuming no change in LEADING for
the next year or so) we can predict the future value of GROWTH even further out on
the horizon.
Lesson 17.1: Ex-Post Forecasts and Forecast Error Statistics

Here is the program predicting economic growth with the Composite Economic
Leading Indicator:
18
This chapter is printed based on the 2004 forecasts, and it is subject to annual revision and
updates. The latest version of forecasts can be found in the e-book copy of this chapter on the
CD-ROM.
272
LEAST SQUARES PREDICTION
/*
** Lesson 17.1: Ex-Post Forecasts and
** Forecast Error Statistics
*/
1 use gpe2;
3 n=181; @ 1959.1 to 2003.4 @

4 load z[n,4] = gpe\gdp96.txt;
5 gdp = z[2:n,2];
6 pgdp = z[2:n,3];
7 leading = z[2:n,4];
8 rgdp = 100*gdp./pgdp;
9 growth = 100*(rgdp-lagn(rgdp,4))./lagn(rgdp,4);
10 xvar = lagn(leading,1)~lagn(leading,5);
/* Model Estimation */
11 call reset;
12 _rstat=1;
13 _dlags=1;
/*
_bgtest=4;
_ebtest=4;
_acf=12;
_acf2=12;
*/
14 _arma={0,4};
15 _iter=100;
16 _method=5;
17 _begin=9; @ 1961Q1 @
18 _end=172; @ 2001Q4 @
19 call estimate(growth,xvar);
/* Forecasting */
20 _fstat=1;
21 _fplot=1;
@ _dynamic=1; @
22 call forecast(growth,xvar);
23 end;
The program is divided into two main sections: estimation and forecasting. Notice
that line 10 assigns the matrix of independent variables to the variable XVAR.
XVAR is then passed to both estimate (line 19) and forecast (line 22).
Modifying the independent variable matrix can quickly be done by editing only line
10.
The distributed lag structure of the model includes the first and fifth quarters lags of
the independent variable LEADING (line 10) and a one quarter lag of the dependent
variable GROWTH (line 13). The first five quarters of data series are lost due to
variable transformation. Through model identification, we determine that the error
structure follows an MA(4) process. Using the QHC method for maximum
likelihood estimation (lines 15-16), the model is estimated from the first quarter of
1961 (or the 9th observation) to the fourth quarter of 2001 (or the 172nd
observation):
_begin = 9;
_end = 172;
273
The _begin statement (line 17) safely blocks out the unusable data series for
estimation, while _end statement (line 18) reserves the rest of the data series (the
last two years of 2002 and 2003) for ex-post forecast evaluation. The output of the
estimated model follows:

------------------------
Dependent Variable = Y

ARMA( 0, 4) Autoregressive Moving Average Process
Tolerance = 0.001
Initial Result:
Parameters = 0.64250 0.32175 -0.33188 1.6991 0.00000
0.00000 0.00000 0.00000

Parameters = 0.65828 0.30965 -0.32007 1.6881 -0.093052
-0.32130 -0.098387 0.49620
…
Parameters = 0.70158 0.29521 -0.30270 1.3044 -0.14822
-0.18625 -0.049189 0.66183
Final Result:
Parameters = 0.70158 0.29521 -0.30270 1.3044 -0.14822
-0.18625 -0.049189 0.66183
Gradient Vector = -0.24025 -5.8682 -5.7777 -0.070409
0.0014879 -0.00099907 -0.00010876 0.0028283

MA(1) -0.14822 0.080464 -1.8421
MA(2) -0.18625 0.078709 -2.3663
MA(3) -0.049189 0.089459 -0.54985
MA(4) 0.66183 0.081899 8.0810


Explained 849.84 3 283.28 459.88 2.0835E-078
Residual 98.559 160 0.61599
Total 917.39 163 5.6281

Y-1 0.70158 0.050990 13.759 6.3428E-029 0.54196
274
X1 0.29521 0.038418 7.6842 1.4572E-012 0.26956

X2 -0.30270 0.038370 -7.8889 4.5030E-013 0.28004
CONSTANT 1.3044 0.35681 3.6557 0.00034744 0.077088

Although the estimated model with MA(4) error structure looks satisfactory, add the
following few lines before the estimation call (line 19):
_bgtest = 4;
_ebtest = 4;
_acf = 12;
_acf2 = 12;
and rerun the model to verify the serial correlation problem in the conditional mean
and variance,19 if any.
The next section of the program (lines 20 to 22), calls for least squares prediction
based on the immediate previously estimated model. Simply calling forecast
specifies the default prediction period, which begins after the last observation used in
the regression estimation, and ends with the end of the sample. We note that the
beginning and end of the prediction period can be controlled by two GPE input
variables, _fbegin and _fend, respectively.
Ex-post forecast error statistics are computed by setting the input control variable
_fstat=1 (line 20). This control variable is similar to its counterpart, _rstat,
used for estimate. In addition, plotting of the forecasts and actuals can provide a
visual clues as to the model’s performance. This is done in line 21 by setting the
input control variable _fplot = 1.
Here is the forecast output:

------------------------
Prediction Range = 173 180
Using Regression Coefficients:
0.70158 0.29521 -0.30270 1.3044 -0.14822 -0.18625
-0.049189 0.66183
Observation Observed Predicted Residual Std.Error

173 1.15838 0.97517 0.18321 0.81319
174 1.90558 2.34159 -0.43601 1.04541
175 2.97348 2.95954 0.01393 1.20889
176 2.73318 3.31610 -0.58292 1.29304
177 2.16124 2.60103 -0.43979 1.30589
178 2.34095 1.81225 0.52870 1.31317
179 3.61795 2.29109 1.32686 1.31651
180 4.29540 3.78341 0.51199 1.31871
R-Square Between Observed and Predicted = 0.59083
19
The dynamic model may be correlated in terms of conditional variance, identifiable with
ARCH or GARCH specification.
275
Mean Error (ME) = 0.13825

Mean Absolute Error (MAE) = 0.50293
Mean Squared Error (MSE) = 0.38241
Root Mean Squared Error (RMSE) = 0.61839
Mean Absolute Percent Error (MAPE) = 19.003
Mean Squared Percent Error (MSPE) = 455.00
Root Mean Squared Percent Error (RMSPE) = 21.331
Theil Inequality Coefficient = 0.11353
Decomposition:
Proportion Due to Bias = 0.049978
Proportion Due to Variance = 0.028836
Proportion Due to Covariance = 0.92119
Proportion Due to Regression = 0.031276
Proportion Due to Disturbance = 0.91875
Each observation in the forecast period is listed, complete with observed and
predicted values. Residuals (or forecast errors) and their standard errors are also
given. Since we have put aside the last two years (eight quarters) of the GROWTH
data series to be compared with the forecasts, ex-post forecast error statistics,
including mean squared error and its components, are computed from the last eight
quarters of GDP growth. Moreover, forecasts in pairs with actuals are plotted
together with the band of two standard errors. We note that the upper and lower
bounds of forecast are the minimal spread of the prediction. In reality, the forecast
interval tends to be much wider due to additional non-sampling errors. Econometrics
texts describe the model evaluation based on this set of forecast error statistics in
detail. We leave judgment of the model’s performance to you.
It can be shown that the method of least squares yields the best, linear, and unbiased
predictor. Since the model is dynamic in nature (with a lagged dependent variable),
we have an option to perform a dynamic forecast. A dynamic forecast is obtained by
using the predicted lagged dependent variable on the right-hand side of the
forecasting equation, instead of the actual lagged dependent variable. Let’s turn on
the dynamic option of least squares prediction:
_dynamic = 1;
Make sure that the dynamic option is added before calling forecast, and run the
program to see the result:

------------------------
NOTE: Dynamic Prediction Computed.

0.70158 0.29521 -0.30270 1.3044 -0.14822 -0.18625
-0.049189 0.66183

173 1.15838 0.97517 0.18321 0.81319
174 1.90558 2.28693 -0.38135 1.04597
175 2.97348 2.99605 -0.02257 1.20865
176 2.73318 3.05532 -0.32214 1.29393
177 2.16124 2.80886 -0.64762 1.30580
178 2.34095 2.26661 0.07434 1.31338
179 3.61795 2.23893 1.37902 1.31661
180 4.29540 2.81592 1.47948 1.32184

276

Decomposition:
As expected, the model performance deteriorates when we forecast farther ahead into
the future. This is because the predicted value of the lagged dependent variable is
used in place of the actual value of the lagged dependent variable. Including the
predicted value of the lagged dependent variable simply means that each forecast
error is compounded over the forecast period. One important characteristic of the
dynamic forecast is that the further in the future we try to predict, the less reliable the
forecasts we get.
Lesson 17.2: Ex-Ante Forecasts

Forecasting is a time sensitive business. At the time of this writing, new information
on the Composite Economic Leading Indicator may become available. We could
modify the data file gdp96.txt, or just add the new data into the program. The latter
is especially helpful to carry out scenario ex-ante forecasts. That is, we extend the
data further by making a scenario assumption about the Leading Indicator to predict
GDP growth in accordance with the assumption. For example, a “constant scenario”
would assume no change (from the last quarter of historical data) in the variable
LEADING for the next year or so. Then, in line 7 of lesson17.2, the hypothesized
observations of LEADING are appended at the end of historical data series as
follows:
leading = z[2:n,4]|115.0|115.0|115.0|115.0|115.0;
Recall that the variable z is the original data matrix read from the data file
gdp96.txt. Similarly, we could create a “pessimistic scenario” similar to the
following, in which the variable LEADING declines at 2 percent over the next year:
leading = z[2:n,4]|115.0|114.4|113.8|113.2|112.6;
Or, we could assume an “optimistic scenario” (2 percent annual growth rate) as well:
leading = z[2:n,4]|115.0|115.6|116.2|116.8|117.4;
In other words, ex-ante forecasts are nothing but a crystal-ball prediction about
uncertain future conditions. To keep the model performance in line with the available
information, we do not use the dynamic features of the least squares prediction
during the ex-post forecast periods. Dynamic forecast is automatic anyway, during
the ex-ante forecast periods, since the value of the lagged dependent variable is not
available and must first be predicted for the period that follows.
The following program is almost identical to that of Lesson 17.1. Pay attention to the
change we made in line 7 assuming a scenario for ex-ante forecast:
277
/*
** Lesson 17.2: Ex-Ante Forecasts
*/
1 use gpe2;
3 n=181; @ 1959.1 to 2003.4 @

4 load z[n,4] = gpe\gdp96.txt;
5 gdp = z[2:n,2];
6 pgdp = z[2:n,3];
/* pessimistic scenario (-2% AGR) */

@ leading = z[2:n,4]|115.0|114.4|113.8|113.2|112.6; @
/* constant scenario (0% AGR) */
7 leading = z[2:n,4]|115.0|115.0|115.0|115.0|115.0;
/* optimistic scenario (+2% AGR) */
@ leading = z[2:n,4]|115.0|115.6|116.2|116.8|117.4; @
8 rgdp = 100*gdp./pgdp;
9 growth = 100*(rgdp-lagn(rgdp,4))./lagn(rgdp,4);
10 xvar = lagn(leading,1)~lagn(leading,5);
/* Model Estimation */
11 call reset;
12 _rstat=1;
13 _dlags=1;
14 _arma={0,4};
15 _iter=100;
16 _method=5;
17 _begin=9; @ 1961Q1 @
18 _end=172; @ 2001Q4 @
19 call estimate(growth,xvar);
/* Forecasting */
20 _fstat=1;
21 _fplot=1;
@ _dynamic=1; @
22 call forecast(growth,xvar);
23 end;
Here is the forecasting output of the “constant scenario”:

------------------------
NOTE: Dynamic Prediction Computed.

0.70158 0.29521 -0.30270 1.3044 -0.14822 -0.18625
-0.049189 0.66183

173 1.15838 0.97517 0.18321 0.81319
174 1.90558 2.34159 -0.43601 1.04541
175 2.97348 2.95954 0.01393 1.20889
176 2.73318 3.31610 -0.58292 1.29304
177 2.16124 2.60103 -0.43979 1.30589
178 2.34095 1.81225 0.52870 1.31317
179 3.61795 2.29109 1.32686 1.31651
180 4.29540 3.78341 0.51199 1.31871
181 . 4.40031 . 1.32010
182 . 4.92272 . 1.32335
183 . 4.86545 . 1.32125
184 . 4.46203 . 1.32145
278
185 . 3.78549 . 1.32217

Decomposition:
Similarly, we run the other two scenarios, pessimistic (LEADING decreases) and
optimistic (LEADING increases), respectively. Instead of listing each of the
forecasting results, we compare the respective ex-ante forecasts in the following
table:
Predicted Pessimistic (low) Constant Optimistic (high)

GDP Growth Scenario Scenario Scenario
(%, annual rate)
2004.1 4.41 4.41 4.41
2004.2 4.93 4.93 4.93
2004.3 4.69 4.87 5.05
2004.4 3.99 4.46 4.94
2005.1 2.92 3.79 4.65
Furthermore, the following graph summarizes the ex-post and ex-ante forecasts for
three different scenarios. The picture can tell a complicate story more clearly.
What you can say about the predictability of the Composite Economic Leading
Indicator? No matter which scenario is used the economy appears to be heading
279
towards a so-called “jobless recovery”. Will the recovery continue? Even under the
optimistic view, the growth may not sustain by the end of 2004. It will certainly
depend on an effective government policy to revive the growth. Only time will tell!
280
Epilogue
This is not the end of Gauss Programming for Econometricians and
Financial Analysts!
It is just the beginning. Many extensions of econometric estimation, testing, and

forecasting techniques discussed in this book can be done with GPE. Taking
advantage of GPE output variables for each call to the main routine estimate or
forecast, you can write add-on programs to do advanced applications as we did
with the GPE application modules in Appendix B. Within the scope of econometric
models we have discussed so far, you can experiment with the following extensions:
• Nonlinear ARCH-M Model Estimation and Prediction

• Qualitative Choice Model with Heteroscedasticity
• Panel Data Analysis with Heteroscedasticity and Autocorrelation
Many important topics in econometrics we did not cover here would certainly be
good candidates for GAUSS implementation. To name a few examples:
• Monte Carlo Simulations and Bootstrapping Methods

• Nonparametric Regression Analysis
• Baysian Estimation and Inference
• Benchmarking Econometric Computation
Beyond GPE, you may feel ready to write your own codes for econometric and
statistical applications. More than 400 GAUSS commands, procedures, and functions
are available at your disposal as part of the GAUSS programming environment. As a
consequence, we have seen powerful procedures being developed over the past
years.
Whatever your eventual goals, you will probably agree that learning econometrics
with GPE is certainly the right first step. We have demonstrated that GAUSS is a
programming environment built on the convenient syntax and operations of matrix
algebra. As you step through each lesson, learning to master GPE, you also learn
GAUSS and econometrics. From here, the next step is up to you!
Appendix A
GPE Control Variables
There are two types of global control variables in GPE: input control variables and
output control variables. For reference purposes, consider the following general
regression equation:
F(Z, β) = ε
where Z is the data matrix of variables and β is the vector of parameters, which
define the functional form F. ε is the error structure of the model. Z can be further
decomposed as Z = [Y, X] with Y denoting the endogenous (dependent) variables
and X the predetermined (independent) variables. If Y consists of more than one
column, it is a system of linear equations. For a classical regression model, Y =
f(X, β) + ε or F(Z, β) = Y - f(X, β) = ε. The simple case of single linear regression
equation is written as:
Y = Xβ + ε
where Y is the left-hand side (LHS) or dependent variable, and X denotes the right-
hand side (RHS) explanatory or independent variables. β is the vector of estimated
parameters, and ε is the vector of estimated residuals.
Input Control Variables

Input control variables must be initialized before calling one of the main econometric
routines: estimate, optimize, or forecast. To initialize all the global
control variables, call reset.
Three categories of input control variables are listed below: general-purpose input
control variables, estimate (and optimize) input control variables, and
forecast input control variables. Unless otherwise specified, setting each variable
to 1 (that is, true or yes) activates or turns on the optional behavior specified. If the
variable is not defined or specified, then its default value is assumed.
General Purpose Input Control Variables

Variable Description
_cmplx Complex number computation.

_cmplx = 0 (default): Do not allow for complex number computation,
therefore negative argument for LN, LOG, and SQRT is not permitted;
_cmplx =1: Allow for complex number computation.
_legend When a graph is requested (_rplot>0, see below),

_legend = 1 (default): Show legends for graph plots;
_legend = 0: No legends will be shown.
_pause Pause the output before displaying graphs.

_pause = 0 (default): No waiting prompt;
_pause = 1: Wait for a keystroke to display graphs.
_print Controls screen output.

_print = 1 (default): Direct full output to screen;
_print = 0: Direct partial output to screen. Verbose iteration outputs
from a nonlinear or iterative model are suppressed;
_print = -1: Suppress all screen output. Suppressing the screen output
will force _rplot = 0 and _fplot = 0 (see below), but it will have no
effect on sending output to a file or printer if requested.
ESTIMATE and OPTIMIZE Input Control Variables

_acf Specify the number of lags for computing autocorrelation and partial
autocorrelation coefficients from the estimated regression residuals.
Useful for testing the ARMA error structure. Display and plot the
functions if _rplot > 0 (see below). In addition, standard errors of
coefficients and Box-Pierece and Ljung-Box portmanteau test statistics
are presented up to the number of lags specified in _acf. For example,
12 lags of autocorrelation and partial autocorrelation functions are
requested by setting:
_acf = 12;
_acf = 0; (default)
As an option for computing autocorrelation coefficients and the
associated standard errors using regression method, the second element
of the vector _acf may be set to a positive value, with the first element
indicating the number of lags requested. For example:
_acf = {12,1};
_acf2 Same as _acf except that the autocorrelation and partial autocorrelation
coefficients are computed from the squared standardized residuals.
Useful for testing the GARCH error structure.
_acf2 = 0; (default)
_ar Specify the order of an autoregressive (AR) error structure. If an

additional moving average process is desired for anautoregressive
moving average ARMA structure, use the variable _arma instead (see
below). Optional initial values of the autocorrelation coefficients may
be appended at the end of _ar for estimation. Providing the initial values
is useful for starting a search from non-zero values of autocorrelation
coefficients. For example:
_ar = 1;
_ar = {1, 0.5}; (with initial value of AR(1) parameter)
_ar = 0; (default)
_arma Specify the orders of an autoregressive moving average (ARMA) error

structure. It is a column vector consisting of at least two elements. The
first element denotes the order of autoregressive portion of the ARMA
process, while the second element is the order of moving average
284
APPENDIX A
portion. If only the autoregressive portion is given, it is exactly the AR

model (see _ar above). The model is estimated using maximum
likelihood method conditional to the initialization of pre-sample series,
which is the sample mean of the error series. Optional initial values of
the autoregressive and moving average coefficients may be appended at
the end of _arma for estimation. Giving the initial values is useful for
starting a search from non-zero values of ARMA coefficients. For
example:
_arma = {1, 0}; (this is identical to: _ar = 1;)
_arma = {0, 1};
_arma = {1, 1, 0.5, -0.5}; (initial values of ARMA(1,1) parameters)
_arma = {0, 0}; (default)
_b A column vector of initial parameter values for nonlinear model

estimation.
_begin Specify the starting observation number for estimation.

_begin = 1; (default)
_bjtest Bera-Jarque test for residual normality.

_bjtest = 0 (default): Skip the test;
_bjtest = 1: Perform the test.
_bgtest Breusch-Godfrey test for higher-order autocorrelation.

_bgtest = 0 (default): Skip the test;
_bgtest = p (>0): Perform the test for autocorrelation up to the p-th
order. The number p (>0) is the highest order tested.
_bptest Breusch-Pagan and White tests for heteroscedasticity.

_bptest = 0 (default): Skip the test;
_bptest = 1: Perform Breusch-Pagan and White tests for general
heteroscedasticity. For the Breusch-Pagan test, all explanatory variables
including constant term (i.e., X) are the RHS variables of the auxiliary
test regression. For the White test, all explanatory variables and their
squares and cross product including constant term are the RHS variables
of the auxilary test regression. Alternatively, _bptest can be set to a data
matrix (a subset of X or including other variables) in place of X for use
with the test for heteroscedasticity.
_const Specify a constant term for a regression model.

_const = 1 (default): Constant term is added in the regression;
_const = 0: No constant is added.
For a system model, this is a column vector of 0 (no constant) or 1 (with

constant) associated with each equation.
_conv Convergence criteria for nonlinear model estimation.

_conv = 0 (default): Convergence in function value and solution;
_conv = 1: Convergence in function value, solution, and zero gradient.
All convergence criteria are checked relative to the tolerance level _tol
(see below).
_corr Compute condition number of explanatory variables and correlation
285
matrix of dependent and explanatory variables, useful for

multicollinearity analysis.
_corr = 0 (default): Do not compute the statistics;
_corr = 1: Compute and show the statistics.
_deriv For a nonlinear optimization problem, if the first analytical derivative

function is defined and used, _deriv is set to the location (address) of the
function. If the second analytical derivative function is also given, its
location is vertically concanated with the first. If analytical derivative
functions are not used, the numerical derivatives are computed instead.
That is,
_deriv = 0; (default)
_dlags A scalar or a 2x1 column vector to specify the use of lagged dependent
variables. As a scalar, it is the order of the regular lagged dependent
variables in use. As a 2x1 column vector, a seasonal lagged dependent
variables model is identified with order _dlags[1] and seasonal span
_dlags[2] (require _dlags[2]>0). Normally, _dlags[2] = 4 for a model
with quarterly data series, while _dlags[2] = 12 for the monthly case.
_dlags[1] or the scalar _dlags is always the order number. For a pure
(regular or seasonal) lagged dependent variables model, set RHS
variable X = 0 in calling ESTIMATE procedure and specify the column
vector _dlags accordingly. For example:
_dlags = q; (or equivalently, _dlags = {q,1};)
_dlags = {q,s};
Where q is the order of autocorrelation and s is the seasonal span.
_dlags = 0; (default)
For a system model, _dlags is a gxg matrix with the value of its entry
indicating the number of lags for each endogenous variable (column) in
each equation (row). A zero ij-element of _dlags signifies that no lag is
used for the j-th variable in the i-th equation. Here, g is the number of
endogenous variables or equations.
_drop Drop the first few observations for model estimation. Depending on the
method of estimation, initial unusable observations may be dropped
automatically.
_drop =1: Drop the first observation or the first seasonal span of
observations for AR model estimation;
_drop =0 (default): Keep the first observation or the first seasonal span
of observations for AR model estimation with appropriate data
transformation.
_ebtest Engle-Bollerslev test for higher-order autoregressive conditional

heteroscedasticity (ARCH).
_ebtest = 0 (default): Skip the test;
_ebtest = q (>0) :Perform the test for ARCH structure up to the q-th
order. The number q (>0) is the highest order tested.
_end Specify the ending observation number for estimation.

_end = rows(y) (default).
_eq Specify the stochastic equation specification matrix for system model
286
APPENDIX A
estimation. This is a gx(gs+ks) matrix with elements -1, 0, and 1

arranged in accordance with the order of endogenous variables followed
by the predetermined variables. Note that g is the number of stochastic
equations, gs is the number of endogenous variables (gs>=g), while ks
is the number of predetermined variables. In the stochastic equation
specification matrix _eq, an element -1 indicates the LHS endogenous
variable. Only one -1 entry is allowed in each equation. An element 1
indicates the use of an endogenous and/or a predetermined variable on
the RHS of an equation. An element 0 indicates the corresponding
unused variable. If _eq is not specified, or _eq=0 by default, a restricted
seemingly unrelated system with common parameters across equations
is assumed. That is, g=gs and -1 in the gs diagonals and 1 in the next
gsxks predetermined variables portion of the matrix. If _eq is a scalar
with value 1 or 2, then an unrestricted seemingly unrelated system is
assumed. If _eq=1, the RHS variables are grouped by variables in the
order of equations. That is, there are (ks/gs) blocks of identity (gsxgs)
matrices concatenated horizontally, which occupy the gsxks
predetermined variables portion of the matrix. If _eq=2, the RHS
variables are grouped by equations in the order of variables. That is,
there are gs blocks of 1x(ks/gs) row vector of ones spanning over the
gsxks predetermined variables portion of the matrix along the diagonal.
Seemingly unrelated system with other forms of arrangements of RHS
variables must be specified with the matrix form of _eq. Normally
constant term is not needed in the equation specification, and it is
automatically included in each equation.
_garch Specify the orders of a generalized autoregressive conditonal

heteroscedasticity (GARCH) error structure. It is a column vector
consisting of at least two elements. The first element denotes the order
of autoregressive (variances) portion of GARCH process, while the
second element is the order of moving average (squared errors) portion.
The model is estimated using maximum likelihood method conditional
to the initialization of pre-sample series, which is the sample variance of
the error series. The optional initial value of GARCH coefficients may
be appended at the end of _garch for estimation. Be reminded that there
is always a constant for the GARCH process. The constant is the last
term of GARCH parameters. Giving the initial values is useful for
starting a search from non-zero values of GARCH coefficients. For
example:
_garch = {1, 0};
_garch = {0, 1};
_garch = {1, 1, 0.5, 0.5, 0.5}; (with initial values of GARCH(1,1))
_garch = {0, 0}; (default)
_garchx Specify additional variables included in the GARCH variance equation

(see _garch above). This may be a data matrix of multiple variables. The
variables must be defined with the same number of rows or observations
as that of the regression residuals.
_garcha Add asymmetry in the GARCH variance equation according to GJR

specification. The asymmetric response is estimated for the negative
errors with the same order specified for the moving average (squared
287
errors) portion of the GARCH variance equation (see _garch above).

_garcha = (default): No asymmetry;
_garcha = 1: Compute asymmetric response. Initial values of the
asymmetry parameters may be appended
_hacv Compute heteroscedasticity and autocorrelation-consistent variance-

covariance matrix and perform adjustment to standard error and t-ratio
of estimated coefficients accordingly. This may be a column vector up
to 3 elements.
_hacv = 0 (default): No adjustment;
_hacv = 1: Compute heteroscedasticity consistent variance-covariance
matrix;
_hacv = {0,p}: Compute p-th order autocorrelation-consistent variance-
covariance matrix with declining weights of autocovariances (Newey-
West estimators), p=1,2,...;
_hacv = {1,p}: Compute heteroscedasticity and p-th order
autocorrelation consistent variance-covariance matrix with declining
weights of autocovariances (Newey-West estimators), p=1,2,....;
_hacv = {1,p,1}: Compute heteroscedasticity and p-th order
autocorrelation-consistent variance-covariance matrix with the equal
weighted autocovariances, p=1,2,....; Therefore, _hacv = {0,p}is the
same as _hacv = {0,p,0} and _hacv = {1,p}is the same as _hacv =
{1,p,0}.
Note: If _hacv is used in conjunction with the instrumental variable

estimation (see, _ivar below) in setting the number of iterations (see,
_iter below) to achieve the convergence of estimated parameters, this is
essentially the Generalized Method of Moments (GMM). _hacv is
meaningful only when there is potential misspecification problem of
autoregressive and/or heteroscedastic error structure.
_hacv may be used in conjunction with the system model estimation. If

the method of 2SLS (see _method below) is requested with non-zero
_hacv, then two stage GMM estimation is performed. Similarly, if the
method of 3SLS is requested with non-zero _hacv, then three stage
GMM estimation is performed. However, the computed variance-
covariance matrix for the system may become non-positive definite due
to excess numerical rounding errors or the improper autocovariance
structures specified.
_id Specify the identity equation specification matrix for a system model.
Similar size and setup as _eq (see above) except that its entries can be
any value as required. If _id is not specified, or _id=0 by default, there
is no identity. Note: gs=rows(_eq|_id) to ensure the system
compatibility.
_iter Maximum number of iterations requested for iterative or nonlinear

model estimation.
_iter = 1; (default)
_ivar Instrumental variable estimation requested with instrumental variables

specified in matrix _ivar. If _dlags[1] > 0, _ivar may be given with a
positive scalar (i.e. _ivar = 1) and perform instrumental variable
288
APPENDIX A
estimation with the internal instrumental variables obtained from the

explanatory variables and their lags. If the matrix form of _ivar is
specified, the external instrumental variables are used for that
cols(_ivar) >= cols(X) + _const and rows(_ivar) >= rows(X). Constant
term is automatically included in _ivar.
_ivar = 0; (default)
For a system model, external instrumental variable estimation may be

requested with the instrumental variables specified in matrix _ivar. The
data matrix _ivar will be combined with all predetermined variables to
form the basis for instrumental variable estimation.
_ivar may be used together with _iter and _hacv (see above) to produce
the GMM estimation.
_jacob Controls the use of Jacobians in deriving the log-likelihood function

from the residuals. Since a correct log-likelihood function may include
non-vanishing log-Jacobian terms, the Jacobian transformation which is
a function of the estimated parameters should be defined. If the
analytical Jacobian function is used, _jacob is set to the location
(address) of the function. If you do not wish to write out the Jacobian
analytically, you may set _jacob = 1. Then the numerical Jacobian is
computed for each sample observation, which is usually a time
consuming process. In case of requesting numerical Jacobians, the first
column of the data matrix used to define the residuals must be the
dependent variable Y (recall that J(Z,ε) = |∂ε/∂Y| and Z = [Y,X]).
_jacob=0 (default): do not use Jacobians.
_ma Specify the order of a moving average (MA) error structure. If an

additional autoregressive process is desired for an autoregressive
moving average ARMA structure, use the variable _arma instead (see
above). Optional initial values of the moving average coefficients may
be appended at the end of _ma for estimation. Providing the initial
values is useful for starting a search from non-zero values of moving
average coefficients. For example:
_ma = 1;
_ma = {1, 0.5}; (with initial value of MA(1) parameter)
_ma = 0; (default)
_method Specify the estimation method for an AR model.

_method = 0 (default): Cochrane-Orcutt iterative LS method;
_method = {0,1}: Cochrane-Orcutt iterative ML method;
method = 1: Beach-MacKinnon iterative ML method (for _ar=1 only,
and _drop=0 is in effect);
_method = 2 or {2,0}: Hildreth-Lu grid search LS method (for _ar=1
only);
_method = {2,1}: Hildreth-Lu grid search ML method (for _ar=1 only).
Note: higher AR order (_ar>1) can only use _method = 0 or method =
{0,1}.
Specify the estimation method for a system model.

_method = 0 (default): Ordinary least squares (biased);
_method = 1: Limited information maximum likelihood;
289
_method = 2: Two-stage least squares;

_method = 3: Three-stage least squares;
_method = 4: Full information maximum likelihood.
Note: LIML and FIML are not true nonlinear maximum likelihood
estimation. Instead they are types of instrumental variables estimation.
Three variants of the FIML method are available:
_method = {4,0} (or 4): FIML instrumental variable method;
_method = {4,1}: FIML linearized method;
_method = {4,2}: FIML Newton method.
Specify the estimation method for a nonlinear model (including a linear

model with nonlinear error structure such as ARMA and GARCH).
_method = 0 (default): Steep-ascent or decent method for mathematical
optimization; Gauss-Newton method for nonlinear least squares
estimation; Berndt-Hall-Hall_Hausman (BHHH) method for maximum
likelihood estimation;
_method = 1: Quasi-Newton BFGS update method;
_method = 2: Quasi-Newton DFP update method;
_method = 3: Greenstadt method;
_method = 4: Newton-Raphson method;
_method = 5: Quadratic hill-climbing method;
_method = 6: Modified quadratic hill-climbing method.
_names Specify a vector of character names for variables (linear model) or

parameters (non-linear model) as appeared in a regression equation.
_nlopt Specify a nonlinear optimization problem.

_nlopt = 0 (default): Mathematical minimization of a scalar-valued
function or nonlinear least squares estimation based on a vector-valued
component error function; For the later case, the minimization is
performed on the sum of squares of the vector-valued error function;
_nlopt = 1: Mathematical maximization of a scalar-valued function or
maximum likelihood estimation based on a vector-valued component
error function. For the later case, the maximization is performed on the
sum of normal log-likelihood of the vector-valued component error
function;
_nlopt = 2 Maximum likelihood estimation based on a vector-valued
component log-likelihood function. The maximization is performed on
the sum of the vector-valued component log-likelihood function;
_nlopt = -1 Minimum distance estimation based on a vector-valued
component distance function (e.g., squares of the vector-valued
component error function). The minimization is performed on the sum
of the vector-valued component distance function.
_pdl Specify a polynomial distributed lag model if _pdl is defined as a

rows(_pdl)x3 matrix. Each row of _pdl consists three elements: {q p r}
where q = lags, p = orders, and r = endpoint restrictions: -1 (beginning),
1 (ending), 2 (both), and 0 (no restriction), for each RHS variable.
Requires rows(_pdl) = cols(X), and cols(_pdl) = 3:
_pdl = 0; (default)
_restart Number of times to restart estimation for iterative or nonlinear models

when function value does not improve. Maximum value of _restart is
290
APPENDIX A
10.
_restart = 0; (default)
_restr Perform restricted least squares estimation with the linear restrictions
defined in accordance with the form: Rb = q, or [R1 R0][b1 b0]’ = q,
where b1 is the vector of slope coefficients and R1 is the restriction
matrix corresponds to b1. Similarly, b0 is the intercept and R0
corresponds to b0. q is the vector of restricted values. Linear restrictions
involving intercept should be specified in conjunction with _const = 0.
If _restr is specified, then _restr = [R1 q]. Requires rows(_restr) =
number of restrictions, and cols(_restr) = cols(X).
_restr = 0; (default)
For a system model, restrictions in the matrix R are stacked horizontally

in accordance with the equations, while the vertical rows indicate the
number of restrictions imposed. Own or cross equation restrictions can
be coded easily. In general restrictions on the constant terms are not
required.
_rlist List regression residual series.

_rlist = 0 (default): Skip listing the series;
_rlist = 1: List observed, predicted, and least squares residual series;
_rlist = 2: In addition to listing least squares residual series, studentized
residuals and leverage information are provided. Useful for checking
influential observations and outliers.
_rplot Plot regression residual series.

_rplot = 0 (default): No plots;
_rplot = 1: Plot residuals only;
_rplot = 2: Plot both observed and predicted, and residuals.
Also for plotting autocorrelation and partial autocorrelation functions if
requested (see _acf above), a positive value of _rplot is needed.
For a nonlinear model, residual plot is meaningful only for the cases
specified with error component functions (i.e., _nlopt=0 or 1). In
defining the error component function, the dependent variable must be
the first column of the data matrix and it has not been transformed
within the definition of error component function.
_rstat Report regression residual statistics.

_rstat = 0 (default): Do not report the statistics;
_rstat =1: Report residual statistics, including DW, DH whenever
appropriate.
For a nonlinear model, residual statistics are meaningful only for the
cases specified with error component functions (i.e., _nlopt=0 or 1). In
defining the error component function, the dependent variable must be
the first column of the data matrix and it has not been transformed
within the definition of error component function.
_rtest Hypothesis testings based on regression residual series. This is a

4-element row vector:
_rtest[1] is the same as _bjtest (see above) for Bera-Jarque test for
291
residual normality;
_rtest[2] is the same as _bptest (see above) for Breusch-Pagan test for
heteroscedasticity;
_rtest[3] is the same as _bgtest (see above) for Breusch-Godfrey test
for higher-order autocorrelation;
_rtest[4] is the same as _ebtest (see above) for Engle-Bollerslev test
for higher-order autoregressive conditional heteroscedasticity.
_rtest = 0 (default): No tests are performed.
_step Specify step size of line search method for iterative or nonlinear model
estimation.
_step = 0 (default): Cut back (half) step size is used;
_step = 1: Quadratic step size is used.
_tol Set the convergence tolerance level for iterative or nonlinear model
estimation.
_tol = 0.001; (default)
_vcov Report the estimated variance-covariance matrix.

_vcov = 0 (default): Do not report the variance-covariance matrix;
_vcov = 1: Report variance-covariance matrix and correlation matrix of
the estimated coefficients.
For a nonlinear model,

_vcov = 1: Variance-covariance matrix is derived from the method
dependent approximated hessian (information matrix);
_vcov = 2: Variance-covariance matrix is derived from the estimated
hessian.
_vcov = 3: Robust variance-covariance matrix, if available, is derived
from the maximum likelihood estimation of component error or log-
likelihood function.
_weight Perform weighted least squares estimation with the weighting variable
defined in _weight. _weight must be a column vector and
rows(_weight) >= rows(X).
_weight = 0; (default)
FORECAST Input Control Variables

In addition to the estimate and optimize input variables which control the
model specification (e.g., _ar, _arma, _dlags, _pdl, etc.), the following are the
FORECAST input variables:
_b Regression parameter estimates for computing the forecasts. Depending

on the model specification, it may require the estimated coefficients of
ARMA or GARCH error structures available from the previously
estimated model.
_dynamic Dynamic forecasts for lagged dependent variables model.

_dynamic = 0 (default): Do not perform dynamic forecasts;
292
APPENDIX A
_dynamic = 1: Perform dynamic forecasts. Dynamic forecast uses

previous predicted lagged dependent variables.
_fbegin Start of forecast observation number.

_fbegin = _end +1; (default)
_fend End of forecast observation number.

_fend = rows(X); (default)
_fplot Plots predicted or forecast series.

_fplot = 0 (default): Do not plots the series;
_fplot = 1: Plot predicted or forecast series.
_fstat Computes ex-post forecast error statistics.

_fstat = 0 (default): Do not compute the statistics.
_fstat = 1: Compute and report the statistics.
_unlog Computes unlogged series of forecasts.

_unlog = 0 (default): Do not compute unlogged series of forecasts;
_unlog = 1: Compute unlogged series of forecasts, assuming the original
series has been log transformed.
_vb Compute and report the estimated variance-covariance matrix of the

basic model (not including the variance-covariance matrix of AR or
ARMA error structure if specified).
Note: forecast is not available for nonlinear models.
Output Control Variables

Output control variables are available after calling the procedure estimate,
optimize, or forecast. They may be used later in the program for further
analysis. Depending on the input variables specified, not all the output variables will
be available. Calling reset assigns all output variables to zero. Each call to
estimate, optimize, or forecast assigns new values to output variables.
ESTIMATE and OPTIMIZE Output Control Variables

__a Estimated coefficients of the autocorrelated error structure. Depending

on the model specification, it may include AR or ARMA, and GARCH
coefficients in that order.
__b Estimated regression coefficients (and possibly including the

coefficients for the autocorrelated error structure, that is __a, if the
model is so specified). For a simultaneous linear equations model, see
also __d and __pi.
__d The structural form parameter matrix of a simultaneous linear equations

system. See also __b and __pi.
293
__dh Estimated Durbin-H statistic.
__dw Estimated Durbin-Watson statistic.
__e Estimated regression residuals; for nonlinear scalar-valued function

optimization, this is the function value at the solution.
__g Gradient vector of nonlinear objective function evaluated at the

solution.
__h Hessian matrix of nonlinear objective function evaluated at the solution.
__hat Diagonal vector of Hat-matrix, X(X’X)-1X’, or leverage.
__ll Maximum log-likelihood function value.
__pi The reduced form parameter matrix of a simultaneous linear equations

system, useful for computing forecasts and multiplier analysis. See also
__b and __d.
__r2 R-square (goodness of fit of the regression).
__r2a Adjusted R-square.
__rss Residual or error sums-of-squares.
__t Estimated t-ratio for each of the regression parameters.
__v Estimated regression variance.
__vb Estimated variance-covariance matrix of the regression coefficients.
__ve Estimated variances of the regression residuals.
__x Data matrix of explanatory variables used in a linear model estimation.
__y Data vector of dependent variable used in a linear model estimation.
FORECAST Output Control Variables

__f Predicted or forecast series.
__mape Mean absolute percent of forecast errors.
__mse Mean sum squares of forecast errors.
__rmspe Root mean absolute percent of forecast errors.
__u1 Theil inequality coefficient (0<=__u1<=1).
294
APPENDIX A
__uc Covariance proportion of mean sum squares of errors.
__ue Disturbance proportion of mean sum squares of errors.
__um Bias proportion of mean sum squares of errors.
__ur Regression proportion of mean sum squares of errors.
__us Variance proportion of mean sum squares of errors.
__vf Variance of predicted or forecast series.
Note: forecast is not available for nonlinear models.
Miscellaneous
A few GAUSS built-in procedures have been modified, and they can be called
throughout the program using GPE package.
Procedure Description
gradp2 Computes the gradient vector or matrix (Jacobian) of a scalar- or vector-

valued function that has been defined in a procedure. The GPE
procedure gradp2 works the same as GAUSS built-in procedure gradp
except that gradp2 takes three input entries and has the format: g =
gradp2(&f,x,b). &f is a pointer to a vector-valued function f(x,b)
defined as a procedure, x is a data matrix, and b is a vector of points at
which to compute gradient. Both x and b are used to define the function
f. The output of gradp2 is the same as that of gradp, which is a vector
or matrix of first derivatives of function f evaluated at b. See GAUSS
Language References or on-line help for more details about GAUSS
built-in procedure gradp.
hessp2 Computes the matrix of second partial derivatives (Hessian matrix)

of a scalar-valued function that has been defined in a procedure. The
GPE procedure hessp2 works the same as GAUSS built-in procedure
hessp except that hessp2 takes three input entries and has the format: h
= hessp2(&f,x,b). &f is a pointer to a vector-valued function f(x,b)
defined as a procedure, x is a data matrix, and b is a vector of points at
which to compute Hessian. Both x and b are used to define the function
f. The output of hessp2 is the same as that of hessp, which is the matrix
of second derivatives of function f evaluated at b. See GAUSS
Language References or on-line help for more details about GAUSS
built-in procedure hessp.
295
Appendix B
GPE Application Modules
Each of the GPE application modules is given AS IS. The user is free to use and to
make changes as needed for different purposes. However, the usual disclaimer
applies. In particular, the following copyright statement must be presented as long as
all or part of the program code is used in your work:
© Copyright 2001-2003 by Kuan-Pin Lin and Applied Data Associates

All Rights Reserved.
THIS SOFTWARE PRODUCT IS PROPRIETARY SOURCE CODE OF APPLIED DATA

ASSOCIATES. THIS FILE HEADER MUST ACCOMPANY ALL FILES USING ANY PORTION,
IN WHOLE OR IN PART, OF THIS SOURCE CODE. THIS SOFTWARE PRODUCT IS DESIGNED
TO BE USED WITH GPE2 AND GAUSS. IF YOU USE THIS SOURCE CODE FOR RESEARCH
AND DEVELOPMENT, A PROPER REFERENCE IS REQUIRED. IF YOU WISH TO DISTRIBUTE
ANY PORTION OF THE PROPRIETARY SOURCE CODE, IN WHOLE OR IN PART, YOU MUST
FIRST OBTAIN WRITTEN PERMISSION FROM THE AUTHOR.
Application Module B-1: GMM.GPE

/*
** GMM.GPE: Nonlinear GMM Estimation
**
** ==> call estimate(&gmmqw,x);
** or
** ==> call estimate(&gmmq,x);
**
** A set of moment functions must be defined as a procedure with the
** name mf(x,b). The result is an nxl matrix of moments. n is the
** number of sample observations; l is the number of moment equations;
** x is the data matrix and b is the parameter vector.
**
** A global variable gmmw is used to define the weighting matrix for
** the GMM criterion function (to be minimized). gmmw is initially
** an identity matrix or 1 for the 1st GMM estimation; gmmw should be
** set to gmmv(x,b), the inverse of the variance-covariance matrix of
** moments functions, for the 2nd GMM estimation.
**
** gmmqw is the objective function with externally defined weighting
** matrix gmmw, the result is a consistent GMM estimation.
**
** gmmq is the objective function with internally estimated weighting
** matrix, the result is an efficient GMM estimation.
**
** GMM estimation is usually called in the following steps:
**
** ==>
** ==> _b=__b;
** ==> gmmw=gmmv(x,_b);
** ==>
** ==> _b=__b;
** ==> call estimate(&gmmq,x);
** ==>
** ==> call gmmout(x,__b);
*/
declare gmmw ?= 1;
/*
Sample average of moments
*/
proc gmmm(x,b);
local m,d;
m=meanc(mf(x,b));
retp(m);
endp;
/*
Covariance matrix of sample averages of moments
considering White-Newey-West autocovariances
depending on global _hacv
*/
proc gmmv(x,b);
local n,m,v,s,j;
n=rows(x);
m=mf(x,b)/n;
v=m'm; @ hetero. variances @
j=1;
do until j>_hacv[2]; @ autocovariances @
s=m'*missrv(lagn(m,j),0);
v=v+(1-j/(_hacv[2]+1))*(s+s');
j=j+1;
endo;
retp(v);
endp;
/*
GMM criterion function: depending on global gmmw
Weighted sum of squared sample averages of moments
*/
proc gmmqw(x,b);
local m;
m=gmmm(x,b);
retp(m'*gmmw*m);
endp;
/*
GMM criterion function: general
Weighted sum of squared sample averages of moments
*/
proc gmmq(x,b);
local m;
m=gmmm(x,b);
gmmw=invpd(gmmv(x,b));
retp(m'*gmmw*m);
endp;
proc (0) = gmmout(x,b);

local m,v,q,g,vb;
m=gmmm(x,b);
v=gmmv(x,b);
q=m'*invpd(v)*m;
g=gradp2(&gmmm,x,b);
vb=invpd(g'*gmmw*g)*g'*gmmw*v*gmmw'*g*invpd(g'*gmmw*g);
print;
print "GMM Estimation Result";
print "=====================";
print " Parameter Std. Error t-Ratio";;
print b~sqrt(diag(vb))~b./sqrt(diag(vb));
298
APPENDIX B
print;
print "Hansen Test Statistic of the Moment Restrictions";
print ftos(rows(m)-rows(b),"Chi-Sq(%*.*lf) = ",4,0);;
print q;
__vb=vb; @ using the GMM var-cov matrix @
endp;
Application Module B-2: JOHANSEN.GPE

/*
** JOHANSEN.GPE - Cointegration test procedure
** based on Johansen's VAR approach
**
** ==> call johansen(z,p,c);
** or
** ==> {lr,lrsum} = johansen(z,p,c);
**
** z is the data matrix for cointegration analysis, p is number of lags
** of VAR structure, c is the model indicator (0=no constant, 1=drift,
** 2=trend drift). outputs are two vectors of maximum eigenvalue and
** trace test statistics, lr and lrsum, respectively.
*/
proc (2) = johansen(z,p,c);

local m,n,j,z1,dz,y1,dy,y,x,u,v,suu,svv,suv,svu;
local r,lr,lrsum,msk,fmt,one,e;
m=cols(z); @ number of variables @

@ maximal lags in the test regression @
z1=lagn(z,1); @ lag of data matrix, at least p=1 @
dz=z-z1; @ construct difference data matrix @
j=1;
do until j>=p; @ use up to p-1 lags of differences @
dz=dz~lagn(z-z1,j);
j=j+1;
endo;
y=packr(z1~dz); @ combined data matrix @

n=rows(y); @ number of usable observations @
y1=y[.,1:m]; @ lag of y data matrix @
dy=y[.,m+1:2*m]; @ difference of y data matrix @
one=ones(n,1);
if p>1; @ VAR(p), p>1 @
x=y[.,2*m+1:cols(y)]; @ RHS x data matrix @
if c>0;
if c==1; @ with drift only @
e=one-x*(one/x);
endif; @ constant regression residuals @
if c==2; @ with trend drift @
x=x~one;
endif;
endif;
@ auxiliary regression residuals @
u=dy-x*(dy/x); @ (1) difference regression @
v=y1-x*(y1/x); @ (2) lag regression @
else; @ p==1, or VAR(1) @
if c>0;
u=dy-meanc(dy)';
v=y1-meanc(y1)';
if c==1; e=one; endif;
else;
u=dy; v=y1;
endif;
endif;
299
if c==1; v=e~v; endif;

suu=u'u/n; @ var-cov matrices @
svv=v'v/n;
suv=u'v/n;
svu=suv';
r=eig(invpd(suu)*suv*invpd(svv)*svu); @ compute eigenvalue @
r=rev(sortc(r,1)); @ sort eigvalues in increasing order @
lr=-n*ln(1-r); @ likelihood ratio test @
lrsum=rev(cumsumc(rev(lr))); @ trace test statistic @
msk={1 1 1 1}; @ print cointegration test results @

fmt={"*.*lf" 8 0,"*.*lf" 5 0,"#*.*lg" 12 5,"#*.*lg" 12 5};
print ftos(c+1,"\lCointegration Test (Model %-*.*lf):",1,0);
print "Cointegrating Eigv. Test Trace Test";
print " Rank DF Statistic Statistic";
call printfm
(real(seqa(0,1,m)~(m-seqa(0,1,m))~lr~lrsum),msk,fmt);
retp(lr,lrsum);
endp;
Application Module B-3: PANEL1.GPE

/*
** PANEL1.GPE - one-way panel data analysis
**
** ==> call panel1u(ys,xs,n,t);
** ys and xs are stacked of dependent and independent variables;
** one-way effects is computed for fixed and random models.
** ys and xs must be arranged in n blocks (cross sections) of t
** observations (time periods). it is used to study the individual
** (cross section) effects; to study the period (time periods)
** effects, re-arrange ys and xs then swap n and t.
**
** make sure to call reset, and define the variable names in _names.
** this version of procedure panel1 allows for unbalanced panels,
** provided that the missing values are coded in ys and xs.
*/
proc (0) = panel1(ys,xs,n,t);
local y,x,ym,xm,yms,xms,k;
local y1,x1,z,zm,ts,i,i1,sem2;
local rssr,rssur,dfr,dfur,bp,wf,v1,v,w,ws,h;
local b1,b2,vb1,vb2,a1,a2,va1,va2,xm1,xm2;
/* panel data processing */

k=cols(xs);
@ ys,xs: stacked data series @
@ use do loop to generate ym,xm,yms,xms,ts @

ts=t*ones(n,1); @ initialzation @
ym=zeros(n,1); xm=zeros(n,k);
i=1;
do until i>n;
i1=(i-1)*t;
y1=ys[i1+1:i1+t];
x1=xs[i1+1:i1+t,.];
z=packr(y1~x1);
ts[i]=rows(z);
zm=meanc(z);
ym[i]=zm[1];
xm[i,.]=zm[2:rows(zm)]';
i=i+1;
endo;
yms=vec(reshape(ym,t,n));
xms=reshape(reshape(xm',k*t,n)',n*t,k);
/* pooled (restricted) regression */
300
APPENDIX B
z=packr(ys~xs);
y=z[.,1];
x=z[.,2:k+1];
call estimate(y,x);
rssr=__rss;
dfr=__df;
@ testing for random one-way effects @

i=1; i1=0; sem2=0;
do until i>n;
sem2=sem2+sumc(__e[i1+1:i1+ts[i]])^2;
i1=i1+ts[i];
i=i+1;
endo;
bp=(sumc(ts)/2)*(1/(meanc(ts)-1))*(((sem2/sumc(__e^2))-1)^2);
/* between-group (mean) regression */

call estimate(ym,xm);
v1=__v.*ts; @ v1 is a nx1 vector @
/* within-group (mean deviation) regression */

@ unrestricted regression @
z=packr((ys-yms)~(xs-xms));
y=z[.,1];
x=z[.,2:k+1];
call estimate(y,x);
rssur=__rss;
dfur=__df-(n-1);
v=__v*(__df/dfur); @ df adjusted variance @
@ testing for fixed one-way effects @
wf=((rssr-rssur)/(dfr-dfur))/(rssur/dfur);
/* fixed effects model */

b1=__b;
vb1=__vb*(__df/dfur);
xm1=xm~(-ones(n,1));
a1=ym-xm1*b1; @ section/period difference @
va1=v/meanc(ts)+xm1*vb1*xm1';
/* random effects model */

w=1-sqrt(v./v1); @ if v>v1 then w=0+, w is a nx1 vector @
w=(w.<=0).*__macheps + (w.>0).*w; @ 0 < w <= 1 @
@ if w=1, it is fixed effects model @
@ if w=0, it is a pooled model (no effects) @
ws=vec(reshape(w,t,n));
z=packr((ys-ws.*yms)~(xs-ws.*xms));
y=z[.,1];
x=z[.,2:k+1];
call estimate(y,x);
b2=__b;
xm2=xm~(-1/w); @ w must be > 0 @
a2=w.*(ym-xm2*b2);
va2=(w^2).*(v/meanc(ts)+xm2*vb2*xm2');
h=(b1[1:k]-b2[1:k])'*inv(vb1[1:k,1:k]-vb2[1:k,1:k])*(b1[1:k]-b2[1:k]);
/* print output */
print;
print "Panel Data Model Estimation Procedure:";
print "(1) Pooled Regression";
print "(2) Between-Groups Regression";
print "(3) Fixed Effects (Within-Groups) Regression";
print "(4) Random Effects (Weighted Within-Groups) Regression";
print;
print "Wald F Test Statistic for Fixed Effects";
301
print ftos(dfr-dfur,"F(%*.*f,",4,0);;
print ftos(dfur,"%*.*f) = ",4,0);;
print wf;
print;
print "Breusch-Pagan LM Test Statistic for Random Effects";
print ftos(1,"Chi-Sq(%*.*f) = ",4,0);;
print bp;
print;
print "Hausman's Test for Fixed and Randon Effects";
print ftos(k,"Chi-Sq(%*.*f) = ",4,0);;
print abs(h);
print;
print "Within-Groups Estimates:";
print " Fixed S.E. Random S.E.";;
print b1~sqrt(diag(vb1))~b2~sqrt(diag(vb2));
print;
print "One-Way Effects:";
print "Section/Period Fixed S.E. Random S.E.";;
print seqa(1,1,n)~a1~sqrt(diag(va1))~a2~sqrt(diag(va2));
endp;
302
APPENDIX B
Application Module B-4: PANEL2.GPE

/*
** PANEL2.GPE - two-way panel data analysis
**
** ==> call panel2(ys,xs,n,t);
** two-way effects is computed for fixed and random models.
** ys and xs must be arranged in n blocks (cross sections) of t
** observations (time periods).
**
** make sure to call reset, and define the variable names in _names.
** this version of procedure panel2 allows for unbalanced panels,
** provided that GAUSS "missing value" (.) is used to identify the
** missing data in ys and xs.
*/
proc (0) = panel2(ys,xs,n,t);
local ymi,xmi,ymis,xmis,ystar,xstar,k;
local ymt,xmt,ymts,xmts,ymm,xmm,ts,ns,ms,missing,e;
local y,x,y1,x1,z,zm,i,j,i1,sem1,sem2; @ temp. var. @
local rssr,rssur,dfr,dfur,bp,wf,h,v1,v2,v3,v,w1,w2,w3,w1s,w2s;
local b1,b2,vb1,vb2,a1i,a1t,a2i,a2t,c1,c2;
/* panel data processing (complicated, do not change) */

k=cols(xs); @ not incl. constant term @
missing=miss(__miss,0);
ms=sumc(((ys~xs).==missing)'); @ index for missing obs. @
@ use do loop to generate ymi, xmi, ymt, xmt, ... @

@ data initialization @
ts=t*ones(n,1); ns=n*ones(t,1);
ymi=zeros(n,1); xmi=zeros(n,k);
ymt=zeros(t,1); xmt=zeros(t,k);
i=1;
do until i>n;
i1=(i-1)*t;
y1=ys[i1+1:i1+t];
x1=xs[i1+1:i1+t,.];
z=packr(y1~x1);
ts[i]=rows(z);
zm=meanc(z);
ymi[i]=zm[1];
xmi[i,.]=zm[2:rows(zm)]';
i=i+1;
endo;
@ swap (n,t) indexing for ys and xs, named y and x @

@ keep ys and xs intact @
y=vec(reshape(ys,n,t));
x1=vec(reshape(xs[.,1],n,t));
i=2;
do until i>k;
x1=x1~vec(reshape(xs[.,i],n,t));
i=i+1;
endo;
x=x1;
i=1;
do until i>t;
i1=(i-1)*n;
y1=y[i1+1:i1+n];
x1=x[i1+1:i1+n,.];
z=packr(y1~x1);
ns[i]=rows(z);
zm=meanc(z);
303
ymt[i]=zm[1];
xmt[i,.]=zm[2:rows(zm)]';
i=i+1;
endo;
@ ymm,xmm: overall means @

ymm=meanc(ymi)';
xmm=meanc(xmi)';
@ ymis,xmis: stacked section means @

ymis=vec(reshape(ymi,t,n));
xmis=reshape(reshape(xmi',k*t,n)',n*t,k);
@ ymts,xmts: stacked time period means @
ymts=reshape(ymt,n*t,1);
xmts=reshape(xmt,n*t,k);
/* pooled (restricted) regression */

z=packr(ys~xs);
y=z[.,1];
x=z[.,2:k+1];
call estimate(y,x);
rssr=__rss;
dfr=__df;
@ first put __e in the original order of data series @

@ with 0 for missing, and not count for sum or average @
e=__e;
i=1;
do until i>n*t;
if ms[i]>0;
e=e[1:i-1]|0|e[i+1:rows(e)];
endif;
i=i+1;
endo;
@ testing for two-way effects @

bp=(sumc(ts)/2)*(
(1/(meanc(ts)-1))*((sumc(sumc(reshape(e,n,t)')^2)/sumc(sumc(e^2))-1)^2)+
(1/(meanc(ns)-1))*((sumc(sumc(reshape(e,n,t))^2)/sumc(sumc(e^2))-1)^2));
/*
bp=(n*t/2)*(
(1/(t-1))*((sumc(sumc(reshape(__e,n,t)')^2)/sumc(sumc(__e^2))-1)^2)+
(1/(n-1))*((sumc(sumc(reshape(__e,n,t))^2)/sumc(sumc(__e^2))-1)^2));
*/
@ between-groups (cross sections) means regression @

call estimate(ymi,xmi);
v1=ts.*__v; @ nx1 vector @
@ between-groups (time periods) means regression @

call estimate(ymt,xmt);
v2=ns.*__v; @ tx1 vector @
/* fixed effects model */

@ within-groups (cross sections and time periods) regression @
ystar=ys-ymis-ymts+ymm;
xstar=xs-xmis-xmts+xmm;
z=packr(ystar~xstar);
y=z[.,1];
x=z[.,2:k+1];
call estimate(y,x);
rssur=__rss;
dfur=__df-(n-1)-(t-1); @ adust df @
v=__v*(__df/dfur);
@ testing for fixed two-way effects @
304
APPENDIX B
wf=((rssr-rssur)/(dfr-dfur))/(rssur/dfur);
b1=__b;
c1=ymm-xmm*b1[1:k]; @ overall effects, note: b1[k+1]=0 @
a1i=(ymi-ymm)-(xmi-xmm)*b1[1:k]; @ cross sections effects @
a1t=(ymt-ymm)-(xmt-xmm)*b1[1:k]; @ time periods effects @
/* random effects model (weights must be computed for each obs (nxt)) */
v3=meanc(v1)+meanc(v2)-v; @ v3 is a scalar @
w1=1-sqrt(v./v1); @ w1 is a nx1 vector @
w1=(w1.<=0).*__macheps + (w1.>0).*w1; @ 0 < w1 <= 1 @
w2=1-sqrt(v./v2); @ w2 is a tx1 vector @
w2=(w2.<=0).*__macheps + (w2.>0).*w2; @ 0 < w2 <= 1 @
w3=maxc((1-sqrt(v./v3))|__macheps);
w3=meanc(w1)+meanc(w2)-w3; @ w3 is a scalar @
w1s=vec(reshape(w1,t,n));
w2s=reshape(w2,n*t,1);
ystar=ys-w1s.*ymis-w2s.*ymts+w3.*ymm;
xstar=xs-w1s.*xmis-w2s.*xmts+w3.*xmm;
z=packr(ystar~xstar);
y=z[.,1];
x=z[.,2:k+1];
call estimate(y,x);
b2=__b;
c2=w3.*(ymm-xmm*b2[1:k])+b2[k+1]; @ overall effect @
a2i=(w1.*ymi-w3.*ymm)-(w1.*xmi-w3.*xmm)*b2[1:k]; @ individual effects @
a2t=(w2.*ymt-w3.*ymm)-(w2.*xmt-w3.*xmm)*b2[1:k]; @ period effects @
h=(b1[1:k]-b2[1:k])'*inv(vb1[1:k,1:k]-vb2[1:k,1:k])*(b1[1:k]-b2[1:k]);
/* print output */
print;
print "Panel Data Model Estimation Procedure:";
print "(1) Pooled Regression";
print "(2) Between-Groups (Cross Sections) Regression";
print "(3) Between-Groups (Time Periods) Regression";
print "(4) Fixed Effects (Within-Groups) Regression";
print "(5) Random Effects (Weighted Within-Groups) Regression";
print;
print "Wald F Test Statistic for Fixed Effects";
print ftos(dfr-dfur,"F(%*.*f,",4,0);;
print ftos(dfur,"%*.*f) = ",4,0);;
print wf;
print;
print "Breusch-Pagan LM Test Statistic for Random Effects";
print ftos(2,"Chi-Sq(%*.*f) = ",4,0);;
print bp;
print;
print "Hausman's Test for Fixed and Randon Effects";
print ftos(k,"Chi-Sq(%*.*f) = ",4,0);;
print abs(h);
print;
print "Within-Groups Estimates:";
print " Fixed S.E. Random S.E.";;
print b1~sqrt(diag(vb1))~b2~sqrt(diag(vb2));
print;
print "Two-Way Effects:";
print " Fixed Random";
print " Overall " c1~c2;
print;
print "Cross Sections Effects:";
print " Sections Fixed Random";;
print seqa(1,1,n)~a1i~a2i;
print;
305
print "Time Periods Effects:";

print " Periods Fixed Random";;
print seqa(1,1,t)~a1t~a2t;
endp;
306
APPENDIX B
Application Module B-5: RANDOM1.GPE

/*
** RANDOM1.GPE - random coefficients model for panel data analysis
**
** ==> call random1(ys,xs,n,t);
** random coefficients model based on Hildreth, Houck, and Swamy is
** estimated. ys and xs must be arranged in n blocks (cross sections)
** of t observations (time periods). in addition to random coefficients
** estimates, the individual parameter prediction for each cross
** section is reported. it is used to study the individual
** (cross section) effects; to study the period (time periods)
** effects, re-arrange ys and xs then swap n and t. it can be used
** for unbalanced panel data analysis, provided that GAUSS
** "missing value" (.) is used to identify the missing data.
**
** before calling the procedure to estimate a random coefficients
** model, make sure to call reset and define the variable names in
** _names. for a large number of individuals (cross sections),
** _print = -1 should be set to surpress printing of the regression
** results.
**
** note: the following GPE variables should not be used with random1:
** _begin, _end, _dlags, _restr, _const
** _weight, _hacv, _ivar, _arma, _garch
*/
proc (0) = random1(ys,xs,n,t);
local b,vb,gb,bstar,vbstar,bw,vbw,swamy,i,k,w,a;
local i1,ik,y1,x1,z1,ginv,vinv,sumgv,sumgvw; @ temp. var. @
/* invidual OLS regressions */

k=cols(xs)+1; @ constant added @
b=zeros(k,n);
vb=zeros(n*k,k);
bw=0;
vbw=0;
i=1;
do until i>n;
i1=(i-1)*t;
ik=(i-1)*k;
y1=ys[i1+1:i1+t];
x1=xs[i1+1:i1+t,.];
@ in case of missing values @
z1=packr(y1~x1);
y1=z1[.,1];
x1=z1[.,2:cols(z1)];
@ ts[i]=rows(z1); @
call estimate(y1,x1);
b[.,i]=__b;
vb[ik+1:ik+k,.]=__vb;
bw=bw+invpd(__vb)*__b;
vbw=vbw+invpd(__vb);
i=i+1;
endo;
/* Swamy Test for Random Coefficients */

bw=invpd(vbw)*bw;
swamy=0;
i=1;
do until i>n;
ik=(i-1)*k;
swamy=swamy+(b[.,i]-bw)'*invpd(vb[ik+1:ik+k,.])*(b[.,i]-bw);
i=i+1;
endo;
307
gb=vcx(b');
sumgv=0;
i=1;
do until i>n;
ik=(i-1)*k;
sumgv=sumgv+invpd(gb+vb[ik+1:ik+k,.]);
i=i+1;
endo;
/* random coefficients estimates */

vbstar=invpd(sumgv);
bstar=0;
i=1;
do until i>n;
ik=(i-1)*k;
w=vbstar*invpd(gb+vb[ik+1:ik+k,.]);
bstar=bstar+w*b[.,i];
i=i+1;
endo;
/* individual parameter vectors prediction */

sumgvw=0;
i=1;
do until i>n;
ik=(i-1)*k;
sumgvw=sumgvw+(vbstar*invpd(gb+vb[ik+1:ik+k,.])'*vbstar');
i=i+1;
endo;
ginv=invpd(gb);
bw=zeros(k,n);
vbw=zeros(k,n); @ variances only @
i=1;
do until i>n;
ik=(i-1)*k;
vinv=invpd(vb[ik+1:ik+k,.]);
a=invpd(vinv+ginv)*ginv;
a=a~(eye(k)-a);
bw[.,i]=a*(bstar|b[.,i]);
sumgv=(sumgvw~vbstar)|(vbstar'~(gb+vb[ik+1:ik+k,.]));
vbw[.,i]=diag(a*sumgv*a');
i=i+1;
endo;
/* output report */
print;
print "Randon Coefficients Model Estimation:";
print "(1) Individual Equation OLS Regression";
print "(2) Generalized Least Squares Regression";
print "(3) Individual Equation Parameters Prediction";
print;
print "Swamy Test Statistic for Random Coefficients";
print ftos(k*(n-1),"Chi-Sq(%*.*f) = ",4,0);;
print swamy;
print;
print "Random Coefficients Estimates:";
print " No. Parameter S.E.";;
print seqa(1,1,k)~bstar~sqrt(diag(vbstar));
print;
i=1;
do until i>k;
print "Individual Parameter Prediction:" i;
print "Section/Period Parameter S.E.";;
308
APPENDIX B
print seqa(1,1,n)~bw[i,.]'~sqrt(vbw[i,.]');
print;
i=i+1;
endo;
endp;
309
Application Module B-6: SYSTEM1.GPE

/*
** SYSTEM1.GPE - systems of regression equations
**
** ==> call system1(ys,xs,n,t);
** a system of regression equations is estimated using single equation
** least squares. ys and xs are stacked of dependent and independent
** variables; ys and xs must be arranged in n blocks (cross sections)
** of t observations (time periods). the model is not limited for
** panel data analysis. seemingly unrelated regression equations with
** varity of restrictions may be estimated. however, missing values
** (or unbalanced panel) can not be used. the alternative approach is
** to estimate the model as a simultaneous linear equations system.
**
** before calling the procedure to estimate a system of regression
** equations, make sure to call reset and define the variable names
** in _names. if the constant term is suppressed for estimation (e.g.,
** due to explicit data construction), _const should be set to 0.
** a more efficient maximum likelihood estimator can be obtained by
** setting the number of iterations (_iter) and convergent tolerlence
** (_tol).
**
** note: the following GPE variables should not be used with system1:
** _begin, _end, _dlags, _restr, _weight, _hacv, _ivar, _arma, _garch
*/
proc (0) = system1(ys,xs,n,t);

local x,e,b,vb,v,vinv,b0,vb0,ll,ll0,fail;
local i,j,k,i1,j1,x1,y1,z1,it,xx,xy;
/* pooled regression */
call estimate(ys,xs);
b=__b;
vb=__vb;
ll=__ll;
e=reshape(__e,n,t)';
v=e'e/t; @ not same as v=vcx(e); @
@ consider both cross section hetero. and serial corr. @
/*
@ consider cross section heteroscedasticity only @
v=diagrv(eye(n),diag(v));
*/
@ take care of no constant, if specified @

if _const==0;
x=xs;
else;
x=xs~ones(n*t,1);
endif;
k=cols(x);
print;
print "System of Regression Equations:";
print ftos(n,"Number of Equations = %-*.*lf",12,0);
print ftos(k,"Number of Parameters = %-*.*lf",12,0);
print ftos(t,"Number of Observations = %-*.*lf",12,0);
print;
it=1; fail=0;
do until it>_iter;
b0=b; vb0=vb; ll0=ll;
/*
@ memory extensive computation @
vinv=invpd(v.*.eye(t));
vb=invpd(x'*vinv*x);
310
APPENDIX B
b=vb*(x'*vinv*ys);
*/
@ memory saving computation @
@ less efficient with double loops @
vinv=invpd(v);
xx=0; xy=0;
i=1;
do until i>n;
i1=(i-1)*t;
j=1;
do until j>n;
j1=(j-1)*t;
xx=xx+vinv[i,j]*x[i1+1:i1+t,.]'*x[j1+1:j1+t,.];
xy=xy+vinv[i,j]*x[i1+1:i1+t,.]'*ys[j1+1:j1+t];
j=j+1;
endo;
i=i+1;
endo;
vb=invpd(xx);
b=vb*xy;
e=reshape(ys-x*b,n,t)';
v=e'e/t;
/*
v=diagrv(eye(n),diag(v));
*/
ll=-0.5*t*(n*(1+ln(2*pi))+ln(det(v)));
format /lds 4,0; print "Iteration = " it;;

format /los 12,5; print "Log-Likelihood = " ll;
print "Parameters = " b';
if ll0>ll;
b=b0; vb=vb0; ll=ll0; fail=1;
endif;
if abs(b-b0)<_tol; break; endif;
it=it+1;
endo;
print;
print "Log-Likelihood Function Value = " ll;
if fail==1;
print "WARNING: Log-Likelihood Fails to Improve!";
endif;
if it>_iter;
print "WARNING: Iteration Limit Exceeded!";
endif;
print;
print " Parameter S.E.";;
format /ros; print b~sqrt(diag(vb));
print;
endp;
311
Appendix C
Statistical Tables
Statistical tables for normal distribution, t distribution, Chi-squared distribution, and
F distribution are available from most statistics references. Durbin-Watson bounds
test statistics are readily available in econometric textbooks. In this appendix, we list
only the not-so-popular statistical tables for testing unit roots and cointegration as
discussed in Chapter XVI.
Table C-1. Critical Values for the Dickey-Fuller Unit Root Test
Based on t-Statistic
Model
Model I: ∆Xt = (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
Model II: ∆Xt = α + (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
Model III: ∆Xt = α + β t + (ρ-1) Xt-1 + Σi=1,2,…ρi ∆Xt-i + εt
Test Statistic
τρ t-statistic (non-symmetric distribution, testing ρ = 1)
τα t-statistic (symmetric distribution, testing α = 0 given ρ = 1)
τβ t-statistic (symmetric distribution, testing β = 0 given ρ = 1)
Source
Fuller (1976, p. 373); Dickey and Fuller (1981).
Model Statistic N 1% 2.5% 5% 10%

I τρ 25 -2.66 -2.26 -1.95 -1.60
50 -2.62 -2.25 -1.95 -1.61
100 -2.60 -2.24 -1.95 -1.61
250 -2.58 -2.23 -1.95 -1.61
500 -2.58 -2.23 -1.95 -1.61
>500 -2.58 -2.23 -1.95 -1.61
II τρ 25 -3.75 -3.33 -3.00 -2.62
50 -3.58 -3.22 -2.93 -2.60
100 -3.51 -3.17 -2.89 -2.58
250 -3.46 -3.14 -2.88 -2.57
500 -3.44 -3.13 -2.87 -2.57
>500 -3.43 -3.12 -2.86 -2.57
III τρ 25 -4.38 -3.95 -3.60 -3.24
50 -4.15 -3.80 -3.50 -3.18
100 -4.04 -3.73 -3.45 -3.15
250 -3.99 -3.69 -3.43 -3.13
500 -3.98 -3.68 -3.42 -3.13
>500 -3.96 -3.66 -3.41 -3.12
II τα 25 3.41 2.97 2.61 2.20
50 3.28 2,89 2.56 2.18
100 3.22 2.86 2.54 2.17
250 3.19 2.84 2.53 2.16
500 3.18 2.83 2.52 2.16
>500 3.18 2.83 2.52 2.16
III τα 25 4.05 3.59 3.20 2.77
50 3.87 3.47 3.14 2.75
100 3.78 3.42 3.11 2.73
250 3.74 3.39 3.09 2.73
500 3.72 3.38 3.08 2.72
>500 3.71 3.38 3.08 2.72
III τβ 25 3.74 3.25 2.85 2.39
50 3.60 3.18 2.81 2.38
100 3.53 3.14 2.79 2.38
250 3.49 3.12 2.79 2.38
500 3.48 3.11 2.78 2.38
>500 3.46 3.11 2.78 2.38
314
APPENDIX C
Table C-2. Critical Values for the Dickey-Fuller Unit Root Test
Based on F-Statistic
Model
Model II: ∆Xt = α + (ρ-1) Xt-1 + Σi=1,2,… ρi ∆Xt-i + εt
Model III: ∆Xt = α + β t + (ρ-1) Xt-1 + Σi=1,2,… ρi ∆Xt-i + εt
Test Statistic
φ1 F-statistic (testing α = 0 and ρ = 1 on Model II)
φ2 F-statistic (testing α = 0, β = 0, and ρ = 1 on Model III)
φ3 F-statistic (testing β = 0 and ρ = 1 on Model III)
Source
Dickey and Fuller (1981).
Model Statistic N 1% 2.5% 5% 10%

II φ1 25 7.88 6.30 5.18 4.12
50 7.06 5.80 4.86 3.94
100 6.70 5.57 4.71 3.86
250 6.52 5.45 4.63 3.81
500 6.47 5.41 4.61 3.79
>500 6.43 5.38 4.59 3.78
III φ2 25 8.21 6.75 5.68 4.67
50 7.02 5.94 5.13 4.31
100 6.50 5.59 4.88 4.16
250 6.22 5.40 4.75 4.07
500 6.15 5.35 4.71 4.05
>500 6.09 5.31 4.68 4.03
III φ3 25 10.61 8.65 7.24 5.91
50 9.31 7.81 6.73 5.61
100 8.73 7.44 6.49 5.47
250 8.43 7.25 6.34 5.39
500 8.34 7.20 6.30 5.36
>500 8.27 7.16 6.25 5.34
315
Table C-3. Critical Values for the Dickey-Fuller Cointegration

t-Statistic τρ Applied on Regression Residuals
Model
Yt = α + Xt β + εt
∆εt = (ρ-1) εt-1 + Σi=1,2,…ρi∆εt-i + ut
K = Numbers of variables in the cointegration tests, i.e., [Yt, Xt].
t = 1,2,…, N (500).
Model 2: E(Yt) = E(Xt) = 0 (both X and Y have no drift).

Model 2a: E(Xt) ≠ 0 (at least one variable in X has drift).
Model 3: E(Yt) ≠ 0 but E(Xt) = 0 (only Y has drift).
Test Statistic
τρ t-statistic (testing ρ = 1)
Source
Phillips and Ouliaris (1990).
Model K 1% 2.5% 5% 10%

2 2 -3.96 -3.64 -3.37 -3.07
3 -4.31 -4.02 -3.77 -3.45
4 -4.73 -4.37 -4.11 -3.83
5 -5.07 -4.71 -4.45 -4.16
6 -5.28 -4.98 -4.71 -4.43
2a 2 -3.98 -3.68 -3.42 -3.13
3 -4.36 -4.07 -3.80 -3.52
4 -4.65 -4.39 -4.16 -3.84
5 -5.04 -4.77 -4.49 -4.20
6 -5.36 -5.02 -4.74 -4.46
7 -5.58 -5.31 -5.03 -4.73
3 2 -4.36 -4.07 -3.80 -3.52
3 -4.65 -4.39 -4.16 -3.84
4 -5.04 -4.77 -4.49 -4.20
5 -5.36 -5.02 -4.74 -4.46
6 -5.58 -5.31 -5.03 -4.73
Note: For the case of two variables in Model 2a, X is trended but Y is not. It is
asymptotically equivalent to ADF unit root test for Model III (see Table C-1, τρ for
N=500). If only Y has drift (Model 3), the cointegration equation can be expressed as
Yt = α + γ t + Xt β + εt. Therefore, the same critical values of Model 2a apply to
Model 3 for one extra variable t (but do not count for K).
316
APPENDIX C
Table C-4. Critical Values for Unit Root and Cointegration Tests
Based on Response Surface Estimates
Critical values for unit root and cointegration tests can be computed from the
equation:
CV(K, Model, N, e) = b + b1 (1/N) + b2 (1/N)2
Notation
Model: 1=no constant; 2=no trend; 3=with trend;
K: Number of variables in cointegration tests (K=1 for unit root test);
N: Number of observations or sample size;
e: Level of significance, 0.01, 0.05, 0.1.
Source
MacKinnon (1991).
317
K Model e b b1 b2
1 1 0.01 -2.5658 -1.960 -10.04
1 1 0.05 -1.9393 -0.398 0.00
1 1 0.10 -1.6156 -0.181 0.00
1 2 0.01 -3.4335 -5.999 -29.25
1 2 0.05 -2.8621 -2.738 -8.36
1 2 0.10 -2.5671 -1.438 -4.48
1 3 0.01 -3.9638 -8.353 -47.44
1 3 0.05 -3.4126 -4.039 -17.83
1 3 0.10 -3.1279 -2.418 -7.58
2 2 0.01 -3.9001 -10.534 -30.03
2 2 0.05 -3.3377 -5.967 -8.98
2 2 0.10 -3.0462 -4.069 -5.73
2 3 0.01 -4.3266 -15.531 -34.03
2 3 0.05 -3.7809 -9.421 -15.06
2 3 0.10 -3.4959 -7.203 -4.01
3 2 0.01 -4.2981 -13.790 -46.37
3 2 0.05 -3.7429 -8.352 -13.41
3 2 0.10 -3.4518 -6.241 -2.79
3 3 0.01 -4.6676 -18.492 -49.35
3 3 0.05 -4.1193 -12.024 -13.13
3 3 0.10 -3.8344 -9.188 -4.85
4 2 0.01 -4.6493 -17.188 -59.20
4 2 0.05 -4.1000 -10.745 -21.57
4 2 0.10 -3.8110 -8.317 -5.19
4 3 0.01 -4.9695 -22.504 -50.22
4 3 0.05 -4.4294 -14.501 -19.54
4 3 0.10 -4.1474 -11.165 -9.88
5 2 0.01 -4.9587 -22.140 -37.29
5 2 0.05 -4.4185 -13.461 -21.16
5 2 0.10 -4.1327 -10.638 -5.48
5 3 0.01 -5.2497 -26.606 -49.56
5 3 0.05 -4.7154 -17.432 -16.50
5 3 0.10 -4.4345 -13.654 -5.77
6 2 0.01 -5.2400 -26.278 -41.65
6 2 0.05 -4.7048 -17.120 -11.17
6 2 0.10 -4.4242 -13.347 0.00
6 3 0.01 -5.5127 -30.735 -52.50
6 3 0.05 -4.9767 -20.883 -9.05
6 3 0.10 -4.6999 -16.445 0.00
318
APPENDIX C
Table C-5: Critical Values for the Johansen’s Cointegration Likelihood

Ratio Test Statistics
Notations
VAR Model: 1=no constant; 2=drift; 3=trend drift
N: Sample Size, 400
M: Number of Variables
r: Number of Cointegrating Vectors or Rank
Degree of Freedom = M-r
Source
Johansen (1988), Johansen and Juselius (1990), and Osterwald-Lenum (1992).
Model M-r 1% 2.5% 5% 10% 20% 50%

λmax 1 1 6.51 4.93 3.84 2.86 1.82 0.58
1 2 15.69 13.27 11.44 9.52 7.58 4.83
1 3 22.99 20.02 17.89 15.59 13.31 9.71
1 4 28.82 26.14 23.80 21.58 18.97 14.94
1 5 35.17 32.51 30.04 27.62 24.83 20.16
2 1 11.576 9.658 8.083 6.691 4.905 2.415
2 2 18.782 16.403 14.595 12.783 10.666 7.474
2 3 26.154 23.362 21.279 18.959 16.521 12.707
2 4 32.616 29.599 27.341 24.917 22.341 17.875
2 5 38.858 35.700 33.262 30.818 27.953 23.132
3 1 6.936 5.332 3.962 2.816 1.699 0.447
3 2 17.936 15.810 14.036 12.099 10.125 6.852
3 3 25.521 23.002 20.778 18.697 16.324 12.381
3 4 31.943 29.335 27.169 24.712 22.113 17.719
3 5 38.341 35.546 33.178 30.774 27.899 23.211
λtrace 1 1 6.51 4.93 3.84 2.86 1.82 0.58
1 2 16.31 14.43 12.53 10.47 8.45 5.42
1 3 29.75 26.64 24.31 21.63 18.83 14.30
1 4 45.58 42.30 39.89 36.58 33.16 27.10
1 5 66.52 62.91 59.46 55.44 51.13 43.79
2 1 11.576 9.658 8.083 6.691 4.905 2.415
2 2 21.962 19.611 17.844 15.583 13.038 9.355
2 3 37.291 34.062 31.256 28.436 25.445 20.188
2 4 55.551 51.801 48.419 45.248 41.623 34.873
2 5 77.911 73.031 69.977 65.956 61.566 53.373
3 1 6.936 5.332 3.962 2.816 1.699 0.447
3 2 19.310 17.299 15.197 13.338 11.164 7.638
3 3 35.397 32.313 29.509 26.791 23.868 18.759
3 4 53.792 50.424 47.181 43.964 40.250 33.672
3 5 76.955 72.140 68.905 65.063 60.215 52.588
319
References
GAUSS for Windows User Guide (Version 5.0), 2002, Aptech Systems, Inc.
GAUSS Language Reference (Version 5.0), 2002, Aptech Systems, Inc.
E. Berndt and D. Wood, 1975, “Technology, Prices, and the Derived Demand for
Energy,” Review of Economics and Statistics, 259-268.
T. Bollerslev, 1986, “Generalized Autoregressive Conditional Heteroskedasticity,”
Journal of Econometrics, 31, 307-327.
T. Bollerslev and E. Ghysels, 1996, “Periodic Autoregressive Conditional
Heteroscedasticity,” American Statistical Association Journal of Business and
Economic Statistics, 14, 139-151.
J. Boot and G. deWitt, 1960, “Investment Demand: An Empirical Contribution to the
Aggregation Problem,” International Economic Review, 1, 3-30.
J. Y. Campbell, A. W. Lo, and A. C. Mackinlay, 1977, The Econometrics of
Financial Markets, Princeton University Press.
R. Davidson and J. G. MacKinnon, 1973, Estimation and Inference in Econometrics,
Oxford University Press.
D. A. Dickey and W. A. Fuller, 1981, “Likelihood Ratio Statistics for Autoregressive
Time Series with a Unit Root,” Econometrica, 49, 1057-1072.
P. J. Dhrymes, 1970, Econometrics, Harper & Row.
R. Engle and C. Granger, 1987, “Co-integration and Error Correction:
Representation, Estimation and Testing,” Econometrica, 35, 251-276.
R. F. Engle, 1982, “"Autoregressive Conditional Heteroscedasticity with Estimates
of the Variance of United Kingdom Inflation,” Econometrica, 50, 987-1006.
R. F. Engle, D. M. Lilien, and R. P. Robins, 1987, “Estimating Time-Varying Risk
Premia in the Term Structure: the ARCH-M Model,” Econometrica 55, 391-407.
R. Fair, 1978, “A Theory of Extramarital Affairs,” Journal of Political Economy, 86,
45-61.
W. A. Fuller, 1976, Introduction to Statistical Time Series, John Wiley.
S. Goldfeld and R. Quandt, 1972, Nonlinear Methods in Econometrics, Chapter 1:
Numerical Optimization, North-Holland, 1-38.
W. H. Greene, 2002, Econometric Analysis, 5th ed., Prentice Hall.
W. H. Greene, 1999, Econometric Analysis, 4th ed., Prentice Hall.
W. H. Greene, 1997, Econometric Analysis, 3rd ed., Prentice Hall.
Y. Grunfeld and Z. Griliches, 1960, “Is Aggregation Necessarily Bad?” Review of
Economics and Statistics, 42, 1-13.
A. Hall, 1993, “Some Aspects of Generalized Method of Moments Estimation,”
Handbook of Statistics, Vol. 11, ed. by G. S. Maddala, C. R. Rao, and H. D.
Vinod, Elsevier Science Publishers, North-Holland, 393-417.
R. Hall, 1978, “Stochastic Implications of the Life Cycle-Permanent Income
Hypothesis: Theory and Evidence,” Journal of Political Economy 86, 971-987.
J. D. Hamilton, 1994, Time Series Analysis, Princeton University Press.
L. P. Hansen and K. J. Singleton, 1982, “Generalized Instrumental Variables
Estimation of Nonlinear Rational Expectations Models,” Econometrica 50, 1269-
1286.
J. Hausman, 1975, “An Instrumental Variable Approach to Full-Information
Estimators for Linear and Certain Nonlinear Models,” Econometrica, 727-738.
F. Hayashi, 2000, Econometrics, Princeton University Press.
S. Johansen, 1988, “Statistical Analysis of Cointegration Vectors,” Journal of
Economic Dynamics and Control, 12, 231-254.
S. Johansen and K. Juselius, 1990, “Maximum Likelihood Estimation and Inference
on Cointegration with Applications to the Demand for Money,” Oxford Bulletin
of Economics and Statistics, 52, 169-210.
G. G. Judge, R. C. Hill, W. E. Griffiths, H. Lutkempohl, and T.-C. Lee, 1988,
Introduction to the Theory and Practice of Econometrics, 2nd ed., John Wiley
and Sons.
G. G. Judge, R. C. Hill, W. E. Griffiths, and T.-C. Lee, 1985, Theory and Practice of
Econometrics, 2nd ed., John Wiley and Sons.
L. Klein, 1950, Economic Fluctuations in the United States: 1921-1941, John Wiley
and Sons.
J. G. MacKinnon, 1991, “Critical Values for Cointegration Tests,” in Long-Run
Economic Relationships: Readings in Cointegration, ed. by R. F. Engle and G.
W. Granger, Oxford University Press, 267-276.
T. C. Mills, 1999, The Econometric Modeling of Financial Time Series, 2nd ed.,
Cambridge University Press.
R. C. Mittelhammer, G. G. Judge, and D. J. Miller, 2000, Econometric Foundations,
Cambridge University Press.
D. B. Nelson, 1991, “Conditional Heteroscedasticity in Asset Returns, A New
Approach,” Econometrica, 59, 347-370.
D. B. Nelson and C. Q. Cao, 1992, “Inequality Constraints in the Univariate GARCH
Model,” Journal of Business and Economic Statistics, 10, 229-235.
M. Ogaki, 1993, “Generalized Method of Moments: Econometric Applications,”
Handbook of Statistics, Vol. 11, ed. by G. S. Maddala, C. R. Rao, and H. D.
Vinod, Elsevier Science Publishers, North-Holland, 455-488.
M. Osterwald-Lenum, 1992, “A Note with Quantiles of the Asymptotic Distribution
of the Maximum Likelihood Cointegration Rank Test Statistics,” Oxford Bulletin
of Economics and Statistics, 54, 461-471.
P. C. B. Phillips and S. Ouliaris, 1990, “Asymptotic Properties of Residual Based
Tests for Cointegration,” Econometrica, 58, 165-193.
R. E. Quandt, 1983, “Computational Problem and Methods,” Handbook of
Econometrics, Vol. I, ed. by Z. Griliches and M. D. Intriligator, Chapter 12, 699-
764, North-Holland.
L. Spector and M. Mazzeo, 1980, “Probit Analysis and Economic Education,”
Journal of Economic Education, 11, 37-44.
J. M. Wooldridge, 2002, Introductory Econometrics: A Modern Approach, 2nd ed.,
Thomson, South-Western.
Index
Coefficient of relative risk aversion, 188
Analysis of variance, AOV, 47 Cointegrating rank, 224
Analytical derivatives, 86 Cointegrating vector, 223
Analytical Jacobian, 108 Cointegration test, 216
ARCH in mean, ARCH-M, 241 Engle-Granger test, 224
ARMA analysis for regression residuals, Johansen test, 227
238 maximal eigenvalue test statistic, 229
ARMAX regression model, 238 trace test statistic, 229
Augmented Dickey-Fuller (ADF) test, 215 Composite economic leading indicator,
Autocorrelation, 143 271
Beach-MacKinnon iterative maximum Concentrated log-likelihood function, 99,
likelihood method, 151 106, 210
Box-Pierce Q test, 147, 155 Constant returns to scale, CRS, 53
Breusch-Godfrey LM test, 147, 155 Correlation matrix, 50, 75
Durbin-H test, 168
Durbin-Watson bounds test, 147 Data generating process, DGP, 215
first order, AR(1), 147 Dickey-Fuller (DF) test, 3, 215
higher order, 154 Distributed lag models, 167
Hildreth-Lu grid search method, 150 Almon lag, 167
Ljung-Box Q test, 147, 155 autoregressive distributed lag, ARDL,
Prais-Winsten modified Cochrane- 167, 176
Orcutt method, 149 geometric lag, 167
Autocorrelation coefficient, 147, 234 Koyck lag, 167
Autocorrelation function, ACF, 148, 233, lagged dependent variable, 167
241 polynomial lag, 167, 173
Autocorrelation-consistent covariance Dummy variable, 65, 115
matrix, 144 additive, 65, 70
Autoregressive conditional multiplicative, 65, 70
heteroscedasticity, ARCH, 233 Dummy variable trap, 66, 69
Box-Pierce Q test, 241
Engle-Bollerslev LM test, 242 Economies of scale, 254
Ljung-Box Q test, 241 Elasticity, 54, 106
Autoregressive moving average, ARMA, long-run, 167
157, 233 short-run, 167
Box-Pierce Q test, 234 Elasticity of subsititution, 206
Breusch-Godfrey LM test, 235, 242 Error correction model, 224
Durbin-Watson bound test, 235 Euler equation, 188
Ljung-Box Q test, 234
Forecasting, 271
Berndt-Wood model, 204 dynamic forecast, 276
Binary choice model, 115 ex-ante forecast, 272
Box-Cox variable transformation, 104, ex-post forecast, 271
114 forecast error statistics, 272
Full information maximum likelihood,
Censored regression model. See Limited FIML, 196
dependent variable model
CES production function, 96, 102, 111 Gamma probability distribution, 90, 179,
Cobb-Douglas production function, 52, 181
70, 161 GAUSS, 5
array operation, 13 White test, 134
characteristic roots or eigenvalues, 30 Heteroscedasticity-autocorrelation-
characteristic vectors or eigenvectors, consistent covariance matrix, 143, 183
30 Heteroscedasticity-consistent covariance
condition number, 31 matrix, 130, 143, 184
cumulative distribution function, 31 Hypothesis testings in nonlinear models
data transformation, 21 Lagrangian multiplier test, LM test, 111
descriptive statistics, 32 Likelihood ratio test, LR test, 111
element-by-element compatibility, 15 Hypothesis testings in nonlinear models,
file input and output, 21 110
gradient, 32 Wald test, 110
hessian, 32
least squares solution, 30 Information matrix, 89
logical operator, 17 Input control variable, 283
matrix operation, 13 _acf, 148, 154, 159, 236
probability density function, 31 _acf2, 241
relational operator, 16 _ar, 150, 158, 237, 238
sequential quadratic programming, 32 _arma, 158, 237
GAUSS for Windows, 5 _b, 84, 275
active file, 18 _begin, 58, 134, 202, 273
Command window, 6 _bgtest, 148, 154, 159, 236
Debug window, 7 _bjtest, 61, 135
Edit window, 6 _const, 55, 70, 203
GAUSS library system, 40 _conv, 87, 158
main file, 18 _corr, 75
main file list, 18 _dlags, 168, 176, 202, 230
Output window, 6 _drop, 150
Generalized autoregressive conditional _dynamic, 276
heteroscedasticity, GARCH, 240 _ebtest, 242
Generalized method of moments _end, 58, 134, 273
Hansen test for moment restrictions, _eq, 196, 266
181 _fbegin, 275
Langrangian multiplier test, 187 _fend, 275
Likelihood ratio test, 188 _fplot, 275
linear model, 192 _fstat, 275
nonlinear model, 182 _garch, 242
quadratic criterion function, 180 _garchx, 243
Wald test, 187 _hacv, 129, 144, 183, 192
White-Newey-West estimator, 180 _id, 196
Generalized method of moments, GMM, _iter, 150, 192, 199
179 _ivar, 170, 192, 196
Global control variables, 41, 283 _jacob, 104, 108
input control variables, 41 _method, 86, 150, 196
output control variables, 41 _names, 46
Goodness of fit, R2, adjusted R2, 47 _nlopt, 84, 105, 118, 125, 158
GPE package, 5, 41 _pdl, 173, 176
_print, 97, 153
Heteroscedasticity, 129 _restart, 87, 138, 201
Breusch-Pagan test, 134 _restr, 54, 68, 72, 173, 203, 218, 266
Goldfeld-Quandt test, 132 _rlist, 49, 61
Koenkar-Basset test, 135 _rplot, 49, 202
maximum likelihood estimation, 136 _rstat, 49, 168, 202
weighted least squares, 132 _step, 87
_tol, 87, 153, 201 line search, 87
_vb, 275 modified quadratic-hill climbing
_vcov, 50, 96, 130, 199 method, 87
_weight, 134 Newton-Raphson method, 86
Instrumental variable, 170, 192 quadratic hill-climbing (QHC) method,
Instrumental variable estimation, IV, 3, 87
167, 185 steepest-ascent method, 86
Nonlinear rational expectation, 179, 189
Jacobian, 211 Normal probability distribution, 89
Jacobian transformation, 103 Numerical derivatives, 83
Numerical Jacobian, 104, 289
Klein Model I, 197
Ordinary least squares, 45
L’Hôspital’s rule, 105 Orthogonality condition, 180
Latent variable, 116 Output control variable, 293, 295
Least squares estimation, 41 __a, 275
Least squares prediction, 41. See __b, 82, 275
Forecasting __e, 225
Likelihood function, 89 __r2, 78
Limited dependent variable model, 115 __rss, 133
Limited information maximum likelihood, __vb, 82, 275
LIML, 196
Linear probability model, 116 Panel data, 251
Linear restriction, 54, 203 Panel data analysis, 251
Wald F-test, 56, 69 between-estimates, 253
Logistic curve, 116 Breusch-Pagan LM test for random
Logit model, 116 effects, 258
Log-likelihood function, 89, 103, 118, deviation approach, 252
123, 210 dummy variable approach, 252
Log-normal probability distribution, 90 fixed effects, 252
Longitudinal data, 251 Hausman specification test for fixed or
random effects, 258
Maximum likelihood estimation, 86, 103 individual effects, 251
Moment function, 179 one-way analysis, 251
Moment restrictions, 180 partial deviation approach, 257
Moving average, 158 random effects, 258
Multicollinearity, 75 SUR method, 264
condition number, 31, 76 time effects, 251
Theil’s measure, 77 two-way analysis, 261
variance inflation factors, VIF, 79 Wald F-test for fixed effects, 253
Multiple regression, 50 within-estimates, 252
Multiplicative heteroscedasticity, 137 Partial adjustment, 167
Partial autocorrelation coefficient, 147,
Newey-West estimator, 143 234
Nonlinear full information maximum Partial autocorrelation function, PACF, 4,
likelihood, 209 148, 233, 241
Nonlinear least squares, 86, 101 Partial correlation coefficient, 75
Nonlinear optimization, 83 Partial regression coefficient, 52
BFGS quasi-Newton method, 86 Perfect collinearity, 66
DFP quasi-Newton method, 86 Permanent income hypothesis, 167, 192,
gradient, first derivatives, 83 223, 227
Greenstadt method, 86 Principal components, 80
hessian, second derivatives, 83 Probit model, 116
P-value, 48 dummy variable approach, 70
Sum-of-squares, 83, 97
Residual analysis, 48 System of simultaneous equations, 195
Bera-Jarque normality test, 61, 135 endogenous variable, 195
Durbin-Watson test statistic, 49 identity equation, 196
first-order rho, 49 predetermined variable, 195
kurtosis, 62 stochastic equation, 196
skewness, 62
Residual diagnostics, 61 Three-stage least squares, 3SLS, 196
DFFITS, 63 Time series analysis, 233
influential observations, 61 serial correlation in the mean, 233
leverage, 63 serial correlation in the variance, 233
outliers, 61 Time series conversion, 37, 39
standardized predicted residuals, 63 Tobit analysis, 123
standardized residuals, 63 Transfer function. See ARMAX
studentized residuals, 63 regression model
Residual sum of squares, RSS, 56 Translog cost function, 204
Restricted least squares, 54, 173, 218, 266 t-ratio, 47
Ridge regression, 80 Two-stage least squares, 2SLS, 196
Seasonality, 65 Unconstrained optimization, 83

Seemingly unrelated regression, SUR, Unit roots test, 215
196, 264
Shepard lemma, 204 Variance-covariance matrix, 50, 96, 110,
Simple regression, 46 131, 143, 180
Spurious regression, 215, 224
Structural change, 57 Weighted least squares, 109
Chow test, 57
ELECTRONIC MEDIA TRANSMITTAL AND GAUSS LightTM SOFTWARE
Terms and Conditions of Use for the CD-ROM Accompanying
COMPUTATIONAL ECONOMETRICS
GAUSS Programming for Econometricians and Financial Analysts
NOTICE: DO NOT UNSEAL THE ENVELOPE ENCLOSING THE CD-ROM (MEDIA)

OR USE THE MEDIA OR ITS CONTENTS BEFORE READING THESE TERMS AND
CONDITIONS OF USE. PUBLISHER AUTHORIZES USE OF THE MEDIA AND
CONTENTS ONLY UNDER THESE TERMS AND CONDITIONS OF USE. IF
RECIPIENT DOES NOT WISH TO BE BOUND BY THESE TERMS AND CONDITIONS
OF USE, RECIPIENT IS PROHIBITED FROM USING THE MEDIA AND CONTENTS,
AND RECIPIENT SHALL RETURN THE MEDIA AND CONTENTS TO PUBLISHER.
RECIPIENT’S UNSEALING OF THE ENVELOPE AND/OR USE OF THE MEDIA OR
CONTENTS SIGNIFIES RECIPIENT’S AGREEMENT TO THESE TERMS AND
CONDITIONS OF USE.
1. ACCEPTANCE OF TERMS. Recipient’s unsealing of the envelope enclosing the Media

and Contents or use of the Media or Contents subjects Recipient to all terms and conditions of
use described in this document and also all terms and conditions of the GAUSS LightTM
License displayed in the Media. Publisher and Aptech authorize use of the Media and
Contents only as set forth herein and within the GAUSS LightTM License.
2. NO WARRANTIES: Publisher and Aptech do not warrant the accuracy of the Contents as
contained in the Media against data corruption, computer viruses, errors in file transfer data,
unauthorized revisions to the files, or any other alterations or data destruction to the file(s).
The Media and its Contents are transmitted as is. Publisher and Aptech shall not have any
liability for Recipient’s use of the Media or its Contents, including without limitation, any
transmittal of bugs, viruses, or other destructive or harmful programs, scripts, applets or files
to the computers or networks of the Recipient. Recipient acknowledges and agrees that
Recipient is fully informed of the possibility of the Media or its Contents being harmful to
Recipient’s computers or networks and the possibility that the Contents may not be an exact
and virus-free copy of masters by Publisher or Aptech. Recipient also acknowledges, agrees,
and warrants that Recipient shall be solely responsible for inspection and testing of the Media
and the Contents for bugs, viruses, or other destructive or harmful programs, scripts, applets
or files, before accessing or using the Media or Contents.
3. NO IMPLIED WARRANTIES. THERE ARE NO IMPLIED WARRANTIES,

INCLUDING WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A
PARTICULAR PURPOSE, WITH RESPECT EITHER TO THE MEDIA OR THE
CONTENTS.
4. LIMITATIONS ON LIABILITY AND REMEDIES. Publisher and Aptech shall have no

liability for any general damages, direct or indirect damages, special damages, exemplary
damages, statutory damages, punitive damages, or consequential damages, including without
limitation, lost profits, interruption of business, for any use of the Media or Contents.
Recipient’s sole and exclusive remedy for any claim based on Recipient’s use of the Media or
Contents shall be either (a) the delivery of another copy of the Contents on replacement
Media, or (b) the costs of a physical replacement Media (for example, the costs of a diskette,
tape or other back-up media) but Publisher shall have sole discretion to elect which remedy to
provide. There shall be no other remedies, whether in law or equity.
5. GAUSS LIGHT SOFTWARE LICENSE: Installation and use of this software is subject to
and governed by the License Agreement displayed in the Media. By installing and using the
GAUSS LightTM Software, Recipient indicates his or her acceptance of, and Recipient is
subject to, all such terms and conditions of the License Agreement. Violation of the License
Agreement is also a violation of the copyright laws.
GAUSS LightTM Software is GAUSS LightTM Software is

furnished by: published by:
Aptech Systems, Inc. (“Aptech”) Tsinghua University Press(“Publisher”)

28304 SE Kent-Kangley Rd Tsinghua University
Maple Valley, WA 98038 Beijing 100084, China
Phone: 425-432-7855 Phone: +86-10-6279-4748
FAX: 425-432-7832 FAX: +86-10-6279-4748
Email: [email protected] Email: [email protected]
www.aptech.com www.tup.tsinghua.edu.cn

Gpebook PDF

Uploaded by

Copyright:

Available Formats

Gpebook PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Gpebook PDF

Uploaded by

Copyright:

Available Formats

GAUSS Programming

for Econometricians and Financial Analysts

Copyright  2001-2003 by K.–P. Lin

Published by ETEXT Textbook Publisher, www.etext.net.

For permission requests or further information, see www.etext.net or email [email protected].

Printed in the United States of America.

Limit of Liability and Disclaimer of Warranty

As you probably know, GAUSS is a programming language designed for matrix-

including single linear equation estimation and prediction, systems of simultaneous

Conventions Used in this Book

Extra notes and additional information are given here.

This warns of common mistakes causing programming errors.

Hints or remarks specific to GAUSS and GPE.3

2SLS Two-stage least squares

QHC Quadratic hill-climbing optimization method

• Click and select File/Exit from the menu bar.

for running single-line commands or program files. It is more commonly referred as

An Introduction to GAUSS Language5

Matrices as Fundamental Objects

Consider the following GAUSS expression:

(0) : error G0036 : matrices are not conformable

or, equivalently (though verbosely):

That is, if an expression has no assignment operator (=), it will be assumed to be an

1.0000000 2.0000000 3.0000000

Making Matrices from Matrices

gives a row vector:

1.0000000 2.0000000 3.0000000 4.0000000

gives a 2×3 matrix:

1.0000000 2.0000000 3.0000000

Note that by putting variables in brace brackets such as:

will not work. It produces a syntax error message.

Using Portions of Matrices

1.0000000 2.0000000 3.0000000

Variables may also be used as subscripts. Thus,

Another useful example is

This is the same as

Try the following:

and see what GAUSS gives you.

Matrix and Array Operations

results in the message:

(0) : error G0036 : matrices are not conformable

will get you a new 3×4 matrix C.

2.0000000 4.0000000 6.0000000

2.0000000 3.0000000 4.0000000

These situations are what we call “array operation” or element-by-element

Matrix multiplication is indicated by an asterisk (*), commonly regarded in

4.0000000 10.000000 18.000000

which is the “dot product” of two row vectors x and y.

1.0000000 4.0000000 9.0000000

GAUSS array operations include multiplication (.*), division (./) and

Logical and Relational Operations on Matrices

• LT or < Less than

• LE or <= Less than or equal to

• GT or > Greater than

• GE or >= Greater than or equal to

Whenever GAUSS encounters a relational operator, it produces a one (1) if the

Relational operators can be used on element-by-element compatible matrices. For

Creating and Editing a GAUSS Program

statement, this is no small accomplishment. However, for many purposes it is